Counterparty Cascade: When Artificial Intelligence Infrastructure Becomes a Network of Interdependent Financial Promises — How Hyperscaler Contracts, Semiconductor Suppliers, New Clouds, Private Credit, and Frontier Models Could Transmit Economic Risk Across the Five-Layer AI Economy Through 2030

Introduction: The Day Artificial Intelligence Revealed Its Financial Architecture

On the morning of November 2, 2026, a document appeared in the electronic archives of the United States Securities and Exchange Commission that deserved attention far beyond Silicon Valley, and far beyond the community of equity analysts who had been waiting for it since the preceding week. Anthropic, PBC, the San Francisco artificial intelligence company founded by Dario and Daniela Amodei, filed Amendment No. 1 to its Form S-1 registration statement, updating the preliminary prospectus for an initial public offering that had first been unveiled on October 26, 2026. The original filing had included financial statements only through the first half of the year; the amendment supplied the third quarter, and with it a figure that reorganized the way the market thought about the economics of frontier artificial intelligence. Anthropic disclosed that, as of September 30, 2026 and in the weeks that followed, it had entered into purchase commitments and lease agreements with an aggregate committed spend of approximately $895.4 billion, substantially all of it related to compute and infrastructure, which is to say chips, datacenters, and the energy required to run them.^[1] ^[2]

The number was extraordinary, and the financial press treated it as such, but it was not necessarily the most revealing part of the document. The amendment also disclosed that approximately $213.8 billion of the aggregate commitments involved arrangements subject to termination rights or to conditions that had to be satisfied before the relevant capacity could commence, including the ability of counterparties to secure financing, obtain facilities, or receive regulatory approvals, and it stated that approximately $69.8 billion of the aggregate contractual obligations was expected to become payable in the twelve months following September 30, 2026.^[3] The company further acknowledged, in the plain language that risk-factor sections tend to adopt when they are describing something that genuinely worries management, that certain computing and infrastructure providers depended on outside financial support to acquire advanced chips and to construct the facilities necessary to fulfill their contracts, and that should such financing become unavailable or more expensive, the promised computing capacity might arrive late, cost more, be reduced in scale, or never become available at all.^[1]

Consider the practical significance of this disclosure for a moment, because its implications extend well past the company that made it. Somewhere in the United States, an infrastructure developer is preparing a multibillion-dollar datacenter whose future revenue depends on a contract with an artificial intelligence company. A lender is evaluating that contract as the primary justification for financing the construction. An equipment supplier is reserving accelerators, networking hardware, cooling systems, and electrical switchgear in anticipation of the project’s completion. A utility is planning transmission upgrades, substations, and additional generation capacity to serve the facility, and it is doing so on a schedule that may extend over a decade. Construction contractors are hiring workers, ordering transformers with lead times measured in years, and negotiating long-term supply agreements. Meanwhile, the prospective computing customer has already incorporated the expected capacity into its model-development roadmap, its revenue projections, and the story it is telling prospective public shareholders.

Now imagine that one participant in this chain encounters an unexpected problem. A lender withdraws its financing commitment because credit conditions deteriorate or because an unrelated portfolio loss forces a reduction in exposure to the technology sector. The developer cannot obtain enough capital to complete construction on schedule. The datacenter misses its delivery date. Equipment that has already been manufactured and delivered cannot be installed or energized. The AI customer must secure replacement computing capacity on the short-term market at a materially higher price per unit. A supplier faces postponed deliveries and an inventory that is depreciating in economic value even while it sits in a warehouse. Other lenders, observing the episode, reevaluate the assumptions underlying similar projects in their own portfolios. The utility discovers that the electricity demand it had projected, and perhaps already begun to build for, may arrive several years later than anticipated, or in a smaller quantity, or not at all.

No single technological failure is required for any of this to occur. A financial decision at one institution can propagate through an entire industrial system.

The sequence is hypothetical, but the underlying dependencies are not. Anthropic’s November 2 disclosure describes multiparty arrangements in which one organization may supply or procure chips, another may finance the equipment, another may provide the physical facility, and yet another may operate the resulting infrastructure, and it identifies the possibility of payment obligations arising before all of these components have been assembled into usable computing capacity.^[1] The filing also names the specific neocloud counterparties, Nscale and Lambda among them, through which a substantial portion of its NVIDIA-based capacity is to be delivered, and in doing so it points the reader toward a second set of documents that tell the same story from the other side of the contract. Nscale’s own registration statement, filed in September 2026, reports approximately $103.4 billion of total contract value against only about $2.6 billion of contracts that were live and generating revenue, and it states plainly that the company still needed to secure binding financing for the GPUs and datacenter infrastructure required to fulfill its agreements with Anthropic.^[4] ^[5]

This is the overlooked financial dimension of the artificial intelligence infrastructure expansion, and it is the subject of this paper.

Why I Choose the Title “Counterparty Cascade”

I choose the title Counterparty Cascade because the next stage of artificial intelligence investment cannot be understood simply by adding together hyperscaler capital expenditures, GPU shipments, datacenter construction budgets, and electricity requirements, as though the industry were a stack of independent spending decisions that happened to point in the same direction. The AI economy is increasingly structured through interdependent contractual promises. A financial institution promises capital; a semiconductor supplier promises equipment; an infrastructure developer promises capacity; a utility promises electricity service; and an AI developer promises future payments supported by expected demand for its models and applications. The word Counterparty identifies the organizations holding these reciprocal obligations, and the word Cascade captures the possibility that a failure, a delay, or a withdrawal of support at one institution could produce consequences throughout the wider system, consequences that are not proportional to the size of the institution where the disruption began.

The title also expresses a deeper economic concern within my Five-Layer AI Economy framework, which I have developed across earlier papers and which organizes the industry into Energy, Chips, Datacenters, Models, and Applications or Agentic Systems. These layers are technologically distinct, and they are studied by different communities of engineers and analysts, but their financial obligations are becoming increasingly interconnected in ways that none of those communities fully observes. I use Counterparty Cascade to describe the process by which financial distress, delayed performance, or the withdrawal of contractual support from one participant transmits operational, financial, and investment consequences to other participants, sometimes in adjacent layers and sometimes in layers that appear to have no direct relationship with the original disruption. The purpose is not to predict an inevitable AI financial crisis. It is to establish a framework for understanding how the industry’s increasingly ambitious infrastructure commitments could create new forms of interconnected risk between 2027 and 2030, and how those risks might be measured, disclosed, and managed before stress reveals their full consequences.

The fundamental research question of this paper can be stated in a single sentence, although answering it will require the remainder of the document:

What happens when an industry whose infrastructure depends on increasingly long-term financial promises must generate sufficient economic value from technologies, customers, and business models that are still evolving?

This question has become unusually urgent in the autumn of 2026, and it is worth pausing to explain why. For the first phase of the generative AI revolution, roughly from the public release of large language models in late 2022 through the end of 2024, investors and corporate executives focused primarily on the capabilities of increasingly sophisticated models. They evaluated parameter counts, inference performance, reasoning benchmarks, GPU availability, and the commercial potential of intelligent applications, and the capital required to pursue these goals was supplied overwhelmingly from the operating cash flows of a small number of extraordinarily profitable technology companies. By 2026, the investment question had become more complicated in at least three respects.

First, developing a competitive frontier model increasingly required access to enormous quantities of computing infrastructure that could not be obtained through ordinary short-term cloud purchases, and accessing that infrastructure required contracts with cloud providers, semiconductor companies, specialist operators, and financial institutions, contracts whose duration frequently exceeded the plausible planning horizon of the technology itself. Second, the aggregate scale of the buildout had outgrown the internal cash generation of even the largest participants. Professor Stijn Van Nieuwerburgh of Columbia Business School, in a September 2026 paper prepared for the Brookings Papers on Economic Activity, estimated that aggregate capital expenditures by Oracle, Microsoft, Amazon, Meta, and Alphabet had risen from roughly $97 billion in 2020 to more than $400 billion in 2025 and were projected to exceed $800 billion in 2026, at which point they would surpass the five firms’ combined operating cash flow for the first time.^[6] Third, and as a direct consequence of the second, the financing of the buildout had migrated toward leases, joint ventures, project debt, private credit, securitization, and special-purpose vehicles, structures that expand funding capacity but that also raise asset-level leverage, obscure contingent obligations, and expose investors to tenant concentration, technological obsolescence, execution delays, and uncertain residual values.^[6]

The central difficulty running through all three developments is that capital commitments can remain contractually fixed while the technological and commercial assumptions supporting them change. A datacenter designed around one generation of accelerators might enter commercial service after its successor has become available. A model developer could discover that an unexpected architectural improvement reduces the amount of computing required for a particular task by an order of magnitude. Conversely, a breakthrough in autonomous agents or scientific reasoning could cause demand for inference to grow faster than any contract anticipated. In each case the physical and financial commitments made in 2025 and 2026 will still be in force in 2029 and 2030, and the institutions that made them will have to live with the consequences of assumptions that were reasonable when they were formed and that may be unrecognizable by the time they mature.

The international financial institutions have recognized related vulnerabilities, and their language has sharpened considerably over the course of 2026. The International Monetary Fund’s October 2026 Global Financial Stability Report, released on October 13 during the Annual Meetings in Bangkok, identified a potential correction in AI-related investment and asset prices as one of three salient near-term risks to global financial stability, alongside a sharp rise in sovereign bond yields and a major cyber incident, and it devoted a background analysis to what it called rising leverage in the AI ecosystem.^[7] The accompanying IMF staff blog, written by Athanasios Vamvakidis and Jason Wu, was unusually direct about the appropriate posture for policymakers.

“There is no room for complacency.” — Athanasios Vamvakidis and Jason Wu, International Monetary Fund ^[8]

The Fund’s analysis matters to this paper for a particular reason. It does not merely warn that AI equities are expensive, which is a familiar observation. It identifies a transmission mechanism, in which equity investment, financing arrangements, and future revenue promises increasingly connect firms across the artificial intelligence industry, so that if a company receiving financing comes under pressure, weaker revenue expectations and valuations could erode the value of related debt and extend stress beyond the AI firms themselves.^[7] The IMF’s companion technical note on the structure and financing of the AI ecosystem, prepared by Yingyuan Chen, Deepali Gautam, and Jason Wu, documented that fourteen publicly disclosed circular-financing transactions since 2023 accounted for roughly $300 billion of equity investments led by hyperscalers in financially weaker AI firms, supplemented by approximately $650 billion in debt, and it identified technological obsolescence as a distinct vulnerability, noting that shortening the assumed asset life from seven to three years reduces modeled operating margins by more than nine percentage points.^[9] The Bank for International Settlements reached a complementary conclusion from a different direction. Its Annual Economic Report, published on June 28, 2026, named the AI capital expenditure cycle as one of four pressure points facing the global economy and flagged circular financing structures as an amplifier of downside risk, and its Bulletin No. 137, published on October 1, 2026, measured for the first time that 55.2 percent of the investment value flowing into AI firms between 2021 and 2025 had come from other AI firms.^[10] ^[11]

These findings provide an important institutional foundation for this paper. However, Counterparty Cascade advances a more specific analytical question than the one the international institutions have so far asked: how does an initial problem at one company or project travel through a network of contractual obligations, and under what conditions does that problem become large enough to affect the financing and operational decisions of otherwise healthy institutions that have no direct relationship with the original participant? Answering that question requires examining the complete industrial architecture, from the power plant to the enterprise software subscription, and it requires treating contracts, rather than companies, as the fundamental unit of analysis.

This paper advances five central propositions, which the sections that follow develop in turn.

First, AI infrastructure commitments should be evaluated as a network of contingent obligations rather than as a collection of isolated investment announcements, because the economic exposure created by a contract depends on the obligations and dependencies it creates for other parties and not only on its nominal value.

Second, the greatest vulnerabilities may arise where the payment obligations of one institution become disconnected from the physical delivery schedules of another, because a system in which individual counterparties can perform and be paid while the complete infrastructure remains unusable is a system in which liquidity stress can accumulate silently.

Third, financing concentration can produce economic contagion even when the underlying projects are located in different states, supported by different utilities, and operated by different cloud providers, because geographic and contractual diversification frequently conceals shared dependencies on the same lenders, the same semiconductor foundries, the same transformer manufacturers, and the same small group of frontier model customers.

Fourth, public electricity infrastructure and regional economic development policies increasingly intersect with private AI financing arrangements, so that the allocation of risk between developers, utilities, and ratepayers has become a question of financial stability as much as a question of regulatory fairness.

Fifth, the industry can materially reduce the likelihood of destructive financial cascades through better contractual design, improved disclosure, diversified financing, independently verified delivery milestones, and more realistic assumptions about technological and commercial change, and it can do so without abandoning the long-term commitments that make large-scale infrastructure financeable in the first place.

The objective throughout is to understand financial interdependence before stress reveals its full consequences, because the history of infrastructure booms, from the canals of the 1830s to the fiber networks of the late 1990s, suggests that the moment of maximum vulnerability is precisely the moment at which the financing structures are newest, the least tested, and the least visible to the people whose capital ultimately supports them.

Section 1: The Contractual Anatomy of the AI Economy

1.1. From Computing Purchases to Multidecade Financial Obligations

For much of the commercial internet’s development, cloud computing was presented, and largely experienced, as an alternative to owning expensive infrastructure. A company could purchase computing capacity from a provider without constructing a datacenter, acquiring servers, operating cooling systems, or directly negotiating electricity contracts, and this arrangement transformed enterprise technology economics in a way that is now so familiar it has become invisible. Organizations converted substantial portions of their computing requirements from upfront capital expenditure into service expenses that scaled with usage. They could expand capacity when demand increased and reduce certain expenditures when demand weakened, and the balance-sheet risk of owning rapidly depreciating hardware was absorbed by a handful of providers whose scale allowed them to manage it efficiently.

Artificial intelligence is partially reversing that model, and the reversal is one of the least appreciated structural changes in the technology industry of the mid-2020s. The largest frontier AI companies now require infrastructure at a scale that cannot always be supplied through ordinary short-term cloud purchases, because the capacity they need must often be reserved years before deployment, in facilities that do not yet exist, powered by electricity connections that have not yet been built. In some circumstances, model developers also have economic reasons to acquire greater control over their infrastructure, since dedicated facilities can offer more predictable access to advanced accelerators, potentially lower long-term operating costs, and improved control over the performance characteristics of massive, tightly coupled computing clusters. However, these advantages require more complex financial arrangements, and the complexity is where the counterparty risk lives.

Anthropic’s registration statement describes its transition toward a hybrid infrastructure strategy combining conventional cloud service agreements with equipment leasing and dedicated facilities, and it describes an internally designated approach, which the filing calls Anthropic Direct, that separates commitments for physical datacenter capacity from commitments for the computing hardware installed inside those facilities.^[1] The financial significance of that separation is substantial and deserves careful statement. A conventional cloud service agreement assigns responsibility for much of the infrastructure delivery process to the provider, who owns the building, owns or leases the equipment, holds the electricity contract, and bears the consequences if any of these components is late. A dedicated infrastructure arrangement may instead divide responsibility among several companies. One entity procures the datacenter facility. Another obtains financing for the hardware. Another sources the specialized accelerators. A separate organization installs and manages the equipment. A utility provides electricity under yet another contractual arrangement. The AI developer, meanwhile, may remain economically exposed to the performance of all of these organizations simultaneously, even though it has a direct contract with only some of them.

Under this structure, the completion of each individual contract does not necessarily mean that the complete computing system is ready to generate revenue. A hardware supplier can perform its obligation in full while a facility developer remains delayed. An equipment lessor can satisfy its financing requirements while a power connection remains unavailable. A construction contractor may complete a building to specification even though an essential networking system has not arrived from a supplier on another continent. Consequently, the AI customer can face substantial expenditures without possessing operational computing capacity, and the institutions that financed each component can each be contractually entitled to payment while the asset that was supposed to generate the cash to pay them sits dark.

This is the first structural foundation of Counterparty Cascade: individually completed contractual obligations can coexist with an economically incomplete infrastructure project. It is a feature of the architecture rather than a defect in any particular contract, and it means that the risk of the system cannot be assessed by reading any single agreement in isolation.

1.2. Understanding Anthropic’s $895.4 Billion Commitment Structure

The November 2 filing permits a more detailed examination than the headline figure alone, and that examination is necessary because the aggregate number has already been used, in the hours since the filing, to support conclusions it cannot bear. Anthropic’s approximately $895.4 billion aggregate commitment figure is composed of several categories, which the filing’s contractual obligations disclosure and its subsequent-events discussion allow us to reconstruct.

Commitment category	Approximate amount
Contractual obligations listed as of September 30, 2026	$320.0 billion
Additional short-term commitments as of September 30	$17.4 billion
Executed leases not yet commenced	$451.2 billion
Additional longer-term commitments entered after September 30	$100.2 billion
Additional short-term commitments entered after September 30	$6.6 billion
Aggregate disclosed commitments	$895.4 billion

Source: Anthropic, Form S-1/A, November 2, 2026, as reconstructed from the contractual obligations table and subsequent-events disclosure; figures rounded.^[1] ^[3]

The categories have different economic characteristics, and the differences are not cosmetic. Some represent minimum purchase payments for cloud services under agreements that are non-cancelable and that apply regardless of actual usage. Others concern leases that have been executed but have not yet commenced, and whose commencement depends on the construction of facilities and the readiness of equipment. The short-term categories, which the filing defines as commitments cancelable without penalty within one year or amounts under agreements with a remaining term of one year or less, carry a fundamentally different liquidity profile from the long-dated leases.^[3] Certain obligations depend on delivery schedules, financing conditions, or other contractual milestones that have not yet been satisfied. Consequently, the total should not be interpreted as a single immediately payable debt obligation, nor should it be represented as money already spent, and the trajectory of the figure is as instructive as its level: the company’s noncancelable purchase commitments stood at approximately $54.6 billion at the end of 2025 and approximately $265.6 billion at June 30, 2026, before the amendment’s broader aggregate was disclosed.^[12]

Equally important, the aggregate does not necessarily represent the maximum amount Anthropic could ultimately spend on computing infrastructure, because the filing discusses additional arrangements without fixed dollar commitments, including estimated future expenditures for certain TPU-based computing resources and AMD systems, which are excluded from the stated aggregate under the filing’s methodology.^[1] These distinctions matter because financial risk depends on the timing, conditionality, enforceability, financing structure, and operating performance associated with each obligation, and a reader who collapses them into a single number will systematically misjudge both the near-term liquidity exposure and the long-term economic exposure. The $895.4 billion figure is therefore best understood as a disclosed aggregate of future spending obligations across different contractual categories, and a rigorous analysis must separate five dimensions of every obligation within it.

Commitment value: the nominal amount associated with an agreement.
Payment timing: when cash must actually be paid, which for leases not yet commenced may be years away.
Delivery dependency: what must occur, and which third parties must perform, before the purchased capacity becomes available.
Termination protection: which obligations can be canceled or modified under specified conditions, and at whose option.
Funding dependency: whether the counterparty can fulfill its obligations without obtaining additional financing that has not yet been secured.

A contract with strong delivery protections may present less immediate risk than another agreement with a smaller nominal value but weaker protections. A lease supported by significant existing capital may have a different risk profile from a proposed facility dependent on future construction financing. The analytical challenge, which this paper attempts to meet, is to distinguish contractual scale from genuine economic exposure.

A particularly important distinction concerns the conditional commitments. Approximately $213.8 billion of Anthropic’s aggregate contractual obligations is associated with termination rights or with conditions that must be satisfied before commencement, which represents approximately 23.9 percent of the disclosed aggregate.^[3] This percentage should not be interpreted as a probability of default, an expected loss, or the share of commitments likely to fail. It simply identifies the portion of disclosed commitments possessing specified forms of conditionality or termination protection, and the economic interpretation is more subtle than either the optimistic or the pessimistic reading would suggest. The existence of a termination right may protect the customer from having to purchase undelivered capacity, which is genuinely valuable. However, the same failed project may still create expenses through replacement procurement at higher short-term prices, delayed revenue from models that could not be trained or served on schedule, contract renegotiation, development reimbursements, or lost commercial opportunities, and the filing itself quantifies the premium involved, sizing approximately $5.0 billion of costlier short-term compute that the company has had to purchase to bridge gaps in longer-term capacity.^[3] Meanwhile, the developer’s lenders and suppliers may encounter losses even if the customer’s contractual payment obligation is successfully terminated, because the termination right protects the customer, not the network.

Termination rights can therefore reduce an individual party’s exposure without eliminating the consequences for the wider system. This distinction will become central to the analysis of financial contagion in Section 3.

1.3. Anthropic, Amazon, Google, and Microsoft: The Cloud Contract Network

Anthropic’s infrastructure strategy demonstrates how frontier model developers are increasingly connected to multiple large technology companies at once, and the amended filing describes major computing commitments involving Microsoft, Google, and Amazon in terms specific enough to permit comparison.

Provider	Disclosed minimum contractual commitment	Relevant period
Microsoft	At least $44.9 billion	November 2026 – mid-2033
Google	At least approximately $121.3 billion	April 2026 – July 2033
Amazon	At least $110.0 billion	May 2026 – April 2036

Source: Anthropic, Form S-1/A, November 2, 2026, as reported by Unite.AI and CNBC. These are disclosed minimum spending commitments under specified agreements that apply regardless of actual usage, subject to their stated terms; they are components of the broader aggregate disclosure and must not be added to it as though they were separate.^[2] ^[3]

These agreements have important differences, but they also illustrate the central importance of minimum spending obligations in the current generation of AI infrastructure contracts. In the Google and Amazon arrangements, the filing describes mechanisms under which insufficient actual spending may require Anthropic to pay the difference between the amount consumed and the applicable contractual minimum, subject to specified adjustments or exceptions.^[1] Such provisions provide greater revenue predictability to the infrastructure supplier, and that predictability can improve the supplier’s ability to justify large capital investments to its own board and its own lenders. However, it also changes the economics for the customer in a way that the customer’s own investors should understand. If demand develops more slowly than expected, the model company may remain obligated to pay for capacity it does not fully utilize, and the result is a transfer of utilization risk from the infrastructure provider to the AI developer.

This transfer is economically rational when the customer values guaranteed access to scarce computing resources more than it values flexibility, and in 2025 and 2026 scarcity has been the dominant fact of the market. The difficulty emerges when expectations change materially after the commitment has been made. Suppose a developer secures ten years of capacity because it anticipates sustained annual increases in inference demand. Several years later, improved model architectures might allow the same workloads to operate using substantially fewer accelerators. The developer would benefit technologically, and its customers would benefit from lower prices, but its infrastructure contract might not automatically decline in proportion to the reduction in computing requirements. Alternatively, the opposite might occur: an unexpected increase in demand for AI agents could make existing contractual capacity insufficient, requiring the developer to purchase expensive short-term resources at exactly the moment when every other developer is trying to do the same thing. In both cases, contractual arrangements, rather than technology alone, determine the relationship between technological progress and financial returns.

