Introduction:
On September 17, 2026, two transactions arrived within hours of each other that captured, with unusual clarity, how dramatically the financing of artificial intelligence has changed in only a few years. CoreWeave, one of the companies at the center of the GPU-cloud expansion, announced its intention to raise $3.0 billion through a private offering of convertible senior notes due 2033, with initial purchasers granted an option to buy up to an additional $500 million.[1] The following day, the transaction was priced at an upsized $3.7 billion — a 2.875 percent coupon, a conversion price of approximately $97.85 per share, and a settlement date of September 22, 2026 — underscoring both the depth of investor demand for AI-linked hybrid securities and the extraordinary capital intensity of the underlying business.[2] On that same September 17, Crusoe, which describes itself as the industry’s first vertically-integrated AI infrastructure provider, announced the initial closing of a $3.9 billion Series F financing at a $30.9 billion post-money valuation, an oversubscribed round co-led by Atreides Management, Mubadala Capital, and Valor Equity Partners, with participation from Founders Fund, GIC, Nvidia, the Qatar Investment Authority, Radical Ventures, and TPG.[3] Two companies serving essentially the same expanding market for artificial-intelligence infrastructure were therefore raising billions of dollars almost simultaneously — but through substantially different financial instruments, aimed at substantially different investor bases, with substantially different implications for risk, dilution, duration, and control.
Neither transaction stood alone, and neither was even the most structurally ambitious financing of its month. A much larger experiment in financial architecture was already emerging around SB Energy, the AI infrastructure and power company majority-owned by SoftBank, which filed its Form S-1 with the Securities and Exchange Commission on September 1, 2026 for a proposed initial public offering on Nasdaq under the ticker SBE.[5] The planned offering combines conventional IPO equity with strategic capital from Nvidia, warrants issued to OpenAI, long-duration data-center leases, power-development commitments, and forms of credit support that have no clean precedent in the history of technology finance. SB Energy’s filing describes Nvidia arrangements that include a $1.5 billion prepaid forward contract entered into on August 17, 2026 and a separate $1.5 billion private placement of non-voting shares closing concurrently with the offering — a combined $3 billion strategic investment by the chipmaker in the company that will build the campuses where its own products are to be deployed.[6] OpenAI, for its part, holds warrants covering approximately 3.99 million SB Energy shares at a nominal exercise price of one cent per share, warrants that were issued in January 2026 as an inducement for OpenAI to enter into its foundational agreements and that were estimated in the filing to be worth roughly $5.5 billion by mid-2026, vesting in stages after the offering as valuation milestones are reached.[10] OpenAI is simultaneously expected to occupy the company’s planned PORTS-Pike Technology Campus in Pike County, Ohio under twenty-year leases covering approximately 8.0 gigawatts of critical IT capacity across seventeen data-center buildings. Nvidia, meanwhile, has agreed to provide conditional credit support — residual-value guarantees with an aggregate cap of $105 billion — associated with an initial approximately 4.25 gigawatts of that capacity, with an option to extend its support to the remainder.[7]
The financing architecture around that Ohio project rewards a moment of careful attention, because it compresses nearly every theme of this paper into a single development. SB Energy secures the land, originates and manages the power infrastructure, and constructs the data-center shells. OpenAI becomes the long-term anchor tenant, converting its own expectations of future compute demand into contractual lease obligations stretching two decades into the future. Nvidia supplies the computing architecture that will occupy the buildings, invests capital into the developer through both a prepaid forward and a private placement, and, under specified circumstances and subject to a declining schedule, stands behind defined portions of the tenant’s lease and power obligations. Nvidia’s own quarterly filing discloses that these guarantees generally become effective as each of nine data centers is placed in service beginning in its fiscal year 2029, with exposure declining as OpenAI fulfills its lease payments.[8] The campus is expected ultimately to support approximately 8 gigawatts of IT capacity, while the underlying development requires enormous investments in generation, transmission, buildings, networking, and computing equipment years before the associated AI workloads generate their eventual economic return. SB Energy’s chief executive, Rich Hossfeld, explained the logic of the supplier’s involvement with disarming directness, telling CNBC that Nvidia’s participation
“helps us to unlock things like investment-grade financing.” — Rich Hossfeld, CEO, SB Energy [9]
That single sentence deserves to be read twice, because it describes something genuinely new in the industrial history of computing: the world’s most valuable semiconductor company lending its balance sheet not to sell chips on credit, but to transform an unbuilt campus in southern Ohio into an investment-grade borrower.
This is not simply a story about technology companies borrowing more money. It is a story about the construction of a new financial architecture around artificial intelligence, an architecture in which the boundaries between supplier and financier, customer and shareholder, landlord and guarantor are dissolving into a dense web of interlocking claims.
For most of the first phase of the AI boom, the enormous infrastructure expenditures involved could be financed primarily by the operating cash flows and balance sheets of Microsoft, Alphabet, Amazon, Meta, and a handful of other highly profitable technology companies, firms that entered the era with pristine credit ratings, decades of accumulated cash, and equity currencies of extraordinary value. But as artificial-intelligence infrastructure expands from clusters containing thousands of GPUs toward campuses measured in gigawatts, the required capital extends far beyond even the largest hyperscalers. Specialized clouds, data-center developers, utilities, independent power producers, networking companies, semiconductor suppliers, nuclear projects, renewable-energy developers, equipment manufacturers, sovereign wealth funds, private-credit managers, and increasingly the public capital markets themselves must participate, each bringing its own risk appetite, its own duration preferences, and its own required return.
The Bank for International Settlements identified this transition with notable precision at the very beginning of 2026. In BIS Bulletin No. 120, published on January 7, 2026, researchers Iñaki Aldasoro, Sebastian Doerr, and Daniel Rees documented that AI-related investment was surging both in nominal terms and as a share of GDP, that the anticipated scale of future investment would require firms to shift their financing from operating cash flows toward debt, and that private credit was already playing a rapidly increasing role, with more than $200 billion of private credit outstanding to AI-related borrowers and information-technology investment reaching approximately 5 percent of United States GDP — a level exceeding even the peak of the dot-com era.[11] The institution returned to the theme in its Annual Economic Report of June 2026, which named the sustainability of the AI investment boom as one of the principal pressure points confronting the global economy, observed that the five largest hyperscalers were on pace to spend more than $1 trillion on AI-related capital expenditure across 2025 and 2026 combined, and warned, in language rarely deployed by the central bank of central banks, that the scale and pace of the boom bear resemblance to historical investment manias whose endings were not gentle:
“These episodes ended with an eventual reversal in investment, inducing economy-wide recessions.” — Bank for International Settlements, Annual Economic Report 2026 [12]
In a subsequent July 2026 working paper, BIS economist Phurichai Rungcharoenkitkul modeled the AI build-out as a dynamic contest in which firms racing for a small number of dominant positions rationally over-commit resources, financing early scale through debt and circular equity ties, and calibrated the resulting over-investment at roughly 1.5 times the efficient level, rising toward three times where demand proves less elastic. His formulation of the central risk is stark in its brevity:
“The larger the boom, the deeper the eventual bust.” — Phurichai Rungcharoenkitkul, Bank for International Settlements [13]
That shift — from a capital-expenditure program conducted inside a few corporate treasuries to a financing system spanning the entire capital markets — changes the nature of what I have elsewhere called the Five-Layer AI Economy. Layer 1, energy, requires power plants, transmission lines, transformers, interconnection rights, storage, cooling, and long-duration power contracts. Layer 2, chips, requires accelerators, memory, networking silicon, advanced packaging, and semiconductor fabrication capacity. Layer 3, datacenters, requires land, buildings, substations, servers, fiber, cooling systems, and years of construction. Layer 4, models, requires enormous and continuous training and inference expenditures. Layer 5, applications and agents, must eventually create the revenue capable of supporting everything underneath. Increasingly, finance does not sit beside these five layers as a supporting service; it runs through all of them simultaneously, binding them together with contracts and claims.
The significant development of 2025 and 2026 is therefore not the creation of any one particular “AI capital stack.” The more important development is that different claims on future artificial-intelligence demand are being systematically converted into financeable instruments. A long-term cloud contract can support borrowing. A data-center lease can support project financing. A GPU fleet can contribute to an infrastructure company’s asset base. A capacity reservation can make construction bankable. A strategic supplier can provide equity, prepayments, or guarantees. Customer commitments can reduce perceived future revenue uncertainty. Warrants can compensate counterparties for assuming early risk. Convertible bonds allow companies to borrow today while providing investors potential participation in tomorrow’s equity value. What ultimately supports these financings is not always “collateral” in the strict legal meaning of secured lending; it is better understood as an emerging economic asset base composed of GPUs, land, power rights, leases, customer contracts, supplier relationships, capacity commitments, and — most abstractly and most importantly — expectations about future demand for tokens, inference, and autonomous machine work.
Artificial intelligence is therefore beginning to acquire something that every previous technology revolution eventually developed as well, from the railroads of the 1880s to the telecommunications networks of the 1990s: its own specialized financial machinery. I call this emerging phenomenon Convertible Compute.
Why I Choose the Title “Convertible Compute”
I choose the term Convertible Compute because artificial-intelligence infrastructure is increasingly being converted from a technological asset into a financial claim, and because that conversion runs in both directions at once. GPUs become the productive equipment standing behind cloud contracts. Electricity rights become the foundation for gigawatt-scale leases. Future AI capacity becomes the basis for customer commitments that stretch across decades. Data-center contracts influence debt capacity and credit ratings. Supplier relationships generate warrants, prepayments, equity investments, and guarantees. In this sense, compute is becoming financially “convertible” in a way that goes far beyond any single instrument: physical infrastructure and expectations of future machine intelligence can increasingly be translated into securities, contractual claims, credit support, and investable cash flows, and those financial claims can in turn be translated back into steel, silicon, and megawatts.
