Introduction: The Death of the Simple Purchase Contract

Drive west out of Abilene, Texas, past the wind turbines that made this stretch of scrubland briefly famous in the last energy boom, and the horizon breaks in a way it did not three years ago. Where mesquite and caliche once ran uninterrupted to the Callahan Divide, a lattice of steel, substations, and switchyards now spreads across more than a thousand acres. This is the flagship campus of the Stargate program, and by the spring of 2026 it was live with roughly 1.2 gigawatts of Oracle Cloud Infrastructure capacity, with the broader program expanding toward nearly seven gigawatts of planned capacity and more than $400 billion in committed investment across sites in Texas, Michigan, Wisconsin, Wyoming, New Mexico, and Pennsylvania.[11] The trucks that queue at the gates carry rack-scale computing systems worth more than the annual budgets of some American states. The electricity that the campus will ultimately draw rivals the peak demand of a major metropolitan area.

Now ask a deceptively simple question: who bought this?

The land is held through developers and joint ventures. The buildings were financed by a labyrinth of project-finance lenders, with some of the largest single-facility technology debt packages ever assembled—including a $16.3 billion financing for a single Michigan campus that closed only after a bond manager anchored roughly $10 billion of it when large U.S. banks stepped back.[14] The computing hardware inside is designed by one set of companies, manufactured by another, and operated by a third. The anchor tenant, OpenAI, has committed roughly $300 billion over five years to Oracle for capacity it does not yet fully need, cannot yet fully pay for out of operating cash flow, and will consume on hardware that its own chip suppliers have partially financed.[10] Oracle, in turn, borrowed tens of billions of dollars to build what its customer promised to rent, prompting Moody’s to reach for an unusual analogy:

“effectively one of the world’s largest project financings” — Moody’s Investors Service, on the Oracle–OpenAI Stargate arrangement [12]

There is no single buyer in Abilene. There is no single seller. There is, instead, an interlocking arrangement: a braid of equity, debt, chips, land, power, software, and political permission in which every party is simultaneously customer, supplier, investor, creditor, and dependent. The simple purchase contract—the clean, arm’s-length transaction in which a buyer pays a seller for a product and the relationship ends at delivery—is dead at the frontier of artificial intelligence. This paper is about what replaced it.

The Multi-Asset Reality

Consider the transaction that, more than any other single deal, crystallizes the new form. On July 22, 2026, Advanced Micro Devices and Anthropic announced a strategic partnership under which Anthropic will deploy up to two gigawatts of AMD Instinct MI450 Series GPUs in AMD Helios rack-scale systems, with the first gigawatt scheduled to come online in the first half of 2027—and under which AMD committed to make a strategic equity investment of up to $5 billion in Anthropic as deployment milestones are met.[1] The agreement covers tens of billions of dollars of hardware. It spans integrated racks—MI455X accelerators, EPYC “Venice” processors, Pensando networking, and the ROCm software stack—rather than standalone chips. It includes a joint engineering program in which Anthropic’s Claude models are used to optimize workloads for AMD’s own silicon and to accelerate AMD’s software ecosystem against Nvidia’s CUDA moat. And, according to reporting at the time of the announcement, AMD was simultaneously in talks to provide a financial backstop for Anthropic’s future data-center leases—the chip vendor as credit enhancer for its own customer’s real-estate obligations.[3]

“Access to compute is central to keeping Claude at the frontier” — Tom Brown, Co-Founder and Chief Compute Officer, Anthropic [1]

“We have very much wanted to be a major part of their infrastructure” — Dr. Lisa Su, Chair and CEO, AMD [3]

Notice how many distinct asset classes and obligations one announcement contains: a hardware deployment measured in gigawatts rather than units; a contingent equity investment running from supplier to customer; deployment milestones that function as covenants; a software co-development program; site-selection collaboration, because, as Su emphasized, gigawatt-scale compute must be planned twelve to twenty-four months in advance; potential lease guarantees; and an implicit long-term capacity reservation that shapes both companies’ balance sheets into the 2030s.[3] One press release; seven instruments. This is the multi-asset reality of frontier AI procurement, and the AMD–Anthropic deal is not an outlier but a genre. AMD’s earlier agreement with OpenAI committed six gigawatts of accelerator capacity and handed OpenAI warrants for up to roughly ten percent of AMD itself, vesting as deployment, commercial, and share-price conditions are met; a subsequent AMD agreement with Meta in February 2026 covered up to six additional gigawatts with performance-based warrants tied to shipment and purchase targets.[5] The equity, remarkably, can run in either direction: with OpenAI, the customer received warrants in the supplier; with Anthropic, the supplier is investing in the customer.[4]

The wider market now consists of a dense web of such arrangements among Nvidia, AMD, Microsoft, Amazon, OpenAI, Anthropic, Oracle, CoreWeave, Meta, Google, and Broadcom—plus the emerging GPU-cloud operators, the sovereign wealth funds of the Gulf, and the governments of the United States and its allies. Microsoft and Nvidia together committed up to $15 billion of investment into Anthropic—a company whose flagship products compete directly with both investors’ own AI offerings—while Anthropic committed roughly $30 billion of Azure cloud purchases and up to a gigawatt of additional Nvidia-based capacity.[21] Google committed up to one million of its own TPUs, silicon traditionally reserved for internal use, to the same rival lab, and then, with Broadcom, expanded that arrangement to roughly 3.5 gigawatts of next-generation TPU capacity beginning in 2027.[26],[29] OpenAI’s total web of potential commitments across Microsoft, Amazon, Oracle, AMD, Nvidia, and Broadcom has been reported at approximately $1.4 trillion over multiple years.[18] No prior technology cycle—not the railroads, not telecoms, not the dot-com buildout—assembled obligations of this scale this quickly among so few counterparties.

Why This Title Was Chosen

The title of this paper is deliberate, and each of its terms carries analytical weight.

“Interlocking Arrangements” is chosen over the more familiar vocabulary of “partnerships,” “alliances,” or “reciprocal commitments” because those terms describe two-party relationships, and the defining feature of the current structure is that it is not bilateral. When AMD invests in Anthropic, the transaction is shaped by Anthropic’s simultaneous obligations to Amazon, Google, Microsoft, Nvidia, and Broadcom; when Oracle borrows to build for OpenAI, its lenders price the debt against OpenAI’s obligations to Microsoft and its financing from SoftBank; when Nvidia backstops CoreWeave’s unsold capacity, the backstop underwrites CoreWeave’s leases, which underwrite CoreWeave’s debt, which funds CoreWeave’s purchases of Nvidia hardware. The arrangements interlock the way the stones of a Roman arch interlock: each is held in place by the pressure of the others, the structure is far stronger than any of its members, and no single stone can be removed without redistributing load across the whole.

“Compute, Capital, and Sovereignty” names the three currencies of power in this ecosystem. Compute—silicon, racks, cooling, and the power to run them—is the scarce physical substrate. Capital—equity, debt, warrants, backstops, and prepayments—is the solvent that moves compute to where it is wanted. Sovereignty—the legal and political permission to build, to export, to import, and to operate—is the layer that both states and, increasingly, corporations contest. A complete account of the AI economy must hold all three in view at once, because the deals themselves do: a single agreement can allocate chips (compute), embed an equity stake (capital), and require national-security assurances (sovereignty).

“The AI Grid” anchors the analysis in physical and financial infrastructure rather than in algorithms. The metaphor is doubly apt. Like an electrical grid, the AI economy is a network of generation (chip fabrication and model training), transmission (cloud platforms and networking), and load (inference and applications), governed by capacity constraints, interconnection queues, and reliability obligations. And like an electrical grid, it is only as strong as its weakest interconnect: a failure at one node—a chip shortage, a credit downgrade, a revoked export license—propagates instantly across the web.

Method, Sources, and Plan of the Paper

This paper synthesizes primary corporate disclosures (press releases, SEC filings, and earnings materials through the second calendar quarter of 2026, including hyperscaler Q1-2026 reports and Nvidia’s results for its fiscal quarter ended April 2026), regulatory documents (most importantly the U.S. Federal Trade Commission’s January 2025 Section 6(b) staff report on cloud–AI partnerships), institutional analyses (the Bank for International Settlements’ Bulletin 120 on financing the AI boom; the International Energy Agency’s 2026 special report on energy and AI; International Monetary Fund commentary on AI-driven market concentration), and the public statements of leading academic economists, including the July 2026 joint statement organized through Stanford University’s Digital Economy Lab and signed by, among others, the MIT Nobel laureates Daron Acemoglu and Simon Johnson.[49]

The argument proceeds in seven movements. Section 1 defines the interlocking arrangement and maps its anatomy, introducing the figure of the chipmaker-as-banker. Section 2 analyzes the financial interlock: equity as currency, the capital recycling loop, equity-for-demand transactions, and lease and credit backstops. Section 3 descends to the physical layer—silicon concentration, the power bottleneck, and grid access as the new barrier to entry—and examines capacity covenants, the contractual technology that welds finance to physics. Section 4 develops the sovereignty matrix: corporate quasi-sovereignty, geopolitical alignment, export controls, and platform lock-in as a form of private governance. Section 5 assesses structural risk and market concentration, confronting the circularity question directly and examining why traditional antitrust frameworks strain against multi-asset, non-merger collaborations. Section 6 proposes a Compute Underwriting Disclosure Standard. Section 7 distills the lessons of the buildout into seven pillars, and the Conclusion offers a forward-looking synthesis.

Throughout, the paper’s ambition is neither cheerleading nor doom-saying. The interlocking arrangement is a rational institutional adaptation to a genuinely unprecedented coordination problem—moving hundreds of billions of dollars of capital into physical infrastructure on eighteen-month timelines under radical demand uncertainty. It is also a structure that concentrates risk, obscures the origin of demand, forecloses entry, and entangles private contracts with national power in ways our disclosure regimes, accounting conventions, and competition laws were never designed to see. Both things are true at once. Holding them together is the work of this paper.

Section 1: Definition and Anatomy of the AI Grid

1.1 Defining the Interlocking Arrangement

An interlocking arrangement, as the term is used in this paper, is a multi-asset, cross-dependent alliance among two or more firms in the AI value chain that satisfies at least three of the following five conditions: (i) it bundles a commercial supply relationship (chips, cloud capacity, power, or networking) with a capital relationship (equity, warrants, prepayments, guarantees, or debt support); (ii) its obligations are contingent on operational milestones—deployment volumes, power delivery, utilization thresholds, or share-price triggers—rather than on simple delivery and payment; (iii) it materially constrains at least one party’s future procurement, architectural, or siting choices; (iv) its economic significance to at least one party is large relative to that party’s balance sheet or revenue base; and (v) its performance depends on the performance of other arrangements to which one or both parties are separately committed.