There is another dimension to this network that distinguishes it from an ordinary supplier relationship. Amazon, Google, and Microsoft are not merely independent infrastructure suppliers to Anthropic. Each is also an investor in the company, each operates competing AI platforms, and each develops or distributes AI services that compete with Anthropic’s own. The filing’s exhibits include a 2026 stock purchase and convertible note facility agreement with Amazon Web Services providing for up to $20 billion of investment, with an option for a further $5 billion in a subsequent equity financing, and Microsoft’s $5 billion investment in November 2025 was announced alongside the original cloud commitment that the amended filing has since modified.^[1] ^[12] Amazon’s own third-quarter 2026 results, released on October 29, remind investors that its reported net income declined year over year in part because the third quarter of 2025 had included $9.5 billion of pre-tax gains on its Anthropic investment, which is to say that the fortunes of the cloud provider’s income statement are already visibly tied to the valuation of its largest AI customer.^[13] These companies can simultaneously be strategic partners, investors, cloud providers, distribution channels, and competitors within the same market, and the IMF has given this configuration a name borrowed from derivatives risk management.

Distress at key investee firms could present significant “wrong-way risk” to the AI ecosystem. — International Monetary Fund, Global Financial Stability Report, October 2026 ^[7]

Such relationships are not inherently destabilizing. They may provide access to financial resources, technical expertise, and global distribution that a standalone model developer could not obtain independently, and the alignment of incentives they create is part of why the infrastructure has been built as quickly as it has. But overlapping roles create governance questions that the contracts do not fully answer. What happens when a strategic investor’s own infrastructure requirements compete with those of its model-development partners during a period of scarcity? How should scarce computing capacity be allocated during an unexpected shortage, and who decides? What happens if a cloud provider changes its capital expenditure priorities in response to pressure from its own shareholders, as Alphabet and Amazon both experienced in the final week of October 2026 when their shares fell sharply on the announcement of larger-than-expected spending?^[14] ^[13] Could a decision made to protect the provider’s balance sheet affect the development schedules of multiple AI companies simultaneously? A contractual network becomes financially important precisely when the same organization occupies several positions within it, because a shock to that organization travels along every position at once.

1.4. NVIDIA, AMD, Broadcom, and the Economics of Semiconductor Dependence

The semiconductor industry creates another layer of contractual interdependence, and in 2026 it has begun to create something more consequential than dependence, which is a financing architecture of its own. Advanced AI accelerators depend on specialized manufacturing, high-bandwidth memory, advanced packaging, high-speed networking, cooling equipment, and increasingly complex rack-scale systems, so that a commitment to purchase accelerators generates demand throughout a much larger supply chain, and a delay anywhere in that chain propagates to the delivery of the finished system. For NVIDIA, the opportunity extends beyond selling individual GPUs. Its infrastructure ecosystem encompasses processors, networking, software, systems integration, and, as of August 2026, partnerships with the largest alternative asset managers in the world.

On August 10, 2026, NVIDIA announced memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish independent compute-financing platforms intended to mobilize more than $500 billion of third-party capital over time for the buildout of AI infrastructure across its ecosystem of frontier labs, enterprises, and AI clouds.^[15] The announcement’s framing was explicit about its ambition, which was nothing less than to make computing hardware an investable asset class alongside pipelines, toll roads, and aircraft.

“In AI, compute is revenue.” — Jensen Huang, Founder and CEO, NVIDIA ^[15]

That $500 billion figure represents an announced financing ambition, subject to execution of final agreements, and not evidence that the entire amount has been raised or deployed. Nevertheless, it illustrates an important transformation. Semiconductor infrastructure is becoming an investment category attractive to large asset managers and alternative financing institutions, and the technology supplier increasingly participates in an ecosystem that combines equipment sales, infrastructure development, customer commitments, and financing. Van Nieuwerburgh’s Brookings paper notes that NVIDIA will provide a 25 percent residual value guarantee under these platforms, exposing it to as much as $125 billion of backstops, and it observes that Google and Broadcom have entered into a similar arrangement to help finance the acquisition of TPU chips.^[6] This arrangement can accelerate industrial investment, because it supplies capital to customers who could not otherwise afford the hardware. However, it also creates dependencies between commercial demand for AI chips and the willingness of financial institutions to fund their deployment, and it places the equipment vendor in the position of guaranteeing the future value of its own products.

Consider two separate risks that this architecture must manage. The first is technological: a new generation of accelerators could deliver superior performance per dollar, changing the attractiveness of existing hardware and the rental price it can command. The second is contractual: even if an older generation remains technically useful, its economic value may not be sufficient to support the financing assumptions under which it was originally acquired. The distinction matters because a functioning GPU is not necessarily a financially adequate GPU. The hardware may continue to perform calculations while generating insufficient revenue to service the debt associated with its acquisition, and the gap between those two states is where lenders, lessors, and guarantors absorb losses.

Broadcom introduces a further dimension that is visible in Anthropic’s own exhibits. The filing describes master equipment lease arrangements involving Broadcom and certain financing entities, with Exhibit 10.11 identified as a Master Equipment Lease Agreement between an entity named AI Infrastructure OpCo II, LLC and Broadcom Inc., dated March 31, 2026, and it describes aggregate rent payments under the disclosed lease arrangements of approximately $161.2 billion excluding interest, with payments generally beginning when equipment becomes ready for service, subject to contractual exceptions.^[1] ^[3] The filing also describes assignments of certain lease rights and obligations to financing entities, guarantees from Anthropic, and provisions involving defaults and potential acceleration.^[1] These arrangements illustrate how semiconductor-related contracts can become financing structures with multiple legal counterparties, in which the company supplying or arranging the computing equipment does not ultimately remain the sole institution bearing the financing exposure. Rights to receive payments can be assigned. Special-purpose entities can hold equipment-related obligations. The economic risks may become distributed among technology companies, financing vehicles, and institutional investors, each of which sees only its own slice of the arrangement.

Such distribution can improve access to capital. But it also raises a question that conventional semiconductor market-share analysis does not address at all: who ultimately bears the loss if the computing equipment is delivered but the economics supporting its long-term lease deteriorate? The answer depends on specific contractual provisions, on the sequence of guarantees, and on which entity holds the residual-value exposure. It cannot be determined merely from the brand of accelerator installed inside a datacenter, and it is rarely disclosed at the level of detail that would allow an outside investor to determine it.

1.5. Specialized Cloud Operators and the Financialization of Computing Capacity

An additional category of participant has become central to the architecture: specialized infrastructure operators, often called new clouds, that provide computing services to AI developers without possessing the diversified business models or the investment-grade balance sheets of the largest hyperscalers. CoreWeave, Nscale, and Lambda are the most prominent examples, and all three appear, directly or indirectly, in the financing chains that this paper examines. These companies serve a valuable economic function. They can develop specialized infrastructure more quickly than traditional enterprise cloud providers, establish close relationships with major semiconductor suppliers, and provide dedicated computing capacity to demanding AI customers who need it faster than the hyperscalers can supply it. But their investment requirements are exceptionally large relative to their existing revenue bases, and their financing structures make the payment–delivery dependencies of the AI economy unusually visible.

CoreWeave’s public disclosures offer the clearest illustration. Its quarterly report for the period ended June 30, 2026 describes infrastructure investment financed through combinations of debt, equity, delayed-draw lending facilities, original-equipment-manufacturer financing, and existing cash resources, and it reports $13.6 billion outstanding under delayed-draw term loan facilities and $16.6 billion of aggregate outstanding senior and convertible notes, with coupons on the senior notes ranging from 9.00 to 9.75 percent.^[16] The company explains that the delayed-draw facilities are collateralized by the assets underlying the contributed customer contracts and by the pledged contractual cash flows, that they are drawn as infrastructure is constructed to satisfy customer requirements, and that they amortize over time as contracted cash flows are generated.^[16] In the same quarter the company reported revenue of $2.58 billion, net interest expense of $640 million, a net loss of $626 million, and a revenue backlog of approximately $104 billion, roughly forty quarters of current-quarter revenue, of which it expected about 41 percent of remaining performance obligations to be recognized within twenty-four months.^[17]

“Our scale began to translate into expanding operating leverage.” — Michael Intrator, Co-founder, Chairman, and CEO, CoreWeave ^[17]

This creates a recognizable financing structure, and it is worth stating its logic explicitly because the same logic recurs throughout the industry. A customer contract supports the economic rationale for an investment. The operator raises capital against anticipated future revenue. The financing enables the purchase and installation of computing equipment. The completed infrastructure begins delivering services. Customer payments support debt service and operating costs. The arrangement works extremely well when contracts, financing, construction, equipment delivery, and commercial utilization remain aligned, and the vulnerability arises when the sequence breaks, because the debt is drawn against the construction schedule while the cash arrives only on the delivery schedule.

Anthropic’s arrangements with Nscale and Lambda provide a particularly useful example of how far down the chain these dependencies now extend. According to Nscale’s own registration statement, filed September 18, 2026, the company entered into a series of GPU services agreements with Anthropic on August 25, 2026, providing for aggregate payments of up to approximately $44.6 billion for dedicated infrastructure at its Monarch Compute Campus to support the deployment of NVIDIA Vera Rubin NVL72 systems, and its financing included a subscription agreement of at least $3.1 billion of convertible loan notes, of which $1.0 billion was to be issued to NVIDIA.^[4] The same filing reported approximately 461,000 GPUs classed as active or contracted, of which only around 25,000 were live, and acknowledged that binding financing for the GPUs and datacenter infrastructure required by the Anthropic agreements had not yet been secured.^[5] Anthropic’s filing, for its part, describes the Nscale fees as beginning only as capacity is delivered and accepted and describes Nscale’s obligation to use best efforts to secure qualifying financing, while the approximately $41.3 billion of agreements with Lambda extending into 2034 include provisions allowing termination under certain circumstances if a facility lease is not secured or delivery milestones are missed.^[1] Reporting on the Lambda arrangement adds a further link: Lambda holds the lease on a datacenter in Nueces County, Texas, developed by Hut 8, while NVIDIA has a separate arrangement with Hut 8 to secure the capacity and is itself an investor in Lambda.^[18]

The contractual details demonstrate two simultaneous realities. First, financing conditions are now directly connected to the physical availability of AI computing resources, so that a frontier model developer’s capacity roadmap depends on whether a neocloud can close a term loan, which in turn depends on whether a bitcoin miner turned datacenter developer can energize a site in South Texas. Second, customers and suppliers are negotiating provisions, including acceptance-based payment triggers, best-efforts financing covenants, and milestone-linked termination rights, that are specifically intended to distribute the risks created by these dependencies. This is why Counterparty Cascade should not be interpreted as evidence that the industry is operating without financial protections. Sophisticated participants frequently negotiate precisely because they recognize the risks. The remaining question, which the contracts cannot answer on their own, is whether the protections are sufficient when multiple dependencies experience stress at the same time.

1.6. The Invisible Counterparties: Utilities, Contractors, and Local Communities

The contractual anatomy does not end with the technology sector, and the parties who appear last in any account of it are frequently the parties with the least ability to protect themselves. A large datacenter requires physical infrastructure that cannot be purchased instantly. Electrical substations, high-voltage transmission connections, transformers, cooling systems, water infrastructure, telecommunications links, and access roads may all require substantial planning and investment with lead times that exceed the duration of many of the commercial contracts they are built to serve. Some of these investments are made by the datacenter developer. Others involve utility companies, equipment manufacturers, local governments, or specialized infrastructure contractors, and the financial consequences of an unsuccessful project can therefore extend beyond its direct investors to institutions that never signed a contract with the AI company at all.

A developer might abandon a facility after a utility has already incurred costs preparing to serve the expected load. A utility could discover that the revenue expected from a major customer will not materialize on schedule, leaving a substation and a transmission line whose cost must be recovered from someone. A supplier might possess specialized equipment manufactured for a particular project and unsuitable for any other. A local government might have approved infrastructure improvements or granted tax abatements based on employment and tax-revenue projections that assumed the facility would operate. The economics depend heavily on contractual cost allocation. A utility that requires substantial upfront financial security from a datacenter developer may be well protected; one that allows considerable infrastructure costs to enter its general rate base may discover that its residential customers have become involuntary creditors of a speculative technology project. A government that conditions tax incentives on verified milestones faces less exposure than one that provides broad benefits before meaningful construction begins.

Virginia’s experience illustrates the scale of what is at stake. According to the State Corporation Commission’s own analysis, Virginia residents had paid $2.8 billion for transmission lines related to datacenters since 2021, a figure that rises to $6.2 billion, more than half of Dominion Energy’s transmission projects, when projects indirectly tied to datacenter growth are included.^[19] Thus the contractual network extends into public policy, and the distinction between private investment and public exposure becomes increasingly important as the scale of AI infrastructure grows, a theme that Section 4 develops at length.

This is the first major conclusion of the contractual analysis, and it frames everything that follows: the Five-Layer AI Economy is not only an industrial supply chain. It is becoming a network of financial obligations in which the performance of one institution increasingly depends on the commitments, financing, and operational capabilities of others, including others it has never met.

Section 2: The Five-Layer AI Economy as a Financial Transmission System

2.1. The Five Layers Are Technologically Distinct but Financially Interdependent

My Five-Layer AI Economy framework identifies five foundational components of modern artificial intelligence, and before the financial argument can proceed it is worth restating them with some care, because the entire thesis of this paper rests on the claim that the boundaries between these layers, which are sharp in engineering terms, are becoming porous in financial terms.

Layer 1 — Energy: electricity generation, transmission, distribution, and the physical resources, including water and land, required to operate computing infrastructure at scale.
Layer 2 — Chips: semiconductor design, fabrication, advanced packaging, accelerators, high-bandwidth memory, networking components, and the manufacturing systems and supply chains that support them.
Layer 3 — Datacenters: physical computing campuses, servers, networking systems, cooling facilities, construction, and the operation of the resulting infrastructure.
Layer 4 — Models: frontier foundation models, specialized scientific models, reasoning systems, and the research organizations that develop them.
Layer 5 — Applications and Agentic Systems: the commercial products, enterprise platforms, autonomous agents, robotics systems, and other applications that translate intelligence into economic activity and, ultimately, into revenue.

The framework is useful for understanding technological development because each layer depends on resources and capabilities supplied by the others, and a bottleneck in any one of them constrains the rest. However, Counterparty Cascade introduces an additional interpretation that the technological reading obscures. Every layer generates contractual obligations. Electricity developers need financing for generation and transmission assets whose useful lives are measured in decades. Semiconductor companies need long-term manufacturing capacity and equipment investments that must be committed years before the chips they produce are sold. Datacenter developers need capital for construction and hardware deployment, and the hardware depreciates faster than the building that houses it. Model companies need sustained access to computing resources on terms that lock in costs for years. Application companies need reliable model access, affordable inference, and customers willing to purchase their products on subscription or consumption terms that can be canceled with a month’s notice. The financial architecture connecting these layers is not visible from the technological architecture alone, and it is frequently not visible from any single company’s financial statements either.

A single financial institution may provide capital to several independent infrastructure developers in Layer 3. A semiconductor supplier in Layer 2 may have commercial relationships, and increasingly equity relationships, with multiple competing cloud operators. A hyperscaler may invest in a model company in Layer 4 while simultaneously receiving infrastructure payments from it and competing with it in Layer 5. An enterprise customer may use the same frontier model indirectly through several different software providers, so that an apparent diversity of application-layer vendors conceals a concentration of model-layer dependence. This creates two types of interconnectedness that behave differently under stress. The first is vertical interconnectedness, where financial obligations connect different layers of the AI economy and a disruption travels up or down the stack. The second is horizontal interconnectedness, where several companies within the same layer share suppliers, lenders, customers, or other dependencies, so that a disruption travels sideways to competitors who had no relationship with the original participant.

An interruption can move vertically, horizontally, or in both directions at once. For example, a semiconductor shortage, or a disruption at an advanced-packaging facility, could initially affect Layer 2. The shortage delays hardware installations in Layer 3. The resulting computing constraints affect model development and inference capacity in Layer 4. Applications and autonomous agents in Layer 5 experience reduced capacity, higher prices, or delayed product releases. The lower application revenue then reduces the cash available to support model infrastructure commitments, and that financial pressure travels backward through the same network it just traveled forward through, arriving at the datacenter lender and the chip financier in the form of reduced utilization and weaker collateral. This creates the possibility of a self-reinforcing cycle, in which the layers that were supposed to insulate one another instead amplify a common shock.

2.2. The Difference Between a Supply Chain and a Counterparty Network

A conventional industrial supply chain describes how inputs become finished products. Raw materials enter manufacturing facilities; manufacturers produce components; suppliers deliver those components to assemblers; finished products reach customers. Financial obligations are present at every step, but the analysis focuses primarily on the movement of physical goods, and the standard tools of supply-chain risk management, such as dual sourcing and safety stock, are designed accordingly. AI infrastructure requires a broader model, because the obligations that connect its participants are no longer dominated by the delivery of goods.

A counterparty network contains several categories of relationships, each with its own legal and economic character, and a disruption may affect all of them at once:

physical supply relationships for chips, servers, transformers, and cooling equipment;
capacity reservation and service contracts with minimum spending commitments;
equity investments and strategic financing between customers and suppliers;
equipment leases and secured lending collateralized by hardware and by contracted cash flows;
guarantees, residual-value backstops, and other credit-support arrangements;
electricity and infrastructure service obligations with utilities and grid operators;
construction contracts with delivery milestones and liquidated damages;
revenue-sharing and distribution agreements between model developers and platforms;
regulatory permissions, tax incentives, and public infrastructure commitments.

These relationships differ in their legal and economic characteristics. An equity investor bears different risks from a secured lender. A utility’s obligation to provide service differs from a semiconductor manufacturer’s obligation to deliver equipment. A model developer’s minimum purchase commitment differs from a customer’s discretionary consumption of model tokens, and the asymmetry between the two, which is examined in Section 3.4, is one of the most important features of the system. Nevertheless, a disruption may affect all of them simultaneously.

Consider a hypothetical example involving five institutions, each of which behaves prudently by its own lights. A model developer signs a long-term agreement to purchase computing services from an infrastructure operator. The operator secures financing from an investment fund. The fund relies on a bank credit facility to support its lending activities. The operator purchases equipment from a semiconductor systems supplier. The equipment is installed in a facility dependent on a new electricity connection. If the electricity connection is delayed, the operator may be unable to deliver computing services. If service delivery is delayed, the model developer’s payment schedule changes, because payments were tied to acceptance. If expected payments are postponed, the operator needs additional liquidity to service debt that was drawn against the construction schedule. If the operator cannot obtain sufficient liquidity, its lender faces a deterioration in credit quality. If the lender experiences significant losses or funding pressure, it reduces financing to other infrastructure projects. The initial problem was an electricity connection. The eventual consequence is reduced availability of financing for unrelated datacenters, and the propagation occurs because financial relationships connect companies that do not share the same physical project. This is the defining characteristic of Counterparty Cascade, and Van Nieuwerburgh’s Brookings paper reaches the same conclusion from the perspective of corporate finance.

“These arrangements transform rather than eliminate risk.” — Stijn Van Nieuwerburgh, Columbia Business School ^[6]

2.3. A Proposed Mathematical Framework for Counterparty Cascade

A useful analytical framework should do more than identify potential problems. It should help researchers, investors, policymakers, and corporate executives evaluate the circumstances under which those problems become significant, and it should do so in a form that can eventually be populated with data. I propose representing the AI infrastructure economy as a network of institutions connected by different types of financial and operational obligations. Let each participating institution be represented by a node, and let the connections between institutions represent payment obligations, financing commitments, delivery requirements, guarantees, or other economically consequential relationships, each carrying a direction and a magnitude.

The initial impact of a disruption on institution i can be represented conceptually as:

D_i = P_i + R_i + L_i + O_i

where D_i represents the direct economic disruption experienced by institution i; P_i represents additional payment or financing requirements triggered by the disruption; R_i represents lost or delayed revenue; L_i represents liquidity pressure associated with unavailable financing or delayed cash receipts; and O_i represents incremental operating and replacement costs, including the premium paid for short-term substitute capacity. This is an analytical decomposition, not an established accounting identity, and the terms must be defined carefully to prevent double-counting; additional borrowing needed to cover a revenue shortfall, for example, should not be counted again as a separate economic loss when measuring the total.

The next question is how the disruption spreads from i to the institutions connected to it. One possible conceptual expression for the stress transmitted to institution j is:

C_j = Σ_i=1ⁿ w_ij · s_i · v_j

where C_j is the transmitted stress reaching institution j; w_ij measures the economic significance of the obligation connecting institution i to institution j, expressed relative to j’s balance sheet or cash flow; s_i represents the proportion of that obligation affected by the disruption at institution i; and v_j measures institution j’s vulnerability to the resulting disruption, which could incorporate liquidity reserves, contractual protections, financing alternatives, customer diversification, collateral quality, and access to replacement infrastructure. A second-round transmission would then treat each affected j as a new source, with s_j determined by the degree to which C_j exceeds j’s capacity to absorb it, and the cascade terminates when no node’s transmitted stress exceeds its absorption threshold.

This model is deliberately simplified, and its limitations should be stated as clearly as its structure. Actual transmission is nonlinear, because contractual protections such as grace periods and materiality thresholds mean that stress below a certain level transmits nothing and stress above it transmits abruptly. The same collateral may support several claims, so that exposures which appear independent in the matrix are in fact correlated through a common asset. Multiple institutions may react to a common market shock simultaneously, which the pairwise formulation does not capture. Liquidity problems may arise before accounting losses are recognized, so that the timing of cash flows matters as much as their ultimate allocation. A more advanced simulation would therefore use different exposure matrices for lending, equipment delivery, leases, electricity, and customer revenue, and it would distinguish the timing of cash flows from the ultimate allocation of economic losses. The purpose is not to claim that a single equation can forecast an AI crisis. The purpose is to make the network’s vulnerabilities measurable, and to make clear which data would need to be disclosed in order to measure them.

2.4. Five Proposed Indicators of Financial Interdependence

I propose five indicators that could form the foundation of a Counterparty Cascade research methodology, and that correspond to the terms of the framework above.

Indicator 1: Counterparty Concentration. This measures how dependent an institution is on a small number of customers, lenders, suppliers, or infrastructure providers, and it corresponds to the distribution of the w_ij weights across j’s connections. A datacenter operator with ten customers generating similar levels of revenue has a different risk profile from one whose revenue depends primarily on two frontier AI companies. Similarly, a model developer using several accelerator architectures possesses greater technical flexibility than one entirely dependent on a single hardware platform. However, contractual diversification must be distinguished from genuine economic independence, because several different cloud providers may rely on the same semiconductor foundry, the same memory supplier, the same transmission system, or the same credit institution. The relevant measure is therefore not simply the number of counterparties but the degree to which those counterparties share underlying vulnerabilities, a point to which Section 2.6 returns.

