Introduction: Financing Tomorrow Before It Exists
Every great investment cycle eventually produces a moment when the market is asked to value something it has never been asked to value before. In the railroad age, that moment arrived when investors were asked to price land grants across territory no surveyor had yet crossed. In the electrification age, it arrived when utility holding companies pyramided claims on generating stations that existed only in engineering drawings. In the dot-com age, it arrived when telecom carriers sold bonds against fiber routes that would not carry commercial traffic for years, and in some infamous cases would never carry it at all. In the opening days of September 2026, the artificial-intelligence infrastructure boom produced its own version of that moment, and it did so with a concentration of headline numbers that would have seemed implausible even eighteen months earlier.
Investors were no longer being asked merely to value operating datacenters, installed GPUs, functioning power plants, or recurring cloud revenues. Increasingly, they were being asked to place a value on something more abstract and more consequential: the contractual promise of infrastructure that had not yet been completed — and, in many cases, had not yet begun construction. The physical asset had not disappeared from the analysis, but it had been displaced from the center of the valuation exercise by something new. The center was now occupied by the contract itself.
SB Energy offered perhaps the clearest example of this inversion. On September 1, 2026, the SoftBank-backed energy and datacenter developer filed publicly for a U.S. initial public offering, seeking to tap the extraordinary investor appetite for companies serving the physical build-out of artificial intelligence. Reuters reported that the company carried a backlog of roughly $439 billion — approximately $430 billion of it attributable to datacenter contracts — while simultaneously disclosing that it had no operational datacenters whatsoever and describing itself as substantially dependent on a single anchor customer, OpenAI [1]. The company’s Form S-1 disclosed 8.8 gigawatts of contracted datacenter capacity, of which only 0.8 gigawatts was actually under construction, with the remaining 8.0 gigawatts contracted but not yet begun; for the six months ended June 30, 2026, SB Energy reported total revenue of $138.7 million — essentially all of it from its legacy solar-and-storage power business — against a net loss attributable to the company of approximately $3.21 billion [2]. The datacenter contracts underlying the backlog carry a weighted average remaining term of 19.6 years, anchored by OpenAI’s twenty-year leases covering approximately 8.0 gigawatts of IT capacity across seventeen planned buildings at the PORTS-Pike Technology Campus in Ohio, with Nvidia supporting the initial 4.25-gigawatt phase through a guarantee valued at $105 billion [3]. The arithmetic is striking on its face: a company whose contracted future is more than three thousand times larger than its most recent half-year of realized revenue was asking public investors to underwrite the difference between the two.
The market commentary surrounding the filing captured the analytical problem precisely. As IPOX research associate Lukas Muehlbauer told Reuters, the entire proposition turns on conversion — on whether
Lukas Muehlbauer, IPOX Research [5]
“hundreds of billions of contracted demand can be turned into cash flow”
over the coming years. That single sentence is, in compressed form, the subject of this entire paper. SB Energy’s own chief executive, Rich Hossfeld, was equally candid about why Nvidia’s guarantee sits inside the structure at all, telling CNBC that the chipmaker’s involvement
Rich Hossfeld, Co-CEO of SB Energy [4]
“helps us to unlock things like investment-grade financing”
— an admission, delivered without embarrassment, that the creditworthiness of the future tenant and the balance sheet of the chip supplier have become load-bearing elements of the financing itself [4].
Less than forty-eight hours after SB Energy’s filing, another extraordinary number entered the AI infrastructure conversation. Nscale, a London-based neocloud provider founded only in 2024, was telling prospective investors that it had accumulated approximately $103 billion of total contracted revenue ahead of a possible initial public offering — a figure that had roughly doubled from about $51 billion in a matter of weeks [8]. The single largest component of that leap was Anthropic’s six-year commitment to spend approximately $45 billion renting AI computing capacity from Nscale’s planned West Virginia campus, a commitment representing roughly 460 megawatts of power capacity, expected to be served by Nvidia’s Vera Rubin generation of accelerators, at a facility not expected to come online until the end of 2027 [9]. The contracts across Nscale’s book average 5.7 years in duration, implying roughly $18 billion of annualized contracted revenue — even as the company’s actual realized revenue was estimated at just over $100 million for the second quarter of 2026, up from approximately $37 million in the first quarter, and even as a source close to the process cautioned that the headline figures were illustrative rather than formal guidance [10]. On September 4, 2026, Reuters reported the natural next step in the sequence: Nscale was seeking roughly $3.5 billion in pre-IPO financing, including up to $1.5 billion in convertible notes led by hedge fund Third Point and approximately $2 billion in proposed investment from Nvidia itself, with Goldman Sachs advising and a public listing contemplated as early as the same month [11]. The mechanism at the center of this paper is visible in that sequence with almost diagrammatic clarity: a large future customer commitment attracts present capital before the productive infrastructure generating those future revenues exists.
Nor were SB Energy and Nscale isolated examples. On September 2, 2026, Olympus-backed datacenter infrastructure manufacturer Accelevation filed for its own U.S. IPO, entering a public-market environment explicitly receptive to businesses serving the AI infrastructure boom, carrying a backlog of approximately $1.1 billion as of June 30, 2026 against roughly $448 million of fiscal 2025 revenue [12]. Accelevation is in one sense the most conventional of the three — it is profitable, it manufactures tangible power-distribution and white-space equipment, and its revenue grew 147 percent year over year — yet even here the investment narrative rests on a projected expansion of its addressable market from roughly $22 billion in 2025 to approximately $80 billion by 2030, a projection that itself depends on the very forward commitments this paper examines [13]. CoreWeave, meanwhile, the first of the publicly traded neoclouds and in many respects the template for the entire category, reported in August 2026 a revenue backlog of approximately $104 billion as of June 30 — up 246 percent year over year — excluding more than $25 billion of additional net new customer commitments secured in the opening weeks of the third quarter, against quarterly revenue of $2.58 billion and a net loss of $626 million driven substantially by surging interest costs [14][15]. CoreWeave’s chief executive Michael Intrator framed the quarter as the moment when
Michael Intrator, Co-founder and CEO of CoreWeave [16]
“our scale began to translate into expanding operating leverage”
— which is another way of saying that the company believes it is climbing, facility by facility and gigawatt by gigawatt, from the contracted stage of its book toward the monetized stage [16].
These developments suggest that the AI infrastructure boom has entered a distinctly new phase, and that the analytical vocabulary developed for the first phase is no longer adequate. During the first phase, running roughly from the launch of ChatGPT in late 2022 through 2025, capital financed physical assets in a recognizably traditional pattern: semiconductor fabrication plants, GPUs, fiber networks, transformers, substations, datacenter shells, cooling plants, and generation capacity. The sums were unprecedented — the four largest hyperscalers alone guided toward roughly $700 billion to $725 billion of combined capital expenditure for calendar 2026, up more than 60 percent from the record levels of 2025, with Goldman Sachs projecting a cumulative $5.3 trillion of hyperscaler capital spending between 2025 and 2030 [18][19] — but the underlying financial logic remained familiar: a company with enormous operating cash flows purchases productive assets and depreciates them against future revenue. Jensen Huang, whose company sits at the center of the entire cycle and which reported record data-center revenue of $75.2 billion in its most recent quarter, described the phenomenon in May 2026 as
Jensen Huang, Founder and CEO of NVIDIA [17]
“the largest infrastructure expansion in human history”
— and whatever one thinks of the superlative, the physical component of the claim is difficult to dispute [17].
In the emerging second phase, however, anticipated demand itself increasingly becomes part of the financing architecture. A long-term lease can support project debt. A multiyear compute commitment can double an IPO candidate’s headline contracted revenue in a single stroke. A power-purchase agreement can justify new generation. A GPU purchase commitment can support factory expansion. A chip supplier’s guarantee can convert a speculative campus into investment-grade collateral. A frontier laboratory’s promise to consume future capacity can influence enterprise valuations years before the associated infrastructure becomes fully productive. The question investors are being asked to answer has quietly changed. It is no longer simply how much AI infrastructure exists. It is increasingly how much financial value can be created from infrastructure that markets expect to exist.
This paper calls that phenomenon Anticipatory Capital.
Anticipatory Capital describes the process through which credible contractual claims on future AI production — future megawatts, future GPU clusters, future datacenter capacity, future compute consumption, and future customer payments — are translated into present-day enterprise value, financing capacity, investment decisions, and physical construction. The consequences extend far beyond Wall Street. Anticipatory Capital can accelerate America’s AI buildout by allowing capital to arrive before assets become operational, compressing development timelines that would otherwise be gated by the slow accumulation of operating history. Yet the same mechanism can magnify forecasting errors, concentrate counterparty risk, encourage overbuilding, distort electricity-demand projections, and create chains of financial dependency connecting AI laboratories, chipmakers, datacenter developers, utilities, lenders, private-equity firms, private-credit funds, and public investors. The AI economy, in short, is learning to finance the future in advance of physically constructing it — and the balance between the productive and the perilous versions of that capability is the central question of the next several years.
The remainder of this paper proceeds in six sections. Section 1 develops the conceptual foundation, tracing the inversion from physical infrastructure to contractual infrastructure and introducing the Anticipatory Capital Conversion Ladder. Section 2 examines the financial architecture through which future capacity is capitalized, including the pre-operational capital stack, the collapse of the customer-investor-supplier distinction, and the problem of duration asymmetry. Section 3 maps Anticipatory Capital across the Five-Layer AI Economy, from energy contracts through chips, datacenters, and models, down to the applications and agents on which every upstream claim ultimately rests. Section 4 analyzes the principal risks: the Delivery Gap, counterparty concentration, technological obsolescence, demand-forecasting error, and the migration of corporate risk into systemic risk. Section 5 proposes measurement and governance frameworks for investors, grid regulators, and federal policymakers. Section 6 distills the argument into seven pillars, and the Conclusion returns to the largest question of all: whether the AI economy is financing the future or borrowing from it.
Why I Chose the Title “Anticipatory Capital”
A brief note on terminology is warranted before the analysis begins, because the choice of name is itself an analytical claim. I chose the term Anticipatory Capital because the conventional vocabulary of finance — backlog, contracted revenue, remaining performance obligations, project finance, infrastructure investment, capital expenditure — describes only individual fragments of what is occurring, and because each of those fragments, taken alone, understates the systemic character of the whole. Backlog is an accounting disclosure. Project finance is a transaction structure. Capital expenditure is a line on a cash-flow statement. None of these terms captures the way in which, across the AI economy of 2025 and 2026, expectations about the future have themselves become a form of productive financial asset, circulating through prospectuses, credit committees, rating-agency models, utility resource plans, and state economic-development negotiations.