The word Convertible also deliberately echoes the convertible securities that have appeared with striking frequency around AI companies, of which CoreWeave’s September 2026 financing is only the most recent and most literal example. Reuters reported in May 2026 that United States convertible issuance had reached approximately $34 billion in the first four months of the year — more than double the same period a year earlier — with roughly half of that issuance tied in some way to artificial intelligence, putting the market on track to surpass the prior full-year record of more than $120 billion set in 2025.[14] By early September 2026, the record had already fallen: companies listed in the United States had raised roughly $131 billion in convertible debt year-to-date, with the month of August alone accounting for $25 billion, while global convertible issuance had reached $186.8 billion across 362 deals, approximately 60 percent of it connected to the AI boom according to Barclays’ head of United States equity strategy, and zero-coupon structures — bonds that pay no interest at all — accounting for roughly 41 percent of United States issuance.[15] [16]
But the concept of Convertible Compute is deliberately broader than convertible bonds, and this breadth is the reason the title fits this paper better than a conventional phrase such as “AI infrastructure finance” or “AI debt.” The title describes a transformation occurring across the entire Five-Layer AI Economy: capital is being converted into compute, while expected compute demand is simultaneously being converted back into finance. A twenty-year lease signed by OpenAI in Ohio, a $105 billion guarantee extended by Nvidia, a $13 billion five-year cloud contract signed between Crusoe and the quantitative trading firm Jane Street,[44] a $300 billion take-or-pay compute commitment between OpenAI and Oracle, and a $3.7 billion convertible note priced by CoreWeave are, in this framework, all instances of the same underlying phenomenon, differing in legal form but identical in economic function: each converts a belief about the future demand for machine intelligence into a claim that can be financed today. That two-way conversion — capital into compute, and expected compute back into capital — is the financial architecture this paper sets out to describe.
Section 1: From Artificial-Intelligence Capex to Artificial-Intelligence Finance
The deepest structural change now underway in the AI economy is easy to miss precisely because it does not announce itself in any single headline. It is the migration of the financing burden — from the internal cash flows of a handful of extraordinarily profitable corporations to the external capital markets, and from the corporate balance sheet to the project, the contract, and the special-purpose vehicle. This section traces that migration in four movements: the cash-financed first phase of the boom; the outward migration of the capital requirement as infrastructure scales toward gigawatts; the transformation of software economics into infrastructure economics; and the reasons why the emerging system is better understood as a financial architecture than as a simple accumulation of debt.
1.1 The First AI Boom Was Financed by Cash
The earliest phase of generative artificial intelligence benefited from a group of sponsors that was, in historical terms, almost uniquely well suited to absorb its costs. Microsoft, Amazon, Alphabet, Meta, and their peers entered the boom with enormous operating cash flows, established borrowing capacity, valuable equity currencies, minimal net debt, and investment-grade balance sheets that had been fortified by a decade of software-margin profitability. When the training costs of frontier models first escalated from millions of dollars into the billions, these companies could simply absorb the expenditure as another line in an already large capital budget, in much the same way that they had absorbed the costs of building conventional cloud infrastructure in the 2010s. They could spend tens of billions of dollars on GPUs and datacenters without constructing a new financing system, without asking permission from credit markets, and without materially altering their capital structures.
The BIS documented this initial condition carefully in its January 2026 bulletin: the firms driving the AI investment boom had historically operated with substantially less debt than comparable large corporations, relying instead on highly profitable operations to generate the cash flows needed to fund investment. But the same bulletin identified the inflection point with equal care: capital expenditures were growing rapidly both in absolute terms and as a share of revenues, and the sheer size of anticipated investments, combined with dwindling free cash flows in some cases, was testing the limits of expansion financed by cash alone.[11] The Annual Economic Report that followed in June 2026 quantified the strain, noting that the commitments of the five largest hyperscalers — more than $1 trillion of AI-related capital expenditure across 2025 and 2026 — were outpacing their earnings and free cash flow, leading some to issue debt to raise additional financing, and warning of the risk that firms were over-committing resources to investment projects with still-uncertain returns.[42]
That condition is now changing visibly, quarter by quarter. In the second-quarter 2026 earnings round, Amazon raised its full-year capital expenditure guidance to approximately $220 billion, Alphabet raised its range to between $195 billion and $205 billion, Meta lifted the floor of its guidance to between $130 billion and $145 billion, and Microsoft’s roughly $175 billion of calendar-year investment was restated under a lease-accounting change that left the underlying spending trajectory unchanged — a combined program exceeding $700 billion for calendar 2026, with each company signaling further acceleration into 2027.[17] Goldman Sachs, revising its estimates upward after the first-quarter reports, projected a combined $5.3 trillion of capital expenditure for the four largest hyperscalers between fiscal 2025 and fiscal 2030, with a baseline aggregate estimate of $7.6 trillion between 2026 and 2031 across compute, data centers, and power.[18] Investors have noticed the strain that such numbers place on even the strongest cash-flow statements: when Alphabet lifted its 2026 guidance in late July 2026, its shares fell roughly 7 percent the following day, dragging Amazon, Meta, and Microsoft down with it, in what CNBC described as increased scrutiny of infrastructure investments that were producing dwindling cash piles with uncertain returns.[19] The market’s message was not that the spending should stop; it was that the spending had become large enough to be a financing question rather than merely an operating one.
The next generation of projects makes the point unavoidable. A single flagship campus can now require multiple gigawatts of electricity, years of construction, specialized liquid cooling, transmission upgrades, networking equipment at unprecedented density, accelerator purchases measured in the millions of units, and long-term commitments extending far beyond any conventional technology investment cycle. AI infrastructure has simply become too large to remain financed solely as an internal capital-expenditure program of any company, however profitable.
1.2 The Capital Requirement Moves Outside the Hyperscaler
As the infrastructure expands, the financing obligation migrates outward through the industrial chain, and each step of that migration creates a new borrower, a new asset class, and a new set of counterparties. A hyperscaler may rent capacity from a specialized neocloud rather than build its own; the neocloud then finances the GPUs. A data-center developer finances the building and the shell. A utility finances the transmission upgrade. An independent power producer finances new generation. Semiconductor suppliers finance fabrication expansion. Private-credit funds finance the infrastructure entities that hold the physical assets. Customers sign long-term commitments that make the projects bankable in the first place. Sovereign wealth funds from Abu Dhabi, Qatar, and Singapore — Mubadala, QIA, and GIC all appeared in Crusoe’s September 2026 round — supply patient equity at the top of the structure.[3]
The clearest single illustration of this outward migration is the joint venture that Meta announced with funds managed by Blue Owl Capital in October 2025 to develop its Hyperion data-center campus in Louisiana. Blue Owl’s funds took an 80 percent interest in the venture and Meta retained 20 percent, with the parties committing to fund their pro rata shares of approximately $27 billion in total development cost; a portion of Blue Owl’s capital was funded through debt issued to PIMCO and other bond investors in a private securities offering for which Morgan Stanley served as sole bookrunner.[20] The transaction was described at the time as both the largest project-finance bond and the largest private-debt securities offering ever completed: roughly $27 billion of fully amortizing debt maturing in 2049, rated A+ by S&P on the strength of Meta’s lease and residual-value support, priced at approximately 225 basis points over Treasuries, with PIMCO anchoring around $18 billion of the paper and BlackRock purchasing more than $3 billion.[21] [22] In exchange for its stake, Blue Owl injected approximately $7 billion of immediate cash into the entity, and the structure allowed Meta to receive a $3 billion distribution at closing while converting what would have been on-balance-sheet capital expenditure into a long-term lease obligation. Days later, Meta separately raised $30 billion of unsecured bonds in the public market — one of the largest corporate bond issues in United States history — demonstrating that the special-purpose structure was a choice about risk allocation and balance-sheet architecture, not a substitute for market access.[21]
Morgan Stanley has estimated that total data-center-related capital expenditure will approach $2.9 trillion between 2025 and 2028, of which roughly $800 billion may need to come from private credit, alongside contributions from public debt, securitization, equity, and internal cash flows.[23] JPMorgan, examining the flow from the other direction, calculated that AI-related investment-grade bond issuance had reached $266 billion in 2026 by early September and projected cumulative financing needs for AI capital expenditure of $2.1 trillion over the following five years.[16] Paul Kedrosky, tracking the same dynamics across a longer historical frame, estimated in mid-2026 that hyperscaler and data-center investment — negligible as recently as 2022 — was tracking toward $1.5 trillion annually by 2030, a level that would surpass, in real terms, the residential construction boom of 2005–2006, the peak year of the 1990s fiber build, and the railroad peak of 1882 combined, and he observed that the share of the build-out financed externally through debt and equity was approaching 90 percent by the end of the decade. His conclusion states the theme of this entire section in a single sentence:
“dependent on capital markets in a way that no prior technology cycle ever was.” — Paul Kedrosky, on the AI infrastructure cycle [38]
The AI economy therefore begins behaving less like traditional software and more like a hybrid of telecommunications, energy infrastructure, project finance, aircraft leasing, semiconductor manufacturing, and commercial real estate — industries in which the balance sheet of the operator was never large enough to carry the assets, and in which the essential financial technology was always the contract that allowed someone else’s balance sheet to carry them instead.
1.3 From Software Economics to Infrastructure Economics
Software historically offered perhaps the most attractive economics in the history of commerce, because the marginal cost of distributing another copy of a program was extremely close to zero. A company could spend heavily once to write the code and then sell it a billion times, with each incremental sale flowing almost entirely to gross profit. That economic structure explains why the software giants of the 1990s and 2000s accumulated the fortress balance sheets described above, and why the capital markets learned to think of technology as an asset-light sector whose principal investments were intangible.
Frontier artificial intelligence complicates — and in important respects inverts — that model. Every additional training run consumes compute, electricity, and cooling. Every inference request consumes accelerator time, memory bandwidth, networking capacity, and power. Every autonomous agent may generate hundreds or thousands of model calls where a human user would have generated one query, multiplying the physical cost of serving intelligence even as the price per token falls. The industry’s own leadership now describes the output of this machinery in explicitly industrial terms; when Nvidia reported the results of its second quarter of fiscal 2027 in August 2026 — revenue of $96.2 billion, up 106 percent from a year earlier, with data-center revenue of $89.0 billion — its founder framed the moment in language that could have come from a nineteenth-century industrialist describing a power plant:
“Its tokens are productive and profitable. Now, compute is revenue.” — Jensen Huang, Founder and CEO, Nvidia [30]
The software may be virtual; its cost structure increasingly is not. And a technology whose cost structure is physical inherits the financing problems of physical industries: long construction lead times, heavy fixed costs, utilization risk, obsolescence risk, energy-price exposure, and — above all — a mismatch between the moment when capital must be committed and the moment when revenue arrives. Those are precisely the problems that project finance, leasing, securitization, and long-duration contracting were invented to solve, which is why all of those tools are now appearing, in rapid succession, around artificial intelligence.