Indicator 2: Payment–Delivery Mismatch. This measures the extent to which cash obligations become payable before infrastructure is ready to generate productive revenue, and it is the principal determinant of L_i. Consider a hypothetical datacenter whose equipment lease payments begin six months before its electricity connection is operational. Even if the contract is fully enforceable and the developer ultimately completes the facility, the project may experience substantial additional financing costs during the gap. The larger the mismatch between cash outflows and productive capacity, the greater the requirement for liquidity reserves, and this indicator is particularly important for dedicated infrastructure arrangements in which equipment and facilities are financed by different parties with different payment triggers.

Indicator 3: Financing Conditionality. This measures the proportion of planned infrastructure capacity that depends on additional financing which has not yet been secured or which remains subject to substantial conditions. A committed construction loan is not equivalent to a preliminary expression of lender interest. A financing platform’s announced capital-raising objective is not equivalent to funded project-level financing. Similarly, a signed customer contract may be insufficient to secure funding if the lender requires additional equity, electricity agreements, regulatory approvals, or equipment guarantees before closing. Nscale’s disclosure that it had not yet secured binding financing for its $44.6 billion of Anthropic agreements is a direct, publicly observable reading of this indicator at the project level.^[5]

Indicator 4: Replacement Capacity. This measures the ability of an institution to replace a failed or delayed counterparty within a commercially acceptable period, and it is the principal determinant of v_j. For a model developer, replacement capacity might involve alternative cloud providers, accelerator architectures, or geographic locations. For a datacenter developer, it could involve alternative equipment suppliers, lenders, or electricity arrangements. Replacement capacity is not theoretical availability. A competing provider may have sufficient computing equipment in aggregate but lack capacity in the required region, with the required architecture, at an acceptable price, and within the necessary schedule, and in a market where Amazon’s chief executive described cloud capacity as largely sold out for the following year, aggregate availability may be close to zero regardless of what any individual contract says.^[13] True replacement capacity must satisfy operational constraints.

Indicator 5: Cross-Default Sensitivity. This measures the potential for a problem under one agreement to activate remedies under other agreements, and it governs the nonlinearity of transmission. Certain debt instruments and commercial contracts contain provisions linking defaults across separate obligations. Such provisions can be important protections for lenders, because they prevent a borrower from selectively neglecting one major obligation while continuing to satisfy others. However, they can also accelerate financial stress. Anthropic’s amended filing explicitly identifies circumstances in which defaults under certain commercial arrangements could activate cross-default provisions in its credit and convertible-note agreements, and the exhibits to the original filing define material indebtedness by reference to a revolving credit facility syndicated among Morgan Stanley, Barclays, Citibank, Goldman Sachs, JPMorgan Chase, Royal Bank of Canada, and MUFG.^[1] A relatively localized contractual problem can therefore become a company-wide liquidity issue under specified conditions, and an important research task is to determine which obligations are interconnected by such legal provisions, information that is rarely available outside the four corners of the agreements themselves.

2.5. The Proposed Counterparty Cascade Exposure Matrix

The five layers can be evaluated using a common set of questions, and the result is a matrix that identifies, for each layer, the principal financial obligations, the disruptions most likely to initiate a cascade there, and the downstream consequences that would follow.

AI economy layer	Principal financial obligations	Potential initiating disruption	Downstream consequences
Energy	Power-purchase agreements, generation financing, transmission construction, utility tariffs	Delayed generation or interconnection; adverse cost-allocation ruling	Datacenter delays, idle equipment leases, unexpected utility and ratepayer costs
Chips	Equipment orders, fabrication capacity, component supply, equipment leasing, residual-value guarantees	Manufacturing or packaging shortage; equipment repricing by a new generation	Deferred installations, increased capital requirements, collateral reassessment
Datacenters	Construction loans, delayed-draw facilities, leases, hosting commitments, secured and high-yield debt	Financing withdrawal or delivery failure; tenant default	Lost revenue, customer replacement costs, lender portfolio repricing
Models	Minimum compute spending, equipment lease guarantees, strategic financing, convertible notes	Revenue shortfall or changed compute requirements	Contract renegotiation, reduced research spending, cross-default pressure
Applications & Agents	Usage contracts, software subscriptions, inference procurement	Lower adoption or pricing pressure; agent reliability shortfall	Reduced demand for upstream computing capacity, weaker support for every layer below

This matrix identifies the point at which technological dependence becomes financial exposure. A model company’s inability to obtain sufficient accelerators is initially an operational problem; it becomes a financial transmission event when the delay changes payment obligations, financing requirements, or expected cash flows for other institutions. Similarly, a reduction in demand for enterprise agents is initially a commercial problem; it becomes a potential Counterparty Cascade when reduced application spending undermines the cash flows supporting infrastructure contracts and the debt written against them. The critical analytical distinction, which recurs throughout this paper, is between an isolated operational difficulty and a disruption capable of changing the financial behavior of multiple organizations.

2.6. The Paradox of Diversification

Diversification is commonly associated with lower risk, and the instinct to diversify is visible throughout the AI economy. A model developer seeks several cloud providers to avoid depending on a single supplier, as Anthropic’s three hyperscaler agreements and its neocloud contracts demonstrate. A lender finances several datacenter projects to avoid excessive exposure to one facility. A semiconductor company sells accelerators to numerous customers rather than concentrating revenue in one institution. Yet diversification can be less protective than it appears, and in the AI economy it may be systematically overstated.

Suppose three infrastructure operators serve different customers and build facilities in different states. Each operator obtains financing from separate investment funds. At first glance, the projects appear independent, and a lender holding exposure to all three would report a diversified book. However, the investment funds may share major institutional investors and may draw on the same bank credit lines. The operators may purchase accelerators from the same supplier, who may also be an investor in each of them. Their customers may compete in the same frontier model market and depend on the same enterprise adoption trends. Their facilities may depend on the same transformer manufacturers with multiyear lead times and on the same regional transmission operator. An unexpected deterioration in AI revenue expectations could therefore affect all three projects simultaneously, and the lender’s apparent diversification would offer no protection at all.

The phenomenon is described in financial economics as correlated exposure, and it was the central lesson of the 2008 crisis, in which mortgage pools that were diversified by geography proved to be undiversified with respect to the national decline in house prices. The Counterparty Cascade framework adds another dimension that correlation analysis does not capture: an institution may become vulnerable not only because it shares an economic risk factor with others, but because its financial obligations depend directly on their continued performance. The distinction between correlation and contractual contagion is essential. Two companies suffering losses during the same market downturn do not automatically demonstrate a cascade. A cascade requires a credible transmission mechanism connecting one institution’s distress or nonperformance to consequences experienced elsewhere, and that mechanism may be a debt covenant, a guarantee, a loss of customer revenue, a collateral requirement, or the withdrawal of a shared financing source. Without identifying the mechanism, a claim of systemic contagion remains speculative; with it, the claim becomes testable.

The BIS measurement that more than half of the investment value flowing into AI firms between 2021 and 2025 came from other AI firms is, in this light, a measurement of how much of the apparent diversity of capital sources in the industry is in fact a single source viewed from different angles.^[11]

2.7. The Fifth Layer as the Ultimate Financial Foundation

The most consequential insight of the five-layer financial reading may concern the final layer of the AI economy, which is the layer that receives the least attention from infrastructure investors and that ultimately determines whether their investments are recovered. Energy, chips, datacenters, and models all involve substantial investment. But the long-term ability to support those investments ultimately depends on economic demand for useful intelligence, and that demand is expressed, measured, and paid for in Layer 5. The fifth layer, Applications and Agentic Systems, is therefore more than the final destination of technological innovation. It is the essential source of the revenue that supports financial commitments throughout the other four layers, and every dollar of debt written against a GPU in Layer 2 or a building in Layer 3 is, in the end, a claim on a Layer 5 customer’s willingness to keep paying.

Consider how the economic sequence operates. An enterprise adopts AI agents to perform customer service, engineering, accounting, software development, or scientific analysis. The enterprise pays for application subscriptions or usage. The application provider purchases model inference. The model provider purchases computing services under the minimum commitments described in Section 1. The infrastructure operator uses that revenue to cover capital costs, debt service, and operations. The semiconductor supplier receives payment for equipment, and the lessor receives rent. The utility receives revenue for electricity services. This description is simplified, and some firms are vertically integrated across several of these steps, but it captures an economic relationship that no amount of integration can eliminate. Ultimately, financial returns must be supported by the value customers receive from AI-enabled products and services. Large investment commitments can anticipate future revenue. They cannot permanently substitute for it.

The scale of the revenue that the fifth layer must eventually generate is now the subject of serious academic estimation. Van Nieuwerburgh calculates that the 183 gigawatts of U.S. capacity his central scenario completes by 2032 would require approximately $3.7 trillion of mature annual revenue to earn a 10 percent unlevered return at a 50 percent operating cash-flow margin, which is equivalent to about 9.2 percent of projected 2032 GDP, and he notes that if the IT equipment has a three-year rather than a six-year economic life the requirement rises to roughly $6.0 trillion; against combined OpenAI and Anthropic revenues of around $100 billion at the time of writing, he observes that revenues would need to grow at roughly 80 percent per year until 2032 to reach the central figure.^[6] Anthropic’s own filing reports approximately $33.2 billion of revenue in the nine months ended September 30, 2026, compared with approximately $2.5 billion in the corresponding period of 2025, and third-quarter revenue of approximately $17 billion, which is growth of a kind that makes the 80 percent figure look less fantastical than it would in any other industry, but which also makes the question of durability more urgent rather than less.^[2]

This explains why the rapid commercialization of autonomous agents matters to financial stability and not only to venture investors. An enterprise agent that produces verifiable productivity improvements generates recurring economic demand that can support decades of infrastructure obligations. A model that achieves impressive benchmarks without corresponding customer willingness to pay may not generate sufficient cash flows to justify every infrastructure investment made in anticipation of its success. The economists who study productivity are divided on which outcome is more likely, and their disagreement is itself informative. Professor Erik Brynjolfsson of Stanford’s Digital Economy Lab argued in a February 2026 Financial Times essay that the productivity J-curve had turned, citing his own estimate that U.S. productivity grew roughly 2.7 percent in 2025, nearly double the prior decade’s average.^[20]

“The AI productivity take-off is finally visible.” — Erik Brynjolfsson, Stanford University ^[20]

Professor Daron Acemoglu of MIT, the 2024 Nobel laureate, has been considerably more cautious about the relationship between the scale of investment and the scale of near-term returns, and his warning, first delivered in 2024 and repeated in interviews through 2026, is the one that infrastructure lenders should keep in view.

“A lot of money is going to get wasted.” — Daron Acemoglu, Massachusetts Institute of Technology ^[21]

The challenge for 2027 through 2030 will be converting technical capability into durable revenue at a scale sufficient to support the rapidly expanding infrastructure base, and the fifth layer may ultimately determine whether the financial promises supporting the first four layers are sustainable. The sections that follow examine what happens when that conversion falls short, even temporarily, and even in a single place.

Section 3: The Transmission Mechanisms: How Counterparty Cascades Begin

3.1. A Cascade Does Not Require the Failure of a Major AI Company

Public discussion of AI financial risk frequently concentrates on the possibility of a dramatic event. A frontier model company could collapse. A hyperscaler’s investment strategy could fail. An AI-related asset bubble could burst, and the IMF’s October 2026 analysis, which placed an AI equity correction among its three salient risks and warned that given sizable cross-border holdings the losses could spread around the world, is framed largely in those terms.^[7] ^[22] Such scenarios attract attention because the consequences appear potentially enormous. However, the initial disruption in a Counterparty Cascade need not originate at a prominent company, and the most instructive historical cascades rarely did.

It could begin with a relatively small infrastructure developer whose project has unusually important contractual dependencies. It could originate with a private lender that suddenly withdraws from project financing because of redemption pressure in an unrelated fund. It might begin with a delayed transmission connection, the financial deterioration of a construction contractor, or the loss of a specialized equipment supplier whose transformers cannot be sourced elsewhere within three years. The initial event could be modest compared with the size of the broader industry. Its significance would depend on where it occurs within the contractual network, because an obscure institution can be systemically important to a particular project if no practical replacement exists. The distinction is between size and centrality. A company may possess a relatively small balance sheet while controlling a service or contractual relationship essential to a much larger investment, and the neocloud filings examined in Section 1.5, in which companies with a few billion dollars of annual revenue hold contracts worth tens of billions, are the clearest contemporary illustration of a small node with high centrality. This suggests that AI infrastructure risk assessment must identify critical dependencies rather than concentrating exclusively on the largest corporations.

3.2. Transmission Mechanism One: Financing Withdrawal

The first major mechanism involves the loss of expected financing, and it is the mechanism that the Federal Reserve and the Financial Stability Board have been watching most closely during 2026. A developer receives a long-term commitment from a computing customer. The developer uses that commitment to seek construction financing. Lenders evaluate projected revenue, the customer’s creditworthiness, construction costs, equipment values, collateral protections, and the likelihood of timely completion, and under favorable conditions financing becomes available. However, the developer may encounter problems before or during construction that have nothing to do with the project itself. Interest rates could rise. Construction costs could exceed expectations. The customer could experience a deterioration in credit quality. A lender might reduce exposure to the technology sector in response to losses elsewhere. A regulatory change could increase project costs or delay permits. An investor might become concerned that the proposed infrastructure will be technologically outdated before it generates sufficient cash flow. These events can affect financing availability even when the customer agreement remains valid and the project remains sound.

The resulting consequences can unfold sequentially. The developer cannot obtain a required financing tranche. Construction slows or stops. Equipment installation is delayed. The promised computing capacity becomes unavailable on schedule. The customer must revise its infrastructure procurement strategy, typically by paying more for short-term capacity. Suppliers and contractors face delayed payments or canceled orders. Lenders reassess the viability of similar projects. The initial financing interruption has now produced consequences across several organizations, and this is a basic Counterparty Cascade. The outcome depends on the availability of replacement capital and on the legal rights of the affected parties. A financially strong developer may replace the lender. A project with substantial sponsor equity may continue operating. A customer may possess adequate alternative computing capacity. Consequently, not every financing withdrawal produces a significant cascade, and the risk becomes greater when multiple projects depend on a limited group of financing institutions, specialized infrastructure developers, or strategically important customers.

The private credit market is the place where this mechanism is most likely to originate, because it has become the largest single source of debt for the buildout and because its funding is the least stable. Morgan Stanley’s estimate, cited by Van Nieuwerburgh, is that of the roughly $2.9 trillion required to meet hyperscalers’ incremental compute needs over 2025 through 2028, more than half will come from outside capital, and that within the debt component private credit accounts for the largest share at about $800 billion.^[6] The Federal Reserve’s May 2026 Financial Stability Report, published on May 8, discussed redemption pressures in private credit funds, noting that outflows from certain semiliquid vehicles had moderately exceeded inflows in the first quarter of 2026 after several of the largest non-traded vehicles restricted investor withdrawals, and it concluded that stability risks from further redemption requests appeared limited and manageable while flagging the growing connections between nonbank lenders and banks.^[23] The Financial Stability Board’s May 6 report on vulnerabilities in private credit sized the asset class at $1.5 to $2.0 trillion and emphasized opacity, leverage, and bank interlinkages.^[24] Governor Michael Barr of the Federal Reserve offered a more vivid characterization of the transmission channel a few days earlier.

Stress in private credit could spark “psychological contagion” leading to a broader credit crunch. — Michael Barr, Governor, Federal Reserve Board ^[25]

Professor Aswath Damodaran of New York University’s Stern School of Business, in his 2026 data update on debt, made the connection to artificial intelligence explicit, arguing that in an all-equity financed AI boom the pain of a correction would fall on shareholders with limited ripple effects, but that because debt has entered AI investment, a correction that marks private credit down to reflect higher default risk would spread pain to the rest of the economy, producing in effect a banking crisis created primarily by nonbank lenders.^[26]

“A shakeout is overdue, which will separate the sloppy lenders from the good ones.” — Aswath Damodaran, New York University ^[26]

3.3. Transmission Mechanism Two: Payment–Delivery Asymmetry

The second mechanism concerns the timing of contractual payments, and it is the mechanism most specific to the current generation of AI infrastructure contracts. Anthropic’s filing describes an especially important possibility: under certain multiparty infrastructure arrangements, a participant may become entitled to payment after performing its own discrete obligation even when failures elsewhere prevent the complete computing system from becoming usable.^[1]

Consider a hypothetical project structured along the lines the filing describes. The semiconductor equipment is delivered. The equipment lessor has fulfilled the contractual requirements that activate the relevant payment obligation, because the equipment is, in the language of the lease, ready for service. However, the datacenter’s electrical substation has not been completed, or the transmission interconnection has been delayed by a regional grid operator’s study queue. The accelerators cannot operate at their intended capacity. The model developer must nevertheless meet the payment obligations that become due under the applicable agreements, because the lessor performed and the lease does not condition rent on the performance of a third party the lessor does not control. The company experiences a cash outflow without the corresponding computing service becoming available for productive use, and if it has guaranteed the lease obligations of a special-purpose vehicle, as the filing describes, the outflow may be larger than the rent itself.

The situation becomes particularly difficult if the developer has already anticipated revenue growth based on the delayed capacity, because it may need to finance the original obligation while simultaneously procuring replacement computing resources at the short-term premium the filing itself quantifies.^[3] This is a form of timing risk that becomes a liquidity problem, and the relevant economic question is not simply whether a project will eventually become operational. It is whether each participant possesses sufficient financial capacity to survive the period before the complete infrastructure begins generating revenue. The difference between a six-month delay and a two-year delay can be financially decisive, because long-term infrastructure investments assume that construction, commissioning, and customer utilization occur within a defined window, and if those schedules diverge, additional interest expense and operating costs accumulate while the original economics deteriorate even though ultimate completion remains technically feasible.

The hyperscalers face a version of the same asymmetry at a different point in the chain. Microsoft disclosed in its fiscal 2026 annual report datacenter leases of $329.1 billion that had not yet commenced, with lease terms of between one and twenty years beginning between fiscal 2027 and fiscal 2033, and on its October 28, 2026 earnings call it attributed a shortfall in first-quarter capital spending relative to guidance to delays in third-party capacity deliveries that it expected to slip into the second half of the fiscal year.^[27] A lease that has not commenced generates no rent, but it also generates no capacity, and a hyperscaler that has sold that capacity forward to a frontier lab under a minimum-commitment contract is carrying the delivery risk on behalf of both its landlord and its customer.

3.4. Transmission Mechanism Three: The Customer Revenue Shortfall

The third mechanism begins at the model or application layer and travels downward. An AI company may secure long-term computing commitments because it expects rapid growth in customer spending, and the expectation may be entirely reasonable. However, not all revenue possesses the same level of contractual durability, and the asymmetry between the duration of the commitments a model company makes and the duration of the commitments its customers make to it is one of the structural facts of the industry.

Anthropic’s filing reports that consumption-based revenue accounted for approximately 79 percent of revenue during the first nine months of 2026, and approximately 80 percent in the first half.^[1] ^[12] These figures demonstrate very rapid commercial expansion, and they demonstrate the kind of customer traction that justifies long-term infrastructure commitments. They also illustrate the importance of distinguishing revenue already earned from revenue expected in future periods. Consumption-based customers may increase spending as their AI usage grows, and in 2026 they have done so at an extraordinary rate. However, some may also reduce consumption, redirect workloads to a competitor, or deploy smaller models internally. Certain customers have minimum contractual commitments, but consumption-based revenue as a category is not equivalent to fully guaranteed multiyear revenue, and the contrast with the ten-year, minimum-commitment, usage-independent structure of the company’s own cloud obligations is the whole point. The developer has bought capacity on terms that resemble a mortgage and sells intelligence on terms that resemble a utility bill.

Suppose an AI developer anticipates $100 billion in annual revenue by a future date and signs computing contracts based partly on that projection. If actual revenue reaches only $80 billion, the company may still be required to meet substantial infrastructure payments. The problem is not that $80 billion represents an economically insignificant business; it would be among the fastest-growing businesses in history. The problem is that financial obligations may have been calibrated to a substantially higher revenue expectation, and the resulting pressure could cause the company to reduce discretionary spending, renegotiate contracts, delay new infrastructure commitments, or seek additional financing. Those decisions affect counterparties. The more widely a model developer’s expected revenue supports other institutions’ financial plans, and Anthropic’s expected revenue now supports the plans of three hyperscalers, at least two neoclouds, a semiconductor lessor, and their respective lenders, the greater the potential transmission. The IMF’s October 2026 World Economic Outlook stated the macroeconomic version of this mechanism in a single sentence.

“A reassessment of AI-related profitability and productivity prospects could sharply reduce investment.” — International Monetary Fund, World Economic Outlook, October 2026 ^[28]

This mechanism is particularly important because AI adoption could remain strong while some investment commitments nevertheless prove excessive. An industry can grow rapidly and still experience financial distress when capital expenditure grows faster than sustainable cash generation, and the third quarter of 2026 supplied the first visible evidence of investor discomfort with that gap at the very top of the capital structure. Alphabet reported third-quarter capital expenditures of $51.4 billion, more than double the prior year, free cash flow of negative $9.1 billion, and bond and equity issuance of approximately $28.7 billion and $12.4 billion respectively, while maintaining full-year guidance of $195 billion to $205 billion and signaling a significant further increase in 2027; its shares fell roughly 8 to 10 percent after hours.^[14] Amazon reported third-quarter capital expenditures of $77.3 billion, nearly $20 billion above analyst expectations, trailing twelve-month free cash flow of negative $42.4 billion, and a commitment to spend meaningfully more than its roughly $220 billion 2026 plan in 2027; its shares fell about 7 percent, even as AWS grew 39 percent and disclosed a contracted backlog of $578 billion.^[13] Meta narrowed its 2026 capital expenditure range to $135 billion to $145 billion, reported capital spending that exceeded operating cash flow by $2.1 billion in the quarter, and announced an equity financing program of up to $95 billion, including a proposed $60 billion share-sale facility, to fund its infrastructure without diluting its founder’s voting control.^[29] Microsoft, which alone among the four reported a positive investor reaction, guided to approximately $170 billion of calendar-2026 capital expenditure and committed to remaining free-cash-flow positive through fiscal 2027.^[27] These are not the results of companies in distress. They are the results of companies whose shareholders have begun to ask, loudly, when the fifth layer will pay for the first three.