The adjective Anticipatory identifies the defining temporal characteristic of the phenomenon: investment and valuation move forward in time, capitalizing expected economic activity before the underlying AI infrastructure becomes operational. The contract does more than promise future revenue in the manner of any ordinary sales agreement. It actively mobilizes present capital around an anticipated future state of the world — a state in which the datacenter is energized, the GPUs are installed, the model is trained, the inference is sold, and the tenant’s own business has grown large enough to honor a twenty-year lease. When SB Energy’s prospectus discloses a backlog whose weighted average remaining contract term is 19.6 years [3], it is asking investors to capitalize a stream of payments that extends to roughly 2046, contingent on a technology whose current commercial form is barely four years old.
The noun Capital matters equally, because these forward commitments increasingly perform functions normally associated with productive assets and with money itself. They support financing, influence enterprise valuation, reduce perceived demand risk, attract lenders, enable initial public offerings, justify power and datacenter construction, and encourage suppliers to expand manufacturing capacity. A signed compute contract can be pledged, referenced, marketed, and — through the structures examined in Section 2 — effectively borrowed against. Anticipatory Capital therefore represents the conversion of future economic expectations into present financial power, and the term is particularly appropriate for a Five-Layer AI Economy increasingly constructed through interdependent promises about future energy, chips, datacenters, models, and applications. The layers are no longer merely a technological stack; they are becoming a financial circuit, and Anticipatory Capital is the current that flows through it.
Section 1: From Physical Infrastructure to Contractual Infrastructure
1.1 The Emergence of Anticipatory Capital
Traditional infrastructure finance generally begins with an identifiable physical asset and a forecast of the cash flows that asset may produce. The toll road exists, or is at least fully permitted and financed, before the traffic study becomes the basis of a bond issue; the pipeline is engineered before the shippers commit; the power plant’s turbines are ordered against a demonstrated load forecast compiled from decades of utility data. The sequence runs from the physical to the financial: asset, then forecast, then capital. Even in classic project finance, where lenders famously look to the contract rather than the sponsor’s balance sheet, the contract is understood as a mechanism for allocating the risks of a defined physical project — not as the primary economic asset in its own right.
Artificial-intelligence infrastructure increasingly reverses that sequence, and the reversal is the founding observation of this paper. Before the datacenter exists, the customer already exists. Before the electrical load appears on any utility’s system, the power commitment has been negotiated, sometimes years in advance and sometimes for quantities that exceed the current peak load of entire metropolitan regions. Before the GPUs arrive — indeed, before the GPU generation in question has entered volume production — the compute capacity has been reserved, priced, and in some cases pledged to support the financing of the facility that will house it. Anthropic’s commitment to Nscale’s West Virginia campus was signed and marketed to investors more than a year before the facility’s expected energization at the end of 2027, and it references Vera Rubin accelerators that Nvidia had only recently begun bringing to market [9]. OpenAI’s Ohio leases with SB Energy extend twenty years into the future at a site where seventeen buildings remain to be constructed [3].
Anticipatory Capital begins at this moment of inversion. The economic asset is initially not the building, the substation, or the GPU. It is a credible claim on future utilization — a legally structured expectation that a specific counterparty will consume a specific quantity of compute, power, or capacity over a specific period, at terms sufficient to service the capital required to build the underlying facility. Everything else in the modern AI capital stack is derivative of that claim: the project debt is sized against it, the equity is valued as a multiple of it, the supplier’s guarantee wraps around it, and the IPO prospectus leads with it.
It is worth pausing on how genuinely novel this is at the present scale. Take-or-pay contracts, capacity reservations, and anchor-tenant leases have existed for decades in energy, shipping, and commercial real estate, and the scholarly literature on project finance has long recognized that contractual cash-flow assignments can substitute for operating history. What is new in the AI economy of 2025–2026 is threefold. First, the magnitudes: single-company forward books in the hundreds of billions of dollars, and an aggregate across the ecosystem that plausibly exceeds one and a half trillion dollars when OpenAI’s disclosed $1.4 trillion of multi-year compute and infrastructure commitments are combined with the contracted backlogs of the neoclouds and integrated developers [34]. Second, the velocity: Nscale’s contracted book doubled from $51 billion to $103 billion in roughly a month [8], a rate of forward-claim formation with no precedent in infrastructure history. Third, the reflexivity: the customers making the commitments are themselves financed, in part, by the suppliers who benefit from the commitments, a circularity examined in detail in Section 2.4 and Section 4.5. The combination of scale, speed, and reflexivity is what elevates Anticipatory Capital from a familiar financing technique into a defining macro-financial phenomenon.
1.2 SB Energy and the Valuation of the Unbuilt
SB Energy serves as this paper’s principal opening case study because it illustrates, more cleanly than any other single company, the distinction between physical capacity and economically anticipated capacity — and because its prospectus, to its credit, discloses that distinction with unusual candor.
Consider what the filing actually describes. On the physical side of the ledger: an operating power portfolio of approximately 2.2 gigawatts of solar and battery storage, an additional 2.5 gigawatts under construction, roughly $4.0 billion of debt outstanding including $999 million of 8.875 percent senior secured notes due 2031, 223 full-time employees, and first-half 2026 revenue of $138.7 million derived almost entirely from selling power — including from a 900-megawatt Texas solar facility serving a Google datacenter [2][7]. On the anticipated side of the ledger: 8.8 gigawatts of contracted datacenter capacity across Texas and Ohio, of which 8.0 gigawatts is contracted but not yet under construction; twenty-year OpenAI leases across seventeen unbuilt buildings; a $105 billion Nvidia guarantee supporting the initial Ohio phase; Nvidia’s committed $1.5 billion private placement at the IPO price; OpenAI warrants valued at approximately $5.5 billion; and the headline figure that organized every news story about the filing — approximately $439 billion of backlog [1][3][7]. Between the two sides of the ledger sits the filing’s own caution: the backlog is a management estimate built on assumptions, it is not recognized revenue, and its realization depends on construction completion, permitting, successful grid interconnection, lease commencement, and tenant acceptance, with the company further flagging long lead times for power equipment and spreading community opposition to datacenters as risks to its growth [2][6].
This creates the paper’s first fundamental question, which every subsequent section will approach from a different angle: what exactly are investors valuing when hundreds of billions of dollars of expected activity are attached to infrastructure that has not yet become productive? The honest answer is that they are valuing a probability-weighted bundle — the probability that the buildings are completed on schedule, multiplied by the probability that the power arrives, multiplied by the probability that the tenant still needs the capacity when it arrives, multiplied by the probability that the tenant can pay for it across two decades, multiplied by the probability that the economics of the underlying technology have not shifted so far that the contract is renegotiated. Reuters Breakingviews and other commentators pressed precisely on the weakest links in that chain: the extreme concentration in OpenAI as anchor tenant, the related-party character of a structure in which SoftBank is simultaneously controlling shareholder, investor, and customer, and the long-duration assumptions embedded in a backlog that equals more than three thousand times the company’s most recent half-year revenue [1][7]. None of this makes the offering irrational. It makes the offering a referendum — arguably the largest yet conducted in a public securities market — on the credibility of anticipated AI demand.
1.3 Nscale, Anthropic, and Contracted Compute
Nscale demonstrates the same phenomenon from the compute side rather than the integrated power-development side, and its trajectory compresses the entire Anticipatory Capital sequence into a period short enough to observe almost in real time.
The company was founded in 2024. By mid-2026 its realized quarterly revenue had only just crossed $100 million [8]. Yet by early September 2026 it was presenting prospective IPO investors with approximately $103 billion of total contracted revenue, anchored by Anthropic’s six-year, approximately $45 billion agreement for roughly 460 megawatts of capacity at the planned West Virginia campus, to be delivered on Nvidia’s Vera Rubin architecture beginning around the end of 2027 [8][9]. Anthropic’s commitment, it should be noted, is itself one strand in a broader capacity-assembly strategy that reportedly includes a $35 billion arrangement with Lambda for a Texas facility developed by Hut 8, approximately $50 billion with Fluidstack, and approximately $45 billion with SpaceX — a portfolio approach to securing future compute that followed a period in early 2026 when surging usage of the Claude models outran available supply [9]. The customer contract therefore begins influencing Nscale’s capital structure before the physical infrastructure completes its own development cycle: within days of the contracted-revenue figure circulating, the company was negotiating $3.5 billion of pre-IPO financing, including convertible notes and a proposed $2 billion investment from Nvidia — the same company whose chips the Anthropic contract obligates Nscale to deploy [11].
The sequence can be expressed compactly, and it is the canonical sequence of Anticipatory Capital:
AI Demand → Long-Term Commitment → Financing Confidence → Capital Formation → Construction → Energization → Compute Delivery → Revenue
Anticipatory Capital exists primarily between the second and fifth stages of that chain — after the commitment is signed but before the infrastructure is energized. It is in that interval that a contract behaves most like capital: it is being valued, marketed, leveraged, and built against, while producing no cash whatsoever. The interval is not short. For a greenfield gigawatt-scale campus, the distance between contractualization and full monetization can span three to five years of permitting, interconnection, construction, and commissioning — and it is precisely because the interval is long that so much financial engineering has grown up inside it.
1.4 Backlog Is Not Revenue: A Taxonomy of Forward Claims
A critical scholarly distinction must be maintained throughout this paper, because the public conversation routinely collapses it. The forward claims circulating through the AI economy are not a single instrument; they form a hierarchy of very different legal and economic objects, and the difference between the top and the bottom of the hierarchy can be — as the reported evolution of one Nvidia-OpenAI arrangement from roughly $600 billion of headline financing discussions to a signed $105 billion guarantee demonstrated in 2026 — half a trillion dollars [3][33].
Table — The Hierarchy of Forward Claims
| Category | Nature of the Claim | Economic Certainty |
| Realized revenue | Cash received or receivable for delivered service | Highest — settled economics |
| Remaining performance obligations | Contracted, unsatisfied obligations under ASC 606 disclosure | High — legally defined, audited |
| Executed take-or-pay leases | Binding leases with payment obligations largely independent of utilization | High — subject to counterparty credit |
| Contracted revenue / backlog | Management aggregation of signed agreements, definitions vary by company | Moderate — assumption-dependent |
| Power-purchase and capacity agreements | Binding but condition-laden energy commitments | Moderate — contingent on delivery milestones |
| Reservations and exclusivity agreements | Priority rights over future capacity, often with exit provisions | Lower — optionality retained by customer |
| Letters of intent and announced partnerships | Non-binding expressions of scale and direction | Lowest — headline value can exceed contract value by multiples |
| Pipeline | Management’s view of addressable future demand | Speculative — not a claim at all |
These categories do not possess equivalent economic certainty, and the sophistication of the AI infrastructure market in 2026 can fairly be measured by how carefully any given participant distinguishes among them. SB Energy’s filing explicitly warns that its backlog is a hypothetical estimate based on management assumptions and is not necessarily indicative of future revenue [2][6]. CoreWeave’s disclosed backlog combines remaining performance obligations with estimated future revenue from committed contracts, and the company is explicit that recognition depends on its actually delivering infrastructure and maintaining service availability [16]. Nscale’s own advisers characterized its $103 billion as illustrative rather than guidance [8][10]. The disclosure practices, in other words, are often more careful than the headlines built upon them. The analytical failure occurs downstream, when a nineteen-year lease, a five-year take-or-pay contract, a conditional guarantee, and a non-binding letter of intent are added together into a single number and that number is treated as a cash equivalent. This paper therefore resists headline backlog figures as valuation inputs and asks instead the question developed formally in Section 5: how much economic credibility should be assigned to each category of forward claim, and how should the market price the distance between categories?