This definition is deliberately operational rather than legal. Antitrust law recognizes mergers, joint ventures, and contracts; securities law recognizes investments, guarantees, and material agreements; accounting recognizes revenue, capital expenditure, and contingent liabilities. The interlocking arrangement is none of these exclusively and all of them simultaneously, which is precisely why—as Section 5 will argue—it slips between the disciplinary boundaries of every regime designed to monitor corporate behavior. The U.S. Federal Trade Commission’s Section 6(b) study of the Microsoft–OpenAI, Amazon–Anthropic, and Google–Anthropic partnerships found exactly this hybridity: more than $20 billion in cumulative financial investment interwoven with “substantial non-monetary value exchange,” including equity and revenue-sharing rights, consultation and control rights, and exclusivity provisions that public announcements had left opaque.[41],[43]

1.2 Beyond Reciprocity: What Makes These Arrangements Different

It is tempting to treat the current wave as merely a scaled-up version of familiar reciprocal commerce—the “you buy from me, I buy from you” arrangements that have existed since firms began trading. Three features distinguish the interlocking arrangement from traditional bilateral reciprocity, and each deserves elaboration because the analytical payoff of the entire paper rests on the distinction.

First, the arrangements are multi-party by construction, not by accident. A traditional reciprocal deal is complete in itself: an airline buys engines from a manufacturer that commits to buying seats for its executives. By contrast, the AMD–Anthropic agreement is legible only within a lattice that includes Anthropic’s arrangements with Amazon (its primary cloud and training partner), Google and Broadcom (up to a million TPUs and then multiple additional gigawatts), Microsoft and Nvidia (up to $15 billion of investment against roughly $30 billion of Azure commitments), and even SpaceX, whose Memphis Colossus 1 capacity Anthropic contracted at a reported $1.25 billion per month through May 2029.[2],[21],[26] Each new arrangement is priced, structured, and publicly justified with reference to the others. The relevant unit of analysis is not the deal but the web.

Second, the consideration exchanged is heterogeneous. Traditional reciprocity trades product for product or product for cash. Interlocking arrangements trade across asset classes: cash for equity, equity for demand, chips for warrants, capacity for credit support, software engineering for silicon roadmap influence, and—at the sovereign layer—market access for security compliance. Because the consideration is heterogeneous, the arrangement’s true economics are extraordinarily difficult to observe from outside. When a supplier invests $5 billion in a customer who then purchases tens of billions of dollars of hardware, what portion of the recognized revenue is “real”? The question is not rhetorical; it is the central disclosure problem this paper addresses in Section 6.

Third, the arrangements are load-bearing for third parties who are not signatories. Oracle’s lenders, CoreWeave’s bondholders, the utilities building substations in West Texas, the municipalities issuing permits in Michigan, and ultimately the index-fund investors whose portfolios are dominated by seven AI-exposed stocks are all exposed to contracts they have never seen. The Bank for International Settlements captured the systemic dimension in early 2026: the financing of AI infrastructure is shifting from internal cash flows to debt—bonds, private credit, and asset-backed structures—which means the arrangement web is being progressively wired into the broader financial system.[46]

1.3 The Core Elements: An Anatomy

Table 1 maps the recurring components of interlocking arrangements as observed in the 2023–2026 deal record. Few arrangements contain every element; most contain at least four.

Table 1. The anatomy of an interlocking arrangement: core elements and representative examples

ElementFunction in the arrangementRepresentative examples (2023–2026)
Hardware allocationReserves scarce accelerator supply, denominated in gigawattsAMD–Anthropic 2 GW MI450; AMD–OpenAI 6 GW; AMD–Meta up to 6 GW; Nvidia–OpenAI 10 GW letter of intent [1],[5],[16]
Equity stake / warrantsAligns incentives; converts supply relationship into capital relationshipAMD up to $5B into Anthropic; OpenAI warrants for ~10% of AMD; Microsoft 27% of OpenAI PBC; Microsoft + Nvidia up to $15B into Anthropic [1],[5],[6],[21]
Cloud consumption commitmentGuarantees demand for the investor-provider’s platformOpenAI’s incremental $250B Azure commitment; Anthropic’s ~$30B Azure commitment; OpenAI’s ~$300B Oracle contract [6],[10],[21]
Capacity covenant / milestoneTies equity vesting or investment tranches to deployment, power, or utilizationAMD–OpenAI warrants vesting on deployment, commercial, and share-price conditions; AMD–Anthropic investment tied to deployment milestones [5],[1]
Lease / credit backstopInvestment-grade party supports weaker party’s real-estate or debt obligationsNvidia’s $6.3B obligation to purchase CoreWeave’s residual unsold capacity through 2032; reported AMD talks to backstop Anthropic data-center leases [22],[3]
Software co-optimizationBinds the buyer’s models to the seller’s silicon and toolchainClaude used to optimize AMD ROCm; Broadcom–Google long-term TPU supply and design agreement through 2031 [1],[28]
Power and site selectionJointly secures energy, land, and interconnectionAMD and Anthropic jointly identifying data-center sites; Stargate campuses in Texas, Michigan, Wisconsin, New Mexico, UAE [3],[11],[51]
Data / IP rightsAllocates model access, IP horizons, and governance triggersMicrosoft’s IP rights to OpenAI models through 2032; independent expert verification of any AGI declaration [6]
Sovereign permission layerEmbeds export-control compliance and government oversightStargate UAE security assurances; the January 2026 Pax Silica declarations [51],[53]

Two observations about this anatomy frame everything that follows. The first is that the elements are fungible with one another at the negotiating table: a party short of cash can pay in warrants; a party short of credit can pay in backstops; a party short of permission can pay in compliance. The second is that the elements compound: each additional element added to an arrangement raises exit costs for every party, which is why arrangements that begin as supply contracts tend, over successive renegotiations, to accrete equity, covenants, and governance rights—the Microsoft–OpenAI relationship being the canonical case, evolving from a $1 billion investment in 2019 into a 27 percent equity stake, a $250 billion consumption commitment, IP rights through 2032, and a negotiated procedure for adjudicating the arrival of artificial general intelligence.[6],[7]

1.4 The Chipmaker Becomes the Banker

The most consequential single mutation in the anatomy is the transformation of the semiconductor vendor from supplier into financier—the chipmaker as banker. The logic is straightforward once stated, though its implications are anything but.

Frontier AI laboratories have revenue trajectories that are steep but young: Anthropic’s annualized revenue run rate passed $30 billion in April 2026, more than tripling from roughly $9 billion at the end of 2025, while OpenAI’s grew from roughly $12 billion in mid-2025 toward $25 billion in early 2026.[26],[11] Impressive as these curves are, they are dwarfed by the capital requirements of the infrastructure the labs have committed to occupy—requirements measured in the hundreds of billions. The labs cannot self-finance the buildout; the capital markets, while enthusiastic, demand credit quality the labs do not yet possess; and the hyperscalers, though deep-pocketed, are competitors as often as patrons. Into this gap steps the party with the strongest incentive to see the buildout completed: the vendor whose chips fill the racks.

The vendor-financier wears at least four hats simultaneously. As equity investor, Nvidia committed up to $10 billion to Anthropic and reportedly executed more than fifty investments across the AI ecosystem in a single year; AMD committed up to $5 billion to Anthropic.[20],[21],[1] As credit enhancer, Nvidia’s obligation to purchase CoreWeave’s residual unsold capacity through April 2032—initially valued at $6.3 billion—functions economically as a co-signature on CoreWeave’s borrowing, allowing a heavily indebted neocloud to raise capital on terms it could never obtain alone.[22],[24] As demand guarantor, the same instrument converts the vendor into the customer of last resort for its own customer’s product. And as market maker, the vendor’s investment announcements move the equity prices of counterparties across the web, which in turn changes the collateral values against which the web borrows.

Why must vendors finance their largest customers? Because in a supply-constrained, capital-intensive, winner-take-most market, the vendor’s greatest risk is not that its customer defaults but that its customer decelerates—that the buildout pauses, the ecosystem consolidates around a rival architecture, and the vendor’s fabrication commitments (Nvidia alone reported supply-related purchase commitments of $119 billion as of April 2026) become stranded.[39] Financing the customer is cheaper than losing the future. This is rational. It is also, as Section 5 will examine, the precise mechanism by which the 1990s telecommunications bubble transmuted vendor optimism into systemic fragility, and the reason the circularity question cannot be waved away.

Section 2: The Frontier Capital Matrix (The Financial Interlock)

2.1 Equity as Currency

In the frontier AI economy, equity has become a medium of exchange—a currency in which compute, loyalty, and alignment are priced and settled. Understanding this requires abandoning the textbook picture in which investment and procurement are separate activities conducted by separate departments for separate reasons. In the arrangements documented here, they are a single activity.

The clearest early template was Microsoft–OpenAI. Microsoft’s cumulative investment of approximately $13.8 billion—much of it delivered not as cash but as Azure cloud credits, that is, as pre-paid claims on Microsoft’s own product—was converted, in the October 2025 restructuring, into a 27 percent stake in OpenAI Group PBC valued at roughly $135 billion, alongside an incremental $250 billion Azure consumption commitment and IP rights through 2032.[6],[7] The structure elegantly reveals the currency function: Microsoft paid substantially in its own capacity, received equity in return, and simultaneously received back a contractual promise that the equity’s issuer would spend a quarter-trillion dollars on that same capacity. Cash, where it appeared at all, was almost incidental to the exchange of claims.

The Amazon–Anthropic relationship followed the same grammar—multibillion-dollar investments paired with Anthropic’s use of AWS as primary cloud and training partner and its large-scale adoption of Amazon’s Trainium silicon—as did Google–Anthropic, which the FTC’s 6(b) study grouped with the other two as partnerships whose equity and revenue-sharing rights, consultation rights, and exclusivity provisions carried potential competitive significance well beyond their headline dollar values.[41],[43] By late 2025 the grammar had spread to direct competitors investing in one another’s ecosystems: Microsoft and Nvidia’s combined commitment of up to $15 billion to Anthropic—announced alongside Anthropic’s roughly $30 billion Azure commitment and its adoption of Nvidia’s Grace Blackwell and Vera Rubin systems—meant that OpenAI’s largest shareholder and OpenAI’s principal chip supplier were now both shareholders of OpenAI’s principal rival.[21] In a currency system, this is unremarkable: everyone holds everyone’s notes. In a competition system, it is extraordinary, and Section 5 returns to why.

The equity currency has a distinctive property: it can be issued contingently, which cash cannot. AMD’s up-to-$5-billion commitment to Anthropic is explicitly future and milestone-linked; AMD’s warrants to OpenAI vest in tranches tied to deployment, commercial, and share-price conditions through 2030; Meta’s warrants from AMD are tied to shipment and purchase targets.[1],[5] Contingent equity is the perfect instrument for an interlocking arrangement because it makes the capital relationship itself a function of the supply relationship’s performance—the covenant and the currency are the same object.