3.5. Transmission Mechanism Four: Credit Covenants and Cross-Defaults

The fourth mechanism involves contractual provisions that connect otherwise separate obligations, and it is the mechanism that converts a linear transmission into a nonlinear one. A cross-default clause allows a default under one agreement to create consequences under another. Such provisions vary substantially by instrument, borrower, and negotiated terms, and they are not automatically activated by an ordinary project delay or a temporary decline in revenue. Nevertheless, they can have important consequences when a qualifying event occurs, and Anthropic’s registration statement specifically discusses potential cross-default consequences involving certain commercial arrangements, its credit agreement, and its convertible promissory notes.^[1]

Consider a hypothetical model developer experiencing significant financial pressure. The developer fails to satisfy a material obligation under a particular equipment lease. That failure constitutes an event of default under the agreement. A related credit facility contains a cross-default provision triggered by the event, and the lender becomes entitled to demand accelerated repayment, subject to the contractual conditions. The demand increases the company’s immediate liquidity requirements. Other creditors reassess their exposures. Additional financing becomes more expensive. Commercial counterparties become concerned about the company’s ability to fulfill future obligations and begin to require prepayment or security. The resulting pressure can intensify even before a formal insolvency proceeding occurs, and this is where contractual risk becomes self-reinforcing: a liquidity problem activates contractual remedies, those remedies increase liquidity requirements, and the additional pressure makes it harder to satisfy other obligations. The system’s behavior becomes nonlinear, and the original economic loss may be relatively small compared with the financial obligations affected by the resulting default provisions.

However, this outcome is not inevitable, and a rigorous analysis must say so. Grace periods, materiality thresholds, negotiated amendments, waivers, contractual remedies, available liquidity, and creditor incentives may interrupt the sequence at any point. Creditors have economic reasons to avoid unnecessarily destroying the value of otherwise viable projects, and a lender who accelerates a loan against a half-built datacenter owns a half-built datacenter. A rigorous Counterparty Cascade analysis must therefore evaluate legal triggers and realistic restructuring options rather than assuming automatic failure, and it must recognize that the same covenants which can accelerate a crisis also give creditors the standing to negotiate one away.

3.6. Transmission Mechanism Five: Equipment Value and Collateral Reassessment

The fifth mechanism involves the value of the equipment supporting financial arrangements, and it is the mechanism most directly exposed to the pace of technological change. AI accelerators are technologically sophisticated and expensive, and they operate in an environment characterized by rapid improvement in which new architectures increase computational throughput, memory efficiency, and energy productivity with each generation. NVIDIA’s successive accelerator generations illustrate the pace, and CoreWeave’s announcement that it had completed the industry’s first bring-up of the Vera Rubin NVL72 system in the same quarter in which much of its installed base consisted of earlier generations illustrates how quickly the frontier moves within a single operator’s fleet.^[17]

This does not mean that previously deployed accelerators immediately become obsolete. Different workloads have different performance requirements, and a processor that is no longer optimal for training the largest frontier model may still provide substantial value for inference, fine-tuning, enterprise applications, or scientific computing. Nevertheless, the financing value of hardware depends on expected future cash flows rather than on physical function. Consider an equipment lease supported partly by the anticipated revenue from a particular generation of GPUs. If the market rental price of that computing capacity declines faster than expected, the economics supporting the lease weaken even though the hardware remains fully operational, and the cash flow available to service the associated financing obligations is lower than projected. The IMF’s scenario analysis, which found that shortening the assumed asset life from seven years to three reduces modeled operating margins by more than nine percentage points, and Van Nieuwerburgh’s finding that the same change raises the industry’s required mature revenue from $3.7 trillion to $6.0 trillion, are two measurements of the same sensitivity.^[9] ^[6]

This issue overlaps with the technological depreciation concerns examined in my earlier paper Collateral Half-Life, but Counterparty Cascade investigates the next stage of the process. The earlier question concerned the declining economic value of hardware collateral. The present question concerns what happens when that change affects lenders, lessors, customers, and related infrastructure projects. A reduction in equipment value may cause a financing institution to demand greater protection, depending on the agreement. The customer may experience higher financing costs. The developer may be required to contribute additional equity. A lender may reduce the amount it is willing to finance against similar equipment across its entire book, and a guarantor of residual value, which under the August 2026 NVIDIA platforms is the equipment vendor itself, may face calls on its guarantee precisely when its own product cycle has caused the decline.^[6] Multiple projects may therefore experience financing pressure even when their equipment remains technically functional, and the essential distinction is between the technological life of an asset and the financial assumptions embedded in the contracts supporting its acquisition.

3.7. Transmission Mechanism Six: Shared Financing Sources

The sixth mechanism can affect projects with no direct commercial relationship, and it operates through the supply of credit rather than through any individual contract. Private credit and other nonbank financing institutions have become increasingly important to the broader economy, and the Federal Reserve’s May 2026 report and the FSB’s May 2026 review both identified the growing connections between nonbank lenders and banks as a concern distinct from the health of any individual fund.^[23] ^[24] The relevance to AI infrastructure lies in the possibility of shared funding sources. Consider three otherwise unrelated datacenter developments, one in Texas, one in Ohio, and one in Virginia. Their customers, operators, and utility providers are different. However, their construction loans could ultimately depend on financing supplied by institutions exposed to the same investors or funding markets, and the six asset managers named in NVIDIA’s August announcement are, as the trade press immediately observed, already shareholders of or lenders to the hyperscalers and AI clouds purchasing NVIDIA hardware, and often hold exposure to the AI labs those clouds serve.^[15]

If losses in one project produce a wider reassessment of risk, financing conditions for the other projects change even though nothing about them has changed. Lenders increase interest margins. They reduce loan-to-value limits. They require more sponsor equity. They shorten available financing commitments or strengthen contractual covenants. Even a financially healthy developer discovers that its next project requires considerably more capital, and projects at the margin of viability are canceled. This is financial transmission through the supply of credit. It differs from a direct customer default, and it may arise from changes in risk appetite, funding costs, portfolio concentration limits, or institutional liquidity rather than from any identifiable loss. Bank of America Global Research estimated that bond issuance tied to AI infrastructure, covering large technology companies, datacenter projects, and chip-financing vehicles, reached $344 billion by early August 2026, an increase of more than $200 billion over 2025, and IMF Managing Director Kristalina Georgieva observed in October that the share of investment-grade issuance accounted for by large technology companies financing AI infrastructure had risen more than tenfold in two years.^[30] ^[22] A market that has grown that quickly has, by construction, a shallow base of investors who have ever seen it under stress.

It is also important not to assume that all private credit institutions behave similarly. Large private credit managers possess different strategies, capital structures, investor commitments, asset exposures, and redemption arrangements. Some have long-duration funding well matched to their assets. Others may encounter more significant liquidity constraints, as the first-quarter 2026 redemption episodes demonstrated. The Counterparty Cascade framework therefore requires institution-specific analysis rather than treating private credit as a single homogeneous market, and it requires the data to perform that analysis, which the private credit market does not currently provide.

3.8. Transmission Mechanism Seven: The Strategic Partner’s Dual Role

A particularly unusual mechanism arises when a major technology company serves multiple financial roles within the same ecosystem, and it is the mechanism that most distinguishes the AI economy from earlier infrastructure booms. A hyperscaler invests in an AI developer. The AI developer commits to purchasing cloud services from the hyperscaler. The hyperscaler uses part of that anticipated demand to justify infrastructure expansion to its own shareholders. The infrastructure expansion creates additional purchases from semiconductor suppliers. The semiconductor supplier may also possess an investment or commercial relationship with the AI developer, and may be guaranteeing the residual value of the hardware the hyperscaler buys. These arrangements create genuine economic benefits. They align incentives, improve access to scarce resources, and accelerate infrastructure investment in a way that arm’s-length contracting could not. But the same financial relationships transmit stress in several directions at once.

Suppose the AI developer experiences slower commercial growth. Its expected ability to purchase computing resources declines. The hyperscaler experiences reduced expectations for future infrastructure revenue, and the value of its strategic investment declines at the same time. If the hyperscaler responds by reducing infrastructure spending, semiconductor suppliers experience lower expected sales, and the equity stakes they hold in the same developer decline too. An organization that appeared diversified across investment and commercial relationships discovers that several exposures are affected by the same underlying business. This is the wrong-way risk the IMF identified, and the feedback works at two levels simultaneously: the firm financing the company that buys its products suffers a loss of sales revenue and at the same time a decline in the value of its equity and credit investment in that company.^[7] Oracle, which is the clearest public example of a company whose fortunes have become tied to a single AI customer, saw the cost of insuring its debt in the credit default swap market rise above its 2008 peak during 2026 as investors priced the possibility that datacenter delays, customer financing shortfalls, rapid chip depreciation, or rising refinancing costs could leave it holding the debt long before the cash arrives.^[31]

Counterparty Cascade extends the IMF’s analysis to the operational obligations embedded in AI infrastructure construction and delivery. The key issue is the possibility that equity investments, credit support, customer contracts, and equipment financing respond simultaneously to the deterioration of one important company, and that the institutions on the other side of each of those relationships are, to a degree none of them individually appreciates, the same institutions.

3.9. A Hypothetical $10 Billion Datacenter: Three Stages of Financial Stress

To understand how these mechanisms interact, consider a simplified infrastructure investment. The figures below are hypothetical, and they are not projections of any identifiable company or actual datacenter, although the capital structure is broadly consistent with the 90 percent project-level leverage that Van Nieuwerburgh documents for Meta’s Hyperion financing.^[6]

Project component	Illustrative amount
Total development cost	$10.0 billion
Secured debt financing	$6.0 billion
Sponsor equity	$2.0 billion
Equipment financing and other capital	$2.0 billion
Planned commercial operations	January 2028
Illustrative annual cash interest rate on secured debt	9%

Assume the project’s revenue depends primarily on long-term computing service commitments from a single frontier model customer, and that the equipment financing is structured so that rent begins when the hardware is ready for service.

Stage One: A Six-Month Electricity Delay. The power connection is delayed by six months because the regional grid operator’s large-load study takes longer than the developer’s consultants projected. If the entire $6 billion secured debt balance were outstanding for that period, the simple illustrative interest cost would be:

$6 billion × 9% × (6 ÷ 12) = $270 million

Actual interest costs would depend on draw schedules, capitalized interest, contractual terms, and other factors, and the 9 percent assumption is roughly in line with the coupons on CoreWeave’s 2031 and 2032 senior notes.^[16] Nevertheless, the calculation illustrates the scale of financing costs associated with a relatively short delay. The project also incurs additional construction expenses, staffing costs, insurance premiums, and equipment storage charges, and if the equipment lease has commenced because the hardware is ready, the developer is paying rent on accelerators it cannot energize. The developer must determine whether existing contingency funds are sufficient.

Stage Two: A Twelve-Month Delay and Financing Reassessment. Suppose the delay extends to twelve months. At the assumed interest rate, a fully drawn $6 billion balance incurs $540 million in annual interest before considering other costs, which is more than a quarter of the sponsor’s entire equity. During the delay, the lender also becomes concerned that newer accelerators announced in the interim will affect the future market value of the planned computing equipment. The lender reduces the amount it is willing to advance against future assets. The developer now requires additional equity, and if existing shareholders cannot provide it, the project needs a new investor or a contractual restructuring. The customer, meanwhile, has exercised its milestone-based termination right on the delayed tranche and procured replacement capacity elsewhere, so that the revenue the lender underwrote has partially disappeared. The original physical delay has become a financing problem.

Stage Three: A Portfolio-Wide Credit Response. Suppose several other developments financed by the same lending group encounter similar problems during the same period, because they depend on the same grid operator, the same transformer supplier, or the same customer. The lender reassesses its infrastructure portfolio. It increases required equity contributions and financing margins for future projects. Developers that have not experienced delays now face higher capital costs, their expected returns decline, and some projects are postponed. Others seek new customers, alternative financing, or lower-cost configurations. Equipment orders are rescheduled. The semiconductor and construction supply chains begin responding to weaker investment demand, and the utilities that had planned generation for the postponed projects revise their forecasts. The initial electrical delay at one project has contributed to a wider adjustment in financing conditions, and the precise magnitude depends on the size of the lender’s exposures, the liquidity available to borrowers, and the degree of correlation across projects. The scenario is illustrative, not a prediction. Its significance is conceptual: the transmission of financial stress can extend well beyond the project where the original disruption occurred, and it does so through channels that no participant in the original project controls.

3.10. Why Not Every Counterparty Disruption Becomes a Cascade

A credible theory must also explain why many disruptions do not produce wider financial instability, because the overwhelming majority of delayed projects in every infrastructure boom have been absorbed without incident. Several protections can interrupt transmission at each stage of the sequence above. A developer may maintain sufficient cash reserves to continue construction despite temporary financing delays. A customer may possess contractual rights to terminate an undelivered tranche of capacity, and may choose not to exercise them because it needs the capacity more than it needs the remedy. A lender may agree to restructure payment schedules rather than realize a loss. A sponsor may provide additional equity. A semiconductor supplier may redirect equipment to another customer in a market where demand exceeds supply. A cloud provider may substitute capacity from a different region. A utility may have required sufficient collateral to prevent stranded infrastructure costs from being transferred to other customers.

These mechanisms matter because financial interconnectedness is not automatically harmful. Contracts distribute risk efficiently when they are well designed. Diversified financing reduces dependence on individual institutions when the diversification is real. Specialized operators deliver infrastructure that would otherwise be unavailable. Large technology companies provide financial support that stabilizes smaller participants, and the hyperscalers’ balance sheets, which remain investment-grade with hundreds of billions of dollars of cash, are the principal reason the IMF concluded that immediate systemic risks remained manageable.^[9] The vulnerability arises when several protections fail or become ineffective simultaneously, and for this reason I propose distinguishing three levels of Counterparty Cascade.

Level One — Localized Disruption. A single institution experiences financial or operational stress, but contractual protections and alternative resources prevent significant transmission. The disruption is absorbed by the sponsor’s equity, the lender’s reserves, or the customer’s flexibility.

Level Two — Network Transmission. The disruption affects multiple directly connected institutions through payment delays, lost revenue, financing requirements, contractual remedies, or replacement costs. Several parties record losses or incur unplanned costs, but the effect remains confined to institutions with a direct relationship to the original project.

Level Three — Systemic Amplification. The disruption changes financing conditions, investment behavior, or operating capacity across a wider market, affecting institutions that have no direct contractual relationship with the original participant, through shared lenders, shared suppliers, shared customers, or a general repricing of the asset class.

The transition between these levels is the central problem for future research, and it is the transition that the indicators of Section 2.4 are designed to anticipate. A successful risk-management system should identify Level Two transmission early enough to prevent unnecessary escalation into Level Three, and it should recognize that the move from Level Two to Level Three is driven less by the size of the initial loss than by the opacity of the exposures surrounding it.

Section 4: From Wall Street to American Power Grids: The Geography of Financial Interdependence

4.1. Why Datacenter Geography Has Become a Financial Variable

The location of an AI datacenter has always influenced its operating costs, and the industry’s historical clustering in Northern Virginia, central Oregon, and the Dallas metropolitan area reflected decades of accumulated advantage in fiber connectivity, land, and electricity pricing. However, the next phase of hyperscale development introduces a consideration that operating-cost analysis does not capture. A project’s location increasingly determines its financial exposure to permitting requirements, electricity infrastructure investment, water availability, utility tariffs, and local political decisions, and in 2026 those exposures have begun to move quickly and in directions that financing models built in 2024 did not anticipate. Two otherwise similar datacenters may face substantially different financing conditions because they are located in different states, and the difference may emerge after the financing has closed.

A facility with an existing high-capacity electricity connection may become operational considerably sooner than one requiring a new transmission line. A project supported by a signed utility agreement may be easier to finance than another dependent on an uncertain interconnection study. A developer operating in a state with clearly defined large-load cost allocation rules may face greater initial financial obligations but less uncertainty over future regulatory treatment, and lenders have begun to price that distinction. Consequently, geography affects not only operating costs but the probability that a contractual commitment will become economically productive on schedule, which is to say that geography has become a determinant of the payment–delivery mismatch indicator described in Section 2.4.

This is particularly relevant because the scale of electricity demand associated with artificial intelligence continues to increase, and the institutions responsible for forecasting it have revised their estimates upward repeatedly. The International Energy Agency’s April 2026 analysis, Key Questions on Energy and AI, found that electricity demand from datacenters had increased by 17 percent in 2025, with demand from AI-focused facilities rising 50 percent, and it projected that total datacenter consumption would roughly double from about 485 terawatt-hours in 2025 to about 950 terawatt-hours by 2030, approaching 3 percent of global electricity demand, even as the energy required per AI task fell by at least an order of magnitude annually.^[32] The IEA also observed that the swelling pipeline of datacenter projects was straining planning and regulatory systems and holding up grid connections, and its executive director summarized the strategic stakes in a phrase that has since become a slogan.

“There is no AI without energy.” — Fatih Birol, Executive Director, International Energy Agency ^[32]

Meanwhile, the North American Electric Reliability Corporation’s 2025 Long-Term Reliability Assessment, released on January 29, 2026, projected that summer peak demand across its assessment areas would rise by more than 224 gigawatts over the next decade, a forecast nearly 70 percent higher than the one it had issued a year earlier, with new datacenters identified as the main source of growth; it found that thirteen of twenty-three assessment areas faced elevated or high resource-adequacy risk over the following five years, and in May 2026 it issued a Level 3 alert, its most serious, over the reliability risks posed by computational loads.^[33] NERC’s director of reliability assessment was careful to note that the forecast did not count every prospective datacenter, which is itself a statement about the financial problem this paper addresses.

“We see real load growth.” — John Moura, Director of Reliability Assessment and Performance Analysis, NERC ^[33]

These projections concern electricity requirements, not guaranteed future consumption. Actual demand will depend on project completion, computing utilization, technology efficiency, electricity availability, and economic conditions, and critics such as Grid Strategies have argued that NERC’s demand assumptions are inflated by speculative interconnection requests that will never be built.^[33] For Counterparty Cascade, the distinction is critical and cuts in both directions. A proposed datacenter is not equivalent to a completed datacenter. A grid interconnection request is not equivalent to a functioning electricity connection. A projected gigawatt of demand does not guarantee that a financially viable customer will eventually consume that electricity, and a utility that builds for the projection bears the risk of the gap. The financial system must distinguish between planned demand, contractually supported demand, and operational demand, and the state actions described below are, at bottom, attempts by governors and regulators to force that distinction into the open.

4.2. The Federal Government: Infrastructure Acceleration and Financial Discipline

Federal policy is becoming an important influence on AI infrastructure development, and it has pursued two objectives that are in some tension with each other. The White House’s July 2025 executive order on accelerating federal permitting of datacenter infrastructure established measures intended to facilitate qualifying datacenter projects and the associated energy development, including certain generation and electrical equipment.^[34] The policy reflects a strategic objective: expanding domestic computing capacity is increasingly viewed as important to American economic competitiveness and national security, and faster permitting could reduce construction delays, improve financing terms, and help the United States attract investment that might otherwise occur overseas. The same strategic logic underlies the Commerce Department’s January 2026 decision to move export licensing for the NVIDIA H200 and AMD MI325X to China from a presumption of denial to case-by-case review, which simultaneously opened a market for older-generation hardware and created a new policy variable in the residual-value calculations of every equipment financier.^[35]

However, accelerating approvals does not eliminate the need for financial discipline, and the federal government has begun to acknowledge this in the one arena where it has direct jurisdiction, which is the interstate transmission grid. A project can receive regulatory approval and still lack sufficient capital. A developer can secure financing and still face equipment shortages. A datacenter can be completed and still encounter a mismatch between customer revenue and infrastructure costs. The relevant policy objective should therefore be broader than simply approving more projects. It should include increasing the proportion of approved infrastructure that becomes operational, financially sustainable, and economically productive, and this creates an important distinction between regulatory efficiency and economic underwriting. Federal agencies can improve permitting processes. They cannot guarantee the commercial success of every AI model, application, or infrastructure investment, and federal policy should avoid treating the announced dollar value of private investment as equivalent to realized economic value. A $20 billion project that becomes operational and earns sustainable returns provides greater long-term economic benefit than a $50 billion proposal that never reaches commercial service, and the Van Nieuwerburgh central scenario, in which 227 gigawatts of the announced U.S. pipeline is never completed at all, suggests that the gap between announcement and realization will be the defining feature of the next five years.^[6]

The Department of Energy’s October 9, 2026 intervention before the Federal Energy Regulatory Commission in support of revisions to PJM’s reliability backstop procurement, urging that large new electricity users pay for the generation and infrastructure needed to serve them and that costs be allocated on the basis of updated, project-level load forecasts, is the clearest federal statement yet that acceleration and accountability are meant to proceed together.^[36] The Counterparty Cascade framework therefore suggests a new analytical priority for federal policy: infrastructure policy should evaluate the credibility and resilience of delivery obligations alongside the scale of announced investment.

4.3. Texas: Governor Greg Abbott and the Financial Credibility of Gigawatt Demand

Texas offers one of the most important case studies for the next phase of artificial intelligence infrastructure, because it combines the largest pipeline of proposed capacity in the country with the most aggressive recent exercise of state authority to verify it. The state’s industrial economy, energy resources, land availability, and technology investment activity make it attractive for large computing projects, and NERC’s assessment projected ERCOT’s summer peak growing from 94.7 gigawatts in 2026 to 154.1 gigawatts by 2035, including some 23 gigawatts of datacenter load.^[33] However, the rapid growth in proposed electricity demand, with ERCOT’s large-load interconnection queue approaching 100 gigawatts, created planning challenges that the state’s leadership concluded could not be resolved by the market alone.

During 2026, Governor Greg Abbott advanced a sequence of directives requiring closer examination of datacenter electricity needs, infrastructure costs, and water consumption. A June 10 directive required datacenters to fully fund the costs of the electric infrastructure needed to serve them. An August 3 directive ordered the Public Utility Commission of Texas and ERCOT to verify and audit every datacenter project advancing through the interconnection process. A September 14 directive required major water users, including datacenters, to comply with water-use reporting and imposed consequences for noncompliance. On September 21, 2026, the governor directed the Texas Commission on Environmental Quality to halt the issuance of all permits related to datacenter projects until ERCOT completed its review, with ERCOT’s report expected by mid-December, and he announced that he would work with the legislature in its next session to eliminate financial incentives for datacenters.^[37]

“Simply put, Texans must come first.” — Greg Abbott, Governor of Texas ^[37]

The implications extend beyond environmental and electricity regulation and reach directly into project finance. Consider a proposed multibillion-dollar Texas AI campus whose financing plan depends on a projected electricity connection date, whose semiconductor equipment agreements assume the facility becomes operational within a particular period, and whose customer contracts are priced according to an expected commencement date. If the project cannot obtain its air or wastewater permit until an audit concludes on a date no agency has fixed, the financing schedule changes. Expected equipment deliveries must be rescheduled, and if the equipment lessor’s rent is triggered by readiness for service rather than by energization, the developer pays for hardware it cannot run. If lenders determine that the project faces increased regulatory risk, they demand additional equity or stronger guarantees. The regulatory outcome therefore affects multiple contractual relationships at once, and the Lambda capacity in Nueces County that forms part of Anthropic’s disclosed commitments is, by virtue of its location, exposed to exactly this sequence.^[18]

At the same time, stricter verification may improve investment quality, and this is the argument that the industry’s more thoughtful participants have made in response. A process that prevents speculative projects from entering construction reduces the risk of financing resources being committed to facilities that lack credible customers, electricity supply, or sufficient capital, and it reduces the queue-clogging phantom demand that inflates NERC’s forecasts and delays the interconnection of real projects. The important distinction is between unnecessary delay and necessary financial verification. An efficient Texas system would identify viable projects early, assign infrastructure costs transparently, and provide credible schedules for developers that meet the applicable requirements, and such a system could make verified projects more attractive to lenders while reducing the possibility that residential customers bear costs associated with infrastructure that is never fully utilized. This suggests a broader consequence for the AI economy: by 2030, access to land may be a less important competitive advantage than access to a financing-ready, regulation-compliant, electricity-supported development site, and the states that can credibly certify such sites will capture the investment.