1.5 The Anticipatory Capital Conversion Ladder
To give the analysis a stable spine, I propose a six-stage framework — the Anticipatory Capital Conversion Ladder — that traces a unit of anticipated AI demand from its first appearance as customer interest to its final settlement as operating cash flow. Each rung of the ladder is a distinct economic event, each transition between rungs carries its own risks, and the aggregate health of the AI infrastructure economy can be read from the distribution of claims across the rungs.
Table — The Anticipatory Capital Conversion Ladder
| Stage | Description | Characteristic Instruments |
| Stage 1 — Reservation | A customer signals credible interest in future capacity | Letters of intent, exclusivity periods, capacity holds |
| Stage 2 — Contractualization | The demand becomes a legally defined commitment | Take-or-pay leases, compute agreements, PPAs |
| Stage 3 — Capitalization | Investors and lenders assign financial value to the commitment | Project debt, convertible notes, pre-IPO rounds, IPOs, guarantees |
| Stage 4 — Construction | Capital is converted into physical infrastructure | EPC contracts, equipment orders, interconnection builds |
| Stage 5 — Energization | Power, cooling, networks, GPUs, and facilities become operational | Commissioning, tenant acceptance, service commencement |
| Stage 6 — Monetization | The anticipated contract becomes operating revenue and cash flow | Recognized revenue, debt service, distributions |
The framework yields the paper’s central measurement principle: the greater the distance between Capitalization and Monetization, the greater the anticipatory character — and potentially the risk — of the capital involved. A company like Accelevation, whose $1.1 billion backlog converts to revenue within quarters as equipment ships [12][13], carries a short conversion distance and correspondingly modest anticipatory risk. A company like CoreWeave, with 1.5 gigawatts of active power against 3.7 gigawatts contracted and a $104 billion backlog recognized over multi-year contract lives [15][16], occupies the middle of the ladder: substantial anticipatory exposure, but a demonstrated record of climbing from Stage 3 to Stage 6 at scale. A company like SB Energy, capitalizing at Stage 3 a book whose weighted average term reaches nearly two decades and whose capacity is more than 90 percent pre-construction [2][3], sits at the maximal anticipatory extreme: nearly the entire enterprise value is a wager on successful ascent through Stages 4, 5, and 6, repeated across seventeen buildings, multiple campuses, and twenty years. The ladder does not tell us which wagers will pay. It tells us where to look, and how hard, before believing any given headline number.
Section 2: The Financial Architecture of Future AI Capacity
2.1 Contracts as Financing Instruments
The deepest structural change documented in this paper is that long-duration AI infrastructure agreements increasingly function as more than commercial sales arrangements. They have become, in practice if not always in legal form, financing instruments — documents whose principal economic effect is to unlock capital rather than merely to allocate goods and services. This is not a metaphor. When a lender’s credit committee evaluates a proposed gigawatt campus, the analysis begins with the contracted cash flows; when a rating agency assesses the securitization of datacenter lease payments, the tenant’s covenant is the collateral; when an underwriter builds an IPO book for a pre-revenue developer, the contracted backlog is the product being sold. The building is, in a very real sense, the residual.
The market has developed an entire vocabulary of structures for converting contracts into capital. Project-level debt is sized against executed leases, with lenders advancing construction funds in tranches keyed to milestones because the lease itself — not the half-built shell — is what secures repayment. Special purpose vehicles hold the contracted asset and issue securities against it, permitting hyperscalers to finance enormous campuses while keeping much of the associated debt off their reported balance sheets; Meta’s $27 billion Hyperion joint venture, in which funds managed by Blue Owl hold 80 percent and Meta retains 20 percent, with Morgan Stanley arranging over $27 billion of debt anchored by PIMCO, has become the acknowledged blueprint for the form [36][37]. Supplier guarantees — of which Nvidia’s $105 billion Ohio commitment is the largest known example — wrap the credit of the equipment vendor around the obligations of the tenant, transforming the risk profile of the entire structure [3]. And customer prepayments, warrants, and equity stakes travel in the opposite direction along the same contracts, so that the paper documenting future consumption simultaneously documents present investment.
A hyperscaler or frontier AI laboratory therefore performs two roles simultaneously: customer and quasi-underwriter of future infrastructure. When OpenAI signs twenty-year leases for 8.0 gigawatts of unbuilt Ohio capacity, it is not merely procuring compute; it is underwriting — in the insurance sense of assuming a defined risk in exchange for priority access — the single largest speculative construction program in the history of digital infrastructure [3]. When Anthropic commits $45 billion over six years to a campus that will not energize until late 2027, it converts Nscale from a two-year-old startup into an IPO candidate with a nine-figure quarterly revenue run rate and an eleven-figure forward book [8][9][11]. The commitment is the collateral; the collateral is the story; the story is the capital.
2.2 From Megawatts to Enterprise Value
The new financial grammar of AI infrastructure translates physical units into financial expectations through a chain that market participants now treat as routine but that deserves to be spelled out, because each link embeds an assumption:
Megawatts → Compute Capacity → Contracted Revenue → Expected Cash Flow → Financing Capacity → Enterprise Value
The first link converts power into compute: a megawatt of critical IT load, populated with a given accelerator generation at a given density, yields a calculable quantity of training and inference throughput. The second link converts compute into contracted revenue: the throughput is priced into a lease or capacity agreement at rates that, in 2025–2026, have been set in a market defined by scarcity. The third link discounts the contracted stream into expected cash flow, applying assumptions about delivery timing, utilization, renewal, and counterparty performance. The fourth link converts expected cash flow into financing capacity — the quantum of debt and equity the stream can support. The final link capitalizes the whole into enterprise value, at multiples that reflect the market’s confidence in every preceding link simultaneously.
This chain explains why datacenter economics now sit at the intersection of technology finance and infrastructure finance, borrowing the vocabulary of both and the discipline, at times, of neither. A 500-megawatt commitment is no longer merely an engineering specification describing transformers and chillers. Once associated with a creditworthy tenant and an enforceable contract, it becomes an input into corporate valuation — and the market’s willingness to run the chain in reverse is equally consequential: a developer that controls land, power, and interconnection positions can back-solve from the enterprise value it seeks to the megawatts it must announce. The chain also explains the market’s obsession with power as the binding constraint. Chips can be manufactured in months; buildings can be erected in a year or two; but generation, transmission, and interconnection operate on utility timescales of three to ten years. Power certainty is therefore the scarcest input into the entire conversion, which is why an integrated developer like SB Energy leads its investment narrative with energy capability, why CoreWeave reports contracted power — approximately 3.7 gigawatts against 1.5 gigawatts active — as a headline operating metric, and why Nvidia’s Ohio partnership commitments extend beyond chips into at least 10 gigawatts of new generation and $4.2 billion of regional grid investment [2][15][3].
2.3 The Pre-Operational Capital Stack
Anticipatory Capital does not arrive from a single source. By 2026 the financing of pre-operational AI infrastructure had become one of the most elaborately layered capital formations in modern markets, and the layering itself is analytically significant: each layer prices the same underlying forward claims differently, and the deeper the stack grows before energization, the more balance sheets depend on the same anticipated demand.
The principal layers include sponsor equity from developers and their private-equity backers; venture capital, which seeded the neocloud generation; infrastructure funds and sovereign capital, which have migrated from toll roads and pipelines into datacenter platforms; private credit, which Morgan Stanley estimates will need to supply approximately $800 billion of datacenter financing through 2028 as part of a $1.5 trillion gap between projected investment and what Big Tech cash flow can fund internally [36]; bank financing and syndicated project debt, with U.S. datacenters and related infrastructure already carrying at least $1.3 trillion of identified debt according to banking-intelligence data — a total the compilers acknowledge is almost certainly understated given the opacity of private structures [38]; investment-grade and high-yield bonds, with the five largest hyperscalers issuing $121 billion of U.S. corporate bonds in 2025 alone — more than four times their 2020–2024 annual average — and global AI-related debt issuance projected by Morgan Stanley to approach $570 billion in 2026 [36][40]; convertible securities and pre-IPO rounds of the kind Nscale was assembling in September 2026 [11]; equipment financing and GPU-backed loans, in which the accelerators themselves serve as collateral; customer prepayments and warrants; strategic corporate investments, exemplified by Nvidia’s $1.5 billion committed placement in SB Energy and its proposed $2 billion in Nscale [1][11]; and finally public equity, the layer to which the entire stack has been migrating through the IPO queue of late 2026.
Two features of this stack deserve emphasis. First, its aggregate scale is set by the physical program it must fund: Morgan Stanley and Moody’s estimate at least $3 trillion of capital spending will be required for datacenters and related infrastructure in the coming years, with JPMorgan projecting more than $5 trillion once power generation is included, and J.P. Morgan separately estimating hyperscaler capital expenditure of $697 billion in 2026 alone [37][39]. Even the largest technology companies cannot fund that program from internal cash, which is precisely why the stack exists. Second, an increasing share of the stack is structurally invisible. Moody’s data compiled by independent researchers indicate that the five largest U.S. hyperscalers held $969 billion of undiscounted future datacenter lease commitments at year-end 2025, of which $662 billion represented leases not yet commenced and therefore absent from reported balance sheets — a shadow obligation exceeding the same companies’ combined adjusted on-balance-sheet debt [40]. The Bank for International Settlements has flagged the migration of financing into off-balance-sheet vehicles as a transparency concern for exactly this reason: when the anticipated demand that supports the stack is revised, no single regulator, rating agency, or investor can currently see the full set of claims that revision would impair [36].
2.4 When Customers, Investors, and Suppliers Become the Same Network
AI infrastructure exhibits a further characteristic that distinguishes it from every prior infrastructure cycle in degree if not in kind: the principal companies are simultaneously customers, investors, technology suppliers, creditors, and strategic partners of one another, frequently within a single transaction.