2.2 The Capital Recycling Loop

The second structural feature of the financial interlock is the closed loop through which invested capital returns, at high velocity, to the investor’s own income statement. The canonical circuit runs as follows:

Table 2. The capital recycling loop: stylized circuit

StepFlowBalance-sheet effect
1Hyperscaler or chipmaker invests cash / cloud credits / chips in AI labInvestor books an equity asset; lab books cash or prepaid capacity
2Lab signs consumption or hardware commitment with the same investorLab books a purchase obligation; investor books backlog / RPO
3Lab spends the invested capital on the investor’s productInvestor recognizes revenue; lab recognizes cost of compute
4Investor’s revenue growth supports its valuation and credit capacityInvestor raises cheaper capital against the enlarged base
5Enlarged capacity is offered to the lab, collateralized by new investmentReturn to Step 1 at larger scale

Every step is individually legitimate; the loop’s significance lies in its aggregate effect on the information content of reported revenue. When Nvidia announced its September 2025 letter of intent to invest up to $100 billion in OpenAI against OpenAI’s deployment of at least ten gigawatts of Nvidia systems—capacity equivalent to the peak electricity demand of New York City—critics immediately observed that the arrangement amounted to Nvidia bankrolling its own future sales.[16],[20] Analysts including Wedbush’s Dan Ives framed investor anxiety explicitly in terms of “circular financing” clouding the true profitability picture of the sector.[19] Goldman Sachs analysts raised the related concern that potential circular revenue from strategic investments could prove dilutive to the quality of Nvidia’s growth.[60]

The loop also runs through the cloud layer. OpenAI’s $300 billion Oracle contract obligates Oracle to deliver approximately 4.5 gigawatts of capacity—roughly the output of more than two Hoover Dams—beginning in 2027; Oracle, holding $19.8 billion of cash against $124.4 billion of debt as of November 2025, announced plans to raise $45–50 billion in calendar 2026 through debt and equity to build what OpenAI has promised to rent.[10],[13] Those data centers are filled with Nvidia systems, completing a triangle in which OpenAI’s promised spending justifies Oracle’s borrowing, Oracle’s borrowing funds Nvidia’s revenue, and Nvidia’s investment intentions support OpenAI’s capacity to promise. The Register memorably described the resulting structure as a trillion-dollar deal wheel with Nvidia at its hub.[20]

Defenders of the structure make three serious points that deserve honest statement. First, end demand is not imaginary: Nvidia’s data-center revenue reached a record $62.3 billion in its fiscal fourth quarter (ended January 2026) and $75.2 billion in the following quarter, up 92 percent year over year, with hyperscaler customers—who are not vendor-financed—still representing roughly half the total; the four large hyperscalers’ capital expenditure is funded overwhelmingly from their own operating cash flows.[37],[39],[34] Second, as the asset manager Acadian argued in a widely read 2026 analysis, cross-corporate holdings and alliances have accompanied industrialization in many countries without causing bubbles; circularity is neither necessary nor sufficient for one.[44] Third, vendor participation solves a genuine coordination failure: someone must move first in committing capital to eighteen-month construction timelines, and the party with the best information about the technology roadmap is arguably the right first mover. The circularity question, in other words, is a question—not a verdict—and Section 5 takes it up in full.

2.3 Equity-for-Demand Transactions

Within the capital matrix, one transaction type deserves its own taxonomy entry: the equity-for-demand transaction, in which a supplier invests in a company that purchases its products, or symmetrically, a customer receives equity claims on its supplier as compensation for committing demand. The 2024–2026 record contains both polarities:

  • Supplier-invests-in-customer: Nvidia’s up-to-$100-billion OpenAI letter of intent; Nvidia’s up-to-$10-billion Anthropic commitment; Microsoft and Nvidia’s joint $15 billion into Anthropic; AMD’s up-to-$5-billion into Anthropic; Nvidia’s additional $2 billion into CoreWeave, a cloud provider that is simultaneously one of its largest customers.[16],[20],[21],[1]
  • Customer-receives-supplier-equity: OpenAI’s warrants for approximately ten percent of AMD at one cent per share, vesting through October 2030; Meta’s performance-based AMD warrants.[5]

The two polarities have different economics but identical strategic function: both convert a price negotiation into a capital-structure negotiation. Instead of competing on unit price—observable, comparable, margin-destroying—the parties compete on the terms of contingent claims, which are opaque, bespoke, and accounting-flexible. Tom’s Hardware noted the elegance of the reversal in the AMD–Anthropic case: where the OpenAI deal handed the customer warrants on the supplier, “here, AMD is putting money into its buyer, becoming both supplier and shareholder.”[4] (The quotation is nine words of description that no paraphrase improves; the structural point is that the instrument bends to whichever party’s paper is more valuable at signing.)

Equity-for-demand transactions create a subtle intertemporal hazard. The supplier’s investment is booked today; the customer’s purchases are recognized as revenue today and over the near term; but the equity’s value depends on the customer’s success many years hence—success that depends, recursively, on the continued willingness of suppliers to finance it. When Nvidia’s OpenAI investment stalled in early 2026—with Jensen Huang emphasizing that the $100 billion framework had been a nonbinding letter of intent and reportedly expressing private concerns about OpenAI’s business discipline and its competition from Google and Anthropic—markets received a live demonstration of the hazard: the withdrawal of contingent capital re-prices not just the two parties but every node whose plans assumed the flow.[16],[17],[19]

“It’s their infrastructure” — Jensen Huang, Founder and CEO, Nvidia, on the first OpenAI gigawatt’s schedule [17]

By March 2026, Huang’s stated rationale for stepping back—that OpenAI and Anthropic were approaching IPOs, closing the private-investment window—coexisted with a broader recognition that the arrangement had become, in one analyst’s framing, more liability than asset as scrutiny of circular structures escalated.[57]

2.4 Lease and Credit Backstops: The Investment-Grade Umbrella

The fourth chamber of the capital matrix is the least visible and, for financial-stability purposes, arguably the most important: the use of investment-grade balance sheets to support the lease and debt obligations of structurally weaker parties. Three mechanisms dominate.

The residual-capacity backstop. Under an order form disclosed in September 2025, Nvidia is obligated to purchase CoreWeave’s residual unsold cloud capacity through April 13, 2032, in an arrangement initially valued at $6.3 billion; the underlying master services agreement dates to April 2023, meaning the backstop silently underpinned CoreWeave’s explosive growth—including its March 2025 IPO—for two years before public disclosure.[22],[23] The instrument functions like a co-signer on a loan: lenders extend credit to CoreWeave’s data-center buildout on materially better terms because the world’s most valuable semiconductor company stands behind utilization.[24] CoreWeave’s own framing celebrated the relationship:

“the scale, trust, and pivotal role CoreWeave plays” — CoreWeave spokesperson, on the Nvidia capacity agreement [22]

The SEC filing’s operative language is more clinical, granting Nvidia

“access to any residual unsold cloud computing capacity” — CoreWeave, Inc., Form 8-K, September 2025 [23]

The leaseback. Nvidia’s reported $1.5 billion arrangement to lease back up to 18,000 of its own GPUs from the cloud provider Lambda over five years converts a hardware sale into a recurring obligation running from the vendor to the customer—revenue recognized on sale, capacity repurchased over time.[25]

The lease guarantee. Reporting around the AMD–Anthropic announcement indicated that AMD was in talks to provide a financial backstop for Anthropic’s future data-center leases, situating AMD within a broader pattern in which large technology companies with investment-grade ratings back the leases or debt of AI companies that could not otherwise raise capital at favorable terms.[3] At sufficient scale, this pattern quietly transforms the credit of the entire AI buildout: facilities are financed against the guarantor’s rating rather than the tenant’s economics, which means the guarantor’s rating is, in substance if not in form, encumbered.

The systemic point was made concrete by Oracle’s experience. Financing Stargate required Oracle to assemble at least $72 billion of data-center partner debt across Michigan, Texas, Wisconsin, and New Mexico—including the record $16.3 billion Saline Township package that closed only after PIMCO anchored roughly $10 billion when U.S. banks retreated—against a balance sheet that both S&P and Moody’s had moved to negative outlook, with performance obligations of $553 billion concentrated heavily in a single counterparty.[14] The convergence analysis of the arrangement observed that Oracle was performing the function of a project-finance vehicle without the ring-fencing that traditional project finance provides: corporate credit risk and infrastructure risk collapsed onto a single investment-grade issuer.[12] Some lenders began declining Stargate-related financings where Oracle was anchor tenant, and at least one major developer, Crusoe, pivoted to Microsoft after its lenders flagged Oracle counterparty concentration—a documented case of financing constraints reshaping the competitive map of AI infrastructure in real time.[12] By March 2026, CNBC could summarize the predicament in a single headline image: Oracle was building yesterday’s data centers with tomorrow’s debt, the only major player funding the buildout primarily with borrowing while chips improved faster than buildings could rise.[15]

2.5 Founder and Executive Cases: The Human Topology of the Capital Matrix

Because interlocking arrangements are negotiated by a remarkably small set of principals, the web has a human topology worth recording. Table 3 summarizes the positions of the key executives as of mid-2026.

Table 3. Principal executives and their positions in the AI Grid (as of July 2026)

ExecutiveFirmPosition in the webRepresentative interlocks
Lisa SuAMDSupplier-financier to three rival labs simultaneouslyUp to $5B into Anthropic + 2 GW; 6 GW + ~10% warrants to OpenAI; up to 6 GW + warrants to Meta [1],[5]
Jensen HuangNvidiaHub of the wheel: supplier, investor, backstop providerUp to $10B to Anthropic; stalled $100B OpenAI LOI; CoreWeave backstop and equity (~7%); 50+ ecosystem investments [20],[16],[22]
Dario AmodeiAnthropicMulti-vendor lab; most diversified compute portfolioAWS Trainium, Google TPUs, Nvidia, AMD, SpaceX Colossus; ~$30B run rate; $380B valuation; public vision of ~100 GW industry capacity by 2028 [26],[2],[59]
Sam AltmanOpenAILargest single web of commitments (~$1.4T reported)27% owned by Microsoft; $250B Azure; $300B Oracle; 6 GW AMD; Broadcom custom silicon; Stargate JV with SoftBank/MGX [6],[10],[18]
Satya NadellaMicrosoftEquity holder in both leading labs’ ecosystems27% of OpenAI PBC (~$135B); up to $5B into Anthropic; ~$190B FY2026 capex trajectory [6],[21],[33]
Andy Jassy / Jeff Bezos (founder)AmazonPrimary patron of Anthropic; largest 2026 capex (~$200B)Multibillion Anthropic investment; Trainium at ~$20B run rate; reported talks toward up to $50B into OpenAI [32],[16]
Mark ZuckerbergMetaSelf-financed integrator with vendor warrants$115–145B 2026 capex; up to 6 GW AMD with warrants; CoreWeave offtake [31],[5],[22]
Larry EllisonOracleDebt-financed landlord of the OpenAI estate$300B OpenAI contract; ~$50B FY2026 capex; $72B+ partner debt; negative ratings outlooks [10],[14]
Michael IntratorCoreWeaveNeocloud intermediary under the Nvidia umbrellaNvidia backstop to 2032; Microsoft, Google, Meta offtake; debt-funded expansion [22],[24]
Hock TanBroadcomCustom-silicon kingmaker across both campsGoogle TPU supply/design through 2031; ~3.5 GW Anthropic routing; OpenAI custom accelerators [28],[30]

The table’s most striking property is its density: nearly every row references at least three other rows. Su finances Amodei while arming Altman and Zuckerberg; Huang backstops Intrator while investing in Amodei and negotiating with Altman; Nadella owns pieces of both leading labs; Tan supplies both. The founders and executives of the AI Grid are not merely counterparties; they are one another’s shareholders, creditors, landlords, tenants, and guarantors—often all at once. This is the financial interlock in its most concentrated human form, and it is the reason the FTC’s observation about consultation and control rights carries weight far beyond any single partnership.[41]

Section 3: The Physical Layer — Compute and Power Realities

3.1 Why the Physical Layer Comes First in Every Negotiation

For two decades, technology strategy could be written as though the physical world were an implementation detail. Software scaled on commodity servers; distribution rode the internet’s declining marginal cost curve; the binding constraints were talent and product-market fit. The AI cycle breaks that intuition comprehensively. Frontier model development is constrained, in descending order of tractability, by capital, by advanced silicon, by high-bandwidth memory, by data-center shells, and—least tractably of all—by electric power and grid interconnection. Every interlocking arrangement documented in Section 2 is, at bottom, a financial technology for allocating these physical scarcities. The finance is the shadow; the physics is the object.