4.4. Pennsylvania: Governor Josh Shapiro and the Allocation of Infrastructure Risk

Pennsylvania presents a different approach, and one that is more explicitly designed around the allocation of financial risk than around the verification of demand. On August 18, 2026, Governor Josh Shapiro signed Executive Order 2026-05, Protecting Pennsylvania Consumers from Data Center Impacts, which directed Commonwealth agencies to require that datacenter proposals with peak demand exceeding 25 megawatts comply with the Governor’s Responsible Infrastructure Development, or GRID, requirements, first announced in February 2026 and published as standards in May.^[38] The order removed all datacenter projects from the state’s fast-track permitting program, barred executive agencies from entering nondisclosure agreements with datacenter projects, required a public map of all proposed projects that had engaged with the Department of Environmental Protection, made local land-use approval a prerequisite to state permitting, and conditioned the state’s sales-and-use tax exemption for datacenter equipment on GRID compliance; it also directed the administration’s Special Counsel for Energy Affordability to engage with the Public Utility Commission to ensure that datacenter development does not shift energy costs onto Pennsylvania ratepayers.^[38] The governor stated at the signing that more than 100 datacenter proposals were under discussion in the Commonwealth, of which roughly twenty had sought permits, and he framed the order as a response to the legislature’s refusal to act.

These measures are particularly relevant to Counterparty Cascade because utility infrastructure investments can create financial exposure even when the datacenter itself is privately financed. Suppose a developer requests a new transmission connection requiring substantial infrastructure expenditure. The utility must determine how the necessary costs will be recovered. If the developer ultimately fails to complete the facility, the expected electricity revenue does not materialize, and whether other customers bear those costs depends on the applicable tariff, regulatory decisions, financial security requirements, and contractual arrangements. A developer’s promise to purchase electricity is therefore not automatically equivalent to a guarantee that all associated infrastructure costs will be recovered, and the Pennsylvania Public Utility Commission’s May 12, 2026 final order establishing a model large-load tariff for electric distribution companies is the regulatory complement to the governor’s executive action.^[38]

Pennsylvania’s approach illustrates the importance of assigning risks to the parties best positioned to evaluate and control them. A developer choosing the location, scale, and timing of a facility should generally be expected to internalize costs attributable specifically to its project, subject to the relevant regulatory framework. A utility should possess mechanisms for protecting customers when proposed loads fail to materialize. Lenders should evaluate whether these obligations affect project returns, and the law firms advising developers were quick to note that financing assumptions would now need to incorporate the GRID compliance strategy and its ongoing commitments.^[38] Local communities should understand which infrastructure costs remain their responsibility. The principle is economically significant: when risks are allocated transparently, investors can incorporate them into financial models before committing capital, and when risk allocation remains uncertain, the consequences emerge only after construction has begun or a project has failed. Pennsylvania also offers an important lesson regarding political accountability. A datacenter development policy is increasingly a statement about who should bear the financial consequences of industrial expansion, and the question is not simply whether a state supports artificial intelligence but whether it can support the industry’s growth while protecting existing electricity customers from becoming its involuntary financiers.

4.5. Virginia: Governor Abigail Spanberger and the Economics of Transmission Cost Allocation

Virginia has long been the center of American datacenter activity, and the concentration of computing infrastructure in Loudoun and Prince William counties creates both considerable economic opportunity and the most acute version of the cost-allocation problem anywhere in the country. In July 2026, Governor Abigail Spanberger’s administration, through a filing by her newly appointed Chief Energy Officer, urged the State Corporation Commission to require large electricity users such as datacenters to pay upfront for transmission lines and substations built specifically to serve their facilities, rather than spreading those costs across residential and small-business customers, and twenty members of the General Assembly wrote in support.^[19] The evidence presented in the proceeding was striking: regulators found that Virginia residents had paid $2.8 billion for transmission lines related to datacenters since 2021, rising to $6.2 billion when indirectly related projects were included. On August 5, 2026, the SCC ordered Dominion Energy to develop a tariff directly assigning the cost of transmission infrastructure built exclusively for datacenters and other large loads to those customers, a decision the governor’s office projected would save ratepayers hundreds of millions of dollars.^[19]

Data centers must “pay their fair share.” — Abigail Spanberger, Governor of Virginia ^[19]

The significance for Counterparty Cascade is the relationship between project finance and regulated utility economics, which operate on fundamentally different time horizons. Transmission infrastructure possesses a useful life measured in decades. The commercial commitments supporting a particular AI datacenter may have substantially shorter durations, and the GPUs inside it shorter still. This creates a potential mismatch: a transmission asset can continue operating long after the original datacenter customer has reduced demand, relocated operations, or entered financial restructuring, and if the infrastructure was built primarily for that customer, the question of cost recovery becomes essential. Should the developer pay the full incremental cost? Should the utility recover some costs through general tariffs on the theory that the network benefits are shared? Should financial security remain available if the customer closes? How should regulators evaluate projects involving several potential customers, as the “but-for” test advocated by the SCC’s own staff and by the state’s electric cooperatives would require? These questions cannot be answered through a single nationwide contractual formula. They depend on utility regulation, actual network benefits, cost-causation principles, and the relationship between dedicated and shared infrastructure. However, the underlying financial issue is consistent: infrastructure investment should not rely on assumptions about future AI demand without examining who absorbs losses if the assumptions prove incorrect.

Virginia’s experience is particularly important because the concentration of datacenters creates the possibility of correlated regional exposure of exactly the kind described in Section 2.6. An adjustment in investment plans by several major customers could affect transmission forecasts, utility capital planning, construction activity, and local development expectations simultaneously, and the governor’s own Chief Energy Officer has told Virginians to expect electric bills to continue rising for the next twenty years.^[19] The resulting economic consequences might be substantial even if no major technology company becomes insolvent, and the SCC’s order is, in effect, an attempt to ensure that the party making the demand forecast is the party that bears the cost if the forecast is wrong.

4.6. California: Governor Gavin Newsom and the Cost of Infrastructure Externalities

California provides another important regulatory perspective, and a notable political reversal. On September 21, 2026, Governor Gavin Newsom signed a package of seven bills that his office described as the most comprehensive datacenter laws in the nation, a year after he had vetoed a water-disclosure measure over concerns that regulation could stifle AI growth.^[39] Three of the laws, including SB 886 and AB 2383, shift electric infrastructure costs away from residential customers and toward datacenter operators, requiring the creation of special datacenter rate classes by 2028 and requiring operators to pay for the grid upgrades their facilities require while preventing cost shifts to low-income customers. Three more, including AB 2469 and AB 2619, require disclosure of water use, water supply assessments, and energy consumption, integrating datacenter demand into urban water-management planning for the first time. The seventh makes datacenters ineligible for certain environmental review exemptions.^[39] The Data Center Coalition, whose members include Google, Microsoft, and OpenAI, warned that the measures singled out datacenters relative to other large industrial users and could push projects to other states.

“Californians remain in the driver’s seat.” — Gavin Newsom, Governor of California ^[39]

California’s relevance extends beyond the number of hyperscale facilities constructed within the state, which is modest relative to Texas or Virginia. The state is home to the major technology companies, venture capital institutions, research organizations, semiconductor designers, and AI developers whose commercial decisions determine where computing infrastructure is built across the country. An AI company headquartered in California may operate infrastructure in Texas, lease computing capacity in Virginia, procure electricity in Pennsylvania, and rely on semiconductor manufacturing in Arizona, and its financial obligations extend across all of those jurisdictions at once. A deterioration in the company’s economic position affects projects located far beyond California, and conversely, regulatory changes in other states alter the company’s capital allocation decisions. California therefore illustrates an important distinction between corporate geography and infrastructure geography. The financial headquarters of an AI ecosystem may not be located near the physical assets supporting its operations, and this separation complicates responsibility, because the people making investment decisions may not face the same electricity costs, construction impacts, or infrastructure risks as the communities hosting the facilities. The Counterparty Cascade framework suggests that the relationship between headquarters and host communities deserves greater scrutiny, and that investment contracts should make clear which party is responsible for infrastructure costs, environmental obligations, and financial consequences if a project is delayed or abandoned.

4.7. Arizona: Governor Katie Hobbs and the Semiconductor Manufacturing Connection

Arizona occupies a strategically important position in the second layer of the AI economy, and its experience illustrates that the contractual network of Counterparty Cascade extends upstream into manufacturing as well as downstream into applications. On July 16, 2026, TSMC announced an additional $100 billion investment to build four more fabs at the 2-nanometer node or beyond in Arizona, bringing its total announced investment in the state to $265 billion across ten fabs, two advanced packaging facilities, and a research and development center; upon completion, around 30 percent of TSMC’s 2-nanometer and more advanced capacity would be located in Arizona, and the state’s economic development agency reported that Arizona had attracted more than 70 semiconductor expansions representing over $314 billion of investment since 2020.^[40]

Arizona is “the nation’s epicenter for advanced semiconductor manufacturing and innovation.” — Katie Hobbs, Governor of Arizona ^[40]

Semiconductor manufacturing differs from datacenter operations in important respects. Fabrication facilities require specialized production equipment, complex chemical processes, highly trained workers, and exceptionally demanding operating conditions, and their capital investment cycles extend over many years with construction timelines that TSMC’s own schedule, which places its second Arizona fab in volume production in 2027 and its third in 2030, makes clear. However, the same principle of financial interdependence applies. A fabrication facility’s economics depend on expectations regarding future semiconductor demand, and in 2026 that demand is overwhelmingly driven by AI accelerators. Its customers depend on the facility’s production schedules. Its suppliers, such as the industrial gas and chemical companies that have announced their own Arizona plants, invest in additional capacity based on anticipated manufacturing activity. Utilities expand infrastructure to support the facility, and the Greater Phoenix Economic Council has estimated that some 12,000 construction trade workers will be required to build the fabs. An unexpected change in demand, production schedules, trade policy, or financing conditions can therefore affect multiple organizations, and the January 2026 export-licensing change and the accompanying 25 percent tariff on imported advanced computing commodities are reminders that trade policy is now a direct input into the economics of every layer.^[35]

For Counterparty Cascade, Arizona is relevant not only because it manufactures chips. It is relevant because manufacturing commitments in Layer 2 support infrastructure promises in Layer 3 and technological development in Layer 4, so that a significant semiconductor production disruption could create consequences across several layers, and conversely a broad reduction in AI infrastructure investment, of the kind the hyperscalers’ shareholders began to demand in the last week of October 2026, would change the demand expectations against which $265 billion of fabrication capacity has been committed. This demonstrates the bidirectional nature of financial transmission, and it demonstrates that the states which have attracted the largest manufacturing investments have thereby acquired the largest exposure to the durability of fifth-layer demand.

4.8. Michigan and Indiana: The Relationship Between Electricity Investment and AI Contracts

Michigan illustrates the importance of reliable generation capacity, and the role of federal credit in supplying it. The Department of Energy’s Loan Programs Office closed a loan guarantee of up to $1.52 billion in September 2024 to support Holtec International’s restoration of the 800-megawatt Palisades Nuclear Plant in Covert Township, the first restart of a commercial reactor from decommissioning in American history, and it continued to disburse funds through 2025 and 2026 as the plant transitioned back to operational status under Nuclear Regulatory Commission oversight; by the sixth disbursement in September 2025, some $491 million had been advanced, and Holtec has announced plans to co-locate small modular reactors on the site after 2030.^[41] The plant’s restoration should not be interpreted as a project exclusively dedicated to AI. Its significance is broader, because reliable electricity generation supports industrial investment, manufacturing, households, and increasingly electricity-intensive computing infrastructure alike. However, generation projects possess their own financial risks. Construction, refurbishment, maintenance, regulatory compliance, and operational readiness require substantial capital, and the restart has faced local opposition and technical scrutiny over steam-generator repairs. If an AI infrastructure developer assumes that new or restored generation will become available by a particular date, a delay in that generation affects the developer’s computing plans, and this creates another possible contractual dependency, one in which the counterparty is a federal loan guarantee program and a nuclear regulator rather than a cloud provider.

Indiana illustrates the industrial-development dimension and offers an instructive contrast in contract design. In November 2025, Amazon announced plans to invest an additional $15 billion in Northern Indiana to build datacenter campuses adding roughly 2.4 gigawatts of capacity, on top of the $11 billion it had committed to St. Joseph County the prior year, and it structured the energy arrangement with Northern Indiana Public Service Company so that Amazon would pay fees to use existing power lines and cover the full cost of any new generation, transmission, or equipment needed to serve the campuses, an arrangement NIPSCO projected would save existing customers approximately $1 billion over fifteen years.^[42]

“This is a massive win for Hoosier ratepayers.” — Mike Braun, Governor of Indiana ^[42]

Whether that projection is realized will depend on the campuses being built and operated on the schedule assumed, but the structure of the agreement, in which the developer bears the incremental infrastructure cost and the utility’s existing customers are contractually protected rather than merely reassured, is precisely the kind of risk allocation that this paper argues reduces cascade exposure. Such investments can create construction employment, demand for electrical equipment, local tax revenues, and opportunities for suppliers, but the long-term benefits depend on projects becoming operational and remaining economically productive. A state should therefore distinguish three separate measures: announced investment commitments; capital actually deployed into completed infrastructure; and sustained economic activity and fiscal benefits produced after completion. The distinction matters because announcements can be revised, financing can change, and development schedules can shift, and for state governments Counterparty Cascade provides a reason to measure project delivery and durability, not only the total value of new investment announcements.

4.9. FERC, NERC, PJM, and the Emerging Federal Framework for Large Electrical Loads

The Federal Energy Regulatory Commission and the North American Electric Reliability Corporation have become central to AI infrastructure planning during 2026, and their actions have begun to define the financial terms on which large loads connect to the interstate grid. On June 18, 2026, FERC issued six simultaneous show-cause orders under Section 206 of the Federal Power Act to every jurisdictional regional transmission organization and independent system operator, including PJM, MISO, SPP, CAISO, ISO New England, and NYISO, preliminarily finding that each tariff appeared unjust and unreasonable because it lacked adequate provisions to address the integration of large loads, especially datacenters supporting AI, and directing the operators to justify or revise their tariffs within sixty days and to explain within thirty days how they would ensure adequate generation to serve existing and new large loads; the orders addressed cost allocation, co-location, behind-the-meter arrangements, and flexible load integration, and all five commissioners concurred.^[43] FERC’s chairman framed the action as a refusal to choose between reliability and growth.

“We can facilitate both [reliability and innovation].” — Laura V. Swett, Chairman, Federal Energy Regulatory Commission ^[43]

The PJM proceedings that followed illustrate how directly these questions now bear on financing. After its most recent capacity auction fell roughly 6.8 gigawatts short of its reliability target, PJM proposed a one-time Reliability Backstop Procurement to secure additional capacity for the 2028–2029 delivery year, filing it on July 31, 2026. On September 29, FERC accepted the procurement in principle but suspended it for five months over flaws in cost allocation and financial rules, recommending that PJM allocate the procurement’s costs based on updated, project-level load forecasts so that the customers who caused the need would pay for it; on October 9, FERC rejected the core of PJM’s companion plan for an interim, curtail-first service for new large loads, and on the same day the Department of Energy intervened to urge PJM to adopt FERC’s recommended revisions and prevent cost shifts onto existing customers.^[36] The practical consequence is that, as of this writing, both of PJM’s answers to a known capacity shortfall remain unresolved, and every datacenter developer in the thirteen-state PJM footprint is financing against a cost-allocation rule that does not yet exist.

NERC has moved in parallel. Its Large Loads Working Group concluded in a March 2026 gap assessment that existing reliability standards, interconnection requirements, and industry practices were inadequate to reliably integrate emerging large loads; it published a voluntary reliability guideline on risk mitigation for emerging large loads in May 2026, issued its Level 3 alert the same month, and has set a 2026 deadline for a new class of registered entity and associated standards addressing computational loads.^[33] These developments are important because electricity availability has both a physical dimension and a financial dimension. The physical dimension concerns whether sufficient generation, transmission, and distribution capacity exists. The financial dimension concerns who funds the necessary infrastructure and what happens if expected demand changes. Consider a proposed datacenter requiring a new high-voltage connection. The utility may need to purchase transformers and undertake engineering work long before the facility begins consuming electricity. If the project is subsequently canceled, those expenditures do not disappear, and someone must absorb them. Strong contractual security and clear regulatory requirements reduce the likelihood that the costs are transferred unfairly, and they improve the quality of financing decisions, because a lender assessing a datacenter project can incorporate known electricity-related obligations into its underwriting, a developer can evaluate whether the project’s expected returns remain attractive, and the utility can plan infrastructure based on more credible demand. The goal should be neither indiscriminate approval nor indiscriminate restriction. It should be credible infrastructure development supported by transparent financial responsibility.

4.10. Georgia: OpenAI, Georgia Power, and the Flexible-Load Alternative

The most constructive development of 2026 in the relationship between AI infrastructure and the electricity system occurred not in a statehouse but in a utility contract. On July 22, 2026, Georgia Power, a subsidiary of Southern Company, announced a 25-year power supply agreement with OpenAI for a new campus in Effingham County, Georgia, about 45 minutes from Savannah, known as Project Camellia. The facility is expected to require approximately 3,200 megawatts of power, phased in between 2028 and 2032, and under the terms of the agreement OpenAI will pay the full infrastructure and electric service costs necessary to serve the complex while committing up to 1,000 megawatts of flexible demand response, meaning that Georgia Power may reduce energy delivered to the facility during periods of peak statewide demand. Georgia Power described the arrangement as among the largest single-facility demand-response commitments in the country. On August 27, 2026, the Georgia Public Service Commission approved the contract, and the utility stated that the resulting growth would increase the expected savings for a typical residential customer to roughly $180 per year beginning in 2029.^[44]

This arrangement highlights a potentially important direction for the entire industry. A datacenter need not be treated as an inflexible electricity consumer, and under appropriate technical and contractual conditions some computing activities, particularly training workloads that are insensitive to latency, may be scheduled, reduced, or shifted during periods of electricity-system stress. NERC’s own reliability staff have made the same point, observing that if datacenters can shift load to backup sites or commit to being absent during peak conditions, they can interconnect faster and reduce the need for near-term system upgrades.^[33] Such arrangements can reduce infrastructure requirements and improve reliability, and they can convert the largest load on a utility’s system from a liability into a reliability asset that functions, in effect, as a virtual power plant without the construction cost.

However, the benefits must be demonstrated rather than assumed. A contractual commitment to provide flexibility is not equivalent to verified performance during a grid emergency. The technical feasibility of reducing electricity consumption depends on the workloads being performed, the computing architecture, operational constraints, and the availability of alternative capacity elsewhere, and inference workloads serving live customers cannot be curtailed in the way training runs can. Financial consequences also require careful evaluation, because a datacenter that agrees to curtail consumption may need to maintain additional computing capacity elsewhere or accept interruptions to particular workloads, and the costs of that flexibility must be incorporated into its commercial arrangements and, ultimately, into the minimum commitments it has made to its own infrastructure providers. For Counterparty Cascade, the opportunity is to reduce one important source of infrastructure risk. If datacenters can reliably provide demand flexibility, some electricity infrastructure investments become less sensitive to peak-demand assumptions, which improves the economics of certain projects and helps utilities manage the consequences of unexpected changes in the pace of AI infrastructure development. The broader lesson is that technological flexibility can serve as a financial risk-management tool, and that the Georgia agreement, by pairing full developer cost responsibility with a curtailment commitment and regulatory approval, is a template for how the energy layer can protect itself from the volatility of the layers above it.

4.11. The Geographic Correlation Problem

A final geographic issue deserves particular attention, because it ties the state-by-state analysis back to the diversification paradox of Section 2.6. Large AI infrastructure projects are typically discussed individually. A company announces a campus in Texas. Another announces a facility in Ohio. A third expands in Michigan or Georgia. Each project receives a separate estimate of investment, electricity demand, employment, and economic impact, and each state’s economic development agency counts it separately. However, the projects share common dependencies that no individual announcement discloses. They require the same advanced accelerators from the same two or three vendors. They rely on the same handful of transformer manufacturers with multiyear backlogs. They seek financing from overlapping institutional investors and, increasingly, from the same six asset managers. Their principal customers compete within the same frontier model market, and their commercial success depends on the same assumptions about enterprise adoption of agentic systems.

Consequently, physical geographic diversification does not automatically produce financial diversification. A lender with exposure to facilities across ten states may still possess concentrated economic exposure to a small number of frontier model customers. A semiconductor supplier serving numerous datacenters may still depend on continued investment by a limited number of hyperscalers. A utility’s revenue projections may depend on a relatively small number of corporate investment decisions made in California boardrooms. This suggests the need for a new form of infrastructure analysis. Instead of asking only where datacenters are being built, policymakers and investors should ask which projects share the same financial dependencies, and how those projects would respond to a common disruption. That question links local infrastructure decisions to the financial stability of the national AI economy, and it is the question that the state actions of 2026, for all their differences in approach, have each begun to ask in their own way.

Section 5: The 2027–2030 Outlook: Modeling Counterparty Cascades Before They Occur

5.1. The Transition From Infrastructure Expansion to Financial Verification

The years 2027 through 2030 may represent a decisive phase in the development of the AI economy, and the reason is structural rather than cyclical. During the early infrastructure expansion, many investment decisions were supported by expectations of rapid growth in model capabilities and commercial adoption, and the scarcity of computing capacity justified substantial financial commitments made under conditions of considerable uncertainty. Companies competed to secure advanced accelerators, electricity access, datacenter sites, and strategic partnerships, and the party that hesitated lost the capacity to a competitor who did not. However, infrastructure investments cannot remain dependent indefinitely on expectations alone, and as more capacity becomes operational investors will increasingly evaluate actual performance against the assumptions embedded in the contracts that financed it.