The September 2026 filings display the pattern in its mature form. In SB Energy’s structure, SoftBank is controlling shareholder, founding investor, and tenant; OpenAI is anchor tenant, strategic investor through a $500 million Stargate-linked infusion, and warrant holder with instruments valued at roughly $5.5 billion; Nvidia is chip supplier to the tenant, guarantor of $105 billion of the initial phase, and committed IPO investor at $1.5 billion [1][3][7]. In Nscale’s structure, Nvidia is chip supplier, existing backer, and proposed $2 billion pre-IPO investor, while the anchor customer’s commitment specifically obligates deployment of Nvidia’s newest architecture [9][11]. Beyond these two, the broader web is well documented: Nvidia’s reported investments include $30 billion into OpenAI and $10 billion into Anthropic alongside stakes in numerous neoclouds that rent out Nvidia-powered capacity; OpenAI’s disclosed commitments distribute roughly $1.4 trillion across Broadcom, Oracle, Microsoft, Nvidia, AMD, Amazon, and CoreWeave; and AMD’s arrangements with OpenAI famously paired chip commitments with equity warrants rather than cash [32][33][34]. Bloomberg’s mapping of the ecosystem found participants recycling capital, commitments, and equity through so many overlapping channels that the venture investor Paul Kedrosky — who covered networking companies as an analyst during the telecom boom — observed that AI capital spending was climbing toward levels last seen at the peak of the late-1990s fiber build-out [30]. Bernstein’s Stacy Rasgon, reacting to the original Nvidia-OpenAI announcement, put the market’s discomfort in a phrase that has organized the debate ever since:
Stacy Rasgon, Bernstein Research [31]
“The action will clearly fuel ‘circular’ concerns”
These arrangements are not automatically economically unsound, and the strongest version of the defense deserves to be stated fairly: building frontier AI is extraordinarily capital-intensive, the most advanced chips remain scarce, and in scarce markets participants rationally lock in supply by pairing long-term purchase commitments with financing — exactly as automakers, airlines, and shipbuilders have done in their own capital-intensive histories [30]. Vendor participation can also genuinely de-risk projects, as SB Energy’s CEO argued in explaining that Nvidia’s guarantee unlocks investment-grade financing for construction that would otherwise price as speculative development [4].
But the network structure creates an analytical challenge that no fair-minded observer can dismiss: demand must be separated into independent market demand and ecosystem-supported demand. Independent demand originates with end customers — enterprises, developers, consumers, governments — whose payments arrive from outside the AI capital circuit. Ecosystem-supported demand originates, at least in part, from capital that the circuit itself supplied: a chipmaker’s investment becomes a laboratory’s compute budget becomes a neocloud’s contracted backlog becomes the collateral for debt that purchases the chipmaker’s products. Analysts had identified more than $800 billion of such circular arrangements across the AI supply chain by 2026 [32][33]. Each individual transaction may be defensible; the aggregate makes it genuinely difficult to measure how much of the demand being capitalized would survive the withdrawal of the circuit’s own financing. The distinction will become steadily more important to investors and regulators, and Section 5 proposes that disclosure regimes begin requiring it explicitly.
2.5 Duration Transformation in the AI Economy
Classical banking theory describes maturity transformation: institutions borrow short and lend long, and the mismatch is both their economic function and their characteristic fragility. The AI infrastructure economy has produced a variant this paper calls duration transformation, and it may prove to be the cycle’s most underappreciated structural risk.
The mismatch runs as follows. On one side, the financial commitments being written are extraordinarily long: twenty-year OpenAI leases in Ohio, 19.6-year weighted average contract terms across SB Energy’s book, fifteen-year triple-net SoftBank leases at the Cosmos campus, multi-decade power-purchase agreements, and thirty-year infrastructure fund horizons [3]. On the other side, the technological assets those commitments finance live fast and depreciate faster. The useful competitive life of a leading-edge AI accelerator generation is measured in a few years — Nvidia has moved from Hopper to Blackwell to Vera Rubin in barely three product cycles — and rack density, cooling architecture, interconnect topology, model efficiency, and inference economics are all evolving on similar clocks. Paul Kedrosky’s formulation of the problem has become canonical in the commentary:
Paul Kedrosky, Venture Investor and Analyst [21]
“These are not railroads—we aren’t building century-long infrastructure”
— AI datacenters, he argues, are short-lived, asset-intensive facilities riding declining-cost technology curves, requiring frequent hardware replacement merely to preserve margins [21].
The defenders of long-duration contracting respond that the building shell, the substation, the transmission upgrade, and the water and fiber connections genuinely are long-lived assets; that leases price the real estate and power envelope rather than any particular chip generation; and that tenants expect to refresh hardware inside a stable physical container. The response has force for the shell. It has much less force for the economics, because the revenue assumptions inside a twenty-year lease were set in a scarcity market defined by one accelerator generation’s cost curve, and nothing in the contract can guarantee that a tenant paying 2026 scarcity prices for capacity delivered in 2029 will still find those prices rational in 2036 — or that a tenant whose own business model has not yet achieved sustained profitability will exist in its current form in 2046. This produces one of the central contradictions of Anticipatory Capital, and it should be stated plainly: long-duration finance is being used to build infrastructure for a short-duration technology cycle. The contracts stretch across four or five accelerator generations, two or three plausible shifts in model architecture, and an entirely unknowable competitive landscape. Duration transformation does not doom the structure. It does mean that the structure’s stability depends on renegotiation, refresh, and repricing mechanisms that most headline backlog figures silently assume will operate smoothly — an assumption that history invites us to examine rather than inherit.
Section 3: Anticipatory Capital Across the Five-Layer AI Economy
The Five-Layer AI Economy — Energy, Chips, Datacenters, Models, and Applications and Agents — was originally conceived as a technological description: each layer supplies the one above it, and intelligence emerges at the top of the stack. The argument of this section is that the framework has acquired a second, financial meaning. Anticipatory Capital operates at every layer, in instruments particular to each, and — most importantly — it binds the layers to one another, so that expectations formed at the top of the stack now mobilize capital at the bottom years before any current flows between them. The section proceeds layer by layer, from the ground up, and then reverses direction to show why the whole structure ultimately hangs from its highest and least proven layer.
3.1 Layer One — Energy: Contracting Tomorrow’s Electricity
At the Energy layer, Anticipatory Capital appears through power-purchase agreements, generation commitments, transmission agreements, behind-the-meter projects, nuclear restart and new-build contracts, battery installations, gas turbine reservations, renewable developments, and long-term utility arrangements. The unifying feature of all of them is temporal: the anticipated AI load justifies power infrastructure years before the actual servers begin consuming electricity.
The magnitudes are without precedent in modern utility planning. Nvidia’s Ohio partnership around the SB Energy campus contemplates at least 10 gigawatts of new energy generation and at least $4.2 billion of regional grid investment through an arrangement with AEP Ohio — commitments announced, negotiated, and financed against a datacenter campus whose first buildings remain under construction [3]. SB Energy’s own identity as an IPO candidate is inseparable from this layer: its operating history is a power company’s history, its 2.2 operating gigawatts of solar and storage and 2.5 gigawatts under construction are the physical collateral beneath its datacenter ambitions, and its prospectus is explicit that long lead times for power equipment are among the gating risks to converting backlog into revenue [2][6]. Utilities across the country face the mirror image of the same phenomenon: interconnection queues filled with speculative and duplicative datacenter requests, resource plans that must decide which contracted loads are credible enough to build generation against, and ratepayer-protection questions that state regulators are only beginning to formalize. The Energy layer thus establishes the paper’s pattern in its purest form: future compute demand capitalizes present energy infrastructure. When the demand forecast is right, the mechanism delivers power ahead of shortage. When it is wrong, the mechanism strands generation, transmission, and rate-base commitments whose costs persist for decades after the anticipated load has evaporated or migrated.
3.2 Layer Two — Chips: Reserving Semiconductor Futures
At the Chips layer, the same process occurs through multiyear accelerator orders, advanced-packaging reservations, high-bandwidth-memory commitments, foundry capacity agreements, networking purchases, and strategic investments in and by suppliers. Future AI demand justifies today’s semiconductor fabrication and supply-chain investment, and the chip itself has become part of an anticipatory production system in which capacity is sold before it is built at nearly every step: model developers reserve accelerators generations in advance, accelerator designers reserve packaging and memory before volume exists, and foundries commit fabrication capacity against order books that are themselves assembled from forward commitments.
Nvidia’s financial disclosures give the layer its scale: record quarterly data-center revenue of $75.2 billion, up 92 percent year over year, in the first quarter of fiscal 2027 [17]. But the more distinctive feature of the layer is the direction in which its capital now flows. The dominant supplier does not merely receive Anticipatory Capital; it manufactures it. Nvidia’s reported $30 billion investment in OpenAI, $10 billion in Anthropic, stakes across the neocloud sector, $105 billion Ohio guarantee, $1.5 billion committed SB Energy placement, and proposed $2 billion Nscale investment collectively make the chip supplier one of the largest single underwriters of its own future demand [3][11][33]. The company’s stated rationale — that these investments grow the ecosystem and offer attractive returns — is not implausible, and the guarantee structure in Ohio demonstrably lowered the financing cost of real construction [4]. Yet the arrangement means that the Chips layer’s order book, the Datacenter layer’s backlog, and the Model layer’s compute budgets are now partially the same money observed at different points in its circulation, which is precisely the measurement problem Section 2.4 identified and Section 4.5 develops.
3.3 Layer Three — Datacenters: Buildings Financed by Future Tenants
The Datacenter layer is where Anticipatory Capital becomes most visible, because it is where the contract and the concrete meet. Land is purchased before tenants arrive; interconnection requests precede construction by years; financing is raised against long-term leases on unbuilt shells; and campuses are valued according to their contracted future capacity rather than their present rent rolls. The datacenter has evolved, in the space of roughly four years, from a specialized category of income-producing real estate into a financial container for future compute demand.
The September 2026 cohort spans the layer’s full spectrum. At one end, Accelevation represents the picks-and-shovels sublayer — power distribution and white-space equipment — where the conversion ladder is short, backlog turns into revenue within quarters, and the company arrives at the public market already profitable, with $448 million of 2025 revenue grown 147 percent and a $1.1 billion backlog [12][13]. In the middle, CoreWeave represents the operating new cloud: 1.5 gigawatts of active power, 3.7 contracted, $104 billion of backlog recognized against delivered service, $35 to $39 billion of planned 2026 capital expenditure, and the scars to show for the climb — $640 million of quarterly interest expense and a $626 million net loss even as revenue grew 112 percent [14][15][16]. At the far end, SB Energy represents the pure form: a $430 billion datacenter backlog resting on 0.8 gigawatts under construction and 8.0 contracted but unbuilt [2]. The layer’s defining risk follows directly from its position in the stack: the datacenter developer is the party that must physically deliver what every other layer has already financed. It absorbs construction risk from below and demand risk from above, and it is the natural home of the Delivery Gap analyzed in Section 4.1.