The scale of the physical commitment is now visible in the aggregate statistics. The International Energy Agency reported in 2026 that the capital expenditure of the largest technology companies exceeded $400 billion in 2025 and was expected to jump by another 75 percent in 2026—such that the capital spending of just five technology companies now exceeds global investment in oil and natural gas production—while the IEA’s satellite-based tracking showed purpose-built “AI factories” more than tripling in capacity over eighteen months.[47] The Q1-2026 earnings season made the corporate arithmetic explicit: Microsoft, Alphabet, Amazon, and Meta guided combined calendar-2026 capital expenditure toward roughly $700 billion—approximately $725 billion on some tallies—up from about $410 billion in 2025, with Amazon at roughly $200 billion, Microsoft near $190 billion, Alphabet at $175–185 billion, and Meta at $115–145 billion; quarterly combined capex reached $129.8 billion in the first calendar quarter alone, up 80 percent year over year.[31],[32],[35],[34] Goldman Sachs raised its 2025–2030 combined hyperscaler capex projection to $5.3 trillion.[35] Nothing in the history of private capital formation—not the railway manias, not the electrification of the 1920s, not the telecom buildout of the late 1990s—matches this concentration of investment among so few firms in so short a window.

3.2 Silicon: Concentration at the Base of the Stack

The silicon layer is not a monopoly, but it is close enough that every participant behaves as though supply were a political question rather than a market one. Nvidia’s data-center revenue—$62.3 billion in the quarter ended January 2026 and $75.2 billion in the quarter ended April 2026, within total fiscal-2026 revenue of $215.9 billion—represents a share of AI accelerator economics without modern precedent, and the company’s disclosure of roughly $500 billion in chip bookings visibility through 2026 (excluding any finalized OpenAI arrangement) indicates how far forward the scarcity has been sold.[37],[39],[20] AMD, the principal challenger, converted its challenger status into precisely the interlocking arrangements this paper documents—6 gigawatts with OpenAI, up to 6 with Meta, 2 with Anthropic—using warrants and equity as the price of anchor tenancy for its Helios rack-scale platform.[5],[1] Broadcom occupies the third position from a different angle: as co-designer and supplier of Google’s TPUs under agreements running through 2031, and as OpenAI’s partner for custom accelerators, it has become the kingmaker of the custom-silicon path, with its CEO projecting AI chip revenue crossing $100 billion in the coming year.[28],[30]

“demand is expected to surge in excess of 3 gigawatts” — Hock Tan, President and CEO, Broadcom, on Anthropic’s 2027 TPU consumption [28]

Three features of the silicon layer feed directly into arrangement design. First, lead times: advanced packaging, high-bandwidth memory, and fabrication slots must be reserved years ahead—Nvidia’s $119 billion of supply commitments and $25.8 billion of inventory as of April 2026 are the balance-sheet shadow of these reservations—so buyers must commit early or queue behind those who did, and a shortage of high-bandwidth memory that emerged in early 2026 is expected to persist through at least the end of 2027.[39],[47] Second, software gravity: Nvidia’s CUDA ecosystem imposes switching costs that hardware price cannot overcome, which is why the AMD–Anthropic arrangement’s most strategically significant clause may be the engineering collaboration using Claude to accelerate ROCm, and why Google’s TPU deals bundle silicon with a full software and networking stack.[1],[28] Third, diversification as strategy: Anthropic deliberately runs Claude across AWS Trainium, Google TPUs, Nvidia GPUs, and now AMD Instinct—the most diversified compute portfolio at the frontier—accepting engineering overhead as the premium on an insurance policy against any single vendor’s pricing power or any single arrangement’s failure.[2],[26]

3.3 The Power Bottleneck: From Megawatts to Gigawatts

The unit of account in AI procurement has changed, and the change is diagnostic. In 2022, capacity was discussed in chips; in 2023–2024, in megawatts; by 2025–2026, every headline arrangement is denominated in gigawatts: 2 GW (AMD–Anthropic), 6 GW (AMD–OpenAI; AMD–Meta), 10 GW (Nvidia–OpenAI letter of intent; the Stargate program’s original scope), 4.5 GW (Oracle–OpenAI’s contracted delivery), 3.5 GW (Anthropic–Google–Broadcom), 5 GW (the Stargate UAE campus).[1],[5],[16],[10],[28],[51] A gigawatt is the scale of a nuclear reactor; ten gigawatts approximates the peak electricity demand of New York City.[16] When commercial contracts are denominated in reactor-equivalents, the counterparty set necessarily expands to include utilities, grid operators, state regulators, and national governments—which is precisely how the sovereignty layer of Section 4 enters through the physical layer’s door.

Dario Amodei has publicly sketched an industry trajectory toward roughly 100 gigawatts of capacity by 2028 and potentially 300 gigawatts by 2029—figures that, whatever their realization probability, define the planning envelope within which arrangements are currently negotiated.[59] The IEA’s 2026 assessment struck a more cautionary register: project pipelines are accelerating dramatically, but bottlenecks across energy supply chains have tightened; planning and regulatory systems are stretched by the application wave; and social acceptability is emerging as a genuine constraint as communities push back and affordability concerns rise.[47] The speed of the AI revolution, the agency observed, increasingly contrasts with the speed of the physical, social, and economic systems that underpin it.[47]

The financial system has begun pricing the mismatch. As Section 2 documented, the largest single-facility technology debt package ever assembled—$16.3 billion for one Michigan campus—required a bond fund’s anchor after banks hesitated over demand sustainability; hyperscaler free cash flow is compressing visibly under the capex wave, with Amazon’s trailing four-quarter capital expenditure exceeding its trailing operating cash flow for the first time since 2022; and the gap between roughly $434 billion of trailing-four-quarter hyperscaler capex and roughly $149 billion of recognized depreciation guarantees that today’s income statements carry only a fraction of today’s buildout, with the remainder arriving as a structural depreciation wave over the next half-decade.[14],[34]

“it’s going to reduce your free cash flow” — Jake Dollarhide, CEO, Longbow Asset Management, on the AI capex wave [36]

3.4 Capacity Covenants: Where Finance Is Welded to Physics

The contractual instrument that fuses the capital matrix of Section 2 to the physical realities of this section is the capacity covenant: a contract term that conditions financial flows on physical accomplishment. The 2024–2026 record displays a rich taxonomy:

  • Deployment-vesting equity. AMD’s warrants to OpenAI vest in tranches as deployment, commercial, and share-price conditions are met through October 2030; AMD’s investment in Anthropic is committed “in the future” against deployment milestones.[5],[1]
  • Power-delivery gates. Stargate financings disburse against construction and energization schedules; Oracle must deploy roughly half of an estimated $135 billion near-term buildout before revenue recognition catches up, creating one of the most acute capital-timing mismatches in modern corporate finance.[12]
  • Utilization triggers. Nvidia’s CoreWeave obligation activates only where capacity is not fully utilized by CoreWeave’s own customers, making the covenant a written option on end demand.[23]
  • Consumption floors. OpenAI’s $250 billion Azure commitment and $300 billion Oracle contract are, economically, take-or-pay floors that convert the labs’ demand forecasts into their partners’ senior obligations.[6],[10]
  • Success-contingent supply. Broadcom’s own filings note that Anthropic’s consumption of TPU capacity depends on its continued commercial success—an unusually candid acknowledgment, inside a disclosure document, that the covenant chain terminates in a variable no contract can guarantee.[26]

Capacity covenants are the reason this paper insists that the frontier AI agreement is a composite financial instrument. A contract that pays equity for gigawatts deployed, gates cash on substations energized, and floors consumption regardless of need is not a purchase order with extra steps; it is a structured product whose underlying is the physical buildout itself.

3.5 Grid Access as the Scarcest Asset

The deepest lesson of the physical layer is a reordering of scarcity. Capital is abundant—the arrangements of Section 2 exist precisely because so much of it is seeking exposure. Talent is scarce but mobile. Silicon is scarce but expanding on known fabrication roadmaps. What cannot be summoned by any amount of money on any timeline shorter than years is energized, interconnected, permitted land: grid connections, transformer capacity, water rights, transmission upgrades, and community consent. Su’s observation that a gigawatt of compute must be planned twelve to twenty-four months in advance understates the constraint for greenfield sites, where interconnection queues alone can run longer.[3] This is why site selection has migrated into the arrangements themselves—AMD and Anthropic jointly identifying data-center locations; Stargate’s geography spanning six U.S. states and the Emirates; Anthropic committing that the vast majority of its new TPU infrastructure will be sited in the United States as part of a $50 billion American infrastructure pledge.[3],[11],[27] Physical real estate with power is the barrier to entry that no software insurgency can code around, and the parties who secured it in 2024–2026 will hold it for decades. Infrastructure, not intellect alone, is the ultimate moat—the proposition to which Section 7 returns as Pillar One.

Section 4: The Sovereignty Matrix

4.1 Corporate Sovereignty: Firms That Behave Like States

Sovereignty, in classical political theory, is the supreme authority to decide within a territory. The firms at the center of the AI Grid do not possess sovereignty in that sense—they hold no monopoly on legitimate force and answer to courts and regulators. Yet across an expanding domain of decisions that were once the exclusive province of states, they now function as sovereignty’s operators: they determine which nations receive frontier compute and on what conditions; they negotiate directly with heads of state; they command capital flows that exceed the fixed-investment programs of mid-sized countries; and they administer the terms on which other companies—and other countries—may participate in the defining technology of the era.

The quantitative markers are stark. The combined 2026 capital budgets of four firms approach $700–725 billion, larger than the defense budgets of every country except the United States and China, and larger, per the IEA, than global investment in oil and gas production.[31],[35],[47] The IMF’s assessment of the U.S. economy in 2026 found growth resting on a strikingly narrow foundation—massive AI investment and elevated equity valuations doing the heavy lifting, with IMF chief economist Pierre-Olivier Gourinchas warning that these forces were masking broader vulnerabilities; the Magnificent Seven stocks alone accounted for roughly a third of the S&P 500’s weight.[48] When the macroeconomic trajectory of the world’s largest economy is a derivative of seven firms’ capex decisions, the vocabulary of private enterprise begins to strain, and the vocabulary of sovereignty—alliances, spheres, dependencies, tribute—begins to fit.