How much computing capacity is being utilized? What revenue does that capacity generate per installed gigawatt? How quickly do infrastructure investments recover their costs, and does the recovery period exceed the economic life of the hardware? What financing obligations remain outstanding when the first generation of leased accelerators is replaced? How do operating margins change as accelerator generations improve and rental prices fall? How much demand originates from economically productive enterprise applications rather than from experimentation, subsidized consumer usage, or the training runs of competitors? What portion of future revenue is supported by enforceable customer commitments rather than by consumption that can be redirected in a month? These questions will become more consequential as the industry’s capital stock expands, and the transition may create financial pressure even if AI adoption continues growing at a historically unprecedented pace.

Suppose infrastructure capacity expands by 60 percent during a particular period while economically sustainable demand increases by 40 percent. The industry would still experience substantial growth by any ordinary standard. However, some investments would remain underutilized. Facilities with comparatively high construction or financing costs would struggle to achieve their originally expected returns. Customers would negotiate lower prices for certain computing services, and the required GPU-hour price that Van Nieuwerburgh calculates the buildout must sustain, roughly $5.5 per installed hour at full utilization and $6.9 at 80 percent utilization, would come under pressure from the supply the buildout itself created.^[6] Lenders would reevaluate future cash flows. The result could be a reduction in infrastructure investment despite continued growth in AI usage, and this illustrates the distinction that organizes the remainder of this section: the economic success of artificial intelligence as a technology does not guarantee that every infrastructure investment supporting its development will be financially successful, and the gap between the two is where cascades begin.

5.2. Hyperscaler Capital Expenditure and the Risk of Synchronized Investment Decisions

The largest technology companies possess substantial financial resources and diversified operating businesses, and their capital expenditure decisions are therefore especially important, because they are simultaneously the largest customers of the chip layer, the largest landlords and tenants of the datacenter layer, the largest investors in the model layer, and the largest distributors in the application layer. During 2026, the scale of their investment plans became the central fact of the macroeconomy. Professor Jason Furman of Harvard calculated that investment in information-processing equipment and software, though only about 4 percent of GDP, accounted for roughly 92 percent of U.S. GDP growth in the first half of 2025, with growth excluding those categories running at just 0.1 percent annualized, a finding he was careful to describe as an accounting decomposition rather than a counterfactual, since absent the AI boom interest rates and electricity prices would likely have been lower.^[45] The third-quarter 2026 results reviewed in Section 3.4 extended that pattern: Alphabet’s full-year guidance of $195 to $205 billion, Amazon’s roughly $220 billion, Meta’s $135 to $145 billion, and Microsoft’s approximately $170 billion for calendar 2026 together imply aggregate spending by the four largest hyperscalers in the vicinity of three-quarters of a trillion dollars in a single year, with every one of them guiding higher for 2027.^[14] ^[13] ^[27] ^[29] These figures have different accounting definitions, since Microsoft’s includes finance leases at commencement and Amazon’s is a cash figure, and they should not be treated as directly interchangeable; they nevertheless demonstrate the magnitude of capital allocation occurring across the technology industry, and the degree to which the rest of the economy has come to depend on it.

The largest concern may not be the investment decision of any single hyperscaler. It may be the possibility that several companies adjust their spending plans simultaneously, because they are responding to the same signals from the same shareholders. Consider a hypothetical 2028 environment. AI applications remain commercially important, but model inference becomes markedly more efficient. Some enterprise customers begin consolidating their AI suppliers. Several hyperscalers determine, independently, that existing computing capacity can support expected demand for longer than previously assumed, and each decides to moderate new infrastructure investment. The decisions are prudent from the perspective of individual shareholders, and the market rewards them. But the aggregate consequences are substantial. Semiconductor suppliers experience slower order growth and their equity stakes in neoclouds decline. Datacenter developers encounter fewer new commitments and find that their delayed-draw facilities, which amortize against contracted cash flows, cannot be refinanced on the terms that were available in 2026. Electrical equipment manufacturers revise expansion plans. Infrastructure lenders reduce expected lending volumes and reprice their existing books. Construction companies discover that anticipated projects are being postponed. Regional utility demand forecasts, and the generation and transmission investments built against them, change.

The result would not necessarily constitute a crisis. It could represent a healthy adjustment between supply and demand, of the kind every capital-intensive industry experiences. However, if financing structures assume uninterrupted expansion, the adjustment creates contractual stress, and this is the difference between ordinary capital expenditure moderation and a Counterparty Cascade. The latter occurs when investment reductions produce financial consequences that force other institutions to change their behavior beyond what the original demand adjustment would otherwise require, and the degree to which the current financing architecture would amplify rather than absorb such a moderation is the central empirical question of the next four years.

5.3. Scenario One: The Orderly Expansion Case, 2027–2030

The first scenario assumes strong commercial AI adoption and broadly successful infrastructure execution, and it is the scenario that the pricing of every long-dated commitment examined in this paper implicitly assumes. Enterprise customers increasingly deploy AI systems that demonstrate measurable productivity improvements, and the productivity statistics begin to confirm the J-curve that Brynjolfsson has argued is already visible.^[20] Autonomous agents perform more complex tasks in software development, customer service, logistics, engineering, and scientific research, and the agent-deployment figures that Amazon reported for its third quarter, with customers deploying twice as many agents in that quarter as in the prior three combined, prove to be the beginning of a curve rather than a plateau.^[13] Demand for inference continues growing faster than efficiency improvements reduce the compute required per task. Model developers improve the efficiency of their systems while expanding their commercial capabilities, and consumption-based revenue proves stickier than its contractual terms would suggest. Infrastructure developers deliver a substantial portion of contracted capacity within acceptable schedules. Utilities successfully support major new electricity loads under the cost-allocation frameworks established in 2026. Financing institutions maintain sufficient access to capital, and the private credit redemption episode of early 2026 proves to have been a contained event.

Under these conditions, the AI economy can sustain considerable investment without experiencing widespread financial instability. Contractual interdependence remains extensive, but the underlying revenue growth supports the obligations. Specialized infrastructure operators achieve stable utilization and refinance their 9 percent notes at lower spreads. Equipment lessors receive contractual payments, and residual-value guarantees are never called. Lenders experience manageable credit losses. Semiconductor suppliers continue investing in improved manufacturing and systems, and the Arizona fabs come online into a market that can absorb their output. Application providers convert model capabilities into recurring economic value at a scale approaching the trillions of dollars of mature revenue that the buildout requires. This scenario demonstrates that financial interconnectedness can support productive economic expansion, that long-term contracts can provide the revenue visibility required to finance expensive infrastructure, that strategic partnerships can distribute technical and financial responsibilities, and that private investment can expand computing capacity faster than companies could achieve independently. The principal requirement is alignment: commercial demand, financing commitments, equipment availability, and infrastructure delivery must remain sufficiently coordinated, and the outcome would validate the economic rationale for much of the industry’s investment. However, even success would not eliminate the need for counterparty risk management, because a larger industry could still possess concentrated vulnerabilities, and because the concentration of the financing architecture would have grown in proportion to the industry.

5.4. Scenario Two: The Selective Infrastructure Adjustment, 2027–2030

The second scenario assumes that AI adoption continues expanding but at a slower or more uneven pace than some infrastructure investment plans anticipate, and it is, in my judgment, the modal outcome. This does not require a technological disappointment. In fact, technological success could contribute to the adjustment, and this is the most counterintuitive feature of the AI economy’s financial structure. Improved model architectures reduce the computing requirements for particular tasks, as the IEA observed they have been doing by an order of magnitude annually.^[32] Specialized models perform some enterprise workloads more economically than general-purpose frontier models. Application developers optimize inference spending once the novelty of deployment gives way to the discipline of budgets. Companies deploy smaller models locally, or on their own hardware, for workloads that do not require centralized frontier infrastructure. Meanwhile, some existing facilities experience lower utilization than expected, and the GPU-hour price that was supposed to sustain the buildout drifts below the level the financing assumed.

This environment creates winners and losers within every layer. Operators with efficient electricity arrangements, modern hardware, strong customers, and manageable financing costs remain competitive and acquire the assets of those that do not. Developers with expensive contracts, delayed construction, or weak financing arrangements face pressure, and the neoclouds whose contract books exceed their live capacity by a factor of forty discover that the gap between the two is where their equity lives. Certain computing service prices decline. Some infrastructure projects are canceled or restructured, and the 227 gigawatts of announced U.S. capacity that Van Nieuwerburgh’s central scenario assumes is never built becomes a visible fact rather than a modeling assumption.^[6] Lenders distinguish more aggressively between financially resilient and vulnerable operators, and spreads on the asset class widen. The financial adjustment remains orderly if losses are absorbed by investors and creditors with sufficient capital, and the hyperscalers, whose balance sheets remain the strongest in the system, acquire distressed capacity at prices that improve their own returns.

However, the risk of Counterparty Cascade increases where several affected projects share financing sources or critical counterparties, and this scenario is the one in which the shared-lender and shared-customer dependencies of Sections 2.6 and 4.11 matter most. The policy implication is that governments should avoid equating project cancellation with a failure of national AI strategy. Canceling an economically weak project before substantial capital is deployed preserves resources for stronger investments, and the Texas audit and the Pennsylvania GRID requirements, whatever their other effects, are mechanisms for making that cancellation happen before the transformers are ordered rather than after. The objective should be sustainable infrastructure capacity, not the completion of every proposed facility regardless of its economics.

5.5. Scenario Three: A Major Counterparty Financing Shock, 2028–2030

The third scenario examines a more severe possibility, and it is the one that the IMF’s three salient risks and the BIS’s four pressure points are designed to anticipate. Assume that a major financing institution or infrastructure operator experiences substantial financial stress, and that several projects depend on its ability to provide capital or deliver computing capacity. The institution encounters a combination of higher financing costs, delayed construction, and weaker-than-expected customer revenue, and it becomes unable to satisfy certain contractual obligations. Affected projects seek replacement financing. Some lenders refuse to increase their exposure. Others demand additional equity. Developers delay equipment purchases. Customers attempt to procure replacement computing capacity, and several contracts are renegotiated or terminated under the milestone provisions that the 2026 generation of agreements included. The consequences spread through the network along the seven mechanisms of Section 3.

At first, the disruption affects a limited number of projects, and it looks like a Level Two event. However, other financing institutions begin reassessing similar exposures, because they cannot determine from public information whether they hold them. They increase required returns. Some reduce lending commitments. Developers across the industry face higher capital costs. The value of certain infrastructure assets declines because expected future revenue no longer supports prior valuations, and the equipment vendors who guaranteed residual values face calls on those guarantees in the same quarter that their own sales decline. Investors become more cautious, and the industry experiences a broader reduction in infrastructure investment. This is a potential Level Three Counterparty Cascade, and the severity depends on the concentration of exposures, the contractual protections in place, the available liquidity, and the willingness of stronger institutions to acquire or refinance distressed assets. The IMF’s own framing of the severe case is that a market correction could cascade through a highly leveraged and circular ecosystem and that, given sizable cross-border holdings of AI assets, the losses could spread around the world.^[22]

There are also important stabilizing mechanisms, and a fair analysis must weigh them. Large hyperscalers possess sufficient financial resources to acquire assets from distressed developers, and they have every commercial incentive to do so. Private credit managers may restructure loans rather than realize losses on collateral they cannot operate. Semiconductor equipment retains substantial value for alternative workloads, particularly inference. Infrastructure assets attract new customers at lower prices, and a datacenter with a power connection is scarce regardless of who built it. Consequently, even a severe financing shock does not imply that the entire AI economy must collapse, and a substantial portion of the underlying computing infrastructure would remain economically useful under new ownership. The critical policy question is whether losses can be absorbed without creating unnecessary disruptions to otherwise viable infrastructure and commercial services, and whether the opacity of the financing architecture would cause the market to treat every exposure as suspect because it cannot identify which ones actually are.

5.6. Scenario Four: The Geopolitical Semiconductor Interruption

A fourth scenario concerns the relationship between American AI infrastructure and global semiconductor production, which is the scenario in which the initiating disruption originates outside the financial system entirely. Advanced accelerators depend on complex international manufacturing systems, and leading-edge fabrication capacity remains geographically concentrated to a degree that no other input to the AI economy approaches. Memory, packaging, substrate manufacturing, networking components, and specialized equipment also involve international dependencies, and Anthropic’s filing specifically identifies semiconductor manufacturing concentration and associated geopolitical and operational risks, including exposure to disruptions affecting Taiwan.^[1]

U.S. semiconductor export controls introduce another important dimension. In January 2026, the Bureau of Industry and Security revised its review policy for certain advanced computing chips exported to China and Macau, establishing case-by-case review for the NVIDIA H200, the AMD MI325X, and less advanced chips, subject to conditions including a cap of 50 percent of the relevant chip’s U.S. sales and a 25 percent federal charge on the exports, replacing a presumption of denial.^[35] Export controls affect product availability, customer markets, and equipment allocation, and the January rule had the specific effect of creating a sanctioned secondary market for the prior-generation hardware that underpins a large share of the equipment financing written in 2024 and 2025. However, a Counterparty Cascade analysis should distinguish regulatory restrictions from actual supply interruptions. An export license requirement does not automatically create a global chip shortage, and a change in market access does not necessarily reduce the ability of domestic datacenter operators to obtain computing equipment. The risk becomes financially significant when a regulatory or geopolitical development materially changes contractual delivery, pricing, or financing assumptions.

Consider a hypothetical interruption affecting an important semiconductor manufacturing or packaging facility. Accelerator deliveries are delayed for a period measured in quarters. Datacenter operators cannot satisfy customer schedules, and their acceptance-triggered revenue does not arrive while their construction debt continues to accrue interest. Model developers experience capacity shortages and bid up the price of whatever short-term capacity exists. Equipment financing arrangements require modification because the collateral has not been delivered. Lenders reevaluate construction projects dependent on the unavailable components. The resulting consequences extend well beyond companies directly purchasing the affected chips, and this scenario demonstrates why semiconductor diversification matters to financial resilience and why the Arizona investments of Section 4.7 are a financial-stability measure as much as an industrial-policy one. However, diversification must include manufacturing processes, advanced packaging, high-bandwidth memory, software compatibility, equipment supply, and geographic exposure. Purchasing chips from multiple designers provides only limited protection if those products depend on the same critical manufacturing facilities, and the relevant issue is whether the industry possesses genuinely independent alternatives at each stage of the chain rather than at the brand level alone.

5.7. The Financial Consequences of More Efficient AI Models

One of the most interesting possibilities, and the one that distinguishes the AI buildout from the railroad and fiber booms it is so often compared with, is that improvements in artificial intelligence could themselves change infrastructure economics. Many investment models assume that AI demand will continue growing rapidly, and the assumption has been correct for four consecutive years. However, computing demand depends on two competing forces. The first is the increasing number and complexity of tasks performed by AI systems, which grows demand. The second is the improving efficiency with which those tasks can be performed, which reduces it. More capable AI systems encourage greater usage. At the same time, improved algorithms, specialized models, quantization, caching, optimized kernels, and better hardware utilization reduce the computing resources required for a given task, and the IEA’s observation that per-task energy consumption has been falling by at least an order of magnitude annually is a measurement of the second force.^[32]

The net effect is uncertain, and intellectually honest analysis must say so. A large reduction in the cost of inference could stimulate so much additional demand that total computing consumption continues rising, which is the pattern that the economist William Stanley Jevons identified for coal in the nineteenth century and that has held for most of the history of computing. Alternatively, efficiency improvements could temporarily outpace demand growth for certain workloads, reducing the value of some reserved capacity, and the episode in early 2025 in which the release of a more efficient open-weight model briefly erased hundreds of billions of dollars of semiconductor market value demonstrated how quickly the market can reprice the second possibility. Both outcomes are plausible, and the financial risk arises when infrastructure commitments depend on a narrow forecast of which one will dominate.

Consider two investment strategies. The first assumes that a particular model architecture will require an approximately fixed amount of computing capacity per unit of commercial activity for several years, and it locks in ten years of capacity accordingly. The second allows for substantial changes in model efficiency and workload composition, and it accepts a higher unit cost in exchange for shorter commitments and the flexibility to adopt new accelerator generations. The second strategy involves more complex planning and, in a market of scarcity, a real risk of being unable to obtain capacity at all, but it is more resilient to technological change. This suggests that infrastructure lenders should examine model efficiency assumptions rather than relying exclusively on projected revenue growth. A lender financing a GPU deployment needs to understand the expected workloads, architecture compatibility, power requirements, and competitive alternatives, which means that technical due diligence becomes part of financial underwriting. By 2030, sophisticated infrastructure financing may increasingly depend on independent evaluations of technological productivity, and the asset managers who announced compute-financing platforms in August 2026 will discover that underwriting a GPU is not the same discipline as underwriting a toll road.

5.8. Agentic Systems and the Durability of AI Revenue

Autonomous agents may become among the most important determinants of long-term AI infrastructure demand, because they are the first category of application in which the computing consumed per unit of economic value is large enough to absorb the capacity being built. Traditional AI applications respond to individual user requests. Agentic systems perform sequences of actions across multiple tools and enterprise systems, analyzing information, generating plans, executing tasks, evaluating results, and interacting with software environments, often for hours at a time without human intervention. This creates new commercial opportunities, and an agent performing valuable work may justify spending that exceeds the cost of a conventional software subscription by an order of magnitude. However, the economics depend on reliability, governance, and measurable productivity. An autonomous system that frequently requires human correction delivers less value than anticipated. A system that consumes large amounts of inference without completing economically useful tasks does not support sustainable enterprise spending, and the agent that burns a million tokens to produce an answer a human must then redo is a liability on both sides of the ledger.

The financial significance is clear. Model providers generate revenue from agent execution. Infrastructure operators benefit from increased inference demand. Semiconductor suppliers experience greater demand for efficient inference hardware. But the commercial willingness to pay originates in the value of completed work, and for the Counterparty Cascade framework I propose evaluating agentic revenue through four questions. First, does the agent perform a task that customers consider economically valuable, measured by what they would otherwise have paid a human or a vendor to do it? Second, can the task be completed reliably enough to support continued adoption, so that the customer expands rather than contracts its deployment after the first renewal? Third, does the customer retain sufficient net benefit after model usage, software integration, security, and human supervision costs, so that the deployment survives a budget review? Fourth, is the resulting revenue recurring and durable enough to justify long-term infrastructure investments, or is it the kind of experimental spending that disappears when a new chief financial officer arrives?

These questions matter because high token consumption is not automatically equivalent to strong economic value creation. An inefficient agent may consume substantial computing resources, and its consumption may show up as revenue in a model provider’s filing and as utilization in a neocloud’s investor presentation, but its customer will eventually reduce usage if the benefits do not justify the costs. The long-term financial stability of the AI economy therefore depends on productivity-adjusted demand rather than on computational activity, and the distinction between the two is the distinction between revenue that can service a twenty-year lease and revenue that cannot.

5.9. OpenAI, Anthropic, and the Economics of Competing Infrastructure Strategies

OpenAI and Anthropic offer useful examples of differing approaches to securing AI infrastructure, and the differences bear directly on how each company’s obligations would transmit stress to its counterparties. OpenAI’s Stargate initiative, announced in January 2025 with Oracle, SoftBank, and MGX, described an ambition to invest $500 billion over four years in American AI infrastructure, and subsequent announcements identified additional sites and development partnerships, including the Abilene, Texas campus that was partially operating by late 2025 and the self-developed Georgia campus described in Section 4.10.^[46] ^[44] The strategy concentrates enormous project-specific financing and construction risk in a small number of development partners, and Oracle’s experience during 2026, in which its five-year credit default swap spread surpassed its 2008 peak, its free cash flow turned deeply negative, reporting indicated that more than half of its $553 billion of performance obligations was tied to a single customer, and it announced a 2026 financing plan of $45 to $50 billion of gross proceeds using a balanced combination of debt and equity, is the clearest public demonstration of what that concentration means for a counterparty.^[31]

Anthropic’s infrastructure strategy, as described in its November 2 filing, combines multiple cloud providers with increasing use of dedicated equipment leases and neocloud capacity, so that its $895.4 billion of commitments are distributed across at least three hyperscalers, at least two neoclouds, a semiconductor lessor, and their respective financing entities.^[1] The approaches are not mutually exclusive, and both companies operate in an environment where access to advanced computing capacity is central to competitive strategy, where technical flexibility must be balanced against financial commitment, and where the future of model architectures, commercial demand, and infrastructure costs is uncertain. However, their contractual arrangements distribute risk differently. An infrastructure strategy relying primarily on cloud service commitments involves different capital requirements and delivery risks from one relying heavily on direct equipment leasing; a strategy involving large development partnerships exposes participants to project-specific financing and construction risks; a vertically integrated approach offers greater operational control but requires larger upfront capital investments. The appropriate comparison is therefore not simply the total number of gigawatts announced. It should examine total committed capital and expected payment timing; financing structures and available liquidity; the degree of dependence on individual providers; contractual delivery protections; flexibility to adopt new accelerator generations; the ability to redirect workloads; expected utilization and revenue per unit of installed capacity; exposure to electricity and permitting delays; and the financial remedies available when counterparties fail to perform. These factors can materially change the risk profile of two companies with similar computing ambitions, and they can materially change the risk profile of the counterparties who have lent against those ambitions.

5.10. The Counterparty Cascade Stress Test

The industry would benefit from a standardized approach to evaluating interconnected infrastructure obligations, analogous to the supervisory stress tests that banking regulators adopted after 2009 but adapted to a system in which the institutions at risk are not banks. I propose a Counterparty Cascade Stress Test organized around five components.

Component A: Financial Exposure Mapping. Identify the material economic relationships connecting each project to customers, lenders, suppliers, lessors, developers, and utilities, and distinguish obligations legally owed by the institution from exposures arising indirectly through commercial dependencies. Van Nieuwerburgh has noted that his own ongoing work with Zhong Sun, under the title Clouds of Credit, attempts exactly this mapping of the financial web behind AI compute, and the exercise would be considerably easier if the parties to the agreements were required to disclose the map themselves.^[6]

Component B: Delivery Milestone Verification. Determine whether financing commitments, equipment deliveries, construction schedules, and electricity availability are supported by credible, independently verifiable evidence rather than by the developer’s representation. A signed contract should not be treated as proof that the associated capacity will become operational on schedule, and the Texas audit, whatever its costs, is the first state-level attempt to perform this verification at scale.

Component C: Liquidity Shock Assessment. Estimate the ability of each institution to absorb delayed revenue, increased interest costs, additional collateral requirements, or replacement procurement expenses over a twelve-month and a twenty-four-month horizon. Liquidity calculations should account for cash restrictions, debt covenants, cross-default triggers, and the actual availability of committed financing under stress rather than under the conditions prevailing when it was arranged.

Component D: Shared Dependency Analysis. Identify concentrations involving common customers, financing institutions, semiconductor manufacturing facilities, transformer suppliers, grid operators, and electricity infrastructure, including geographic and technological dependencies that are not apparent from direct contractual relationships. This is the component that converts a collection of individually prudent exposures into a portfolio view.