1.4 Why This Is Not Simply “AI Debt”
It would be a serious analytical mistake to compress this transformation into the phrase “AI companies are taking on debt,” because debt is only one instrument among many, and in several of the most important transactions it is not even the dominant one. CoreWeave’s convertible notes illustrate one approach: borrowed capital with embedded equity optionality, sold to institutional investors who want AI exposure with a partial floor.[2] Crusoe’s Series F illustrates another: pure equity, priced by sophisticated growth investors and sovereign funds willing to absorb business risk directly in exchange for uncapped participation in the upside.[3] SB Energy demonstrates something considerably more complicated: public equity, strategic private placements, a prepaid forward, penny warrants held by the anchor tenant, twenty-year leases, power-development commitments, and supplier credit support with a nine-figure-per-gigawatt guarantee cap, all coexisting around the same underlying stream of expected AI demand.[5] [6]
The emerging system is therefore better understood as an AI financial architecture rather than as AI borrowing — an architecture in the proper sense of the word, in which different instruments occupy structurally different positions, bear different risks, mature on different schedules, and are held by different classes of investors, yet are all anchored, directly or indirectly, to the same foundation: the expectation that the world’s demand for machine intelligence will grow enough, and soon enough, to service all of the claims that have been written against it. The remainder of this paper examines that architecture instrument by instrument, layer by layer, and risk by risk.
Section 2: The Instruments of Convertible Compute
If Section 1 described why the financing of artificial intelligence is migrating outward from corporate treasuries into the capital markets, this section examines the specific instruments through which that migration is occurring. Each instrument answers a different question about risk: who bears it, for how long, in exchange for what compensation, and with what recourse if expectations disappoint. The genius — and the fragility — of the emerging architecture lies in how these instruments are being combined around individual projects, so that a single campus in Ohio or Louisiana may simultaneously support convertible notes at one company, warrants at another, guarantees from a third, and lease obligations from a fourth. Table 1 summarizes the principal instruments before the subsections examine each in depth.
Table 1. The Instrument Set of Convertible Compute, with Representative 2025–2026 Transactions
| Instrument | Economic Function | Representative Transaction | Approximate Scale |
| Convertible debt | Borrow now; share upside later; defer dilution | CoreWeave 2.875% notes due 2033 (Sept. 2026) | $3.7 billion |
| Growth equity | Absorb business risk directly; fund pre-revenue scale | Crusoe Series F at $30.9B valuation (Sept. 2026) | $3.9 billion |
| Strategic equity & prepayments | Supplier finances the market for its own products | Nvidia prepaid forward + private placement in SB Energy | $3.0 billion |
| Warrants | Compensate counterparties for anchoring demand | OpenAI warrants in SB Energy; AMD warrants to OpenAI | ~$5.5 billion; up to 160M shares |
| Capacity contracts & leases | Convert expected demand into contractual cash flow | OpenAI–Oracle compute contract; OpenAI 20-year Ohio leases | ~$300 billion; 8.0 GW-IT |
| Guarantees & credit support | Transfer tail risk to the strongest balance sheet | Nvidia residual-value guarantees at PORTS-Pike | $105 billion cap |
| SPVs & private credit | Move assets off balance sheet; tap institutional debt | Meta–Blue Owl Hyperion joint venture (Oct. 2025) | $27 billion debt + $2.5 billion equity |
2.1 Convertible Debt: CoreWeave
CoreWeave offers one of the clearest examples of why convertible financing has become attractive within the AI economy, and its financing calendar across 2025 and 2026 reads like a case study in the instrument’s uses. An infrastructure company of this kind needs enormous amounts of capital immediately — to purchase accelerators, secure data-center capacity, and sign power agreements — while much of the expected economic value of that capacity lies years into the future, embedded in contracts that have been signed but not yet performed. Straight debt at such a company’s stand-alone credit quality is expensive; CoreWeave’s existing senior notes carry coupons of 9.25 percent and 9.0 percent. Common equity issuance at scale, meanwhile, dilutes existing holders at whatever price the market offers on the day of the sale. The convertible note threads between these constraints: investors begin as creditors, entitled to interest and principal, while retaining exposure to appreciation in the company’s equity through the conversion feature; the issuer, in exchange, borrows at a fraction of its straight-debt cost — 2.875 percent in the September 2026 offering, and as low as 1.75 percent in the two offerings that preceded it — while deferring dilution until the share price clears the conversion premium.[2]
The September 2026 transaction displays the machinery in full. CoreWeave announced a $3.0 billion offering of convertible senior notes due April 2033 on September 17, priced it the following day at an upsized $3.7 billion with an initial conversion price of approximately $97.85 per share — a premium above the $79.88 closing price on the day of pricing — and simultaneously entered into capped call transactions designed to reduce potential dilution upon any future conversion.[1] [2] It was the company’s third major convertible financing in ten months, following an upsized $2.25 billion offering of 1.75 percent notes due 2031 in December 2025 and a $3.5 billion offering of 1.75 percent notes due 2032 in April 2026 that ultimately closed at $4.0 billion once the purchasers’ option was exercised in full. Each successive offering was announced at $3.0 billion or less and each was enlarged at pricing — a pattern that testifies to sustained institutional demand for AI-linked hybrid paper even as the same investors grew visibly more cautious toward AI-linked common equity.
The extraordinary growth of AI-linked convertible issuance suggests that this is not an isolated financing technique but a durable feature of the new architecture. The $131 billion of United States convertible issuance recorded by early September 2026 — surpassing the full-year record set only the year before, with roughly half of the total tied to artificial intelligence, and with zero-coupon structures accounting for two-fifths of the market — represents, in effect, the capital markets inventing a standardized wrapper for a specific and novel risk: exposure to companies whose asset bases are appreciating rapidly in expectation but whose cash flows cannot yet service conventional debt at conventional scale.[14] [15] [16] The convertible bond, an instrument dating to the railroad financings of the nineteenth century, has found in the GPU cloud its most natural application in a century, and the historical rhyme is not accidental: in both cases, investors were asked to fund physical networks ahead of proven demand, and in both cases the instrument that succeeded was the one that paired downside protection with a claim on the upside if the network filled.
2.2 Equity: Crusoe
Crusoe represents another pathway through the same capital requirement, and the contrast with CoreWeave is instructive precisely because the two companies serve overlapping markets. Crusoe’s September 17, 2026 financing raised $3.9 billion of equity at a $30.9 billion post-money valuation in an oversubscribed round whose investor list — Atreides, Mubadala Capital, Valor, Founders Fund, GIC, Nvidia, QIA, Radical Ventures, TPG — spans growth funds, sovereign wealth, and a strategic supplier.[3] Equity absorbs more direct business risk than any form of debt: there is no coupon, no maturity, no covenant, and no recourse, only a claim on whatever residual value the enterprise ultimately creates. In exchange, equity investors receive uncapped participation in the upside if the underlying infrastructure becomes significantly more valuable — and the disclosures accompanying the round explain why sophisticated investors were willing to accept that trade. Crusoe reported more than $140 billion in total contracted value across its vertically-integrated platform, more than 6 gigawatts of gross contracted capacity across its data centers and cloud with 1 gigawatt delivered and operational, year-over-year growth of more than twenty-fold in cloud bookings, and more than $100 million of contracted annual recurring revenue in a managed-inference product launched less than a year earlier; the company had also recently signed a five-year cloud contract worth approximately $13 billion with the quantitative trading firm Jane Street.[3] [44] Gavin Baker of Atreides Management, co-leading the round, articulated the investment thesis in a sentence that doubles as a theory of the entire sector:
“the economics flow to the lowest-cost producer of intelligence.” — Gavin Baker, Managing Partner and CIO, Atreides Management [41]
Equity of this kind is particularly relevant during an infrastructure expansion in which future utilization, model economics, electricity prices, hardware generations, and customer concentration all remain deeply uncertain, because equity can absorb varieties of uncertainty that debt markets are structurally unwilling to price. A lender must be able to model the downside; an equity investor need only believe in the distribution. The coexistence of Crusoe’s equity round and CoreWeave’s convertible offering in the same twenty-four hours is therefore not a curiosity but a signal that the AI capital stack is stratifying, with different investors self-selecting into different positions on the same underlying bet.
2.3 Strategic Equity and Prepayments: The Supplier as Financier
The relationship between Nvidia and SB Energy introduces a third model, one with the fewest precedents in conventional corporate finance: the supplier who finances the environment in which its own products will eventually operate. A traditional equipment vendor waits for a customer to construct infrastructure and then sells into it. Nvidia has inverted the sequence. Its $3 billion commitment to SB Energy divides between a $1.5 billion prepaid forward contract for non-voting shares entered into on August 17, 2026 and a $1.5 billion private placement at the IPO price closing concurrently with the offering — capital delivered to the developer before the campus exists, before the first lease payment is due, and before a single accelerator has been installed.[6] Announcing the arrangement, Nvidia’s founder placed it within an explicitly infrastructural vision of the industry’s future:
“land, power and shell have become vital in the age of AI.” — Jensen Huang, announcing the PORTS-Pike partnership [7]
The strategic logic extends well beyond financial return, and it is worth walking through the causal chain deliberately, because the chain is the structure. Capital helps SB Energy secure land in Pike County. Land, combined with SB Energy’s power-development expertise, helps secure the ten gigawatts of gross power the campus is designed to draw. Power enables the construction of seventeen data-center buildings. Those buildings, under the exclusivity provisions disclosed in the filing, will host Nvidia’s full-stack compute architecture and essentially no one else’s. That architecture will serve OpenAI’s workloads under twenty-year leases. Each link in the chain increases the probability that the next link forms, and the supplier’s investment at the first link therefore helps create the future market into which its primary products will be sold at every subsequent link. Nor is SB Energy an isolated case: Nvidia participated in Crusoe’s Series F, committed as much as $100 billion in support of OpenAI’s deployment of at least 10 gigawatts of its systems (ultimately contributing $30 billion to OpenAI’s early-2026 funding round), and has made a series of strategic investments across the AI ecosystem.[3] [27] [29] The pattern has drawn criticism as a form of vendor financing that risks circularity — a concern taken up in Section 5 — but its function within the architecture is unambiguous: the supplier’s balance sheet has become a load-bearing element of the industry’s capital formation.
2.4 Warrants: Binding the Customer to the Asset
Warrants add a fourth dimension, and the SB Energy filing provides the richest disclosed example to date. OpenAI holds warrants covering approximately 3.99 million SB Energy shares at an exercise price of one cent per share, issued in January 2026 under a warrant agreement that describes the grant as a material inducement for OpenAI to enter into the foundational agreement and the associated leases; portions vest in connection with the public offering, additional tranches remain subject to future vesting conditions tied to valuation milestones, and the package — issued at an estimated value of roughly $3.6 billion — was carried at approximately $5.5 billion by June 30, 2026, with the remeasurement of the warrant liability accounting for $2.57 billion of SB Energy’s $3.21 billion first-half net loss.[5] [10] OpenAI also maintains a board seat while its ownership remains above 5 percent, and invested $500 million alongside SoftBank, meaning the tenant of the Ohio campus is simultaneously an investor in, a warrant holder of, and a director-appointing shareholder of its own landlord.[6]
Nor is the structure unique to real assets. When AMD announced its multi-year partnership to supply OpenAI with 6 gigawatts of Instinct GPU capacity in October 2025 — an arrangement its chief executive suggested could generate well over $100 billion of revenue — it issued OpenAI warrants for up to 160 million shares of AMD common stock, approximately a tenth of the company, vesting as deployment, commercial, and share-price milestones are achieved.[29] The economic function in both cases is identical. Instead of maintaining a purely transactional relationship — customer buys capacity from supplier, supplier books revenue — the relationship becomes partially financial: the customer’s commitment creates the demand that makes the supplier’s expansion valuable, and the warrant hands the customer a share of the value its own commitment created. Warrants thus solve a coordination problem that no amount of ordinary contracting can solve, aligning the long-term interests of counterparties whose fortunes have become mutually constitutive. They also, of course, deepen the interdependence that regulators and rating agencies have begun to scrutinize, converting what would once have been an arm’s-length supply chain into a lattice of cross-held equity claims.