3.4 Layer Four — Models: Frontier Laboratories as Infrastructure Underwriters
At the Model layer sit the frontier laboratories — OpenAI, Anthropic, xAI, Meta, Google DeepMind, and their peers — and the argument of this subsection is that they have become the most consequential allocators of Anticipatory Capital in the entire economy, exercising an underwriting function far below their own layer. A laboratory’s decision to contract tens of billions of dollars of compute initiates, in sequence, investment in GPUs, networking equipment, datacenter shells, substations, generation, transmission, cooling, land, construction labor, and financial securities. The Model layer effectively underwrites portions of Layers One through Three.
The disclosed magnitudes justify the strong language. OpenAI has acknowledged approximately $1.4 trillion of infrastructure commitments over roughly eight years — spanning Broadcom, Oracle, Microsoft, Nvidia, AMD, Amazon, and CoreWeave — against an annualized revenue run rate that crossed $20 billion at the end of 2025, and HSBC analysis estimates the company would require approximately $207 billion of new financing by 2030 to fulfill its commitments as modeled [34][35]. Anthropic’s capacity assembly — approximately $45 billion with Nscale, $35 billion with Lambda, $50 billion with Fluidstack, $45 billion with SpaceX, alongside its multi-year CoreWeave arrangement — makes a single laboratory the anchor demand beneath several independent infrastructure platforms simultaneously [9][14]. Sam Altman, confronting the obvious question of what happens if the commitments outrun the revenue, gave an answer notable for its bluntness:
Sam Altman, CEO of OpenAI [34]
“If we screw up and can’t fix it, we should fail”
— and while the sentiment is admirably capitalist, Section 4 will examine why the failure of a Layer Four underwriter would not, in the current architecture, remain a Layer Four event [34].
3.5 Layer Five — Applications and Agents: The Ultimate Assumption
Yet every infrastructure contract in the preceding four layers ultimately rests on Layer Five, for a reason that no financial engineering can amend: someone must eventually pay for the intelligence being produced. The deepest assumption beneath a $45 billion compute contract is not merely that GPUs will operate, or that buildings will energize on schedule. It is that enterprises, consumers, developers, governments, robots, autonomous systems, and AI agents will generate sufficient economic value — and route sufficient spending — to pay for enormous quantities of training and inference at prices that service the entire capital stack beneath them.
The complete chain therefore runs backward through the stack: Expected Agentic Demand → Model Revenue → Compute Commitments → Datacenter Development → Chip Procurement → Power Infrastructure → Capital Formation. Anticipatory Capital connects the entire Five-Layer AI Economy backward from expected future applications to present physical investment, and the honest state of the evidence at Layer Five is mixed in ways this paper has an obligation to record. On the encouraging side: application revenue is real and compounding — OpenAI’s run rate crossed $20 billion, Anthropic’s reported run rate expanded dramatically through 2026, enterprise adoption broadened visibly in CoreWeave’s customer disclosures, and HSBC’s own skeptical analysis still projects billions of regular users of AI products by 2030 [14][34][35]. On the cautionary side stands the academic literature. Daron Acemoglu — the MIT Institute Professor and 2024 Nobel laureate whose task-based framework remains the most-cited formal treatment of AI’s macroeconomics — estimates that only about 5 percent of tasks will be profitably automated within the decade, translating into total factor productivity gains on the order of 0.55 to 0.7 percent and a GDP boost of roughly 1.1 to 1.6 percent over ten years, characterizing the effect as
Daron Acemoglu, Institute Professor, MIT, 2024 Nobel Laureate [24]
“nontrivial, but modest”
— a fraction of the projections embedded, implicitly, in trillion-dollar forward commitment structures [23][24][25]. The gap between Acemoglu’s decade-scale 1.1 percent and the revenue trajectories required to service $1.4 trillion of Layer Four commitments is not a rounding error; it is the entire question. If the optimists are right, Layer Five demand validates every claim beneath it and Anticipatory Capital will be remembered as the financial innovation that built the intelligence age ahead of schedule. If the task-based skeptics are right, the stack has capitalized a future that will arrive smaller and later than contracted — and the settlement of that difference is the subject of the next section.
Section 4: The Risks of Capitalizing the Future
Every financing innovation carries a characteristic failure mode, and the failure modes of Anticipatory Capital deserve systematic treatment rather than the episodic attention they receive in market commentary. This section identifies five, ordered from the most concrete to the most systemic: the Delivery Gap, counterparty concentration, technological obsolescence, demand-forecasting error, and the migration of corporate risk into systemic risk. They are not independent. The final subsection argues that their interaction — rather than any one of them alone — is what could transform a repricing of anticipated AI demand from a sector event into a macro-financial one.
4.1 The Delivery Gap
A signed contract does not produce electricity. A lease does not construct a substation. A financing commitment does not install GPUs, cure a concrete slab, or move a 765-kilovolt transmission line through a contested permitting docket. Between contractualization (Stage 2 of the Conversion Ladder) and monetization (Stage 6) sits what this paper calls the Delivery Gap — the accumulation of physical, regulatory, and operational conditions that must all be satisfied before anticipated revenue becomes actual revenue.
SB Energy’s own prospectus, read as a risk document rather than a marketing document, is among the best available maps of the Gap. Realization of its contracted capacity depends on construction completion across seventeen planned buildings, successful permitting, grid interconnection, lease commencement, and tenant acceptance; the company separately flags long procurement lead times for transformers and power equipment, and notes spreading community opposition to datacenter development as a factor that could hinder growth [2][6]. Independent reporting on the filing adds the temporal dimension: bringing new capacity online can take 24 to 36 months even after construction begins, first datacenter revenue is not expected until at least the end of 2026, and the Texas transmission upgrades supporting the near-term pipeline — including more than 400 miles of newly approved 765-kV infrastructure — remain exposed to legal challenges and ERCOT interconnection processes [3][6]. CoreWeave’s financial statements quantify what crossing the Gap costs even for an experienced operator: the company incurs lease, power, and hardware costs one to two months before each facility generates revenue, a structural drag that held adjusted operating margins to 5 percent in a quarter when revenue grew 112 percent, while $13.5 billion of quarterly debt issuance financed the climb [15][16].
The Delivery Gap should therefore become one of the principal risk measures in any serious analysis of forward AI claims. Its width can be expressed in time (months from capitalization to monetization), in money (capital consumed before first revenue), and in conditionality (the number of independent approvals, interconnections, and acceptances that must all succeed). By all three measures, the 2026 cohort of forward claims is historically wide — and the width is being financed, month by month, with borrowed money whose interest clock runs regardless of construction progress.
4.2 Counterparty Concentration
Anticipatory Capital becomes more fragile when an enormous infrastructure pipeline depends on one or two customers, and the flagship structures of 2026 are concentrated to a degree that would be remarkable in any other infrastructure asset class. SB Energy states in its own filing that it is substantially dependent on OpenAI, whose twenty-year leases constitute the overwhelming majority of the $430 billion datacenter backlog; commentators pressing on the filing noted that OpenAI is simultaneously tenant, investor, and warrant holder, that SoftBank occupies a similar multiplicity of roles, and that the related-party density of the structure makes the backlog harder, not easier, to evaluate [1][3][7]. Nscale’s transformation into a $103 billion forward book rests principally on two counterparties, Microsoft and Anthropic, with the Anthropic commitment alone accounting for nearly half the total [8][9]. Even CoreWeave’s admirably broadening enterprise base still layers $35 billion of Meta commitments atop concentrated hyperscaler and laboratory relationships [14].
The concentration problem compounds with the credit problem. The anchor tenants of the AI buildout are extraordinary companies, but several are not yet profitable at operating scale: OpenAI recorded a reported operating loss of $20.9 billion in 2025, with losses continuing into 2026 even as revenue compounded [3]. A twenty-year lease from a counterparty that has never generated sustained positive cash flow is a genuinely novel credit instrument, and the market’s solution — wrapping supplier guarantees around tenant obligations — resolves the immediate bankability question at the cost of deepening the circularity examined below. A trillion-dollar infrastructure ecosystem can therefore appear diversified geographically — campuses in Texas, Ohio, West Virginia, and beyond — while remaining concentrated economically in a handful of balance sheets whose own health depends on the very demand projections the infrastructure was built to serve.
4.3 Technological Obsolescence
AI infrastructure can suffer obsolescence far faster than traditional infrastructure, and the asymmetry cuts directly across the duration structure documented in Section 2.5. A transmission line might operate for sixty years. A well-built datacenter shell may serve for thirty. But rack density, cooling architecture, interconnect technology, accelerator generations, model efficiency, and inference economics can shift within a few years — and each shift reprices, at the margin, every long contract written under the previous regime’s assumptions.
Three vectors of obsolescence deserve individual attention. Hardware-generational risk is the most familiar: facilities engineered for one accelerator’s power draw and cooling profile face costly retrofit when the next generation arrives at higher densities, and contracts referencing specific architectures — as the Nscale-Anthropic agreement references Vera Rubin [9] — embed a particular moment of the technology curve into a six-year obligation. Efficiency risk is subtler and potentially larger: sustained improvements in model efficiency, quantization, distillation, and custom silicon reduce the compute required per unit of delivered intelligence, and while the historical pattern has been that efficiency gains expand total consumption rather than shrink it — the Jevons dynamic on which the entire bull case rests — a contract priced at scarcity rates does not automatically participate in that expansion. Architectural risk is the deepest: a genuine shift in how frontier systems are built or served — toward smaller specialized models, edge inference, or techniques not yet published — could redraw the map of where and how compute is consumed while two decades of lease payments remain outstanding on the old map. Anticipatory Capital must therefore price not merely construction risk but technological displacement risk, and the honest observation is that no one currently knows how to do so: the discount rates visible in 2026’s structures price counterparty credit and execution, while displacement risk is absorbed, unpriced, by whoever holds the longest claims.
4.4 The Demand Forecasting Problem
Beneath every layer of the analysis sits a forecast, and the forecast is the true underlying asset of the entire Anticipatory Capital complex. AI infrastructure developers increasingly extrapolate enormous future workloads from the extraordinary consumption growth of 2023–2026, and the extrapolation may prove correct — but it is worth being precise about how wide the plausible range remains, because the range itself is the risk.