The interlocking arrangement is the constitutional instrument of this corporate quasi-sovereignty. Where states conclude treaties, the AI Grid concludes capacity covenants; where states exchange ambassadors, the Grid exchanges board observers and consultation rights; where states form defensive alliances, the Grid forms backstops and consumption floors. The FTC’s 6(b) study documented exactly the treaty-like features—control rights, exclusivity, information access—that distinguish these arrangements from ordinary commerce.[41],[43]

4.2 Geopolitical Alignment: The State Re-Enters Through the Power Cable

If Section 3 showed physics dragging utilities and regulators into the arrangements, geopolitics drags in the state itself—and the state, once inside, restructures the web around national interest. Three developments define the 2025–2026 sovereign layer.

First, the White House made the buildout a national project. The Stargate initiative—OpenAI, Oracle, SoftBank, and Abu Dhabi’s MGX, with a headline $500 billion commitment—was announced at the White House in January 2025 as a national priority, and subsequent policy discussion extended to federal loan-guarantee concepts for domestic chip fabrication; state governments layered on tax incentives and grants, collectively lowering the effective cost of capital for the entire ecosystem.[58],[11] Private arrangements now form under an explicit canopy of public sponsorship.

Second, export controls turned chip access into diplomacy. Under the U.S. tiered export-control regime, access to the highest-performance accelerators became a licensed privilege calibrated to alliance behavior: Tier 1 partners enjoy broadly unrestricted access, while Gulf states negotiated government-to-government arrangements exchanging U.S. oversight provisions for chip access.[56] The institutional architecture thickened in January 2026 with the Pax Silica declarations—Qatar signing on January 12, the UAE on January 15—an American-led technology-security framework that formalizes the terms on which allied states may host frontier compute.[53] A commercial data-center campus is now, simultaneously, an instrument of alliance management.

Third, sovereign wealth entered the capital matrix. Abu Dhabi’s MGX became a Stargate principal; SoftBank completed a $41 billion OpenAI investment in December 2025; Saudi Arabia’s Public Investment Fund launched HUMAIN as a national AI champion; and the UAE’s G42—operating under the chairmanship of national security adviser Sheikh Tahnoon bin Zayed, at the head of a reported $1.5 trillion sovereign wealth complex—partnered with OpenAI, Oracle, Nvidia, Cisco, and SoftBank to build Stargate UAE, a five-gigawatt campus whose first gigawatt facility and initial 200-megawatt cluster were on track for 2026 delivery.[58],[51],[55] The announcement rhetoric made the sovereignty framing explicit:

“sets a new standard for digital sovereignty” — Larry Ellison, Chairman and CTO, Oracle, on Stargate UAE [52]

Yet the fine print of Gulf sovereignty illustrates this paper’s core thesis better than any Western case. As the International Institute for Strategic Studies documented, G42 secured access to advanced Nvidia chips and OpenAI models only after accepting conditions that made American permission a precondition for building ostensibly sovereign infrastructure—while hedging visibly, with MGX investing in both the American stack and the European alternative through Mistral, and the UAE committing $35–59 billion to French data-center expansion.[51] G42’s chief executive Peng Xiao conceded that despite months of effort to diversify suppliers, the machines inside the UAE’s flagship campus would be

“mostly” — Peng Xiao, CEO, G42, on the share of Nvidia hardware in Stargate UAE’s first stage [54]

Nvidia. Sovereignty at the national layer, it turns out, is itself an interlocking arrangement: chips for compliance, capital for access, hedges against the hegemon—the same grammar the corporations wrote, now spoken by states.

4.3 Platform Lock-In as Governance: The Corporate-Feudal Ecosystem

The third face of the sovereignty matrix operates downward, on the firms and users inside the web. Every element of the arrangement anatomy—consumption floors, software co-optimization, custom silicon, decade-long IP rights—functions, from the buyer’s side, as a constraint on future autonomy. A lab that has committed $250 billion to one cloud, $300 billion to another, six gigawatts to one chip architecture and multiple gigawatts to a second, has not merely purchased inputs; it has constitutionally bound its own future architecture, siting, and cost structure. The FTC staff report identified precisely this dynamic in its competitive analysis: the partnerships may affect access to computing resources and engineering talent, increase switching costs for AI developers, and grant cloud providers access to sensitive information—the classic instruments by which a superior power governs a dependent one.[41],[42]

“we must guard against tactics that foreclose this opportunity” — Lina M. Khan, Chair, Federal Trade Commission, announcing the 6(b) inquiry [40]

The feudal metaphor, used advisedly, captures the structure’s political form: a hierarchy of protection and obligation rather than a market of arm’s-length exchange. The hyperscaler-suzerain provides capital, capacity, and credit enhancement; the lab-vassal provides consumption commitments, exclusivity concessions, and the prestige of frontier models flying the suzerain’s banner; the neocloud-knight holds capacity granted under the chipmaker’s backstop and owes utilization in return; and the smaller startup enters the system only by pledging fealty—cloud credits accepted, compute allocated through the web’s own channels—to one of the great houses. Even the mightiest participants accept constraint: Microsoft surrendered its right of first refusal over OpenAI’s compute in exchange for the certainty of the $250 billion commitment, and OpenAI accepted an independent expert panel’s verification over any future AGI declaration—private parties contracting over the governance of a technology with civilizational stakes, a function that in any earlier era would have belonged to public law.[6],[8] The subsequent April 2026 renegotiation capping revenue-share payments showed the constitution remains amendable—but only by the parties themselves, behind closed doors.[9]

Neutrality, meanwhile, has ceased to be a viable strategy anywhere in the stack. Chipmakers must choose which labs to arm and finance; labs must choose whose capital to take and whose silicon to marry; clouds must choose which rivals to host; nations must choose whose declarations to sign. The market structure itself now forces alignment from the silicon layer to the application layer—the proposition Section 7 records as Pillar Three.

Section 5: Structural Risk and Market Concentration

5.1 The Fragility of the Grid: How Failure Propagates

An interlocking structure distributes load; it also distributes shock. The AI Grid’s fragility is not the fragility of any single balance sheet—most of its members are extravagantly solvent—but the fragility of correlation: the web has been engineered, arrangement by arrangement, so that its members succeed together, which mathematically entails that they can only fail together. Consider the propagation channels the 2025–2026 record has already stress-tested in miniature.

The vendor-capital channel. When reports emerged in late January 2026 that Nvidia’s $100 billion OpenAI investment plan had stalled—amid internal doubts and Huang’s insistence that the letter of intent was never binding—Nvidia’s own shares fell, commentary immediately gamed the domino sequence (less Nvidia capital, less OpenAI demand, less CoreWeave utilization, fewer Nvidia chip orders), and the episode became a live referendum on whether the loop was load-bearing.[16],[17],[19] The system absorbed the shock, in part because the arrangement had been explicitly non-binding and the ecosystem had diversified—but the channel’s existence was demonstrated, not refuted.

The credit channel. Oracle’s experience showed lenders repricing an entire construction ecosystem around a single counterparty’s concentration: banks stepping back from the Michigan financing until PIMCO anchored it; developers like Crusoe redirecting to Microsoft because lenders flagged Oracle exposure; S&P and Moody’s moving to negative outlooks against $553 billion of performance obligations concentrated in OpenAI.[14],[12] A downgrade at one node reprices debt at every node that shares its tenant.

The technology-obsolescence channel. Chips improve faster than buildings rise. CNBC’s March 2026 analysis of OpenAI’s decision to route expansion toward next-generation Nvidia hardware at new sites rather than extending its flagship Stargate campus identified the mismatch squarely: multi-year construction timelines against two-year silicon generations put every debt-financed facility at risk of hosting yesterday’s compute at tomorrow’s rates.[15] Because the web’s covenants are written in gigawatts and years, an acceleration in silicon cadence is itself a systemic shock.

The depreciation channel. The hyperscalers’ roughly $434 billion of trailing capex against $149 billion of recognized depreciation defers the income-statement reckoning of the buildout; the reckoning arrives on schedule regardless of whether AI revenue does.[34] Two-thirds of Microsoft’s quarterly capex went to short-lived assets—GPUs and CPUs on five-to-six-year schedules—meaning the wave crests soon.[33]

The political channel. Export-control revisions, blacklistings, tariff-driven component inflation (Microsoft attributed roughly $25 billion of its 2026 capex to component pricing), and shifting alliance frameworks can revalue arrangements overnight, because the sovereign permission layer is embedded in the contracts themselves.[33],[56],[53]

The BIS framed the aggregate concern with institutional restraint: the shift from cash-flow financing to debt financing of the AI buildout wires these correlations into the banking system, private credit markets, and asset-backed securities structures—the same amplification channels that turned previous investment booms’ endings into financial events.[46]

5.2 The Circularity Question: Genuine Demand versus Vendor-Supported Demand

The single most contested empirical question about the AI Grid is deceptively simple to state: how much of the demand that justifies the buildout would exist without the buildout’s own financing? This paper has deliberately reserved the question for this point, after the mechanisms have been laid bare, because both the alarmist and the dismissive answers fail on the evidence.

The case for concern rests on structure and history. The structural point was put with unusual clarity by the market analyst Wayne Tassie: the most obvious warning sign is when

“a company’s customers are also its investors or its lenders” — Wayne Tassie, on circular financing in the AI sector [45]

—and that description is not an accusation but a literal account of how the Grid is wired. The historical point is the memory of the late-1990s telecommunications cycle, when vendor financing of customers who bought capacity created an illusion of demand whose correction erased trillions; economists and market watchers drew the parallel explicitly and repeatedly through 2025–2026.[60] Andrew Odlyzko, the University of Minnesota emeritus professor who has studied financial bubbles across two centuries, offered the discriminating version of the worry: the deepest-pocketed participants—Nvidia, with its more than fifty deals in a year, and Microsoft—can afford to lose their investments if AI falls short of the hype, but the same cannot be said of every node, and markets duly punished Oracle with a 30 percent quarterly drawdown on doubts about its capacity to deliver and OpenAI’s capacity to pay.[21] A Yale Cowles Foundation analysis of 2023–2025 data found many AI-related firms trading well above what realized performance supported.[45]

The case for calm rests on three genuinely strong observations. First, the largest purchasers are not vendor-financed: the hyperscalers fund the overwhelming bulk of their capex from operating cash flow, and their AI revenue disclosures—Microsoft’s AI run rate above $37 billion, AWS growing 28 percent, Google Cloud’s backlog above $460 billion, Anthropic’s tripling to a $30 billion run rate with more than a thousand customers spending over $1 million annually—describe end demand that no loop conjured.[32],[26],[28] Second, the explicitly circular flows, while headline-grabbing, remain a modest fraction of total financing: Huang’s consistent rebuttal has been that Nvidia’s investments represent a small portion of what the ecosystem must raise, and the stalling of the largest single circular arrangement—the $100 billion OpenAI plan—did not, in the event, topple the demand it was accused of fabricating.[16] Third, as the Acadian analysis argued from historical cases running from DuPont–General Motors through East Asian industrial alliances, cross-holdings and reciprocal deals have often accompanied durable, mutually beneficial industrialization; the bubble episodes involved leverage pyramids and fictitious capacity, not alliance structure per se.[44]

The honest synthesis is that circularity is neither proof of fraud nor guarantee of health; it is an information problem. Vendor-supported demand and genuine demand produce identical revenue lines, identical backlogs, and identical press releases. The difference between them is observable only in data the current disclosure regime does not require: the contractual linkage between investment tranches and purchase obligations, the share of a vendor’s revenue attributable to counterparties it finances or backstops, the utilization of capacity underlying take-or-pay floors, and the recourse structure of guarantees. Where the dot-com era’s failure was fraudulent accounting, the AI era’s exposure is lawful opacity—every individual disclosure accurate, the systemic picture nonetheless invisible. That diagnosis motivates Section 6.