2.5 Capacity Contracts, Leases, and Guarantees: The Contracts That Make Everything Else Possible
Potentially the most consequential instruments in the entire architecture are not securities at all. They are contracts — twenty-year leases, minimum-capacity commitments, take-or-pay arrangements, power-purchase agreements, credit guarantees, and reservations of future compute — and their consequence lies in a simple fact of credit analysis: a project with a creditworthy long-term counterparty can borrow against the contract, while an identical project without one cannot borrow at all. Long-duration commitments transform uncertain future demand into contractual cash-flow expectations, and contractual cash-flow expectations are the raw material from which every other instrument in this section is manufactured.
The scale these contracts have reached is without precedent in the history of commercial agreements. OpenAI’s five-year compute contract with Oracle, reported at approximately $300 billion beginning in 2027 and associated with roughly 4.5 gigawatts of capacity under the Stargate program, single-handedly transformed Oracle’s reported backlog: remaining performance obligations surged 359 percent year-over-year to $455 billion in the quarter of the announcement and reached $523 billion by the third quarter of Oracle’s fiscal 2026, while Moody’s warned that the capital spending required to serve the contract would push Oracle’s leverage toward four times EBITDA and characterized the broader program in terms that capture the industrial scale of the undertaking:
“effectively one of the world’s largest project financings.” — Moody’s, on the Stargate program [26]
OpenAI’s aggregate commitments across its suppliers — Oracle, Microsoft, Amazon Web Services, CoreWeave, Nvidia, AMD, Broadcom, and Cerebras — were reported to exceed $1.4 trillion by early 2026, against planned compute spending that the Wall Street Journal reported at $750 billion through 2030, supported by a funding round that closed with $122 billion of committed capital at an $852 billion post-money valuation.[24] [25] [27] [28]
The Nvidia–SB Energy guarantee arrangement then illustrates how contracts and credit support interlock. Under the guarantees disclosed in Nvidia’s quarterly filing, the chipmaker provides credit support on the land, power, and shell build-out covering leases for approximately 4.25 gigawatts of IT load at PORTS-Pike, with its obligation capped at $105 billion in the aggregate; each guarantee generally becomes effective upon commencement of the applicable lease, guarantee amounts increase as each of nine data centers is placed in service beginning in fiscal 2029, exposure declines as OpenAI fulfills its lease payments, and the support is limited to defined portions of lease and power payments rather than the full cost of the site.[8] The guarantee does not eliminate risk; it relocates it, moving the tail risk of tenant default from the project’s lenders — who could not have absorbed it at investment-grade pricing — onto the balance sheet of the one participant strong enough to carry it. In such structures the contract itself becomes part of the financial architecture, as load-bearing as any bond indenture, and the analyst who reads only the securities filings while ignoring the lease schedules and guarantee terms will systematically misunderstand where the risk in the system actually resides.
2.6 Special-Purpose Vehicles, Private Credit, and Securitization: The Off-Balance-Sheet Frontier
A final family of instruments deserves separate treatment because it changes not who bears the risk but where the risk appears — or fails to appear — in the financial statements the market reads. The Meta–Blue Owl Hyperion structure described in Section 1.2 is the template: a special-purpose joint venture holds the physical asset; the hyperscaler retains a minority stake, operational control, and a long-term lease; institutional debt investors fund the majority of the cost through a private placement rated on the strength of the lease and the sponsor’s residual-value support; and the sponsor converts capital expenditure into operating expense while receiving cash at closing.[20] [21] [22] S&P’s A+ rating on $27 billion of debt yielding 6.58 percent at issue — a yield closer to high-yield territory than to the single-A corporate curve — captured the market’s ambivalence about the novelty: the rating agencies priced the lease, while the bond buyers priced the technology risk behind it.[21]
Private credit’s role in this frontier is expanding at a pace the official sector is watching closely. The BIS documented that private-credit funds originated more than $40 billion of loans to AI-related companies in 2025, compared with roughly $3 billion in 2010, with AI-related direct lending reaching approximately 4 percent of the market and total private credit outstanding to AI-related borrowers exceeding $200 billion.[11] Morgan Stanley’s estimate that roughly $800 billion of private-credit capital will be required between 2025 and 2028 — about a third of the $2.9 trillion of expected data-center capital expenditure — implies that an asset class built over two decades on middle-market corporate lending is being rapidly repurposed into the project-finance engine of the AI economy.[23] The securitization of GPU-backed loans, the emergence of asset-backed finance structures that KKR projects could reach $9 trillion globally by 2029, and the migration of data-center debt into insurance balance sheets and retirement portfolios all extend the same logic: the physical assets of artificial intelligence are being converted, tranche by tranche, into the standardized instruments through which institutional capital can hold them.[23] This is Convertible Compute in its most literal financial-engineering form — and it is also the channel through which any future disappointment in AI economics would propagate beyond the technology sector into the broader financial system, a possibility examined in Section 5.
Section 3: What Actually Supports the Financing?
Every financing rests on something. A mortgage rests on a house; a corporate bond rests on a stream of earnings; a project-finance loan rests on the cash flows of the completed project. The question this section asks is deceptively simple and analytically central to everything else in this paper: what, precisely, supports the hundreds of billions of dollars of claims now being written against the artificial-intelligence economy? The answer, examined carefully, is a layered stack of asset types that ranges from the concrete to the almost purely expectational — GPUs, power rights, leases, supplier support, and, at the foundation, beliefs about future token demand — and the character of the entire architecture is determined by the fact that the most important layer is the least tangible one.
3.1 GPUs as Productive Assets
The accelerator is no longer merely a semiconductor component; inside an AI cloud it becomes a revenue-producing capital asset, closer in economic function to an aircraft in a leasing fleet or a locomotive on a nineteenth-century railroad than to a chip inside a consumer device. Its economic value depends on utilization rates, rental pricing, technological obsolescence, electricity costs, maintenance, the networking fabric that binds it into clusters, and the operator’s ability to migrate workloads toward newer architectures as they arrive. Nvidia’s fiscal second-quarter 2027 results give a sense of the flow of new productive assets entering the system: $89.0 billion of data-center revenue in a single quarter, up 117 percent from a year earlier, with the company guiding to $108 billion of total revenue in the following quarter and its next-generation Vera Rubin platform already in production with purchase orders from every major customer.[30] [31]
This creates an unusual and genuinely difficult financing problem, one that recurs throughout this paper because it cannot be engineered away. A building may last forty years. A power plant may operate for sixty. A data-center lease may run for twenty. But a particular generation of accelerator may become economically less competitive within several years — not because it stops functioning, but because a successor generation delivers so much more computation per watt and per dollar that workloads, and therefore rental revenue, migrate away from it. The depreciation schedule assigned to GPUs is consequently among the most contested accounting judgments in the entire sector, since a change of two years in assumed useful life can swing reported profitability by billions of dollars. Convertible Compute therefore finances assets operating on radically different clocks, and Section 4 returns to this duration mismatch as the central structural tension of the Five-Layer AI Economy.
3.2 Power Rights as Financial Assets
For a gigawatt-scale AI campus, obtaining GPUs without obtaining electricity accomplishes nothing at all; a rack of accelerators without power is inventory, not capacity. Interconnection rights, generation agreements, transmission access, power-purchase contracts, backup generation, on-site storage, and sometimes entire energy-development programs therefore become economically intertwined with compute, and in many markets they have become the binding constraint on the entire build-out. The scarce asset is frequently not the accelerator itself; it is the right to energize it, and that right — a queue position with a grid operator, a signed PPA, a permitted substation — has quietly become one of the most valuable classes of property in the American economy.
The transactions of 2025 and 2026 make the point empirically. Crusoe’s entire competitive thesis, as articulated in its Series F announcement, is an energy-first strategy in which the company originates and manages power directly at the source before designing the data center around it, inverting the sequence followed by conventional developers who treat power as a constraint to be navigated after site selection.[3] SB Energy is, at its core, a renewable-power developer — essentially all of its revenue to date, roughly $138.7 million in the first half of 2026, derives from its power business, and it operates a standalone solar and battery portfolio of 5.5 gigawatts — that is converting its energy-development capability into data-center campuses precisely because the power expertise, not the building expertise, is the scarce input.[5] [6] The PORTS-Pike campus is designed to draw approximately 10 gigawatts of gross power from a site in southern Ohio selected for its energy characteristics. When power rights can anchor a $439 billion contracted backlog at a company whose data centers have yet to come online, power rights have become financial assets in every meaningful sense of the term.
3.3 Leases and Customer Commitments
A datacenter without a customer represents construction risk, speculative real estate of an unusually expensive and unusually specialized kind. A datacenter with a highly creditworthy twenty-year tenant represents something categorically different: a bond in the shape of a building. Long-duration commitments transform uncertain future demand into contractual cash-flow expectations, and this transformation explains why the identity and creditworthiness of the tenant increasingly matter to lenders and investors almost as much as the physical quality of the facility itself. S&P’s treatment of the Hyperion financing is the cleanest evidence: the A+ rating on $27 billion of project debt reflected Meta’s lease and residual-value support — the rating agency described that support as the linchpin of the rating — rather than any property of the Louisiana real estate.[21] [23]
The same logic, operating at even larger scale, explains both the promise and the fragility of the SB Energy structure. OpenAI’s twenty-year leases across approximately 8.0 gigawatts constitute the overwhelming majority of the company’s contracted backlog and are the foundation on which the IPO, the Nvidia investment, and the guarantee architecture all rest; the S-1 mentions OpenAI more than three hundred times and states plainly that the company is substantially dependent on it.[6] [9] The lease converts OpenAI’s expectations about its own future compute demand into an obligation that others can finance — but it also concentrates the campus’s contracted revenue in a single tenant whose delays, disputes, or credit deterioration would propagate through every layer of the structure at once. The contract that makes the financing possible is the same contract that makes the financing concentrated. This duality is not a flaw to be engineered out of the system; it is the system.