On the downside, improvements in model efficiency, custom silicon displacing merchant GPUs, inference optimization, smaller task-specific models, edge computing, algorithmic breakthroughs, or simply slower-than-expected application monetization could materially reduce compute requirements relative to contracted trajectories. The institutional warnings accumulated steadily through 2025 and 2026: the Bank of England stated in October 2025 that
Bank of England, Financial Policy Committee [26]
“The risk of a sharp market correction has increased”
with equity valuations for AI-focused technology companies appearing stretched by measures comparable to the 2000 peak, and its July 2026 Financial Stability Report reiterated that AI-related valuations had become further stretched by historical standards while market concentration in a narrow set of AI names left indices exposed to any repricing of expectations [26][29]. IMF Managing Director Kristalina Georgieva, observing valuations approaching dot-com-era levels, warned investors bluntly that sentiment can reverse:
Kristalina Georgieva, Managing Director, IMF [27]
“History tells us this sentiment can turn abruptly”
and the IMF’s April 2026 Global Financial Stability Report devoted sustained attention to the household exposure channel — the concentration of AI-driven gains inside benchmark indices held through retirement accounts, which transmits any correction directly onto household balance sheets [27][28]. The academic demand skeptics, led by Acemoglu’s task-based estimates, supply the fundamental version of the same caution: if only a modest share of economic tasks is profitably automatable within the decade, the revenue base available to service the stack is correspondingly modest [23][24][25].
On the upside — and the upside is equally real — agentic AI, robotics, autonomous systems, video generation, scientific discovery models, and persistent always-on inference could produce demand far exceeding today’s projections, exactly as every previous general-purpose technology ultimately generated consumption categories its early forecasters never imagined. The hyperscalers’ own behavior encodes this view: four companies with different business models arrived simultaneously at the conclusion that being short compute is the one unaffordable mistake, and backed the conclusion with roughly $725 billion of 2026 capital expenditure [18][19]. Anticipatory Capital is therefore, at bottom, a wager on the future elasticity of intelligence consumption — on whether demand for cognition behaves like demand for bandwidth after 2000 (which eventually consumed the overbuilt fiber, but a decade later and after the equity that financed it was destroyed) or like demand for electricity after 1900 (which grew for a century and rewarded nearly every early kilowatt). The uncomfortable truth is that both precedents are available, both are being cited, and the contracts now being capitalized will pay out very differently depending on which one history selects.
4.5 From Corporate Risk to Systemic Risk
The largest concern emerges when the risks above stop being independent — when one company’s future commitment finances another company’s expansion, which supports another supplier’s capacity, which underwrites another company’s valuation. Each individual contract may be rational. Collectively, the ecosystem becomes financially interdependent, and interdependence is the raw material of systemic risk.
The channels of transmission are now well mapped. The circularity channel: more than $800 billion of arrangements in which vendors finance customers who purchase from those vendors means that a demand disappointment anywhere in the loop impairs balance sheets everywhere in it, a dynamic that CNBC’s Jim Cramer — reacting in July 2026 to reports of a proposed $250 billion Nvidia backstop for OpenAI’s Ohio expansion — explicitly compared to the vendor-financing arrangements that preceded the dot-com crash, when telecom equipment makers financed customer purchases that evaporated together [32][33]. The credit channel: at least $1.3 trillion of identified datacenter debt, projected global AI-related issuance approaching $570 billion in 2026, an $800 billion private-credit build-out, and $662 billion of not-yet-commenced lease obligations sitting off hyperscaler balance sheets collectively mean that the claims on anticipated AI demand are distributed across banks, bond funds, insurers, private-credit vehicles, and securitization structures whose aggregate exposure no single supervisor can currently observe [36][38][40]. The equity channel: AI-related companies’ weight in benchmark indices transmits any repricing directly to household wealth, as both the IMF and the Bank of England have documented [28][29]. And the macroeconomic channel, which is the most striking of all: Harvard’s Jason Furman calculated that investment in information-processing equipment and software — roughly 4 percent of U.S. GDP — accounted for approximately 92 percent of U.S. GDP growth in the first half of 2025, meaning the real economy’s measured expansion had itself become substantially a function of the AI buildout [20]. Kedrosky, in conversation with Paul Krugman, located the scale historically:
Paul Kedrosky, in conversation with Paul Krugman [22]
“somewhere between rural electrification and World War Two rearmament”
— with AI-related capital expenditure plausibly exceeding one trillion dollars annualized and constituting more than half of first-half U.S. GDP growth, arguably keeping the economy out of recession from a single sector’s spending [22].
The relevant question is therefore no longer merely whether Company A can pay Company B. It has become: how many balance sheets — corporate, financial, and household — have been constructed around the assumption that the same future AI demand materializes on schedule? That is where Anticipatory Capital potentially becomes a systemic-finance issue rather than a sector story, and it is why the governance frameworks of the next section are addressed not only to investors but to regulators whose mandates were written for a world in which infrastructure was financed after demand was demonstrated, not before.
Section 5: Governing and Measuring Anticipatory Capital
Diagnosis without measurement is commentary, and measurement without governance is trivia. This section therefore proposes, in ascending order of institutional ambition: a quality framework for evaluating individual forward claims; disclosure standards for public markets; allocation principles for utilities and state regulators; and a monitoring agenda for federal policymakers. The unifying premise is that Anticipatory Capital is neither to be celebrated nor suppressed but to be seen — priced, disclosed, and allocated with full knowledge of what it is.
5.1 The Anticipatory Capital Quality Framework
I propose evaluating forward AI infrastructure commitments along five dimensions, each answering a distinct question that headline backlog figures conflate:
Table — The Anticipatory Capital Quality Framework
| Dimension | Central Question | Illustrative Evidence |
| A — Contract Strength | How binding is the customer’s obligation? | Take-or-pay terms, termination rights, milestone conditions, remedy caps |
| B — Counterparty Quality | Can the customer realistically honor the commitment? | Profitability, funding runway, guarantee support, concentration of the customer’s own revenue |
| C — Delivery Readiness | Are land, permits, chips, construction, and cooling actually available? | Site control, permit status, equipment procurement, EPC capacity |
| D — Power Certainty | Is sufficient generation and interconnection secured? | Signed PPAs, interconnection position, transmission timeline, utility commitments |
| E — Technology Durability | Will the planned infrastructure remain economically competitive when completed? | Density and cooling headroom, refresh provisions, architecture-neutrality of pricing |
The product of the five dimensions yields what might be called an Anticipatory Capital Quality Score — a disciplined alternative to headline backlog. Applied to the September 2026 cohort, the framework immediately differentiates what the headlines merge. Nvidia’s $105 billion guarantee dramatically raises SB Energy’s score on dimension B while leaving dimensions C and D — seventeen unbuilt buildings, transformer lead times, contested transmission — exactly where the prospectus admits they are [3][6]. CoreWeave scores strongly on C and D by demonstration — it has repeatedly converted contracted power into active power — while carrying concentrated exposure on B and the sector’s shared uncertainty on E [15][16]. A remedy cap, a non-binding letter of intent, and an executed twenty-year lease occupy entirely different positions on dimension A even when journalism renders all three as “deals.” The framework’s purpose is not to generate a single ranking but to force the decomposition: any forward claim that cannot be scored on all five dimensions has not yet been analyzed, only admired.
5.2 Investors Need Better Forward-Capacity Disclosure
Public markets price what issuers disclose, and the current disclosure regime for forward AI capacity is a patchwork of company-defined terms. The reform agenda is straightforward and requires no new theory, only standardization. Issuers marketing forward capacity should be required to distinguish clearly among: operating capacity; capacity under construction; contracted capacity not yet under construction; capacity under exclusivity or reservation; speculative pipeline; contracted revenue meeting the definition of remaining performance obligations; estimated backlog beyond that definition; and pure management projection. Each category should carry its counterparty concentration, its weighted average remaining term, its termination and remedy provisions in summary form, and — following the logic of Section 2.4 — a statement of how much of the associated demand is financed, guaranteed, or invested in by parties within the issuer’s own supply chain.
The best current filings already approach this standard voluntarily: SB Energy’s segmentation of 0.8 gigawatts under construction from 8.0 contracted, its disclosure of the 19.6-year weighted term, and its explicit backlog caveats are genuinely informative [2][3]; CoreWeave’s separation of remaining performance obligations from estimated committed revenue is likewise a model [16]. The problem is that voluntary best practice coexists with headline arithmetic that adds letters of intent to leases, and the larger the AI infrastructure market becomes, the more expensive that ambiguity becomes for price discovery. A market that will be asked, over the coming eighteen months, to absorb the IPO queue now forming behind SB Energy, Nscale, and Accelevation deserves a common vocabulary before the queue arrives, not after the first major backlog restatement.
5.3 Governors and Grid Regulators Must Separate Contracts From Credible Loads
State governments and utilities face a parallel problem with higher physical stakes. A proposed datacenter may possess a customer contract yet still lack land readiness, interconnection position, power supply, financing, or final permitting — and a utility that builds generation against a load which never materializes will recover the cost from ratepayers for decades. The interconnection queues of every major U.S. grid region now contain anticipated load far exceeding any plausible near-term realization, partly because developers rationally file duplicative and optional requests, and partly because the same anticipated demand is being marketed simultaneously to multiple jurisdictions competing for the investment.
Utilities and governors therefore need methods — analogous to the Quality Framework above — for determining which future loads deserve scarce grid capacity: financial commitment requirements that scale with requested capacity, milestone-based queue positions that expire when delivery readiness stalls, tariff structures that assign the cost of speculative capacity to those who request it rather than to the general rate base, and ratepayer-protection provisions of the kind Nvidia and SB Energy asserted their Ohio arrangement was designed to include [3]. The policy objective should not be to obstruct AI development, which would merely relocate it. It should be to allocate infrastructure according to credible execution probability rather than headline megawatts alone — to ensure, in the vocabulary of this paper, that public infrastructure commitments are made at Stage 4 confidence, not Stage 1 enthusiasm.
5.4 Federal Policymakers Should Watch Cross-Layer Financial Concentration
The federal debate over AI concentrates on chips, export controls, model safety, energy production, and technological competition with China. Anticipatory Capital introduces a question the current debate largely omits: could financial concentration become an infrastructure chokepoint of its own? If a small number of AI laboratories, hyperscalers, chip suppliers, infrastructure funds, and banks collectively underwrite most future U.S. AI capacity, their private decisions — where to guarantee, which campus to anchor, which developer to finance — will determine where America’s power plants, datacenters, chip purchases, and transmission upgrades are built, with consequences for regional development, grid architecture, and national resilience that no public process currently reviews.
The monitoring agenda follows from the transmission channels of Section 4.5. Financial-stability supervisors should map aggregate exposure to anticipated AI demand across banks, insurers, private credit, and securitization markets — the $1.3 trillion of identified datacenter debt is a floor, not an estimate, precisely because so much of the stack is private [38]. The off-balance-sheet lease overhang — $662 billion not yet commenced at the five largest hyperscalers alone — should enter standard leverage analytics rather than remaining a footnote reconstructed by independent researchers from Moody’s data [40]. Macroeconomic agencies should institutionalize what Furman’s calculation demonstrated ad hoc: measured U.S. growth has become substantially a function of one investment category, which means a deceleration in that category is a macroeconomic event and should be modeled as one in advance [20]. And competition authorities should recognize that the customer-investor-supplier networks of Section 2.4, whatever their efficiency justifications, are also structures of mutual dependence among the small number of firms on which the entire national AI trajectory currently runs. Financial structure, in short, has become part of national AI strategy, whether or not policy chooses to treat it as such.