Component E: Recovery and Resolution Planning. Evaluate how affected participants would respond to a serious disruption, including refinancing, contract amendments, asset transfers, replacement service providers, restructuring, or orderly project cancellation, and identify in advance which institutions have the balance sheet to acquire distressed capacity. The objective is to identify options before severe financial stress eliminates flexibility, because the parties with the resources to stabilize a cascade are the same parties whose incentives to do so are weakest once it has begun.

5.11. A Proposed Counterparty Cascade Index

To support comparative research, I propose an exploratory Counterparty Cascade Index (CCI) that would measure a project’s vulnerability across five dimensions corresponding to the indicators of Section 2.4.

Dimension	Illustrative weight
Counterparty concentration	25%
Payment–delivery mismatch	20%
Financing conditionality	20%
Shared dependencies	20%
Liquidity and cross-default sensitivity	15%
Total	100%

These weights are proposed research assumptions, not empirically validated risk coefficients, and their calibration against realized project outcomes between 2027 and 2030 would itself be a valuable contribution. Each dimension would receive a standardized score from zero to 100, with higher scores indicating greater vulnerability, and the overall score would be the weighted average of the five dimensions:

CCI = 0.25K + 0.20T + 0.20F + 0.20S + 0.15L

where K is the counterparty concentration score; T is the payment–delivery mismatch score; F is the financing conditionality score; S is the shared dependency score; and L is the liquidity and cross-default sensitivity score. The index would not represent the probability of default, and it would not determine the expected financial loss. Its preliminary function would be comparative, identifying projects that deserve closer analysis, in the way that an early-warning indicator in banking supervision directs examiner attention rather than determining capital. A more sophisticated version could incorporate contractual loss-given-default estimates, probabilistic construction delays, contingent liabilities including residual-value guarantees, and financial network simulations of the kind sketched in Section 2.3.

The value of the index lies in its ability to distinguish projects whose risks differ in kind even when their nominal investment values are similar. A project might possess relatively low direct credit risk but substantial exposure to a shared semiconductor bottleneck. Another might possess adequate infrastructure but rely on a highly concentrated customer base, as every neocloud examined in this paper does. A third might face significant payment obligations before electricity becomes available, as any Texas project caught in the permitting pause now does. These risks differ, and the CCI would provide a structured way to compare them and to direct the limited attention of lenders, regulators, and researchers to the places where it matters most.

5.12. The Need for Better Disclosure

A major obstacle to evaluating interconnected AI infrastructure risk is the limited public visibility into many private contractual arrangements, and the obstacle is growing in proportion to the share of the buildout that is financed off balance sheet. Public companies disclose certain material commitments and risks through regulatory filings, and the Anthropic amendment, the Nscale registration statement, and the CoreWeave quarterly report are, whatever their limitations, the reason this paper could be written at all. However, private infrastructure developers, special-purpose financing entities, investment funds, and commercial counterparties provide much less detailed public information, and the confidential treatment granted to the dollar thresholds in Anthropic’s own credit agreement exhibits is a reminder that even public filings redact the terms that matter most for cascade analysis.^[1] This does not necessarily indicate misconduct. Confidentiality can be commercially appropriate. Nevertheless, limited visibility complicates assessments of aggregate financial exposure, and Van Nieuwerburgh’s observation that publicly rated debt represents only a relatively small share of datacenter financing, with much larger volumes provided through syndicated bank loans and private-credit facilities for which little information is publicly available, identifies the scale of the blind spot.^[6] The Wall Street Journal’s tally, cited in the same paper, that the four largest hyperscalers carry $2.4 trillion of off-balance-sheet obligations against $604 billion on balance sheet, so that only one-fifth of their datacenter-related obligations are currently recognized, suggests that the blind spot extends to the strongest credits in the system.^[6]

Improved disclosure could help distinguish sustainable industrial investment from excessive financial concentration, and useful information might include aggregate and annualized infrastructure obligations; financing commitments that remain conditional; material delivery milestones and acceptance requirements; the share of contracted capacity dependent on a limited number of customers; significant guarantees, residual-value backstops, and contingent reimbursements; cross-default provisions that could materially affect liquidity; exposure to shared infrastructure bottlenecks; expected equipment replacement and depreciation assumptions; and the treatment of canceled or undelivered capacity. Not every item would need to be disclosed publicly at the individual-contract level. Regulators could collect confidential information where appropriate, as banking supervisors do. Lenders could establish standardized due diligence requirements, as the rating agencies have begun to do for the small share of the market they rate. Investors could seek more consistent presentation of material exposures. The objective would be to improve the visibility of financial interdependence without unnecessarily revealing commercially sensitive information, and the IMF’s October 2026 recommendation that policymakers use existing tools, including capital and concentration requirements, to contain the buildup of risk and to encourage broader participation in financing the AI ecosystem presupposes that the exposures can be seen.^[22]

5.13. Policy Recommendations for 2027–2030

A successful response to Counterparty Cascade should preserve the benefits of infrastructure investment while reducing unnecessary financial fragility, and it should do so without substituting public guarantees for private discipline. I propose five policy priorities.

First, improve the verification of large infrastructure commitments. State governments, utilities, and financing institutions should distinguish financially supported developments from speculative proposals, and verification should focus on actual contractual commitments, financing availability, project milestones, and infrastructure readiness. The Texas audit and Pennsylvania’s binding GRID commitments are early, imperfect versions of this discipline, and their results should be studied rather than dismissed.

Second, improve the allocation of costs associated with dedicated electricity infrastructure. Regulators should evaluate whether large-load customers provide appropriate financial security for infrastructure constructed primarily to serve their projects, following the direction set by FERC’s June 2026 orders, Virginia’s August 2026 SCC decision, and the Georgia and Indiana utility agreements, and the objective should be to prevent unreasonable transfers of project-specific costs to households and smaller businesses.

Third, encourage transparency in interconnected financing structures. Financial supervisors should monitor concentrations involving major AI customers, infrastructure developers, equipment collateral, and shared nonbank financing sources, and they should extend to private credit the exposure reporting that the Federal Reserve began requesting from banks in April 2026.^[23]

Fourth, strengthen contractual resilience. Commercial participants should evaluate milestone-based payments, completion protections, financing contingencies, replacement rights, and mechanisms for addressing prolonged infrastructure delays, and they should recognize that a termination right which protects one party may simply relocate the loss to another.

Fifth, support technological flexibility. The ability to redirect computing workloads, improve hardware utilization, curtail load during grid stress, and adapt infrastructure to new accelerator generations reduces the financial consequences of unexpected technological change, and the Georgia demand-response commitment demonstrates that flexibility can be contracted for rather than merely hoped for.

These priorities do not require the government to guarantee AI infrastructure investments. Indeed, excessive guarantees could encourage imprudent risk-taking if investors expect public institutions to absorb losses, and the history of every prior infrastructure boom suggests that the expectation of rescue is itself a cause of the overbuilding that eventually requires it. The more appropriate objective is to support transparent markets in which private participants understand and bear the risks they choose to undertake.

5.14. What Could Make Counterparty Cascade a Major Research Subject by 2030?

The relevance of the concept will depend on several observable developments, and it is worth stating them in advance so that the framework can be judged against them. One would be a substantial increase in project-level financing involving specialized AI infrastructure, which the neocloud filings of 2026 suggest is already underway. Another would be greater use of equipment leases, asset-backed structures, and private credit to finance computing capacity, which the NVIDIA platforms and the Broadcom lease arrangements make nearly certain. A third would be more extensive contractual relationships between frontier model companies and several competing infrastructure providers, which the Anthropic filing documents. A fourth would be greater regulatory attention to the financial security of large electricity customers, which four governors and a federal commission supplied in a single summer. A fifth would be an increase in the number of infrastructure projects facing refinancing, delivery, or utilization challenges, which is the one development on this list that has not yet occurred at scale.

These developments would not prove that a systemic crisis is approaching. They would indicate that financial interdependence has become sufficiently important to warrant more systematic measurement, and the research should also examine stabilizing trends with equal rigor. Greater enterprise AI adoption could strengthen revenue. Improved computing efficiency could increase returns on installed infrastructure. More diversified financing could reduce dependence on individual institutions. Better regulation could lower construction uncertainty. Stronger contractual protections could contain losses. The Counterparty Cascade framework is therefore not simply a theory of financial fragility. It is a theory of how financial obligations transmit both stress and resilience through the AI economy, and the same network that can carry a shock from a Texas substation to a New York credit fund can carry a hyperscaler’s balance sheet to a neocloud that would otherwise fail.

Section 6: What Have We Learned? Seven Pillars of the Counterparty Cascade

The November 2, 2026 Anthropic disclosure provides a revealing perspective on the financial structure supporting the expansion of artificial intelligence, and its significance extends beyond the extraordinary value of its contractual commitments. The filing demonstrates how computing infrastructure investment increasingly involves multiple institutions whose financial obligations, operational schedules, and commercial expectations must remain sufficiently aligned for the system to function effectively, and it arrived in the same fortnight as the third-quarter results that showed the four largest hyperscalers spending beyond their cash flows, the IMF and BIS analyses that named the ecosystem’s circularity as a systemic risk, and the academic work that measured, for the first time, how large the buildout is relative to every previous American infrastructure boom. The analysis also reveals that financial interdependence is not inherently undesirable. Long-term commitments provide the revenue predictability necessary to finance infrastructure. Equipment leasing improves access to advanced computing resources. Strategic partnerships distribute risks and accelerate construction. Private credit mobilizes capital that traditional financing arrangements would not provide at comparable speed. However, these arrangements become vulnerable when the financial commitments of different institutions depend on assumptions that are not realized simultaneously. From these findings, I propose seven fundamental pillars.

Pillar 1 — Contractual Interdependence Has Become a Defining Feature of the AI Economy

The first lesson is that the AI infrastructure industry can no longer be understood exclusively through semiconductor shipments, datacenter capacity, or hyperscaler capital expenditures. It must also be analyzed through the contractual relationships connecting the organizations responsible for financing, constructing, supplying, operating, and consuming computing infrastructure. The largest frontier model developers increasingly depend on several cloud providers, specialized infrastructure operators, semiconductor suppliers, and financing institutions at once, and the resulting relationships connect organizations that would traditionally have operated within clearly separated industries. A semiconductor company has become a financing partner and a residual-value guarantor. A cloud provider has become an equity investor in its customer. An infrastructure operator depends on a frontier model developer’s long-term purchasing commitments for the collateral behind its term loans. An institutional lender finances projects whose ultimate revenue depends on commercial AI adoption it has no means of forecasting. Consequently, a decision made by one participant influences several others, and the central lesson is that an AI infrastructure contract should be evaluated not only for its individual economic value but also for the obligations and dependencies it creates throughout the broader network. This does not require assuming that every contractual relationship creates systemic risk. It requires distinguishing manageable commercial interdependence from exposures capable of transmitting significant financial distress, and by 2030 this distinction could become an essential component of AI infrastructure investment analysis.

Pillar 2 — Delivery Timing May Be as Important as Investment Size

The second lesson concerns the relationship between financial commitments and physical infrastructure delivery. Large computing systems require numerous components to become available within compatible schedules. Accelerators must be manufactured and delivered. Buildings must be completed. Electricity must be connected, which in 2026 means surviving a grid operator’s study queue, a state audit, and a cost-allocation proceeding. Networking systems must be installed. Cooling infrastructure must be operational. Software environments must be commissioned. Only then can the complete system generate its intended economic value, and yet individual contracts create payment obligations before every necessary component becomes operational. This introduces timing risk, and a financially viable project can experience distress when payments become due before productive revenue begins. The risk becomes more significant when project delays cause additional financing expenses, require replacement computing capacity at the short-term premium that Anthropic’s own filing quantifies, or activate contractual remedies. For lenders, the relevant question is therefore not simply whether a project possesses a large customer commitment but whether the schedule of available cash flows is consistent with the project’s financing obligations. For model developers, the question is whether committed computing capacity will become operational within the period required by their technical and commercial strategies. For governments and utilities, the question is whether supporting infrastructure can realistically be delivered within the assumptions underlying private investment commitments. The central principle is that financial commitments become more resilient when payment schedules, delivery milestones, and revenue generation are coordinated through realistic contractual arrangements, and this pillar suggests that the industry should develop consistent methods for measuring the financial consequences of infrastructure delays, as it has long measured the consequences of credit losses.

Pillar 3 — Financial Diversification Must Extend Beyond the Number of Counterparties

The third lesson is that apparent diversification can conceal concentrated economic exposure. An AI developer may contract with several cloud providers. A datacenter operator may serve multiple customers. A lender may finance projects across numerous states. However, these relationships share underlying dependencies: several cloud providers rely on the same semiconductor manufacturing facilities; different datacenter developers depend on overlapping financing institutions, which the August 2026 NVIDIA platforms are explicitly designed to consolidate; multiple lenders possess exposure to the same frontier model companies; and several regional infrastructure projects depend on similar assumptions regarding future enterprise AI demand. Consequently, diversification must be evaluated at the level of underlying economic dependencies, and it is insufficient to count the number of contracts. Investors should understand whether those contracts are exposed to the same financial, technological, or operational shocks. The Federal Reserve, the FSB, the IMF, and the BIS have each highlighted concerns associated with financial concentration, private credit, and interconnected AI investments during 2026, and the BIS measurement that more than half of the capital flowing into AI firms came from other AI firms is the single most important statistic in this paper, because it quantifies how much of the industry’s apparent diversity of funding is an illusion.^[23] ^[24] ^[7] ^[11] The central principle is that true diversification requires independent sources of financing, technology, infrastructure, and commercial demand, not merely a larger number of contractual relationships.

Pillar 4 — State Infrastructure Policy Is Becoming a Form of Financial Risk Management

The fourth lesson is that American state governments increasingly influence the financial structure of AI infrastructure investment, and that they have begun to do so deliberately. Governors and utility regulators make decisions affecting electricity connections, cost allocation, construction approvals, water resources, public incentives, and infrastructure requirements, and these decisions determine the economics of private projects more directly than most project-finance models acknowledge. Texas under Governor Greg Abbott has emphasized electricity demand verification, developer cost responsibility, and a permitting pause pending audit. Pennsylvania under Governor Josh Shapiro has established binding requirements intended to strengthen developer accountability and remove speculative projects from the queue. Virginia under Governor Abigail Spanberger has secured a regulatory decision assigning the cost of dedicated transmission infrastructure to the datacenters that cause it. California under Governor Gavin Newsom has enacted measures addressing electricity cost allocation, water disclosure, and environmental review. Georgia, Indiana, Arizona, and Michigan are navigating the relationship between technology investment and industrial infrastructure expansion through utility contracts, manufacturing incentives, and federal credit. These approaches differ, sometimes sharply, but they reflect a common policy concern: the economic benefits of AI infrastructure development should not depend on transferring inappropriate financial risks to households, smaller businesses, or public institutions. The same principle applies to regional transmission planning, where FERC’s June 2026 orders and its subsequent handling of the PJM backstop procurement have made large-load cost allocation and financial security a federal question.^[43] ^[36] The central lesson is that responsible infrastructure development requires clear financial accountability for the costs and risks created by large computing projects, and that such accountability can improve investment quality while reducing the likelihood that an unsuccessful private development generates long-lasting financial consequences for the surrounding community.

Pillar 5 — Private Credit and Off-Balance-Sheet Structures Have Relocated Leverage Rather Than Eliminated It

The fifth lesson, which the research of 2026 has made impossible to ignore, is that the migration of AI infrastructure financing toward leases, joint ventures, special-purpose vehicles, and private credit has expanded the industry’s funding capacity while relocating its leverage to places where it is harder to see. Van Nieuwerburgh’s analysis of the Hyperion transaction, in which a project financed roughly 90 percent with debt at a 1.12 times debt-service coverage ratio was structured so that none of its lease obligations or residual-value guarantees appeared on the tenant’s balance sheet, and his citation of Moody’s estimate that hyperscalers carry roughly $970 billion of lease commitments of which about $660 billion is not yet reflected on balance sheet, describe a system in which reported leverage systematically understates economic leverage.^[6] The Federal Reserve’s and FSB’s attention to private credit redemptions and bank interlinkages, the BIS’s warning that private credit has become a fast-growing and lightly disclosed source of AI financing, and Damodaran’s argument that debt has transformed an equity-holder’s problem into everyone’s problem all point to the same conclusion.^[23] ^[24] ^[10] ^[26] This is not an argument against off-balance-sheet financing, which serves legitimate economic purposes and which has allowed the buildout to proceed at a pace that balance-sheet financing alone could not have supported. It is an argument that the obligations such financing creates should be mapped, disclosed, and stress-tested with the same seriousness that the industry applies to the hardware it finances. The central principle is that leverage which has been moved is still leverage, and the institutions that cannot see it are the institutions most likely to be surprised by it.

Pillar 6 — The Scale of the Buildout Makes Its Financial Architecture a Macroeconomic Question

The sixth lesson is that the AI infrastructure cycle has grown large enough that its financing structure is no longer a sector question. Furman’s finding that technology investment accounted for the overwhelming majority of U.S. GDP growth in the first half of 2025, Van Nieuwerburgh’s estimate that the buildout would average 3.6 percent of GDP annually through 2032 and thereby exceed the canal, railroad, electrification, highway, and telecommunications booms as a share of the economy, the IMF’s identification of an AI correction as a salient risk to global financial stability, and the BIS’s placement of AI capital expenditure among its four pressure points together establish that the mechanisms examined in this paper operate at a scale where their failure would be felt far outside the technology industry.^[45] ^[6] ^[7] ^[10] The NPR summary that a single bank’s analysis attributed roughly two-thirds of the S&P 500’s gains since early 2024 to some forty-two AI-related companies is a measurement of how much household wealth now depends on the fifth layer paying for the first four.^[47] This does not mean that a cascade is imminent, and every institution cited here has been careful to say so. It means that the margin for error in the design of the industry’s contracts has narrowed, because the economy that would absorb a cascade is the same economy whose growth has come to depend on the investment that would be cascading. The central principle is that financial architecture becomes public infrastructure when it grows large enough, and that the design of AI infrastructure contracts is now a matter of macroeconomic consequence.

Pillar 7 — Sustainable Agentic Revenue Will Ultimately Support the Infrastructure Commitments of the AI Economy

The seventh and final lesson concerns the long-term source of financial stability. The AI economy requires substantial investment in electricity, semiconductors, datacenters, and frontier models, but these investments must ultimately generate economic returns, and the strongest foundation for those returns is recurring demand for useful intelligence. Enterprise applications, scientific systems, autonomous agents, robotics platforms, and other commercial uses of AI will play an increasingly important role, yet commercial success must be measured through economic value rather than computational activity alone. A model may consume enormous quantities of computing capacity without generating sufficient revenue. An application may achieve rapid user growth without developing sustainable margins. An autonomous agent may perform technically sophisticated operations without producing savings or productivity improvements that justify its operating costs. Conversely, reliable AI systems that measurably improve business outcomes create durable demand, and that demand supports model revenue, which supports computing contracts, which support infrastructure financing, which supports semiconductor investment and electricity development. The financial system becomes more resilient when its obligations are supported by demonstrable economic activity, and the disagreement between Brynjolfsson, who sees the productivity take-off already in the data, and Acemoglu, who expects much of the investment to be wasted, is a disagreement about whether that activity will arrive in time to service the obligations already incurred.^[20] ^[21] The central principle is that the long-term sustainability of the first four layers of the AI economy will increasingly depend on the economic productivity and commercial durability of the fifth layer, and this may become the most important financial test of the AI industry’s expansion through 2030.

Conclusion: Why Counterparty Cascade Defines the Next Financial Frontier of Artificial Intelligence

On November 2, 2026, Anthropic’s amended registration statement offered an unusually detailed view of the financial architecture supporting frontier artificial intelligence. The disclosure of approximately $895.4 billion in future computing and infrastructure commitments attracted attention because of its extraordinary magnitude, and the commentary that followed within hours concentrated, understandably, on the size of the number and on what it implied for the valuation of the company that disclosed it. But the deeper significance lay in the contractual relationships underlying those commitments. The filing described long-term cloud purchasing obligations with minimum spending floors, dedicated infrastructure arrangements that separate the building from the hardware, equipment leases assigned to financing entities and guaranteed by the customer, neocloud contracts whose counterparties had not yet secured the financing to perform them, contingent commitments representing nearly a quarter of the total, cross-default provisions linking commercial agreements to bank credit, and risks associated with the timely delivery of computing capacity in a system where each participant could perform and be paid while the whole remained incomplete. It also revealed that these obligations involve multiple organizations performing separate functions within the same infrastructure project, and that the performance of the whole depends on the performance of each.

The implications extend far beyond Anthropic, and the sources assembled in this paper make clear that they extend far beyond any single company. As artificial intelligence infrastructure expands, technology companies increasingly depend on financial arrangements connecting semiconductor manufacturers, cloud providers, specialist datacenter operators, equipment lessors, institutional investors, utilities, grid operators, state governments, and frontier model developers. The resulting system can mobilize enormous quantities of capital, as the $500 billion NVIDIA platforms, the $27 billion Hyperion financing, and the hundreds of billions of dollars of hyperscaler bond issuance demonstrate. It can accelerate technological innovation. It can create productive infrastructure with economic benefits extending across multiple industries and, as Furman’s accounting shows, across the entire economy. However, it also generates new forms of financial interdependence. A construction delay affects equipment payments. An equipment delay affects computing capacity. A computing shortage affects model revenue. A revenue shortfall affects financial commitments. A financing withdrawal affects projects beyond the institution where the disruption originated, and the consequences can travel in either direction through the Five-Layer AI Economy, from the substation to the subscription and back again.

This is why I choose the title Counterparty Cascade. The word Counterparty captures the economic relationships through which institutions make enforceable or conditional promises to one another. The word Cascade captures the possibility that a problem originating within one institution can transmit financial and operational consequences to others, and that the magnitude of the consequences depends on the structure of the network rather than on the size of the institution where the problem began. Together, the words describe an increasingly important dimension of the AI infrastructure industry, and one that the existing tools of equity analysis, credit analysis, and supply-chain analysis each capture only in part.

The term does not presume that financial distress is inevitable, and nothing in this paper should be read as a prediction that it is. Nor does it imply that large capital commitments are inherently unsustainable; the railroads were built, the grid was electrified, and the fiber was lit, and in each case the infrastructure outlived the financial structures that built it. Instead, Counterparty Cascade provides a framework for asking whether the commitments supporting the AI economy are sufficiently transparent, appropriately financed, realistically scheduled, and supported by durable economic demand. The distinction is essential. A large financial obligation can be manageable when supported by strong revenue, adequate liquidity, and credible delivery arrangements, and Anthropic’s reported revenue growth, from $2.5 billion to $33.2 billion in nine months year over year, is the kind of growth that makes very large obligations manageable. A much smaller obligation can become destabilizing when it depends on uncertain financing, weak contractual protections, or several other institutions whose performance cannot be independently assured, and a neocloud with $2.6 billion of live contracts and $103.4 billion of total contract value is the kind of institution whose smaller obligations deserve the closer look. Financial risk is therefore not determined by headline investment size alone. It depends on the relationships among obligations, cash flows, financing conditions, and operational capabilities.