3.4 Supplier Support
The traditional technology supplier sells equipment and collects payment, and its involvement with the customer ends at the loading dock. The emerging AI supplier may do considerably more: it may invest in its customers, guarantee portions of their infrastructure commitments, participate directly in their financings, reserve future capacity, provide strategic equity, and issue or receive warrants. Nvidia alone, across 2025 and 2026, invested in Crusoe’s Series F, committed up to $100 billion in support of OpenAI’s build-out before contributing $30 billion to its funding round, invested $3 billion in SB Energy across the prepaid forward and private placement, extended $105 billion of guarantee capacity over the Ohio leases, and made a string of further strategic investments across the ecosystem.[3] [6] [8] [27] [29] Nvidia’s widening involvement across AI infrastructure is therefore financially significant in addition to being technologically significant: the strongest balance sheet in the industry has voluntarily made itself a source of credit enhancement for the industry’s expansion, accelerating the build-out while binding the supplier’s fortunes to the solvency of its customers in ways that a pure equipment vendor never experiences. Whether this represents rational market-making by the party with the best information about future demand, or a channel through which one company’s eventual disappointment becomes everyone’s, is among the most consequential open questions in the entire architecture.
3.5 Future Token Demand: The Assumption Underneath Everything
The most abstract component of the asset base is future AI usage itself. No lender possesses a warehouse containing “future tokens” as collateral; no court can foreclose on inference that has not yet been demanded. Yet almost every major infrastructure investment described in this paper ultimately depends on the assumption that companies, consumers, governments, robots, vehicles, and autonomous agents will demand substantially more machine intelligence in the future than they demand today — enough more to fill 8 gigawatts in Ohio, 4.5 gigawatts under the Oracle contract, 6 contracted gigawatts at Crusoe, and the hundreds of additional gigawatts implied by more than $700 billion of annual hyperscaler capital expenditure.[17] Future token demand is therefore not legal collateral. It is something arguably more important: the economic assumption underneath much of the financing, the invisible foundation on which every visible instrument rests.
If token consumption, agent adoption, and inference demand grow dramatically, the large infrastructure commitments of 2025 and 2026 will look, in retrospect, like the disciplined early moves of a historic expansion. If those expectations prove excessive, the financing structures built upon them will face pressure in a particular sequence — utilization first, then rental pricing, then covenant compliance, then refinancing capacity, then guarantees — that Section 5 traces in detail. What deserves emphasis here is the epistemic asymmetry at the heart of the system: the instruments are precise, contractual, and legally enforceable, while the assumption supporting them is a forecast. The Bank for International Settlements has made exactly this point the center of its analysis, concluding that while the macroeconomic and financial-stability risks of the boom currently appear moderate, sustainability ultimately depends on whether firms generate earnings sufficient to justify today’s expectations — a condition that no contract, however well drafted, can guarantee.[11] [42]
Section 4: Convertible Compute Across the Five-Layer AI Economy
The Five-Layer AI Economy — energy, chips, datacenters, models, applications and agents — was originally a framework for describing technological dependence: each layer requires the ones beneath it and enables the ones above. What the developments of 2025 and 2026 reveal is that the layers are now bound together financially as well as technologically, and that each layer imports a characteristic financing horizon, risk profile, and instrument set from the physical properties of its assets. This section walks the stack from bottom to top, and Table 2 summarizes the duration structure that emerges — the multiple financial clocks on which Convertible Compute simultaneously runs.
Table 2. The Financial Clocks of the Five-Layer AI Economy
| Layer | Representative Assets | Economic Life | Characteristic Financing |
| 1 — Energy | Generation, transmission, storage, interconnection rights | 30–60+ years | Project finance, PPAs, utility debt, infrastructure funds |
| 2 — Chips | Accelerators, memory, networking silicon, fabs | 3–7 years (accelerators); 10–20 (fabs) | Supplier cash flow, prepayments, vendor support, equipment finance |
| 3 — Datacenters | Land, shells, substations, cooling, fiber | 20–40 years (shell); leases 15–20 years | SPVs, private credit, lease-backed bonds, REIT structures |
| 4 — Models | Training runs, model weights, research talent | Months to ~3 years per generation | Venture and growth equity, strategic investment, compute contracts |
| 5 — Applications & Agents | Products, workflows, agent deployments | Months to years | Equity, revenue-based expectations; ultimately the repayment engine |
4.1 Layer 1 — Energy
Energy projects require perhaps the longest financing horizons anywhere in the AI system. Nuclear plants, gas generation, renewable portfolios, transmission corridors, substations, transformers, and grid-scale storage can require commitments measured in decades, permitting processes measured in years, and capital that must be patient in a way that no other layer demands. What has changed is that AI customers have begun to make some of these investments economically viable through long-duration power demand that traditional load growth never supplied: a campus designed to draw 10 gigawatts of gross power, as PORTS-Pike is, represents a demand commitment that can anchor generation and transmission development across an entire region.[6] [7] PwC’s 2026 outlook estimated that the global data-center build-out will absorb $31.6 trillion of cumulative capital expenditure through 2050, with annual spending rising from roughly $800 billion in 2026 to $1.1 trillion in 2030 and $1.8 trillion by mid-century — figures in which power and energy infrastructure constitute a large and growing share.[40] The financing consequence is that electricity, the oldest and most conservatively financed of industries, is being pulled into the AI capital stack, and the AI capital stack is inheriting, through that connection, the multi-decade obligations of the power sector.
4.2 Layer 2 — Chips
Semiconductors operate on a much shorter technology cycle, and the tension between the layer’s enormous near-term revenue and its obsolescence risk defines its financing character. Accelerators generate revenue at a pace without precedent in industrial history — Nvidia’s $89.0 billion of data-center revenue in a single quarter, growing 117 percent year-over-year, with guidance implying roughly 70 percent growth into fiscal 2028 — while simultaneously carrying the risk that each generation will be economically superseded within a handful of years by its own successor.[30] [31] Financing this layer therefore requires assumptions about utilization, resale value, upgrade cycles, and the continued competitiveness of particular computing architectures, and it explains why the chip layer, almost uniquely, is financed principally from operating cash flow and forward commitments rather than from long-duration debt: no lender wants to hold a fifteen-year claim against a three-year asset. It also explains the emergence of the prepayment and the supplier guarantee as characteristic Layer 2 instruments — devices by which the chip layer, rich in current cash flow but exposed to demand risk, exchanges financial support today for locked-in deployment tomorrow.[8] Even the layer’s input costs have become entangled with its own success: Nvidia warned in August 2026 that memory scarcity, driven in large part by the AI build-out itself, would compress its gross margins into the low seventies by the fourth quarter of fiscal 2027.[31]
4.3 Layer 3 — Datacenters
Datacenters occupy the structural middle of the stack, connecting the long duration of real estate and power infrastructure with the short duration of semiconductor technology, and this position makes Layer 3 the site of the most fundamental financial tension in the AI economy. A twenty-year building may contain computing equipment that is refreshed three or four times during a single lease term. Financiers must therefore learn to distinguish, with precision, the value of the site, the power rights, the shell, the network, and the customer contract from the value of whatever generation of accelerator happens to occupy the racks at a particular moment — because these components have different owners, different depreciation schedules, different residual values, and different claims written against them. The Hyperion structure is best understood as exactly this decomposition executed in legal form: Blue Owl’s funds and the bond investors hold the long-duration components — land, shell, lease — while Meta retains the technology risk through its residual-value support and its ownership of the equipment inside.[20] [21] [23] The Nvidia guarantee at PORTS-Pike performs the same decomposition with the roles rearranged: the guarantee covers the land, power, and shell build-out, precisely the long-lived components, while the technology risk remains with the parties best positioned to manage it.[7] [8] Layer 3 is where Convertible Compute’s financial engineering is most visible, because Layer 3 is where the clocks collide.
4.4 Layer 4 — Models
Model companies convert compute capacity into intelligence, and their economics ultimately determine how much of the underlying infrastructure the market can support. Their revenues, margins, subscription economics, API demand, and future agent businesses are the variables to which every lower layer’s financing is, directly or indirectly, exposed. The concentration of this exposure is extraordinary and is worth stating numerically: OpenAI alone stands behind more than $1.4 trillion of reported infrastructure commitments, the $300 billion Oracle contract that transformed that company’s backlog and leverage trajectory, twenty-year leases across 8 gigawatts in Ohio, the demand cases of CoreWeave and Crusoe, and the deployment schedules of Nvidia, AMD, and Broadcom — while itself remaining a private company financed by successive equity rounds, most recently $122 billion of committed capital at an $852 billion post-money valuation.[24] [25] [27] [28] The financing of Layer 3 therefore increasingly depends on expectations about Layer 4, and the credit quality of Layer 4 — unrated, unlisted, and evolving quarterly — has become, through the lease and the capacity contract, an embedded input to the pricing of investment-grade bonds. This is a genuinely novel feature of the financial system, and the rating agencies’ treatment of tenant-dependent AI project debt will be one of the quiet battlegrounds of 2027–2030.
4.5 Layer 5 — Applications and Agents
Layer 5 ultimately becomes the repayment engine for the entire structure. Millions or billions of useful AI tasks — code written, claims processed, documents reviewed, deliveries routed, experiments designed, customers served — must eventually generate economic value sufficient to justify the extraordinary capital invested beneath them. An autonomous coding agent, a customer-service system, a scientific discovery platform, a warehouse robot, an advertising engine, an enterprise workflow, or an autonomous vehicle is therefore not disconnected from a power plant in Ohio or a GPU-backed bond in Louisiana; it is the eventual economic reason those assets were financed, the final purchaser of the tokens whose anticipated demand supports every claim in the stack.
The academic literature counsels both patience and humility about how quickly this repayment engine will reach full power. Erik Brynjolfsson and his co-authors, in the productivity J-curve research program spanning 2020 through 2026, demonstrated that general-purpose technologies require large complementary investments in processes, business models, and human capital before their benefits appear in measured output — investments that depress measured productivity before raising it:
“We call this phenomenon the Productivity J-curve.” — Erik Brynjolfsson, Daniel Rock, and Chad Syverson [33]
Their 2026 work with Census Bureau microdata found causal evidence of exactly this pattern in American manufacturing: short-term performance losses preceding longer-term gains as firms absorb industrial AI.[34] Daron Acemoglu of MIT, working from a task-based framework, reached a more conservative assessment still, characterizing AI’s ten-year macroeconomic effects as
“nontrivial but modest” — Daron Acemoglu, MIT [32]
with total-factor-productivity gains bounded around two-thirds of a percentage point over a decade and GDP effects on the order of one percent — figures dramatically below the growth assumptions embedded in the more aggressive infrastructure financings.[32] The gap between Acemoglu’s arithmetic and the capital markets’ behavior is the single most important number in this paper, even though it cannot be stated precisely: it is the spread between what the economics profession can currently document and what the financing architecture currently assumes. Convertible Compute makes the dependency chain explicit and visible — applications generate demand, demand supports models, models require compute, compute requires datacenters, datacenters require power, and contracts across those layers support financing — and it thereby also makes explicit where the chain would break first if Layer 5 matures more slowly than the leases require.