5.3 The Monopolization Matrix: Barriers That Compound

Market concentration in the AI Grid is not a single barrier but a matrix of mutually reinforcing ones, and the interlocking arrangement is the loom on which they are woven together. An independent entrant seeking to train and serve a frontier model in 2026 must simultaneously overcome: a capital barrier (training-scale clusters priced in the tens of billions); a silicon barrier (multi-year allocation queues already reserved by incumbents’ capacity covenants); a power barrier (interconnection and permitting timelines measured in years, with prime sites already optioned by the web); a software barrier (CUDA gravity and the co-optimization moats being dug by Claude-on-ROCm and TPU-stack integration); a distribution barrier (enterprise channels owned by the same hyperscalers who hold equity in the incumbent labs); and a credit barrier (the investment-grade backstops available to web members but not to outsiders, which translate directly into financing cost differentials).[24],[41] Each barrier alone is surmountable; their conjunction, contractually welded, approaches closure. The FTC staff report’s understated formulation—that the partnerships “may impact access to certain inputs, such as computing resources and engineering talent”—names the mechanism; the deal record of 2025–2026 supplies its scale.[41]

Concentration also propagates upward into public markets. Seven AI-exposed names accounted for roughly a quarter of S&P 500 market capitalization on some estimates, with the Magnificent Seven generating over 40 percent of the index’s total return; the IMF explicitly conditioned its U.S. growth outlook on AI delivering the productivity and profits that markets have already priced.[48] The Grid’s concentration is thus no longer a sector question; it is a question about the risk profile of household savings and national growth accounting.

5.4 Antitrust and Regulatory Blindspots

Why have competition authorities, historically vigilant against far smaller concentrations, struggled to engage the AI Grid? The answer is not inattention—the FTC opened its 6(b) inquiry into the Microsoft–OpenAI, Amazon–Anthropic, and Google–Anthropic partnerships in January 2024 and published its staff report in January 2025—but doctrinal mismatch.[40],[41] Merger review attaches to acquisitions of control; the arrangements convey influence—board observation, consultation rights, revenue shares, IP options—while scrupulously avoiding control’s legal definition. Section 1 collusion doctrine attaches to agreements among competitors to restrain trade; the arrangements are vertical, investment-shaped, and efficiency-justified on their face. Essential-facilities and refusal-to-deal doctrines presume a monopolist withholding access; the Grid’s members grant access promiscuously—to allies, on terms. Public-utility regulation presumes stable technology and measurable cost-of-service; the Grid’s assets depreciate on silicon time. Each doctrine sees a slice; none sees the web.

Chair Khan’s parting statement upon the report’s release identified the stakes in the vocabulary this paper has adopted—partnerships that

“create lock-in, deprive start-ups of key AI inputs” — Lina M. Khan, on the FTC’s cloud–AI partnership findings [42]

—but the report bound no one, and the succeeding administration’s enforcement posture toward the partnerships remained, through mid-2026, an open question analyzed extensively in the antitrust bar.[61] Meanwhile the academic economics profession executed a notable pivot: the July 2026 joint statement organized through Stanford’s Digital Economy Lab—signed by Erik Brynjolfsson, Ajay Agrawal, and, most strikingly, the former AI-skeptic MIT Nobel laureates Daron Acemoglu and Simon Johnson—called for urgent institutional action on AI’s economic transformation.[49]

“guide AI to complement humans rather than simply imitate them” — Erik Brynjolfsson, Director, Stanford Digital Economy Lab [49]

“There’s been a notable change in the profession” — Erik Brynjolfsson, on economists’ shifting assessment of AI’s disruptive potential [50]

The statement addressed labor-market disruption rather than market structure, but its institutional significance for this paper’s argument is direct: the scholarly consensus that AI’s economic effects demand new governance instruments is now broad enough to include the field’s most prominent former skeptics. The instruments, however, remain to be designed. Section 6 proposes one.

Section 6: A Compute Underwriting Disclosure Standard

6.1 The Case for Disclosure Rather Than Prohibition

The analysis of Sections 2 through 5 supports a specific regulatory conclusion: the first-order problem posed by interlocking arrangements is informational, not behavioral. Prohibiting vendor investment, capacity backstops, or consumption commitments would forfeit their genuine coordination benefits and would in any case be evaded through structuring. What cannot currently be done by any investor, counterparty, lender, or regulator—and what could be done at modest cost—is to see the web whole: to measure, for any reporting entity, how much of its revenue, backlog, and asset value depends on counterparties it finances, guarantees, or is financed and guaranteed by. Securities regulation solved an analogous problem once before: related-party disclosure exists because transactions between affiliated entities carry information risks that arm’s-length transactions do not. The interlocking arrangement is the related-party transaction of the AI era—conducted among formally unaffiliated parties whose economic affiliation runs through contract rather than ownership.

This paper therefore proposes a Compute Underwriting Disclosure Standard (CUDS): a supplemental reporting framework, implementable by securities regulators through existing disclosure authority or adoptable voluntarily as an industry standard, comprising five schedules filed annually and updated for material changes.

6.2 The Five Schedules

Table 4. The Compute Underwriting Disclosure Standard: proposed schedules and required fields

ScheduleTitleCore required disclosures
CUDS-1Financed-Counterparty RevenueRevenue and backlog/RPO attributable to counterparties in which the registrant holds equity, warrants, or convertible instruments, or to which it has extended credit, prepayments, or guarantees; stated in dollars and as a percentage of totals
CUDS-2Capacity Covenant RegisterAll agreements denominated in deployment capacity (MW/GW), with milestone structure, vesting or disbursement triggers, take-or-pay floors, and remaining term; aggregate committed capacity versus energized capacity
CUDS-3Backstop and Guarantee ExposureMaximum contractual exposure under residual-capacity purchases, lease guarantees, leasebacks, and credit support to customers, suppliers, or infrastructure partners; current utilization of backstopped capacity
CUDS-4Cloud and Procurement InterdependencyConsumption commitments given and received, by counterparty category; share of compute costs owed to parties that are also investors; share of cloud revenue received from parties in which the registrant invests
CUDS-5Circularity ReconciliationA single reconciliation, analogous to a cash-flow statement, tracing capital provided to counterparties in the period against revenue recognized from those counterparties in the same and subsequent periods

Three design principles govern the standard. Symmetry: both directions of every flow are reported, so the AMD–Anthropic arrangement appears in AMD’s CUDS-1 and CUDS-3 and in Anthropic’s CUDS-2 and CUDS-4, allowing cross-filing verification. Denomination in physics as well as dollars: because the arrangements are written in gigawatts, the disclosure must be too, or the covenant structure remains invisible. Aggregation: each schedule requires web-level totals, not merely deal-level line items, because the systemic exposure documented in Section 5 lives in the aggregate.

6.3 What the Standard Would Have Revealed

The proposal is best evaluated retrospectively. Under CUDS, Nvidia’s 2023 CoreWeave backstop would have been visible from inception rather than surfacing in a 2025 filing two years and one IPO later.[22] Oracle’s investors and lenders would have seen, in one schedule, the full $553 billion concentration of performance obligations, the partner-debt structure across four states, and the timing mismatch between disbursement and revenue recognition that Moody’s later flagged.[14],[12] Market participants debating the Nvidia–OpenAI letter of intent would have known precisely what fraction of Nvidia’s backlog assumed the arrangement’s completion—information the CFO ultimately supplied piecemeal at an investor conference.[20] And the FTC’s 6(b) team would not have required compulsory process to learn the equity, exclusivity, and consultation terms of the cloud–lab partnerships; the terms would have been on file.[41] Disclosure does not deflate genuine demand; it prices vendor-supported demand accurately. In a structure whose central risk is lawful opacity, that is the entire remedy this paper claims—necessary, insufficient alone, and available now.

Section 7: What Have We Learned? The Seven Pillars of the AI Grid

The preceding sections have moved from anatomy to finance to physics to sovereignty to risk to remedy. This section compresses the argument into seven pillars—propositions that, on the evidence of 2020–2026, now function as the operating constitution of the frontier AI economy. Each pillar is stated, grounded in the record, and pushed to its uncomfortable implication.

Pillar 1: Infrastructure Is the Ultimate Moat

Wealth and talent are secondary to securing physical power sites, long-term silicon capacity reservations, and data-center real estate. The record is unambiguous: the scarcities that determined competitive position in 2025–2026 were interconnection queues, transformer lead times, high-bandwidth memory supply, and fabrication allocation—none of which respond to capital on timelines shorter than years.[47],[3] The uncomfortable implication is that the AI industry’s hierarchy for the 2030s was substantially fixed by land, power, and silicon decisions made in 2024–2026, before the technology’s ultimate economics were known. Whoever holds the energized acre holds the future’s option; everyone else rents it from them.

Pillar 2: Capital Flows in Circular Loops

Modern AI financing relies on a closed-loop system in which investments are structurally tied to immediate infrastructure consumption commitments: the hyperscaler’s equity returns as cloud spending, the chipmaker’s investment returns as hardware orders, the backstop returns as lease-secured debt capacity.[6],[16],[22] The loop is a rational solution to a first-mover coordination problem, and the largest flows in the system—hyperscaler capex funded from operating cash—remain outside it.[34] The uncomfortable implication is informational: because looped and unlooped demand are indistinguishable in reported financials, the system’s true demand signal degrades precisely as the loop grows, and confidence in the boom becomes progressively harder to distinguish from the boom’s own financing. That is not a prophecy of collapse; it is a measurement crisis—the one Section 6’s disclosure standard exists to resolve.

Pillar 3: Neutrality Is No Longer Feasible

Labs, chipmakers, clouds, new clouds, and nations must pick alliances; the market forces integration from silicon up to deployment. Google sold a million of its own TPUs—silicon traditionally reserved for internal use—to the lab most directly threatening its flagship model, because refusing the arrangement would have ceded the custom-silicon ecosystem to rivals.[29] Microsoft invested in its own investee’s chief competitor.[21] States sign Pax Silica or negotiate around it.[53] The uncomfortable implication is that every participant’s strategic map is now drawn by others’ alliances as much as its own choices: alignment is not a strategy but a tax, levied by the structure on all who enter it.

Pillar 4: Vertical Co-Dependencies Replace Vertical Integration

Instead of one company owning the entire stack, a tightly knit matrix of sovereign corporate entities locks together to dominate the market collectively. No firm in the Grid—not even Nvidia at its hub, not even Microsoft with its 27 percent of OpenAI—owns a full stack from fab to application; each instead holds contractual claims on the layers it lacks.[6],[20] Co-dependency achieves integration’s coordination benefits while distributing its capital costs and, crucially, while evading the legal categories—merger, monopoly, single-firm conduct—through which integration has historically been governed.[41] The uncomfortable implication is that the twentieth century’s answer to concentrated industrial power presumed that power would take the form of ownership; the Grid demonstrates that it need not.