5.5 The 2027–2030 Question: How Much of the AI Economy Will Be Built on Forward Claims?
The next phase of the AI economy will test whether present expectations become future productivity, and the test has a schedule. The Nscale West Virginia campus is due to energize around the end of 2027 [9]. SB Energy’s first datacenter revenue is targeted from late 2026, with the Ohio program phasing in across the years beyond [6][3]. CoreWeave’s contracted 3.7 gigawatts must become active gigawatts quarter by quarter [15]. OpenAI’s $1.4 trillion of commitments amortize against a revenue trajectory that must, per its own bankers’ models, close a nine-figure financing gap before 2030 [34][35]. The window from 2027 through 2030 is when the widest cohort of forward claims in the history of infrastructure finance is scheduled to cross the Delivery Gap — and when the market will learn, empirically, what fraction of Anticipatory Capital converts.
If revenues from models, agents, enterprise automation, robotics, scientific discovery, autonomous transportation, and consumer applications grow rapidly, Anticipatory Capital will have proved an extraordinarily efficient mechanism for accelerating infrastructure construction — capital arriving years ahead of shortage, timelines compressed, and the United States entering the 2030s with the physical substrate of the intelligence economy already built. If demand disappoints, technological change renders projects obsolete, or customers cannot honor increasingly enormous commitments, the same mechanism will have exposed the degree to which present valuations incorporated revenues that had not yet arrived — and the unwinding will run backward through every layer the forward claims ran forward through. The AI infrastructure boom is therefore best understood as an experiment: the capitalization of technological expectations at unprecedented physical scale, with the results scheduled to report between 2027 and 2030.
Section 6: What Have We Learned? Seven Pillars
The argument of this paper can be distilled into seven pillars — five carried forward from the framework’s original formulation, and two added because the evidence assembled above demands them. Together they constitute what I take to be the durable lessons of the 2025–2026 phase of the AI investment cycle, whatever the coming years decide about the fate of any individual company or contract.
Pillar 1 — Future AI Demand Is Becoming a Present Financial Asset
The first lesson is that AI demand no longer needs to be realized before it influences capital formation. A sufficiently credible future commitment now affects financing, valuation, construction decisions, supplier capacity, and investor behavior years before the associated compute is delivered. Anthropic’s signature moved $52 billion onto Nscale’s forward book and moved $3.5 billion of pre-IPO capital toward the company within days [8][11]; OpenAI’s leases made a company with zero operating datacenters one of the largest IPO candidates of the year [1]. Anticipation itself has acquired economic value — it can be contracted, disclosed, marketed, guaranteed, and financed. This is the founding fact of Anticipatory Capital, and it is not going away: once markets learn to monetize expectation at this scale, the capability persists across cycles, in AI and beyond.
Pillar 2 — Contract Quantity Matters Less Than Contract Quality
A $100 billion headline is not necessarily worth $100 billion economically. Its value depends on enforceability, duration, counterparty strength, cancellation rights, construction milestones, power availability, technology assumptions, and the probability that the underlying infrastructure becomes operational — the five dimensions of the Quality Framework, compounded. The 2026 record supplies the proof by demonstration: a reported $600 billion of discussed financing resolved into a signed $105 billion guarantee [33]; illustrative contracted-revenue figures were expressly distinguished from guidance by the very bankers circulating them [8]; and prospectuses warned in plain language that backlog is a hypothesis about the future, not a claim on it [2]. The critical metric is therefore not how much has been contracted, but how much of the contract can realistically be converted into cash flow — and at what date, against what conditions, and net of whose optionality.
Pillar 3 — The Five Layers Are Beginning to Finance One Another
The Five-Layer AI Economy should no longer be understood only as a technological stack; it has become a financially interdependent stack. Expected applications justify models. Models justify compute contracts. Compute contracts justify datacenters. Datacenters justify chip purchases. Chip deployments justify electricity infrastructure. And the capital flows in the opposite direction from the dependency: the chip layer invests in the model layer, the model layer underwrites the datacenter layer, the datacenter layer anchors the energy layer. Capital travels backward through the Five Layers before revenue travels forward through them — and the interval between the backward journey of capital and the forward journey of revenue is precisely where Anticipatory Capital lives, and where its risks accumulate.
Pillar 4 — Physical Reality Remains the Ultimate Settlement Mechanism
Financial markets can anticipate infrastructure; they cannot eliminate physics. Transformers must still be manufactured on multi-year lead times. Transmission lines must still survive contested approvals. Electricity must still be generated at the hour the training run demands it. Datacenters must still be built by finite construction labor forces; GPUs must still be delivered, installed, cooled, and networked; permits must still be granted by institutions that answer to communities whose opposition SB Energy’s own filing acknowledges is spreading [6]. Every forward claim eventually encounters physical reality, and physical reality is the settlement layer of Anticipatory Capital — the place where the contract either becomes a building or becomes litigation. The wisest participants in the 2026 market are distinguishable by exactly this test: they underwrite the physics first and the finance second.
Pillar 5 — The Macroeconomy Itself Has Become a Counterparty
The fifth pillar is new to this formulation, and it elevates a finding from Section 4.5 to the rank it deserves: the AI buildout is no longer merely exposed to the macroeconomy; the macroeconomy is exposed to the AI buildout. When 4 percent of GDP accounts for roughly 92 percent of measured growth [20]; when a single sector’s annualized capital expenditure plausibly exceeds one trillion dollars and stands, in Kedrosky’s historical placement, between rural electrification and wartime rearmament [22]; when central banks and the IMF publish standing warnings that index-level household wealth is concentrated in the sector’s equities [27][28][29] — then national economic performance, employment in construction and utilities, regional development, and household balance sheets have all become, functionally, counterparties to the forward claims this paper describes. A boom that becomes a load-bearing element of GDP cannot decelerate privately. This does not argue for slowing the buildout; it argues for macroeconomic institutions treating AI-investment trajectories with the same standing analysis they devote to housing, energy prices, and credit conditions.
Pillar 6 — Transparency Is the Cheapest Systemic Insurance Available
The sixth pillar is likewise new, and it converts Section 5 into a principle: nearly every pathological outcome available to Anticipatory Capital is made worse by opacity, and nearly every productive outcome is made cheaper by disclosure. The $662 billion of not-yet-commenced leases invisible to balance-sheet analysis [40]; the private-credit and SPV structures the BIS flags as hard to trace [36]; the definitional drift by which letters of intent, reservations, and executed leases merge into single headline numbers — none of these creates risk by itself, but each ensures that when risk arrives it will be mispriced, misallocated, and discovered late. The dot-com era’s most expensive lesson was not that expectations were wrong but that the financial system could not see where the wrong expectations had been distributed until they failed simultaneously. Standardized forward-capacity disclosure, circularity reporting, and supervisory mapping of AI-linked credit exposure are, by the standards of what they insure against, nearly free.
Pillar 7 — The Central Question Is Shifting From “How Much Will Be Built?” to “How Much of the Future Has Already Been Capitalized?”
The final lesson is the most important, and it reframes the entire research agenda. The AI infrastructure boom is no longer adequately described by annual capital expenditures alone — by the $700 billion hyperscaler headlines or the gigawatt announcements that dominate coverage [18][19]. Researchers, investors, and policymakers must now examine a compound object: future commitments plus financial claims plus counterparty relationships plus infrastructure readiness plus expected monetization. The decisive question for 2027 and beyond may not be how many gigawatts, GPUs, or datacenters companies intend to construct. It may be how much financial value has already been assigned to those future assets before society knows whether the underlying assumptions will be realized — and who, exactly, is holding that value when the answer arrives.
Conclusion: Anticipatory Capital and the Financial Colonization of the Future
Artificial intelligence is producing one of the largest infrastructure investment cycles of the modern era — by mid-2026, an annual program approaching three-quarters of a trillion dollars from four companies alone, a multi-trillion-dollar cumulative trajectory through 2030, and a physical footprint reshaping regional grids from Texas to Ohio to West Virginia [18][19][37]. Yet the most consequential transformation of this cycle may not be visible in semiconductor fabs, datacenter campuses, power plants, or GPU racks. It may be occurring inside contracts, financing agreements, investor presentations, IPO prospectuses, and corporate balance sheets — in the quiet conversion of tomorrow’s anticipated intelligence production into today’s financial power.
SB Energy’s extraordinary backlog alongside an absence of operating datacenter capacity provides the vivid illustration with which this paper opened [1][2]. Nscale’s approximately $103 billion of contracted revenue, Anthropic’s roughly $45 billion commitment to a campus that will not energize until late 2027, CoreWeave’s $104 billion book, Accelevation’s arrival in the same IPO queue, and OpenAI’s $1.4 trillion of layered commitments reinforce the same underlying pattern [8][9][12][14][34]. The AI economy has learned how to convert expectations of future compute consumption into capital available today — and it has learned to do so at a scale, a speed, and a degree of reflexivity that no prior infrastructure cycle achieved.
That is precisely why the term Anticipatory Capital fits. It does not simply describe speculative investment, ordinary project finance, or corporate backlog; it identifies a more specific economic process: the capitalization of credible future claims on AI productive capacity before the underlying infrastructure becomes fully operational. The adjective Anticipatory matters because the financing arrives ahead of production. The noun Capital matters because those expectations do not remain predictions — they influence valuations, mobilize debt and equity, initiate construction, reserve scarce electricity, secure semiconductor capacity, reshape corporate strategy, and redirect regional economic development.
Within the Five-Layer AI Economy, Anticipatory Capital also reveals something deeper. Energy, Chips, Datacenters, Models, and Applications and Agents are no longer connected only through technology; they are increasingly bound together by forward financial claims. The anticipated success of an AI agent several years from now helps justify a model company’s compute commitment today; that commitment supports a datacenter developer’s financing; that financing generates orders for accelerators; those accelerators justify new substations and generation projects — and, in the circuit’s most distinctive feature, the accelerator maker’s own capital flows backward through the chain to strengthen the commitments that justify its orders. Future intelligence has begun pulling physical capital toward itself before the intelligence has generated the revenue required to pay for it.