Section 5: The 2027–2030 Financial Architecture of Artificial Intelligence
Forecasting is a hazardous business in a sector that has repeatedly outrun its own most aggressive projections, and this section therefore proceeds not by predicting particular outcomes but by identifying the structural features of the financial architecture that are already locked in — features that will shape the 2027–2030 period regardless of whether the underlying technology delights or disappoints. Five such features stand out: the institutionalization of AI infrastructure as an asset class; the dissolution of the boundary between supplier and financier; the concentration of counterparty exposure; the centrality of duration mismatch; and the migration of the relevant risk category from technology risk toward financial-system risk.
5.1 AI Infrastructure Becomes an Asset Class
Between 2027 and 2030, AI infrastructure is likely to complete its transformation into a recognizable institutional-investment category rather than an extension of ordinary technology investing, because the quantities involved leave no alternative. If the four largest hyperscalers alone deploy the $5.3 trillion that Goldman Sachs projects through fiscal 2030, if total data-center capital expenditure approaches Morgan Stanley’s $2.9 trillion estimate for 2025–2028, if annual global build-out spending rises toward PwC’s projected $1.1 trillion by 2030, and if — as Kedrosky calculates — external debt and equity finance approaches ninety percent of the total, then AI infrastructure will absorb a share of global institutional capital comparable to entire established asset classes such as commercial real estate or public infrastructure.[18] [23] [38] [40] Capital at that scale does not flow through bespoke, hand-crafted transactions; it flows through standardized structures with recognized risk categories, benchmark indices, rating methodologies, and dedicated fund vehicles, and the standardization is already visibly underway in the convertible market’s AI cohort, the private-credit data-center funds, the lease-backed bond structures, and the sovereign-wealth allocations documented throughout this paper.
Different investors will seek different exposures within the category, and the stratification observable in 2026 offers a preview of the mature ecosystem. Some will finance power, accepting utility-like returns for utility-like duration. Some will finance data-center shells through SPVs and lease-backed paper, holding single-A risk with a technology tenant behind it. Some will buy convertible securities, trading yield for optionality. Some will provide private credit against contracted cash flows. Some will finance GPUs and networking equipment on short amortization schedules matched to technology cycles. Some will hold equity in model or application companies, absorbing the residual risk of the entire stack in exchange for its residual upside. The Five-Layer AI Economy will consequently develop a corresponding five-layer financial ecosystem, with capital sorted by duration, risk tolerance, and required return into the layer whose physical characteristics match its liabilities — insurance capital into power and shells, credit funds into contracted infrastructure, growth equity into models and applications.
5.2 The Supplier Becomes Financier
The distinction between supplier, customer, investor, lender, landlord, and guarantor — a distinction on which most of financial analysis quietly relies — is already becoming unreliable in the AI economy, and by 2030 it may be largely obsolete for the sector’s central firms. A chip company invests in an infrastructure provider and guarantees its tenant’s leases. A model developer is simultaneously tenant, warrant holder, and board-seat-holding shareholder of its landlord. A hyperscaler finances a power developer. An infrastructure company owns generation and datacenters at once. A cloud provider’s largest customer is also its supplier’s largest investment. The SB Energy prospectus, read carefully, is less a description of a company than a map of one neighborhood of this network: SoftBank as controlling shareholder and customer, OpenAI as tenant, investor, and warrant holder, Nvidia as strategic investor, exclusive compute supplier, and guarantor.[5] [6] [10]
This does not necessarily indicate improper circularity, and the point deserves emphasis because the word “circular” has become an epithet in commentary on the sector. In many capital-intensive industries, strategic counterparties routinely finance one another: aircraft manufacturers support their airline customers’ fleet financings, energy majors take equity in the pipelines that carry their product, and equipment vendors have extended trade credit since the invention of trade. The economically meaningful questions are narrower and more technical: whether the intercompany flows create revenue recognition that overstates independent demand; whether equity stakes and guarantees transform what appears to be diversified counterparty exposure into correlated exposure; and whether the network as a whole could absorb the failure of any single node. The BIS working-paper literature of 2026 flagged precisely these mechanisms — debt-financed racing and circular equity ties as amplifiers of a potential bust — and the correct analytical response is the one this paper has attempted throughout: analysts must examine the entire network rather than any one company’s balance sheet in isolation, because in a network of cross-held claims, the balance sheet of the system is not the sum of the balance sheets of its members.[13]
5.3 Counterparty Concentration Becomes Important
The same handful of names appears on both sides of a remarkable share of the transactions described in this paper. OpenAI anchors the Oracle contract, the SB Energy leases, material portions of CoreWeave’s and Crusoe’s demand, and the deployment schedules of Nvidia, AMD, and Broadcom. Nvidia appears as supplier to essentially everyone, investor in Crusoe, SB Energy, and OpenAI, and guarantor in Ohio. Oracle’s remaining performance obligations of more than half a trillion dollars are dominated by a single customer.[24] [25] [43] A single project may depend simultaneously on one model company as tenant, one accelerator supplier, one power developer, several lenders, and a few large infrastructure investors — and those same entities recur, in permuted roles, across the neighboring projects.
This creates a question that the AI economy has not yet had to answer and that its financiers, by 2027–2030, will be forced to price explicitly: how much infrastructure can ultimately depend on the creditworthiness and future demand of a relatively small group of frontier-model companies and hyperscalers? The precedents from other industries are only partially reassuring. Aircraft finance survived the concentration of airframe supply in two manufacturers because the customer base was diversified across hundreds of airlines; AI infrastructure finance exhibits the opposite geometry, with diversified suppliers of capital converging on a customer base that can be counted on two hands. Gita Gopinath, the former First Deputy Managing Director of the IMF, has quantified the macro-financial stakes of this concentration, warning that an AI-driven market correction could erase as much as $35 trillion of global asset value, with AI-exposed companies having grown to represent 44 percent of S&P 500 capitalization by late 2025 against 22 percent in 2022, and offering a caution about what the boom’s very success may be concealing:
“The AI boom could be masking a slowdown in the traditional US economy.” — Gita Gopinath, former First Deputy Managing Director, IMF [37]
5.4 Duration Mismatch Becomes a Central Risk
Perhaps the deepest problem in Convertible Compute is time, and everything in Table 2 converges on it. Debt matures in several years. GPUs become technologically outdated in several years. Model leadership can change within months. Datacenter leases run twenty years. Transmission infrastructure operates for decades. Nuclear plants operate for generations. The economic life of the financing therefore does not naturally match the technological life of the asset, and the financial system must bridge the mismatch through some combination of refinancing, residual-value support, contractual cash-flow matching, and — where all else fails — risk-bearing equity.
The bridging mechanisms visible in 2026 each carry their own vulnerabilities, and it is worth naming them precisely because they define the stress scenarios of 2027–2030. Refinancing works until credit conditions tighten, at which point an industry that Kedrosky describes as approaching ninety percent external financing discovers that its expansion has a single point of failure in the capital markets themselves.[38] Residual-value guarantees work as long as the guarantors — today, principally Nvidia and the hyperscalers — retain the earnings and market confidence that make their support valuable; a guarantee is only as strong as the guarantor on the day it is called. Twenty-year leases match the duration of the building but embed the credit of tenants whose own business models are being reinvented on eighteen-month cycles. And equity absorbs whatever remains, which is why the valuation of that equity — Crusoe at $30.9 billion, OpenAI at $852 billion, the AI complex at 44 percent of the S&P 500 — has become, in effect, the shock absorber of the entire architecture.[3] [28] [37] The system works smoothly as long as each clock can be reset before it strikes; the risk scenario is the one in which several clocks strike at once.
5.5 From Technology Risk to Financial-System Risk
The official sector’s assessment of the boom has evolved with instructive speed across 2026, and its trajectory tells us where the analytical frontier now lies. In January, the IMF raised its global growth forecast to 3.3 percent for 2026 largely on the strength of AI-driven investment — the technology boom acting, in its chief economist’s phrase, as a tailwind offsetting trade disruption — while flagging the mirror-image risk in language every reader of this paper should hold in mind:
“There is a risk of a correction — a market correction” — Pierre-Olivier Gourinchas, Chief Economist, IMF, on unrealized AI productivity expectations [35]
By July, after an energy shock emanating from the Middle East, the Fund’s updated outlook described a world economy balanced between two opposing forces — the war’s drag and the AI investment boom’s lift — with the net effect varying across countries according to their position in the technology value chain.[36] The BIS moved along a parallel path: its January bulletin judged macro-financial risks “moderate” while establishing the empirical groundwork on debt migration and private credit; its June annual report elevated the boom to one of four global pressure points and invoked the recessionary endings of historical investment manias; its July working paper supplied the formal mechanism, a racing model in which over-investment of 1.5 to 3 times the efficient level emerges from rational competition for dominant positions.[11] [12] [13] [42]
The distinction underlying all of these assessments matters enormously and is the correct note on which to close this section. A failed AI project is corporate risk: painful for its investors, instructive for its competitors, and contained. A large number of heavily financed projects whose obligations depend on overlapping customers, suppliers, lenders, and assumptions about future AI demand is a different object entirely — a financial-system question, in which the transmission channels run through private-credit funds, insurance balance sheets, pension allocations, bank exposures to non-bank intermediaries, and the household wealth embedded in equity indices now dominated by AI-exposed firms. By 2027–2030, central banks, securities regulators, rating agencies, private-credit managers, pension funds, infrastructure investors, utilities, state regulators, and corporate boards will consequently need to understand not merely artificial intelligence, but the contractual architecture financing it — the leases, guarantees, warrants, prepayments, and capacity commitments that this paper has attempted to catalogue, because those contracts, far more than the models themselves, will determine how any disappointment propagates.
Section 6: What Have We Learned? Seven Pillars
The argument of this paper can be distilled into seven pillars — five carried forward and expanded from the framework’s original formulation, and two added in light of the evidence assembled in Sections 2 through 5. They are offered not as predictions but as structural observations: statements about how the AI economy is now organized that should remain true across a wide range of technological outcomes.