Pillar 5: Technical Debt Is Anchored in Long-Term Contracts

Software optimization and hardware commitments lasting a decade strip labs of architectural flexibility, binding future AI progress to present infrastructure choices. Claude is being engineered into ROCm; OpenAI’s models into Azure’s topology and Broadcom’s custom silicon; every gigawatt covenant is also an architecture covenant, because the racks it reserves embody specific memory hierarchies, interconnects, and software assumptions.[1],[28],[30] The obsolescence channel sharpens the bind: chips now improve faster than the facilities contracted to house them, as OpenAI’s redirection away from extending its flagship campus demonstrated.[15] The uncomfortable implication is that the research frontier of the 2030s will be shaped not only by scientific possibility but by the depreciation schedules of the 2020s: some architectures will win because the contracts already paid for them.

Pillar 6: The Chipmaker Is Now the Banker—and the Underwriter of Last Resort

The vendor-financier role documented in Sections 1 and 2 is not a transitional anomaly but a durable structural position: whoever supplies the scarce input in a capital-intensive boom is inevitably drawn into financing its purchase, guaranteeing its utilization, and backstopping its housing—because the supplier’s downside from a stalled buildout exceeds the cost of underwriting it.[22],[3],[16] The uncomfortable implication is prudential: the entities performing bank-like functions in the AI economy—extending credit, guaranteeing obligations, maturity-transforming between chip sales and decade-long backstops—are regulated as chipmakers and clouds, not as underwriters, and their exposures are disclosed accordingly, which is to say partially.

Pillar 7: Sovereignty Is Contractual All the Way Up

The same instrument that binds a startup to a hyperscaler binds an emirate to Washington: access to frontier compute is everywhere exchanged for alignment, oversight, and consumption commitments, whether the dependent party is a lab accepting consultation rights or a state accepting security preconditions on “sovereign” infrastructure that runs, in its CEO’s own word, “mostly” on one vendor’s chips.[41],[51],[54] National AI sovereignty, on the 2026 evidence, is not the opposite of dependence but a negotiated position within it. The uncomfortable implication is for democratic governance: the terms of technological sovereignty—for firms and nations alike—are being written in private agreements, disclosed selectively, and amended bilaterally, while the publics whose economies now lean on the Grid’s success hold no seat at any of its tables.

Conclusion: Synthesis of Findings

This paper began with a question posed on a Texas plain—who bought this?—and the answer, developed across seven sections, is that the question itself belongs to a vanished era. The AI Grid is not bought; it is arranged. Its capital structure is a web of equity-for-demand transactions, contingent warrants, consumption floors, and credit backstops in which every major participant is simultaneously customer, supplier, investor, and guarantor of the others. Its physical structure is a portfolio of gigawatt-denominated capacity covenants welding finance to land, silicon, and power. Its political structure is a lattice of permissions in which corporate arrangements carry sovereign conditions and sovereign ambitions are executed through corporate arrangements. The frontier AI agreement—the AMD–Anthropic deal with which this paper opened, combining two gigawatts of deployment, up to $5 billion of milestone-contingent equity, joint site selection, software co-optimization, and reported lease backstop discussions—is the representative institution of this economy, as the joint-stock company was to the mercantile era and the vertically integrated corporation to the industrial one.[1],[3]

The evidence assembled here supports three summary judgments. First, the interlocking arrangement is a rational adaptation: it solved, faster than any alternative institution could have, the coordination problem of moving roughly $700 billion of annual capital formation into physical infrastructure under radical uncertainty, and the demand it serves is substantially real—attested by hyperscaler AI revenue, lab run rates tripling within quarters, and data-center revenue at the silicon layer compounding at rates without industrial precedent.[31],[26],[39] Second, the arrangement is a risk concentrator: it correlates the fortunes of its members by construction, wires that correlation into credit markets through debt, backstops, and guarantees, and defers its income-statement reckoning through a depreciation wave that arrives on schedule regardless of revenue.[46],[14],[34] Third, the arrangement is a transparency evader—not by fraud but by form: its heterogeneous consideration, contingent structure, and multi-party topology defeat the deal-level disclosure regime through which markets and regulators are accustomed to seeing, which is why this paper’s central prescriptive contribution is informational: the Compute Underwriting Disclosure Standard of Section 6.

The Future of the Market: Oligopoly, Unwinding, or Something Stranger

Will the interlocking structure solidify into a permanent oligopoly or collapse under its own weight? The honest answer is that both trajectories are live, and the record of 2026 contains the leading indicators of each. The consolidation path is visible in the compounding barrier matrix of Section 5.3, in the $5.3 trillion of projected hyperscaler capex through 2030, and in Huang’s characterization of an AI factory buildout

“accelerating at extraordinary speed” — Jensen Huang, on the state of the buildout, May 2026 [38]

The unwinding path is visible in the stalled Nvidia–OpenAI megadeal, in bank retreat from anchor-tenant concentration, in negative ratings outlooks at the Grid’s most leveraged node, in compressing free cash flows, and in the IMF’s warning that a narrow set of AI assumptions now carries the U.S. growth outlook.[16],[14],[36],[48]

But the most probable future is neither pole; it is differentiation within the web. The Grid’s nodes are not equally fragile: hyperscalers financing from operating cash occupy a different risk universe than a debt-financed landlord with a single anchor tenant; a lab with five compute suppliers is differently exposed than a lab with $1.4 trillion of concentrated commitments.[34],[26],[18] A correction, when and if it comes, will not dissolve the Grid; it will reorganize it, transferring capacity and covenants from over-extended nodes to resilient ones—very likely increasing concentration rather than relieving it, as corrections in network industries historically have. The policy window for shaping the Grid’s governance is therefore now, while the structure remains plural, and the instruments that matter are the unglamorous ones: disclosure, cross-filing verification, guarantee-exposure reporting, and competition doctrine capable of seeing contractual integration as clearly as it sees ownership.

Final Thesis Reaffirmation

Understanding the frontier AI industry requires looking beyond the algorithm. The model weights that draw the world’s fascination are the visible tip of a structure whose mass lies elsewhere: in the equity that suppliers hold in their customers, in the gigawatts that covenants reserve against substations not yet built, in the backstops that quietly re-rate a neocloud’s debt, in the export licenses that price an emirate’s alignment, in the depreciation schedules that will bill the 2030s for the 2020s’ conviction. Compute, capital, and sovereignty are not three subjects that happen to intersect in this industry; they are three denominations of a single currency, exchanged through the interlocking arrangements that constitute the AI Grid. The paper’s thesis, reaffirmed: the defining institution of the artificial intelligence era is not the neural network but the arrangement—and until our disclosure regimes, our competition law, and our public deliberation learn to read arrangements as fluently as the Grid’s principals write them, the most consequential economic structure of the age will remain, in plain sight, unseen.

Footnotes / Endnotes

[1] Advanced Micro Devices, Inc. (Press Release; statements of Dr. Lisa Su and Tom Brown) — “AMD and Anthropic Announce Strategic Partnership to Deploy Up to 2 Gigawatts of AMD Instinct MI450 Series GPUs,” July 22, 2026. https://ir.amd.com/news-events/press-releases/detail/1292/amd-and-anthropic-announce-strategic-partnership-to-deploy-up-to-2-gigawatts-of-amd-instinct-mi450-series-gpus

[2] CNBC (Ashley Capoot et al.) — “AMD to invest up to $5 billion in Anthropic as part of computing power deal,” July 22, 2026. https://www.cnbc.com/2026/07/22/amd-anthropic-ai-chip-investment.html

[3] The Wall Street Journal via Yahoo Finance — “AMD and Anthropic Sign Major Chips-and-Investment Deal,” July 2026. https://finance.yahoo.com/technology/ai/articles/amd-anthropic-sign-major-chips-123000630.html

[4] Tom’s Hardware — “AMD to supply Anthropic with 2 gigawatts of Instinct MI450 GPUs — will invest up to $5 billion in the Claude developer,” July 2026. https://www.tomshardware.com/tech-industry/amd-to-supply-anthropic-with-2-gigawatts-of-instinct-mi450-gpus

[5] AI News (Artificial Intelligence News) — “AMD to invest up to $5 billion in Anthropic under AI infrastructure deal,” July 2026. https://www.artificialintelligence-news.com/news/amd-anthropic-ai-infrastructure-deal/

[6] GeekWire (Todd Bishop) — “Microsoft gets 27% stake in OpenAI, and a $250B Azure commitment,” October 28, 2025. https://www.geekwire.com/2025/microsoft-secures-27-stake-in-openai-in-new-deal-with-commitment-for-250b-in-azure-usage/

[7] CNBC — “OpenAI completes restructure, solidifying Microsoft as a major shareholder,” October 28, 2025. https://www.cnbc.com/2025/10/28/open-ai-for-profit-microsoft.html

[8] Data Center Dynamics — “OpenAI completes for-profit move, Microsoft given 27 percent stake and $250bn Azure contract but no longer has cloud right of first refusal,” 2025–2026. https://www.datacenterdynamics.com/en/news/openai-completes-for-profit-move-microsoft-given-27-stake-and-250bn-azure-contract-but-no-longer-has-cloud-right-of-first-refusal/

[9] CNBC — “OpenAI shakes up partnership with Microsoft, capping revenue share payments,” April 27, 2026. https://www.cnbc.com/2026/04/27/openai-microsoft-partnership-revenue-cap.html

[10] CNET via AOL — “OpenAI Needs Data Centers So Much, It Signed a $300B Deal With Oracle,” 2025. https://www.aol.com/articles/openai-needs-data-centers-much-212345856.html

[11] IntuitionLabs — “Oracle–OpenAI $300B Deal Explained: 2026 Update,” April 2026. https://intuitionlabs.ai/articles/oracle-openai-300b-deal-analysis

[12] Convergences (analysis citing Moody’s Investors Service) — “The Oracle–OpenAI Crisis Files: The $300 Billion Bet That Wall Street Can’t Quite Hold,” April 2026. https://convergences.substack.com/p/the-oracleopenai-crisis-files-the

[13] The Next Platform (Timothy Prickett Morgan) — “Oracle’s Financing Primes The OpenAI Pump,” February 2, 2026. https://www.nextplatform.com/2026/02/02/oracles-financing-primes-the-openai-pump/

[14] The Next Web — “Oracle’s $16.3B data centre financing required PIMCO to anchor $10B after US banks retreated,” May 2026. https://thenextweb.com/news/oracle-data-centre-16-billion-financing-stargate

[15] CNBC — “Oracle is building yesterday’s data centers with tomorrow’s debt,” March 9, 2026. https://www.cnbc.com/2026/03/09/oracle-is-building-yesterdays-data-centers-with-tomorrows-debt.html

[16] Bloomberg via Business Standard — “Nvidia halts plan to invest $100 billion in OpenAI after talks collapse,” January 31, 2026. https://www.business-standard.com/technology/tech-news/nvidia-halts-plan-to-invest-100-billion-in-openai-after-talks-collapse-126013100211_1.html