This mechanism can be extraordinarily productive. It compresses infrastructure-development timelines, allows society to build capacity before shortages become overwhelming, and may prove to be one of the financial engines that makes the next stage of AI scaling possible at all — the institutional answer to the coordination problem of building energy, chips, buildings, and models on synchronized schedules that no single balance sheet could fund alone. But its strength is also its vulnerability. When hundreds of billions of dollars of prospective value depend upon forecasts extending across multiple technology generations, unproven counterparties, contested construction schedules, regulatory approvals, power systems, and yet-unproven application revenues — and when the institutions warning about the resulting concentration include the IMF, the Bank of England, and the most-cited economist of his generation [24][27][29] — the distinction between financing the future and borrowing from the future becomes the most important distinction in the entire AI economy. Financing the future means present capital builds capacity that future demand fills. Borrowing from the future means present valuations consume returns that future demand was supposed to deliver, leaving the shortfall to be settled — as such shortfalls always are — by whoever holds the claims when the music changes.
That distinction is ultimately what this paper seeks to illuminate, and what the Conversion Ladder, the Quality Framework, and the seven pillars are designed to help investors, regulators, and researchers keep in view as the great cohort of forward claims approaches its 2027–2030 settlement window. The defining financial innovation of the next phase of the AI economy may therefore be neither a new security, nor a new cryptocurrency, nor a new form of venture capital. It may be the ability to convert a promise about tomorrow’s intelligence infrastructure into economic power today.
That is Anticipatory Capital.
Endnotes:
[1] Jaiveer Shekhawat and Manya Saini, Reuters — “SoftBank-backed SB Energy files for U.S. IPO as AI turbocharges infrastructure demand,” September 1, 2026. https://www.bnnbloomberg.ca/business/artificial-intelligence/2026/09/01/softbank-backed-sb-energy-files-for-us-ipo-as-ai-turbocharges-infrastructure-demand/
[2] Securities.io Editorial — “SoftBank-Owned SB Energy Files S-1 for Proposed Nasdaq IPO” (summarizing SB Energy Form S-1, U.S. Securities and Exchange Commission, September 1, 2026). https://www.securities.io/softbank-owned-sb-energy-files-s-1-for-proposed-nasdaq-ipo/
[3] Yahoo Finance Technology — “SoftBank’s SB Energy IPO filing offers 8.8 GW of contracted AI capacity, with most still unbuilt,” September 2026. https://finance.yahoo.com/technology/ai/articles/softbank-8217-sb-energy-ipo-145941881.html
[4] Jordan Novet, CNBC — “SoftBank’s SB Energy files for IPO, says it’s ‘substantially dependent’ on OpenAI,” September 1, 2026. https://www.cnbc.com/2026/09/01/sb-energy-ipo-softbank-open-ai-nvidia.html
[5] Lukas Muehlbauer, IPOX Research, quoted by Reuters — “IPOX Research Associate Lukas Muehlbauer Comments on SB Energy’s IPO and AI Infrastructure Demand,” September 2026. https://www.ipox.com/ipox/reuters-muehlbauer-sbenergy
[6] Latitude Media — “Under the hood of SB Energy’s IPO filing,” September 2026. https://www.latitudemedia.com/news/under-the-hood-of-sb-energys-ipo-filing/
[7] Maria Deutscher, SiliconANGLE — “SoftBank’s SB Energy AI infrastructure unit files to go public,” September 1, 2026. https://siliconangle.com/2026/09/01/softbanks-sb-energy-ai-infrastructure-unit-files-to-go-public/
[8] Investing.com / The Information — “Nscale doubles contracted revenue to $103 billion after Anthropic win,” September 2, 2026. https://finance.yahoo.com/technology/ai/articles/nscale-doubles-contracted-revenue-103-222526129.html
[9] Quartz — “Nscale touts $103 billion in contracts ahead of IPO,” September 3, 2026. https://qz.com/nscale-ipo-103-billion-contracted-revenue-090326
[10] Reuters — “Nscale touts $103 billion contracted revenue ahead of potential IPO, The Information reports,” September 2, 2026. https://www.aol.com/articles/nscale-touts-103-billion-contracted-214955000.html
[11] Crypto Briefing / Reuters — “Nscale seeks $3.5B in financing ahead of planned US IPO,” September 4, 2026. https://cryptobriefing.com/nscale-seeks-billions-financing-ahead-ipo/
[12] Reuters — “Olympus-backed Accelevation files for US IPO,” September 2, 2026. https://www.aol.com/articles/olympus-backed-accelevation-files-us-222235000.html
[13] TradingView News — “Accelevation, Data Center Infrastructure Solutions Provider, Files for Nasdaq IPO” (summarizing Accelevation Holdings Corp. Form S-1, September 2, 2026). https://www.tradingview.com/news/tradingview:af69b1b20deb7:0-accelevation-data-center-infrastructure-solutions-provider-files-for-nasdaq-ipo/
[14] CoreWeave, Inc. — “CoreWeave Reports Strong Second Quarter 2026 Results,” Investor Relations, August 11, 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx
[15] Ari Levy, CNBC — “CoreWeave (CRWV) Q2 earnings report 2026,” August 11, 2026. https://www.cnbc.com/2026/08/11/coreweave-crwv-q2-earnings-report-2026.html
[16] William Foxley, Blockspace — “CoreWeave Q2 revenue reaches $2.58 billion as backlog grows to $104 billion,” August 11, 2026. https://blockspace.media/insight/coreweave-q2-revenue-backlog-grows-2026/
[17] NVIDIA Corporation — “NVIDIA Announces Financial Results for First Quarter Fiscal 2027,” Form 8-K exhibit, U.S. Securities and Exchange Commission, May 20, 2026. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000051/q1fy27pr.htm
[18] Ari Levy, CNBC — “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
[19] Yahoo Finance / Goldman Sachs Research — “Meta, Microsoft, Amazon, and Alphabet are about to spend a shocking amount of money to dominate the AI era,” June 3, 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
[20] Nick Lichtenberg, Fortune — “Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says” (on Jason Furman’s calculation), October 7, 2025. https://fortune.com/2025/10/07/data-centers-gdp-growth-zero-first-half-2025-jason-furman-harvard-economist
[21] Paul Kedrosky — “Honey, AI Capex is Eating the Economy,” July 2025, as excerpted by Simon Willison. https://simonwillison.net/2025/Jul/19/paul-kedrosky/
[22] Paul Krugman — “Talking With Paul Kedrosky,” Paul Krugman Substack, December 2025. https://paulkrugman.substack.com/p/talking-with-paul-kedrosky
[23] Daron Acemoglu — “The Simple Macroeconomics of AI,” NBER Working Paper No. 32487, National Bureau of Economic Research, 2024; published in Economic Policy 40(121), 2025. https://www.nber.org/papers/w32487
[24] MIT Sloan School of Management — “A new look at the economics of AI” (on Daron Acemoglu’s estimates), January 2026. https://mitsloan.mit.edu/ideas-made-to-matter/a-new-look-economics-ai
[25] Fortune — “Nobel Laureate Daron Acemoglu on the ‘brainless’ AI discourse, the myth of capitalism and the Gen Z revolution risk,” June 21, 2026. https://fortune.com/2026/06/21/nobel-laureate-daron-acemoglu-ai-productivity-capitalism-democracy/
[26] Associated Press — “Is there an AI bubble? Financial institutions sound a warning” (Bank of England Financial Policy Committee and IMF statements), October 8, 2025. https://seekingalpha.com/pr/20259584-is-there-an-ai-bubble-financial-institutions-sound-a-warning
[27] Kristalina Georgieva, IMF, via Scottish Financial News — “IMF and Bank of England warn of AI-driven market correction,” October 2025. https://www.scottishfinancialnews.com/articles/imf-and-bank-of-england-warn-of-ai-driven-market-correction
[28] International Monetary Fund — Global Financial Stability Report, April 2026, Chapter 1. https://www.imf.org/-/media/files/publications/gfsr/2026/april/english/ch1.pdf
[29] Bank of England, Financial Policy Committee — Financial Stability Report, July 2026. https://www.bankofengland.co.uk/-/media/boe/files/financial-stability-report/2026/financial-stability-report-july-2026.pdf
[30] Bloomberg Graphics — “AI Circular Deals: How Microsoft, OpenAI and Nvidia Keep Paying Each Other,” January 22, 2026. https://www.bloomberg.com/graphics/2026-ai-circular-deals/
[31] Stacy Rasgon, Bernstein Research, via Bloomberg / Business Standard — “Nvidia-OpenAI deal sparks concerns over circular financing in AI boom,” September 24, 2025. https://www.business-standard.com/amp/companies/news/nvidia-openai-deal-sparks-concerns-over-circular-financing-in-ai-boom-125092401589_1.html
[32] NBC News — “The AI boom’s reliance on circular deals is raising fears of a bubble,” October 2025. https://www.nbcnews.com/business/economy/openai-nvidia-amd-deals-risks-rcna234806
[33] CNBC — “Jim Cramer warns AI’s circular financing frenzy echoes the dot-com bubble,” July 27, 2026. https://www.cnbc.com/2026/07/27/jim-cramer-warns-ai-circular-financing-echoes-dot-com-bubble.html
[34] Rashi Shrivastava, Forbes — “Here’s What Happens If OpenAI Can’t Pay For Its $1.4 Trillion AI Deals,” November 7, 2025. https://www.forbes.com/sites/rashishrivastava/2025/11/07/why-sam-altman-wont-be-on-the-hook-for-openais-massive-spending-spree/
[35] Data Center Dynamics — “OpenAI must find $207bn to meet AI data center spending commitments — HSBC,” July 2026. https://www.datacenterdynamics.com/en/news/openai-must-find-207bn-to-meet-ai-data-center-spending-commitments-hsbc/
[36] Robert Szczerba, Forbes — “Bond Investors Push Back As AI Debt Heads Toward $570 Billion” (Morgan Stanley and Bank for International Settlements analysis), July 17, 2026. https://www.forbes.com/sites/robertszczerba/2026/07/17/bond-investors-push-back-as-ai-debt-heads-toward-570-billion/
[37] Yahoo Finance — “AI Data Center Boom Confronts $3 Trillion Bill as Debt Markets Step In” (Morgan Stanley, Moody’s Ratings, and JPMorgan estimates), February 2, 2026. https://finance.yahoo.com/news/ai-data-center-boom-confronts-195502219.html
[38] Bisnow / AtriumData.ai — “These 15 Lenders Are Leading The $1.3T Data Center Debt Surge,” August 2026. https://www.bisnow.com/news/national/data-center-capital-markets/these-15-companies-leading-trillion-dollar-data-center-lending-boom
[39] J.P. Morgan Investment Banking — “Financing AI infrastructure and U.S. data centers,” August 2026. https://www.jpmorgan.com/insights/banking/capital-markets/financing-ai-infrastructure-data-centers
[40] Axis Intelligence Research — “AI Data Center Financing Statistics 2026: Bonds, Private Credit, and the $1.5 Trillion Gap” (analysis of Moody’s Ratings data), August 2026. https://axis-intelligence.com/ai-data-center-financing-statistics/