Pillar 1 — AI Is Moving From a Capex Boom to a Financing System
Artificial intelligence began as a technological investment cycle dominated by companies capable of paying for infrastructure out of enormous operating cash flows, and it is becoming a financing system in which equity markets, debt markets, private credit, infrastructure funds, strategic suppliers, sovereign wealth, and project-level contractual structures each carry a designated share of the load. The evidence is quantitative and cumulative: information-technology investment at five percent of United States GDP; more than a trillion dollars of hyperscaler capital expenditure across two years; a convertible market whose annual record fell before September; private credit to AI-related borrowers growing from $3 billion to more than $200 billion in fifteen years; and external financing approaching ninety percent of the build-out by decade’s end.[11] [16] [17] [38] The AI economy is therefore becoming financially institutionalized, and institutionalization is a one-way door: once an industry’s expansion depends on the capital markets, the state of the capital markets becomes a permanent input to the industry’s trajectory.
Pillar 2 — Contracts Can Matter as Much as Capital
A billion-dollar investment is visible; it appears in funding announcements, cap tables, and league tables. A twenty-year lease can be equally consequential while remaining nearly invisible to conventional analysis, because capacity commitments, guarantees, customer contracts, power agreements, and prepayments determine whether billions of dollars of infrastructure become financeable in the first place. The $105 billion guarantee cap in Ohio, the $300 billion take-or-pay commitment behind Oracle’s backlog, the $140 billion of contracted value supporting Crusoe’s valuation, and the leases that earned Hyperion’s bonds their single-A rating are all contracts, not capital — yet each one conjured more financing capacity than any plausible equity injection could have.[8] [3] [21] [24] The financial architecture of AI therefore cannot be understood by counting debt and equity; it must be read through its contracts, and the most important documents in the sector are increasingly lease schedules, guarantee agreements, and vesting conditions rather than balance sheets.
Pillar 3 — Compute Operates on Multiple Financial Clocks
Energy infrastructure can last for generations; datacenters for decades; customer contracts for twenty years; GPUs for a handful of years; model leadership for months; and application economics for even less. The greatest financial-engineering challenge of the era is matching long-duration capital obligations to rapidly changing technological assets, and every major structure examined in this paper — the Hyperion decomposition, the PORTS-Pike guarantee schedule, the convertible note’s maturity-plus-optionality design — is best understood as an attempt to bridge between clocks. The bridges hold in fair weather. Pillar 3’s warning is that the defining stress event of Convertible Compute, whenever it arrives, will take the form of clocks striking together: a hardware transition, a refinancing window, and a tenant renegotiation arriving in the same year.
Pillar 4 — The Five Layers Are Becoming Financially Interdependent
The Five-Layer AI Economy is not merely technologically connected; it is increasingly contractually and financially connected, and financial claims now transmit expectations vertically through the entire system with a speed and force that technological dependence alone never possessed. A Layer 5 expectation about autonomous agents shapes a Layer 4 model company’s capacity commitment, which supports a Layer 3 datacenter lease, which justifies Layer 2 GPU purchases and Layer 1 power infrastructure — and the transmission runs downward through signatures, not merely through sentiment. The corollary is that disappointments will also transmit vertically, and faster than in previous cycles, because the layers are bound by enforceable claims rather than by mere commercial relationships.
Pillar 5 — Finance Is Not a Sixth Layer; It Is the Connective Tissue
The purpose of Convertible Compute as a framework is emphatically not to add a sixth layer to the Five-Layer AI Economy, because finance does not behave like a layer. It has no fixed position in the stack; it passes through every layer simultaneously. Capital finances generators, semiconductor plants, datacenters, model companies, and applications alike, and the contracts between the layers then help finance still more infrastructure. Finance is therefore best understood as the connective tissue surrounding the Five-Layer AI Economy — the medium through which future expectations about machine intelligence become physical infrastructure in the present, and through which the present’s physical infrastructure becomes a claim on the future’s machine intelligence.
Pillar 6 — The Guarantor Is the New Systemically Important Institution
The transactions of 2025 and 2026 reveal a new category of load-bearing participant: the entity whose credit support, rather than whose capital, makes the system function. Meta’s residual-value support was the linchpin of $27 billion of single-A bonds; Nvidia’s guarantees are the mechanism by which an unbuilt Ohio campus reaches toward investment-grade financing; and the hyperscalers’ leases stand behind an expanding universe of SPV debt.[8] [9] [21] [23] These guarantors are performing, in miniature and by private contract, the function that governments and monoline insurers performed in earlier infrastructure eras — and the concentration of that function in a handful of technology balance sheets means that the market’s assessment of those few companies now propagates into the cost of capital for the entire stack. Watching the guarantors — their earnings, their aggregate contingent exposure, their disclosure practices — will tell analysts more about the system’s health between 2027 and 2030 than watching any individual borrower.
Pillar 7 — The Payoff Question Is an Empirical Race Between the J-Curve and the Debt Schedule
The final pillar joins the academic literature to the financing calendar. Brynjolfsson’s productivity J-curve implies that AI’s measured economic payoff will arrive with a lag, after complementary investments in process redesign, organizational change, and human capital are absorbed — and his 2026 microdata work confirms the pattern of short-term losses preceding longer-term gains.[33] [34] Acemoglu’s task-based arithmetic implies that even the eventual payoff may be far more modest than the market assumes.[32] The financing architecture, meanwhile, has its own unforgiving calendar: coupons that begin in 2027, leases that commence as buildings reach service from fiscal 2029, guarantees that phase in on the same schedule, and maturities clustering at the turn of the decade.[2] [8] The economic question of the next five years is therefore a race — between the upward slope of the J-curve and the amortization schedule of the claims written against it. If productivity and revenue arrive on the financiers’ timetable, Convertible Compute will be remembered as the machinery that built the intelligence economy; if the J-curve proves longer or shallower than the leases assume, the same machinery will transmit the disappointment with contractual efficiency. Both branches are still open, and honest analysis requires holding both.
Conclusion: Why “Convertible Compute” Fits the Emerging AI Economy
On September 17, 2026, CoreWeave and Crusoe provided two snapshots of the same historical transition. CoreWeave turned expectations about future AI infrastructure into billions of dollars of convertible financing, priced the following day at $3.7 billion by institutional investors willing to lend cheaply today for a claim on tomorrow’s equity.[1] [2] Crusoe turned similar expectations into $3.9 billion of new equity at a $30.9 billion valuation, sold to growth funds and sovereign investors willing to absorb the business risk directly.[3] Around SB Energy, the transformation went further still: public equity, prepaid forwards, penny warrants, twenty-year leases, ten gigawatts of power development, customer commitments, and $105 billion of conditional credit support became interconnected around the construction of AI capacity that does not yet exist, on behalf of demand that has not yet materialized, financed by institutions convinced — on substantial but necessarily incomplete evidence — that it will.[5] [6] [7]
These transactions suggest that artificial intelligence has entered a distinctly different stage of its development, and the succession of governing questions marks the path. The first question, in the years surrounding the arrival of the transformer architecture, was whether machines could become intelligent enough to generate text, software, images, scientific insight, and increasingly autonomous action; that question has been answered affirmatively enough to move markets, governments, and household expectations. The second question became whether the world could manufacture enough GPUs, and an industry that shipped $89 billion of data-center silicon in a single quarter is answering it.[30] The third became whether enough electricity and datacenter capacity could be constructed, and the gigawatt campuses now rising in Ohio, Texas, and Louisiana are the provisional answer. The next question — the question that will organize the period from 2027 through 2030 — is increasingly and unavoidably financial: how will all of it be paid for, by whom, through what instruments, and at what allocation of risk?
The answer, as this paper has argued throughout, will not come from any single instrument, because no single instrument can span the full range of durations, risks, and investors that the build-out requires. Debt will finance part of it, as the $266 billion of AI-linked investment-grade issuance recorded by September 2026 already demonstrates.[16] Equity will finance part of it, as Crusoe’s round and OpenAI’s $122 billion of committed capital demonstrate.[3] [28] Convertible securities will finance part of it, as a record-shattering $131 billion market demonstrates.[16] Private credit will finance part of it, on its way from $200 billion of AI exposure toward the $800 billion that Morgan Stanley projects the sector will require.[11] [23] Strategic investment will finance part of it, as Nvidia’s $3 billion in SB Energy and its investments across the ecosystem demonstrate.[6] [27] And warrants, prepayments, leases, capacity commitments, guarantees, and long-duration customer contracts will make the remaining portions possible, by converting expectations into claims that the other instruments can then fund.
This is why Convertible Compute is the appropriate title, and why its meaning operates at two depths. “Convertible” first describes the financial instruments themselves; CoreWeave’s convertible notes are the most literal example, and the record convertible issuance of 2025 and 2026 makes the label empirically apt. But the deeper meaning matters more. Artificial-intelligence infrastructure is becoming convertible between two worlds. In one direction runs the physical conversion: Capital → Power → Chips → Datacenters → Models → Applications and Agents. In the other direction runs the financial conversion: Expected AI Demand → Customer Contracts → Capacity Commitments → Financial Claims → Capital. Future intelligence helps finance today’s infrastructure; today’s infrastructure creates tomorrow’s intelligence. That feedback loop — virtuous if the demand arrives, vicious if it does not — is the financial architecture now assembling itself underneath the Five-Layer AI Economy, and it is the reason the central banks, the rating agencies, and the economics profession have all converged on the same conditional verdict: the structure is sound if and only if the earnings come.
And this, finally, is why Convertible Compute is not merely another name for AI debt. It describes the moment when compute itself becomes financeable, contractual, securitizable, investable, and inseparably intertwined with the capital markets that must fund the artificial-intelligence economy of 2027, 2030, and beyond — the moment when the question “can the machines think?” gives way, in the world’s largest financing programs, to the older and harder question that every infrastructure revolution has eventually faced: will the traffic come to the road that has been built for it? The railroads, the electric grids, the telephone networks, and the fiber backbones all answered that question eventually, and in every case the traffic did come — though rarely on the timetable the financiers had written, and rarely to the benefit of the financiers who wrote it. The intelligence economy will answer it too. The purpose of this paper has been to describe, precisely and without prejudice, the remarkable machinery through which that answer is now being financed.