[17] Yahoo Finance (statements of Jensen Huang) — “Pledge to Invest $100 Billion in OpenAI Was ‘Never a Commitment,’ Says Nvidia’s Huang,” February 2026. https://finance.yahoo.com/news/openai-investment-never-commitment-nvidia-100650678.html

[18] Yahoo Finance — “Is the Stalled Nvidia–OpenAI Megadeal AI’s First Domino to Fall?,” January 31, 2026. https://finance.yahoo.com/news/stalled-nvidia-openai-megadeal-ai-131959187.html

[19] CNBC (analysis citing Dan Ives, Wedbush) — “Nvidia shares are down after a report that its OpenAI investment stalled,” February 2, 2026. https://www.cnbc.com/2026/02/02/nvidia-stock-price-openai-funding.html

[20] The Register — “AI’s trillion dollar deal wheel bubbling around Nvidia, OpenAI,” November 2025. https://www.theregister.com/special-features/2025/11/04/nvidia-openai-and-the-trillion-dollar-loop/353799

[21] Global Finance Magazine (citing Andrew Odlyzko, Professor Emeritus, University of Minnesota) — “AI’s Financial Circle Game,” February 2026. https://gfmag.com/technology/the-circle-game/

[22] CNBC — “CoreWeave’s stock rallies on disclosure of $6.3 billion order from Nvidia,” September 15, 2025. https://www.cnbc.com/2025/09/15/coreweave-stock-jumps-on-disclosure-of-6point3-billion-order-from-nvidia.html

[23] CoreWeave, Inc. — Form 8-K, U.S. Securities and Exchange Commission, September 9, 2025. https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm

[24] The Motley Fool via Yahoo Finance — “CoreWeave’s $6.3 Billion Backstop Deal With Nvidia: What It Means for Each Company,” October 2025. https://finance.yahoo.com/news/coreweaves-6-3-billion-backstop-103000258.html

[25] Data Center Dynamics — “Nvidia to purchase unsold compute capacity from CoreWeave for $6.3bn,” September 2025. https://www.datacenterdynamics.com/en/news/nvidia-to-purchase-unsold-compute-capacity-from-coreweave-for-63bn/

[26] Silicon Republic (statements of Krishna Rao, CFO, Anthropic) — “Anthropic, Google, Broadcom announce 3.5GW TPU deal,” April 7, 2026. https://www.siliconrepublic.com/machines/anthropic-google-broadcom-announce-3-5gw-tpu-deal

[27] Anthropic — “Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute,” April 2026. https://www.anthropic.com/news/google-broadcom-partnership-compute

[28] CNBC (statements of Hock Tan, CEO, Broadcom) — “Broadcom agrees to expanded chip deals with Google, Anthropic,” April 6, 2026. https://www.cnbc.com/2026/04/06/broadcom-agrees-to-expanded-chip-deals-with-google-anthropic.html

[29] LightSource — “Google Sold a Million Chips to Its Own Competition,” June 2026. https://lightsource.ai/blog/google-sold-anthropic-a-million-tpus

[30] Tom’s Hardware — “Broadcom to supply Anthropic with 3.5 gigawatts of Google TPU capacity from 2027,” April 2026. https://www.tomshardware.com/tech-industry/broadcom-expands-anthropic-deal-to-3-5gw-of-google-tpu-capacity-from-2027

[31] Gotrade News — “Big Tech Q1 2026 Earnings Power $700B AI Capex Spree,” April 30, 2026. https://www.heygotrade.com/en/news/big-tech-q1-2026-earnings-ai-capex-spree/

[32] Yahoo Finance — “Hyperscalers Hit $700 Billion in 2026 AI Spending Plans,” May 1, 2026. https://finance.yahoo.com/sectors/technology/articles/hyperscalers-hit-700-billion-2026-111243744.html

[33] Om Malik — “What I Learned about Hyperscalers’ AI Spend,” April 30, 2026. https://om.co/2026/04/30/what-i-learned-about-hyperscalers-ai-spend/

[34] Silicon Analysts — “Hyperscaler AI Capex 2026: $434B Trailing Four Quarters, D&A Lag, Debt Wave,” July 2026. https://siliconanalysts.com/analysis/hyperscaler-ai-capex-depreciation-wall-2026

[35] Quartz via Yahoo Finance (citing Goldman Sachs Research) — “Meta, Microsoft, Amazon, and Alphabet are about to spend a shocking amount of money to dominate the AI era,” June 2026. https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html

[36] CNBC (statements of Jake Dollarhide, Longbow Asset Management) — “Tech AI spending approaches $700 billion in 2026, cash taking big hit,” February 6, 2026. https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html

[37] NVIDIA Corporation — CFO Commentary on Fourth Quarter and Fiscal 2026 Results, Form 8-K, U.S. Securities and Exchange Commission, February 2026. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000019/q4fy26cfocommentary.htm

[38] CNBC (statements of Jensen Huang) — “Nvidia earnings takeaways: Data center revenue nearly doubles,” May 20, 2026. https://www.cnbc.com/2026/05/20/nvidia-nvda-earnings-report-q1-2027.html

[39] StockTitan — “Record $81.6B Q1 revenue as NVIDIA boosts dividend” (NVIDIA Form 8-K summary, fiscal Q1 2027), May 20, 2026. https://www.stocktitan.net/sec-filings/NVDA/8-k-nvidia-corp-reports-material-event-56086a88bbb4.html

[40] U.S. Federal Trade Commission (statement of Chair Lina M. Khan) — “FTC Launches Inquiry into Generative AI Investments and Partnerships,” January 2024. https://www.ftc.gov/news-events/news/press-releases/2024/01/ftc-launches-inquiry-generative-ai-investments-partnerships

[41] U.S. Federal Trade Commission — “FTC Issues Staff Report on AI Partnerships & Investments Study,” January 2025. https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study

[42] TechCrunch (statement of Chair Lina M. Khan) — “FTC says partnerships like Microsoft-OpenAI raise antitrust concerns,” January 18, 2025. https://techcrunch.com/2025/01/18/ftc-says-partnerships-like-microsoft-openai-raise-antitrust-concerns

[43] U.S. Federal Trade Commission, Office of Technology — “Partnerships Between Cloud Service Providers and AI Developers: Staff Report on AI Partnerships & Investments 6(b) Study,” January 2025. https://www.ftc.gov/system/files/ftc_gov/pdf/p246201_aipartnerships6breport_redacted_0.pdf

[44] Acadian Asset Management (Owen Lamont) — “Straight Talk About Circular Deals in AI Today,” 2026. https://www.acadian-asset.com/investment-insights/owenomics/straight-talk-about-circular-deals-in-ai

[45] Built In (citing Wayne Tassie; Cowles Foundation, Yale University) — “How Circular Financing Is Fueling the AI Boom,” 2026. https://builtin.com/articles/ai-circular-financing

[46] Aldasoro, I., S. Doerr, and D. Rees, Bank for International Settlements — “Financing the AI boom: from cash flows to debt,” BIS Bulletin No. 120, 2026. https://www.bis.org/publ/bisbull120.pdf

[47] International Energy Agency — “Key Questions on Energy and AI” (World Energy Outlook Special Report), Executive Summary, 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary

[48] TheStreet via AOL (citing Pierre-Olivier Gourinchas, Chief Economist, International Monetary Fund) — “IMF drops blunt warning on US economy,” 2026. https://www.aol.com/finance/imf-drops-blunt-warning-us-153300892.html

[49] Tech Times (statement of Erik Brynjolfsson, Stanford University; signatories Daron Acemoglu and Simon Johnson, MIT) — “Nobel Economists Who Doubted AI Job Fears Now Sound the Alarm on White-Collar Displacement,” July 14, 2026. https://www.techtimes.com/articles/320398/20260714/nobel-economists-who-doubted-ai-job-fears-now-sound-alarm-white-collar-displacement.htm

[50] Platformer (Casey Newton, citing Erik Brynjolfsson and The New York Times) — “The loudest warning about AI and jobs yet,” July 2026. https://www.platformer.news/ai-jobs-warning-brynjolfsson-acemoglu/

[51] International Institute for Strategic Studies — “Gulf AI infrastructure and the limits of technological sovereignty,” Strategic Comments, June 2026. https://www.iiss.org/publications/strategic-comments/2026/06/gulf-ai-infrastructure-and-the-limits-of-technological-sovereignty/

[52] G42 (statements of Larry Ellison, Jensen Huang, and Masayoshi Son) — “Global Tech Alliance Launches Stargate UAE,” 2025. https://www.g42.ai/resources/news/global-tech-alliance-launches-stargate-uae

[53] Foreign Affairs Forum — “Silicon Sovereignty: How Washington’s Elevation of the UAE to Trusted Chip Status Is Redrawing the Geopolitical Map of Artificial Intelligence,” July 15, 2026. https://www.faf.ae/home/2026/7/15/silicon-sovereignty-how-washingtons-elevation-of-the-uae-to-trusted-chip-status-is-redrawing-the-geopolitical-map-of-artificial-intelligence

[54] Rest of World (statements of Peng Xiao, CEO, G42) — “The Gulf has billions to spend on AI. It still needs Nvidia,” July 2026. https://restofworld.org/2026/gulf-ai-investment-nvidia-monopoly-blackwell/

[55] Center for Strategic and International Studies — “The United Arab Emirates’ AI Ambitions,” May 2026. https://www.csis.org/analysis/united-arab-emirates-ai-ambitions

[56] Vamsi Talks Tech — “Sovereign AI and the Geopolitics of Compute: Export Controls, National Chip Programs, and the Fracturing Global AI Stack,” June 2026. https://www.vamsitalkstech.com/ai-infrastructure/sovereign-ai-and-the-geopolitics-of-compute-export-controls-national-chip-programs-and-the-fracturing-global-ai-stack/

[57] iNetanel — “Someone Caught the AI Money Loop — Here’s What They Found,” March 2026. https://inetanel.com/articles/the-ai-money-loop-nvidia-openai-big-tech

[58] IntuitionLabs — “Stargate Project: OpenAI’s $500B AI Data Center Plan” (White House announcement, January 2025; SoftBank funding completion, December 2025), February 2026. https://intuitionlabs.ai/articles/openai-stargate-datacenter-details

[59] Futurum Group (Brendan Burke) — “Anthropic’s Gigawatt-Scale TPU Deal with Broadcom Creates a Structural Advantage,” April 2026. https://futurumgroup.com/insights/anthropics-gigawatt-scale-tpu-deal-with-broadcom-creates-a-structural-advantage/

[60] Cainz Research (citing Goldman Sachs and Reuters analysis) — “The AI boom’s reliance on circular deals,” May 2026. https://www.cainz.org/14237/

[61] Concurrences / Troutman Pepper Locke (Brad Weber and Taylor Levesque) — “Cloud and Competition Policy, Part VIII — Cloud Service–AI Partnerships: The FTC’s Section 6(b) Report and its Antitrust Implications in the Trump 2.0 Administration,” August 2025. https://www.concurrences.com/en/review/issues/no-8-2025/on-topic/cloud-and-competition-policy/cloud-and-competition-policy-part-viii-cloud-service-ai-partnerships-the-ftc-s