Between 2027 and 2030, the industry will increasingly be required to demonstrate that its infrastructure investments can generate sustainable economic returns. The largest technology companies may continue expanding their computing capacity, as every one of them told investors in late October that they intend to do. Semiconductor manufacturers may introduce increasingly powerful accelerators, and the Arizona fabs may come online to build them. Infrastructure developers may construct new datacenter campuses under the cost-allocation rules that governors and regulators wrote in 2026. Utilities may invest in additional electricity capacity, some of it nuclear, some of it flexible. Frontier model developers may achieve further advances in reasoning, scientific discovery, and autonomous systems, and those advances may generate the agentic revenue on which everything upstream depends. But every investment will remain connected to economic assumptions about future utilization, pricing, customer demand, and technological progress, and the critical question is whether those assumptions can support the contracts already being signed.

The financial system should be prepared for several possible outcomes, and it should prepare for them now, while the contracts are still being written and the exposures can still be mapped. The most favorable outcome would involve rapid AI commercialization, efficient infrastructure deployment, and sufficient revenue to sustain investment commitments. A more selective adjustment could involve project cancellations, changing equipment values, and refinancing challenges without creating widespread instability, and this is the outcome that a well-designed system of verification, cost allocation, and disclosure would make most likely. A more severe outcome could emerge if concentrated financial exposures, contractual defaults, and shared infrastructure dependencies amplify the consequences of an initial disruption that, on its own, would have been absorbed. The purpose of this paper is not to predict which outcome will occur. It is to provide a framework capable of distinguishing among them, and of identifying, in advance, the features of the network that would push the industry from the second outcome toward the third.

For private corporations, Counterparty Cascade encourages a more comprehensive approach to capital allocation and contractual risk, one that treats the obligations a contract creates for others as part of its cost. For startups and neoclouds, it highlights the importance of understanding the financing and operational resilience of infrastructure providers before building a business on their delivery schedules. For lenders and institutional investors, it suggests that underwriting should examine shared dependencies rather than individual customer contracts, and that a diversified book of AI infrastructure loans may be considerably less diversified than it appears. For federal policymakers, it identifies the intersection of AI investment, private credit, technological concentration, and financial stability as a single problem rather than four. For state governors, utility commissions, and local governments, it emphasizes the importance of credible development commitments and transparent infrastructure cost allocation, and it suggests that the states which solve this problem well will attract the investment that the states which solve it badly will eventually regret. For researchers, it offers a foundation for studying how financial stress propagates through complex technological systems, and a set of indicators, a stress test, and an index against which that propagation might eventually be measured.

Above all, Counterparty Cascade identifies a central tension in the next stage of artificial intelligence development. The industry is investing in a future whose economic possibilities are extraordinary, but whose commercial structure remains under construction. Many of its financial obligations extend years or decades into the future; Anthropic’s extend to 2045. Meanwhile, its technologies can improve dramatically within months, and the cost of performing a given task falls by an order of magnitude each year. This mismatch between the duration of financial commitments and the pace of technological change creates both opportunity and risk, and it is a mismatch that no previous infrastructure boom has faced in quite this form, because no previous boom financed assets whose economic life was shorter than the construction period of the buildings that housed them.

The Five-Layer AI Economy helps explain the industrial structure of artificial intelligence. Counterparty Cascade helps explain how financial obligations connect that structure, and how stress within one part of the system could affect the others. The lesson is not that investment should slow simply because financial interdependence exists. The lesson is that investment quality will increasingly depend on understanding it. The next defining question of the AI economy may not be how much capital the industry can mobilize, but how reliably its interconnected institutions can fulfill the financial promises supporting that capital.

That is the economic challenge, and the research significance, of Counterparty Cascade.

Endnotes and Research References:

[1] Anthropic, PBC. Amendment No. 1 to Form S-1 Registration Statement (Form S-1/A). U.S. Securities and Exchange Commission, filed November 2, 2026 (CIK 0002130357). Contractual obligations, Anthropic Direct infrastructure arrangements, equipment lease and guarantee disclosures, cloud commitments, cross-default risk factors, and operating results. SEC EDGAR — Anthropic, PBC filings; original S-1 exhibits (October 26, 2026): Exhibit 4.4 — Credit Agreement definitions; Exhibit 4.3 — 2026 Stock Purchase and Convertible Note Facility Agreement with Amazon Web Services.

[2] Ashley Capoot. “Anthropic reports $17 billion in Q3 revenue, purchase commitments of $895.4 billion in revised S-1.” CNBC, November 2, 2026. cnbc.com.

[3] Theo Nash. “Anthropic Details Long-Term Compute Commitments in Amended IPO Filing.” Unite.AI, November 2, 2026. unite.ai; and Cape Fear Advisors. “Anthropic S-1/A: Short-Term Compute and the SpaceX Line.” November 2, 2026, citing the S-1/A contractual obligations disclosure ($69.8 billion payable within twelve months; $17.4 billion and $6.6 billion short-term commitments; $213.8 billion conditional; approximately $5.0 billion of costlier short-term compute; Exhibit 10.11 Master Equipment Lease Agreement with Broadcom Inc.). capefearadvisors.substack.com.

[4] Nscale Ltd. Registration Statement on Form S-1. U.S. Securities and Exchange Commission, September 18, 2026 (Anthropic Services Agreements of August 25, 2026 providing aggregate payments of up to approximately $44.6 billion; $3.1 billion convertible loan note subscription including $1.0 billion to NVIDIA). sec.gov.

[5] “Nscale’s S-1 shows it still needs financing for $44.6B Anthropic deal.” The Stack, September 2026 ($103.4 billion total contract value; $2.6 billion live; approximately 461,000 GPUs active or contracted, around 25,000 live). thestack.technology.

[6] Stijn Van Nieuwerburgh (Columbia Business School). “Financing the AI Buildout.” Brookings Papers on Economic Activity, Fall 2026 conference draft, September 24–25, 2026 (1 GW campus cost of about $41 billion; 183 GW completed by 2032; investment averaging 3.63 percent of GDP; hyperscaler capex exceeding operating cash flow in 2026; Hyperion/Beignet structure; required mature revenue of $3.7 trillion; NVIDIA 25 percent residual value guarantee; Moody’s lease commitment estimates). brookings.edu.

[7] International Monetary Fund. Global Financial Stability Report: Resilience, Risk, and Transformation, Chapter 1, “Beneath the Resilience: Rising Risks, Unresolved Vulnerabilities,” including the background analysis “Rising Leverage in the AI Ecosystem.” October 13, 2026. imf.org.

[8] Athanasios Vamvakidis and Jason Wu, IMF Blog accompanying the October 2026 GFSR, as quoted in Vera Syn, “IMF Warns AI Market Correction Could Trigger Global Financial Turmoil,” Unite.AI, October 2026. unite.ai.

[9] Yingyuan Chen, Deepali Gautam, and Jason J. Wu. The AI Ecosystem: Structure, Financing, and Financial Stability Considerations. IMF Technical Notes and Manuals, October 2026, as summarized in “The Hidden Risks of AI Boom: IMF Warns Trillions in Investment Could Shake Global Markets,” Devdiscourse, October 2026. devdiscourse.com; asset-life sensitivity reported in “AI Boom Hidden Risks: IMF Warns of Trillion-Dollar Investments,” Hakky, October 2026. book.st-hakky.com.

[10] Bank for International Settlements. Annual Economic Report 2026, June 28, 2026 (AI capital expenditure as one of four pressure points; circular financing; private credit as a lightly disclosed funding source). bis.org; Iñaki Aldasoro, Sebastian Doerr, and Daniel Rees, “Financing the AI boom: from cash flows to debt,” BIS Bulletin No. 120, January 15, 2026; coverage: “BIS warning: Asia’s data centre hosts sit downstream,” TechWire Asia, June 30, 2026. techwireasia.com.

[11] Jon Frost, Kansal, Rishabh, Shreeti, and Zhang. “Circular relationships among AI firms.” BIS Bulletin No. 137, October 1, 2026 (55.2 percent of investment value into AI firms, 2021–2025, from other AI firms), as reported in “Who’s paying whom? The circular money machine behind the AI boom (or bubble),” Phys.org / The Conversation, October 2026, and in the Reserve Bank of Australia’s October 2026 Financial Stability Review. phys.org.

[12] CNBC Staff. “Anthropic IPO live updates: AI lab’s S-1 filing.” CNBC, October 26, 2026 (noncancelable purchase commitments of $54.6 billion at year-end 2025 and $265.6 billion at June 30, 2026; consumption-based revenue of 80 percent for the six months ended June 30, 2026; Microsoft commitment of at least $44.9 billion). cnbc.com.

[13] Amazon.com, Inc. Third Quarter 2026 Results and earnings call, October 29, 2026 (net sales $201.9 billion; AWS revenue $46.0 billion, up 39 percent; cash capital expenditures $75.9 billion, purchases of property and equipment $77.3 billion; 2026 cash capex of approximately $220 billion; AWS backlog $578 billion; $9.5 billion prior-year Anthropic gain; CEO Andy Jassy on capacity being “largely sold out” for 2027). Transcript highlights: MarketBeat. marketbeat.com; Implicator.ai. implicator.ai.

[14] Alphabet Inc. Third Quarter 2026 Results and earnings call, October 28, 2026 (revenue $127.6 billion; capital expenditures $51.4 billion; free cash flow negative $9.1 billion; 2026 capex guidance $195–205 billion maintained; significant increase expected in 2027; CFO Anat Ashkenazi). Summaries: FourWeekMBA, “Google Cloud Revenue Doubles to $30.7B as Capex Hits $51.4B.” fourweekmba.com; XTB, “Alphabet disappoints Wall Street despite record results” (bond issuance $28.7 billion; equity issuance $12.4 billion). xtb.com.

[15] NVIDIA Corporation. “NVIDIA Partners with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital.” Press release, August 10, 2026. nvidianews.nvidia.com; commentary: StorageNewsletter, August 12, 2026. storagenewsletter.com.

[16] CoreWeave, Inc. Quarterly Report on Form 10-Q for the quarter ended June 30, 2026 (delayed draw term loan facilities of $13.6 billion; senior and convertible notes of $16.6 billion; OEM and software license financing arrangements). sec.gov.

[17] CoreWeave, Inc. “CoreWeave Reports Strong Second Quarter 2026 Results.” Press release, August 11, 2026 (revenue $2.575 billion; net loss $626 million; interest expense $640 million; revenue backlog approximately $104 billion; Vera Rubin NVL72 bring-up). coreweave2025ipo.q4web.com.

[18] “Anthropic has signed a $35bn cloud deal with Nvidia-backed Lambda.” The Next Web, September 1, 2026, citing The Wall Street Journal and Reuters (Lambda lease on Hut 8’s Nueces County, Texas facility; NVIDIA’s separate arrangement with Hut 8). thenextweb.com.

[19] Office of Virginia Governor Abigail Spanberger. July 23, 2026 news release on the administration’s SCC filing. governor.virginia.gov; Josh Janney, “SCC orders Dominion to develop policy shifting transmission costs to data centers,” Virginia Business, August 5, 2026. virginiabusiness.com; “Virginia to require data centers to pay for new transmission infrastructure,” WHSV/WDBJ, August 5, 2026. whsv.com; Adam Aton, “Virginia mulls shifting major grid costs to data centers,” E&E News, July 16, 2026 ($2.8 billion and $6.2 billion ratepayer figures). eenews.net; WJLA, “Spanberger: Data centers should pay for major electrical infrastructure,” August 2026. wjla.com.

[20] Erik Brynjolfsson (Stanford University, Digital Economy Lab). “The AI productivity take-off is finally visible.” Financial Times, February 15, 2026, as reported by Jason Ma, “One of Stanford’s original AI gurus says productivity liftoff has begun after doubling in 2025,” Fortune, February 15, 2026. fortune.com.

[21] Daron Acemoglu (Massachusetts Institute of Technology), interview with Bloomberg Businessweek, as reported in “MIT Professor Daron Acemoglu Sounds the Alarm on AI Frenzy,” Open Data Science, October 3, 2024. opendatascience.com; and Nick Lichtenberg, “Nobel Laureate Daron Acemoglu on the ‘brainless’ AI discourse,” Fortune, June 21, 2026. fortune.com.

[22] Kristalina Georgieva and IMF staff. “The World Can’t Afford to Get AI Wrong.” IMF Blog, October 15, 2026. imf.org; see also “IMF warns AI investment boom could threaten global financial stability,” Telangana Today, October 16, 2026 (investment-grade issuance share rising more than tenfold). telanganatoday.com.

[23] Board of Governors of the Federal Reserve System. Financial Stability Report, May 8, 2026 (private credit redemption pressures described as “limited and manageable”; nonbank–bank interconnectedness; April 2026 requests to banks for private credit exposure data). federalreserve.gov; coverage: “Fed Sees Private Credit Redemptions as ‘Manageable’ Risks,” Bloomberg Government, May 8, 2026. news.bgov.com; “Federal Reserve asks U.S. banks for exposure to private credit firms,” Fortune, April 10, 2026. fortune.com.

[24] Financial Stability Board. Report on Vulnerabilities in Private Credit. May 6, 2026. fsb.org.

[25] Michael S. Barr, Governor, Federal Reserve Board, interview with Bloomberg News, as reported by Reuters, “Stress in private credit could spark ‘psychological contagion,’ Fed’s Barr tells Bloomberg News,” May 3, 2026. Reuters via SoCast.

[26] Aswath Damodaran (New York University, Stern School of Business). “Data Update 7 for 2026: Debt and Taxes.” Musings on Markets, 2026, as summarized by Edward Conard’s Macro Roundup. edwardconard.com.

[27] Microsoft Corporation. Fiscal Year 2027 First Quarter Results and earnings call, October 28, 2026 (Azure revenue $33.2 billion, up 48 percent; capex including finance leases $45.2 billion; calendar 2026 capex of approximately $170 billion; third-party capacity delays; free-cash-flow-positive commitment). investinglive.com; and Microsoft Fiscal Year 2026 Form 10-K disclosure of $329.1 billion of datacenter leases not yet commenced, as reported by Reuters, “Microsoft sees strong cash flow, lower capex,” The Star, July 31, 2026. thestar.com.my.

[28] International Monetary Fund. World Economic Outlook, October 2026, Executive Summary, as quoted in “IMF Holds 2026 Growth at 3.1% as AI Investment Props It Up,” FourWeekMBA, October 2026. fourweekmba.com.

[29] Meta Platforms, Inc. Third Quarter 2026 Results, October 21, 2026 (revenue $64.6 billion; capex $34.91 billion; 2026 capex range narrowed to $135–145 billion; capex exceeding operating cash flow by $2.1 billion) and October 26, 2026 financing announcement (equity financing of up to $95 billion). FourWeekMBA, “Meta’s Q3, Explained.” fourweekmba.com; Beinsure, “Meta stock price forecast amid $95bn AI spending.” beinsure.com.

[30] Bank of America Global Research estimate of $344 billion of AI-infrastructure-linked bond issuance by early August 2026, and Bloomberg reporting on NVIDIA-disclosed financing arrangements exceeding $540 billion in 2026, as compiled in “AI build-out financing,” Wikipedia, accessed November 2, 2026. en.wikipedia.org.

[31] “Oracle’s debt and OpenAI: the key risk in the hyperscaler market.” Fortune, February 3, 2026 (2026 financing plan of $45–50 billion; Microsoft disclosure that 45 percent of future cloud growth is tied to OpenAI). fortune.com; “Oracle default risk reached a record high, and CDS surpassed the 2008 peak,” Zhitongcaijing via Webull, August 5, 2026. webull.ca; “Oracle and the AI Boom’s Hidden Debt Bomb,” Fast Company, May 23, 2026 ($553 billion of performance obligations, 55 percent tied to OpenAI). pulse.bot.

[32] International Energy Agency. Key Questions on Energy and AI. April 2026 (datacenter electricity demand up 17 percent in 2025; AI-focused facilities up 50 percent; consumption projected to double from 485 TWh to 950 TWh by 2030; per-task energy falling by an order of magnitude annually); Fatih Birol remarks. Coverage: Jonathan Spencer Jones, Enlit, April 17, 2026. enlit.world; W.Media, April 20, 2026. w.media.

[33] North American Electric Reliability Corporation. 2025 Long-Term Reliability Assessment, released January 29, 2026 (summer peak demand growth of more than 224 GW over ten years; 13 of 23 assessment areas at elevated or high risk; ERCOT and PJM datacenter load); NERC Level 3 Alert on computational loads, May 2026; Large Loads Working Group gap assessment, March 2026. Coverage: POWER Magazine, “NERC Warns Long-Term Grid Reliability Risks Mounting.” powermag.com; Power Engineering, “NERC warns reliability risk is rising as load growth outpaces infrastructure.” power-eng.com; EE Power, “NERC Warns AI Data Centers Threaten Grid Reliability,” May 20, 2026. eepower.com.

[34] The White House. Accelerating Federal Permitting of Data Center Infrastructure. Executive Order, July 23, 2025. whitehouse.gov.

[35] U.S. Department of Commerce, Bureau of Industry and Security. “Department of Commerce Revises License Review Policy for Semiconductors Exported to China.” January 13, 2026; Revision to License Review Policy for Advanced Computing Commodities, 91 Fed. Reg. 1684 (January 15, 2026). bis.gov; govinfo.gov; Covington & Burling client alert, January 2026 (25 percent tariff and 50 percent cap). cov.com.

[36] Federal Energy Regulatory Commission. Order Accepting Reliability Backstop Procurement and Establishing Further Procedures, PJM Interconnection, L.L.C., Docket Nos. ER26-3380-000 and EL26-108-000, 196 FERC ¶ 61,245, September 29, 2026. pjm.com; U.S. Department of Energy, “Energy Department Supports FERC’s Call for PJM to Protect Ratepayers and Build More Reliable Power,” October 9, 2026. energy.gov; Ethan Howland, “DOE urges PJM to revise backstop procurement to prevent cost shifts,” Utility Dive, October 2026. utilitydive.com; “FERC Rejects PJM’s Curtail-First Plan for New Large Loads,” mGrid, October 10, 2026. mgrid.org.

[37] Office of the Texas Governor. “Governor Abbott Directs TCEQ to Halt Data Center Permits.” September 21, 2026, as reported by BIC Magazine. bicmagazine.com; Beveridge & Diamond, “Texas Pauses All Environmental Permitting for Data Centers” (June 10, August 3, and September 14 directives; ERCOT report due December 2026). bdlaw.com; DLA Piper, “Governor Abbott expands data center permitting pause,” September 25, 2026. dlapiper.com.

[38] Commonwealth of Pennsylvania, Office of Governor Josh Shapiro. “Governor Shapiro Signs Executive Order Demanding Data Center Developers Comply with Strict Requirements and Blocking Speculative, Irresponsible Data Center Projects.” Executive Order 2026-05, Protecting Pennsylvania Consumers from Data Center Impacts, August 18, 2026. pa.gov; Greenberg Traurig, “Gov. Shapiro Issues Executive Order 2026-05,” August 2026. gtlaw.com; Polsinelli, “Pennsylvania Executive Order Changes the Regulatory Landscape for Data Center Development” (PUC model large-load tariff, May 12, 2026). polsinelli.com.

[39] Office of California Governor Gavin Newsom. “Governor Newsom signs most comprehensive data center laws in the nation, providing communities more control on water, electricity, and land use.” September 21, 2026 (AB 2383, AB 2469, AB 2619, AB 1577, SB 886, SB 887, SB 1168). gov.ca.gov; CalMatters, “New California laws target data centers,” September 2026. calmatters.org; Bloomberg Law, “Newsom Signs Bills Regulating Data Centers’ Water, Energy Use,” September 21, 2026. news.bloomberglaw.com.

[40] Arizona Commerce Authority. “TSMC Announces Additional $100 Billion Investment In Arizona.” July 16, 2026 (total Arizona investment of $265 billion; Governor Katie Hobbs remarks; 30 percent of 2-nanometer and more advanced capacity in Arizona). azcommerce.com; City of Phoenix, “TSMC Announces Additional $100 Billion Investment In Arizona.” phoenix.gov; AZFamily, “TSMC adds $100B to Arizona plans as leaders call it a chipmaking ‘gold rush’,” July 27, 2026. azfamily.com.

[41] U.S. Department of Energy, Loan Programs Office. “DOE Approves Sixth Loan Disbursement to Restart the Palisades Nuclear Plant.” September 16, 2025 (up to $1.52 billion loan guarantee; 800 MW; $491 million disbursed to date). energy.gov; Nuclear Engineering International, “Funding continues for Palisades restart.” neimagazine.com; Power Engineering, “DOE closes on $1.5B loan to reopen Palisades nuclear plant.” power-eng.com.

[42] Amazon. “Amazon to invest $15 billion in Northern Indiana for new data centers.” November 24, 2025 (2.4 GW of capacity; NIPSCO energy agreement with projected $1 billion in customer savings over 15 years; Governor Mike Braun remarks). Coverage: The Journal Gazette. journalgazette.net; ABC57. abc57.com.

[43] Federal Energy Regulatory Commission. “FERC Launches Aggressive Targeted Action to Speed Large Load Integration.” News release and six Section 206 show-cause orders to PJM, MISO, SPP, CAISO, ISO-NE, and NYISO, June 18, 2026; Chairman Laura V. Swett remarks. Summaries: Day Pitney, June 23, 2026. daypitney.com; Vinson & Elkins, “FERC Institutes Six Simultaneous Section 206 Proceedings.” velaw.com; NAES summary with FERC quotations. naes.com.

[44] Georgia Power Company (Southern Company). “Georgia Power to serve OpenAI project in Effingham County.” Press release, July 22, 2026 (25-year agreement; approximately 3,200 MW; up to 1,000 MW of flexible demand response; OpenAI to pay full infrastructure costs). southerncompany.mediaroom.com; Georgia Public Service Commission approval, August 27, 2026: “Georgia Power approves OpenAI contract in Effingham County,” FOX 5 Atlanta. fox5atlanta.com.

[45] Jason Furman (Harvard University), analysis posted September 27, 2025, as reported by Nick Lichtenberg, “Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says,” Fortune, October 7, 2025. fortune.com; see also James Knightley, “How much is AI contributing to US economic growth?” ING THINK, August 19, 2026. think.ing.com.

[46] OpenAI. “Announcing The Stargate Project.” January 21, 2025. openai.com.

[47] Greg Rosalsky. “What if an AI bubble bursts — and that accelerates AI job losses?” NPR Planet Money, October 20, 2026 (J.P. Morgan analysis attributing roughly two-thirds of S&P 500 gains since early 2024 to 42 AI-related companies). npr.org

Featured Posts