Endnotes:
[1] Cris Tolomia, “CoreWeave raises $3 billion in convertible notes offering,” Quartz, September 17, 2026. https://qz.com/coreweave-convertible-notes-offering-3-billion-091726
[2] Benzinga Newsdesk, “CoreWeave Expands Convertible Notes Offering To $3.7 Billion,” Benzinga, September 18, 2026. https://www.benzinga.com/markets/large-cap/26/09/61874056/coreweave-expands-convertible-notes-offering-to-3-7-billion
[3] Crusoe, “Crusoe Raises $3.9B Series F at a $30.9B Valuation for AI Infrastructure,” Crusoe Newsroom, September 17, 2026. https://www.crusoe.ai/resources/newsroom/crusoe-announces-series-f-funding
[4] TechCrunch Staff, “Crusoe raises $3.9B to build massive data centers and small modular ‘AI factories,’” TechCrunch, September 17, 2026. https://techcrunch.com/2026/09/17/crusoe-raises-3-9b-to-build-massive-data-centers-and-small-modular-ai-factories/
[5] SB Energy, Inc., Form S-1 Registration Statement, U.S. Securities and Exchange Commission, September 2026. https://www.sec.gov/Archives/edgar/data/0002133037/000162828026059639/sbenergy-sx1.htm
[6] Yahoo Finance Technology Desk, “SoftBank’s SB Energy IPO filing offers 8.8 GW of contracted AI capacity, with most still unbuilt,” Yahoo Finance, September 2026. https://finance.yahoo.com/technology/ai/articles/softbank-8217-sb-energy-ipo-145941881.html
[7] NVIDIA Corporation, “NVIDIA Guarantees SB Energy’s PORTS-Pike Technology Campus in Ohio to Exclusively Host NVIDIA AI Compute,” NVIDIA Newsroom, August 17, 2026. https://nvidianews.nvidia.com/news/nvidia-guarantees-sb-energy-s-ports-pike-technology-campus-in-ohio-to-exclusively-host-nvidia-ai-compute
[8] NVIDIA Corporation, Form 10-Q for the quarter ended July 26, 2026, U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000075/nvda-20260726.htm
[9] CNBC Staff, “SoftBank’s SB Energy files for IPO, says it’s ‘substantially dependent’ on OpenAI,” CNBC, September 1, 2026. https://www.cnbc.com/2026/09/01/sb-energy-ipo-softbank-open-ai-nvidia.html
[10] Benzinga Newsdesk, “OpenAI Holds $5.5B In SB Energy Warrants,” Benzinga, August 2026. https://www.benzinga.com/markets/prediction-markets/26/08/61532442/openai-sb-energy-warrants
[11] Iñaki Aldasoro, Sebastian Doerr, and Daniel Rees, “Financing the AI boom: from cash flows to debt,” BIS Bulletin No. 120, Bank for International Settlements, January 7, 2026. https://www.bis.org/publ/bisbull120.pdf
[12] Nick Lichtenberg, “The central bank of central banks just released its flagship annual report — and it sees a $1 trillion AI investment boom headed for a reckoning,” Fortune (on the BIS Annual Economic Report 2026), June 29, 2026. https://fortune.com/2026/06/29/bis-central-bank-warning-hyperscaler-data-center-1-trillion-gamble-recession/
[13] Phurichai Rungcharoenkitkul, “The AI investment race,” BIS Working Papers No. 1367, Bank for International Settlements, July 2026. https://www.bis.org/publ/work1367.pdf
[14] Chibuike Oguh, “AI financing fueling a surge in U.S. convertible bond sales,” Reuters, May 20, 2026. https://finance.yahoo.com/markets/options/articles/ai-financing-fueling-surge-u-100326400.html
[15] Axios Markets Desk, “AI frenzy means that debt can be interest free,” Axios, September 2, 2026. https://www.axios.com/2026/09/02/ai-nvidia-zero-bonds
[16] Crypto Briefing Markets Desk, “US convertible bond sales hit record high as AI spending devours capital markets” (reporting Bloomberg, Dealogic, and JPMorgan data), September 2026. https://cryptobriefing.com/us-convertible-bond-record-ai-spending/
[17] TMT Finance, “2026 hyperscaler capex tops US$700bn — analysis,” August 18, 2026. https://www.tmtfinance.com/intel/2026-hyperscaler-capex-tops-us700bn-analysis
[18] Yahoo Finance Technology Desk, “Meta, Microsoft, Amazon, and Alphabet are about to spend a shocking amount of money to dominate the AI era” (reporting Goldman Sachs estimates), Yahoo Finance, 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
[19] CNBC Staff, “Amazon, Meta and Microsoft face skeptical investors this week after Google report sparked sell-off,” CNBC, July 28, 2026. https://www.cnbc.com/2026/07/28/hyperscalers-face-higher-capex-scrutiny-after-alphabet-report-panned.html
[20] Meta Platforms, Inc., “Meta Announces Joint Venture with Funds Managed by Blue Owl Capital to Develop Hyperion Data Center,” Meta Investor Relations, October 2025. https://investor.atmeta.com/investor-news/press-release-details/2025/Meta-Announces-Joint-Venture-with-Funds-Managed-by-Blue-Owl-Capital-to-Develop-Hyperion-Data-Center/default.aspx
[21] Dow Jones Newswires, “Meta, Blue Owl and AI: Here Are the Details of Wall Street’s Biggest Private-Credit Deal Ever,” October 22, 2025. https://www.itiger.com/news/1142559708
[22] Private Equity Insights, “Blue Owl and Meta close record $30bn financing for AI data centre expansion in Louisiana,” 2025. https://pe-insights.com/blue-owl-and-meta-close-record-30bn-financing-for-ai-data-centre-expansion-in-louisiana/
[23] Global Data Center Hub, “Meta + Blue Owl’s $27B Bet: Is This the US Blueprint for Financing AI Data Centers?” (reporting Morgan Stanley and KKR estimates), October 2025. https://www.globaldatacenterhub.com/p/meta-blue-owls-27b-bet-is-this-the
[24] Data Center Dynamics, “OpenAI signs $300bn cloud deal with Oracle — report,” 2025–2026. https://www.datacenterdynamics.com/en/news/openai-signs-300bn-cloud-deal-with-oracle-report/
[25] IntuitionLabs, “Oracle-OpenAI $300B Deal Explained: 2026 Update” (reporting Oracle RPO of $523 billion and Moody’s leverage analysis), April 2026. https://intuitionlabs.ai/articles/oracle-openai-300b-deal-analysis
[26] AI CERTs News, “Oracle-OpenAI $300B Cloud Deal Explained” (reporting Moody’s commentary on Stargate), December 2025. https://www.aicerts.ai/news/oracle-openai-300b-cloud-deal-explained/
[27] Jordan Novet et al., “OpenAI chip deal with Cerebras adds to roster of Nvidia, AMD, Broadcom,” CNBC, January 16, 2026. https://www.cnbc.com/2026/01/16/openai-chip-deal-with-cerebras-adds-to-roster-of-nvidia-amd-broadcom.html
[28] Blockspace via Yahoo Finance, “OpenAI lifts planned compute spending to $750 billion through 2030: WSJ,” July 22, 2026. https://finance.yahoo.com/technology/ai/articles/openai-lifts-planned-compute-spending-144917731.html
[29] Fierce Network, “Fierce Network’s Encyclopedia of AI deals,” November 2025. https://www.fierce-network.com/cloud/fierce-networks-encyclopedia-ai-deals
[30] NVIDIA Corporation, “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027,” Form 8-K Exhibit 99.1, U.S. Securities and Exchange Commission, August 26, 2026. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm
[31] Kif Leswing et al., “Nvidia earnings takeaways: Huang forecasts 70% fiscal 2028 revenue growth, far above estimates,” CNBC, August 26, 2026. https://www.cnbc.com/2026/08/26/nvidia-nvda-earnings-report-q2-2027-live-updates.html
[32] Daron Acemoglu, “The Simple Macroeconomics of AI,” NBER Working Paper No. 32487 (2024); published in Economic Policy 40(121), pp. 13–58 (2025). https://www.nber.org/papers/w32487
[33] Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies,” American Economic Journal: Macroeconomics 13(1), pp. 333–372 (2021). https://www.aeaweb.org/articles?id=10.1257%2Fmac.20180386
[34] Kristina McElheran, Mu-Jeung Yang, Zachary Kroff, and Erik Brynjolfsson, “The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s),” American Economic Association Annual Meeting, January 2026. https://swlb1.aeaweb.org/conference/2026/program/1160
[35] SFG Media, “The IMF Warns of a Risk to the Global Economy if the AI Investment Boom Reverses” (reporting the IMF January 2026 World Economic Outlook Update and remarks by Pierre-Olivier Gourinchas), January 19, 2026. https://sfg.media/en/a/imf-warns-ai-investment-boom-reversal-global-growth/
[36] Straight Arrow News, “Will AI drive a global boom or a bust? IMF report warns of economic threats” (reporting the IMF July 2026 World Economic Outlook Update and remarks by Petya Koeva Brooks), July 2026. https://san.com/cc/will-ai-drive-a-global-boom-or-a-bust-imf-report-warns-of-economic-threats/
[37] Cointribune, “Former IMF Chief Warns Of Looming Global Downturn” (reporting remarks by Gita Gopinath and JPMorgan data), October 2025. https://www.cointribune.com/en/former-imf-chief-warns-of-looming-global-downturn/
[38] Paul Kedrosky, “AI Set to be Largest CapEx Cycle Ever … and Soon Majority Externally Financed,” Paul Kedrosky — Applied Complexity, June 2026. https://paulkedrosky.com/ai-set-to-be-largest-capex-cycle-ever-and-soon-majority-externally-financed/
[39] Derek Thompson, “This Is How the AI Bubble Will Pop” (interview with Paul Kedrosky), October 2025. https://www.derekthompson.org/p/this-is-how-the-ai-bubble-will-pop
[40] PwC, “Where $31.6 trillion of capex flows in the era-defining AI build-out,” Global Data Centre Outlook, 2026. https://www.pwc.com/gx/en/1/services/consulting/technology/data-centre-outlook.html
[41] Crusoe via GlobeNewswire, “Crusoe Raises $3.9 Billion Series F for its Vertically-Integrated AI Infrastructure Platform” (including remarks by Gavin Baker, Atreides Management), September 17, 2026. https://www.globenewswire.com/news-release/2026/09/17/3364326/0/en/crusoe-raises-3-9-billion-series-f-for-its-vertically-integrated-ai-infrastructure-platform.html
[42] Carter Pape, “AI boom could cause the next economic crash, BIS warns,” American Banker, June 29, 2026. https://www.americanbanker.com/news/ai-boom-could-cause-the-next-economic-crash-bis-warns
[43] Gokhshtein Media, “Oracle’s $300B OpenAI Contract Inflates RPO 359%, But 2028 Delays Approach,” September 8, 2026. https://gokhshtein.com/news/2026-09-08-oracles-300b-openai-contract-inflates-rpo-359-but-2028
[44] TechCrunch Staff, “Crusoe reportedly raises $3B at a $30B valuation” (reporting the Jane Street cloud contract), TechCrunch, September 3, 2026. https://techcrunch.com/2026/09/03/crusoe-reportedly-raises-3b-at-a-30b-valuation/
