Introduction: When the Chip Seller Starts Backstopping the System
For most of Nvidia’s history, the basic economic relationship between the company and the rest of the world was easy to describe and even easier to model. Nvidia designed increasingly powerful processors; customers bought them; contract manufacturers fabricated them; cloud companies installed them in datacenters that somebody else financed; software developers eventually figured out what to do with the resulting computing power. Revenue moved toward Nvidia whenever customers wanted more chips, while nearly all of the financial risks associated with land, electricity, leases, construction debt and underutilized computing capacity remained somewhere else in the system, distributed among hyperscalers, datacenter developers, utilities, lenders and the customers of customers. Nvidia was, in the classical sense, a supplier of a component, and a supplier of a component does not ordinarily worry about whether the building that houses its product will be able to pay its mortgage.
By the summer of 2026, that description had become not merely incomplete but misleading, and the purpose of this paper is to explain why. Consider what happened during a remarkable eight-week stretch between the beginning of July and the last days of August. On July 1, Nvidia and its chief financial officer, Colette Kress, unveiled what the company called an AI Compute Partnership, a new business model under which AI cloud providers could purchase Nvidia systems with credit support from Nvidia itself, while Nvidia earned not only its traditional hardware revenue but also a share of the cloud revenue generated by that capacity [15]. More unusually, the arrangement allowed Nvidia to rent back unused GPU capacity at a fixed rate when participating cloud companies could not place that capacity with outside customers, providing them with a guaranteed buyer of last resort and thereby making it far easier for those companies to borrow the capital needed to buy Nvidia’s systems in the first place [16]. The first participants were not household names but companies such as Firmus, deploying 170,000 GPUs in Batam, Indonesia, and Sharon AI, deploying up to 40,000 Grace Blackwell GB300 systems in Australia under six-year agreements [16]. By July 26, Nvidia’s quarterly filing disclosed that its commitments under this program, typically six years in duration, already totaled $36 billion [11].
On August 10, a second and much larger piece of the architecture became visible. Nvidia announced memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish what it called independent compute financing platforms intended to mobilize more than $500 billion of third-party capital for AI infrastructure over time [3]. The language of the announcement matters as much as the number: Nvidia described its compute as an investable asset that provides the lowest token cost, highest revenue and longest life, supported by a rich ecosystem of offtakers built upon its CUDA platform [3]. The subtitle of the press release promised to turn Nvidia compute and full-stack AI infrastructure into an investable asset class for global capital while enabling long-duration, usage-linked revenue [3]. Reuters reported the next day that Jensen Huang, writing on X, had said Nvidia held the option to backstop up to $125 billion, or 25 percent, of the potential deals [4]. The president of Blackstone, Jon Gray, told CNBC that AI compute would come to be seen as a financeable asset class in much the same way that mortgage lenders assess homes [69].
One week later, on August 17, Nvidia filed a Form 8-K with the U.S. Securities and Exchange Commission disclosing that it had entered into multiple residual value guaranties with SB Energy, the SoftBank affiliate that will build, own and operate the PORTS Technology Campus in Pike County, Ohio, relating to leases for approximately 4.25 gigawatts of IT load at what the filing calls the Portsmouth Site, where an affiliate of OpenAI will be the tenant [5]. The filing states that, through the partnership and the credit support, Nvidia has secured land, power and shell capacity at the site to host Nvidia AI compute infrastructure [5]. Nvidia’s aggregate payment obligation under the initial guaranties is cumulatively capped at $105 billion, the guaranties become effective as individual leases commence beginning in 2028, Nvidia retains discretion to provide credit support for roughly 3.8 additional gigawatts, and OpenAI has agreed to reimburse and indemnify Nvidia for amounts actually paid [5][6]. The full campus is planned at roughly 8 gigawatts of IT capacity under a twenty-year lease, sitting on the reindustrialized grounds of a Cold War uranium-enrichment plant, with SB Energy and SoftBank planning at least 10 gigawatts of new generation including 9.2 gigawatts of natural gas, and with SB Energy and AEP Ohio committing at least $4.2 billion to new 765-kilovolt transmission and four substations [7].
Then came Nvidia’s August 26 earnings report for the second quarter of fiscal 2027, which ended on July 26. Huang declared that AI had reached its inflection point, that it was doing useful work, and that its tokens were productive and profitable [1]. Revenue reached $96.2 billion, up 18 percent sequentially and 106 percent from a year earlier; Data Center revenue alone reached $89.0 billion, up 117 percent; GAAP and non-GAAP gross margins were both 75.0 percent; the company forecast approximately $108.0 billion of revenue for the following quarter while assuming no Data Center compute revenue from China; and Nvidia returned approximately $26.0 billion to shareholders during the quarter through repurchases and dividends [1]. Huang’s framing of the quarter was characteristically compressed into a four-word formulation that will recur throughout this paper.
“Now, compute is revenue.”
— Jensen Huang, Founder and CEO, NVIDIA [1]
But perhaps the more consequential figures in that report were buried deeper inside the industrial machinery required to sustain such growth. In its quarterly filing, Nvidia disclosed that it had increased its supply commitments from $119 billion at the end of the prior quarter to $279 billion as of July 26, 2026, describing these commitments as covering data center infrastructure systems, primarily memory and manufacturing facilities, to produce its products for long-term demand across current and future architectures [12]. The company said the increase was primarily related to the procurement of memory, and it guided its gross margin down to 74 percent for the following quarter as component prices rose [13]. On the earnings call, Kress said the memory shortage would act as a bottleneck to the company’s growth at least through fiscal 2028 [14]. In other words, Nvidia was no longer only forecasting how many chips customers would want; it was placing hundreds of billions of dollars of forward bets on how much artificial-intelligence infrastructure would exist years from now, and it was doing so at a scale that reshaped the entire memory industry around its own order book.
And then, on the morning after the earnings, the limits of the new model came into view. Reuters reported on August 27, citing the Wall Street Journal, that Nvidia had paused some deals in the AI Compute Partnership after some Nvidia employees expressed concerns to current and potential customers that the initiative could draw antitrust scrutiny [9]. The Journal reported that in the early weeks of the program Nvidia had irked some potential partners with the extent of control it sought, including restrictions on which customers participating clouds could rent capacity to and a preference for spreading capacity across smaller firms [10][11]. Nvidia’s spokesperson responded that the new business model introduced in July remained in place and continued to evolve because of high demand [9]. Nvidia itself, in a statement to Tom’s Hardware, denied that the initiative had been paused at all, and Kress on the earnings call had already described the mechanics in the plainest terms available.
“a minimum revenue guarantee that gives lenders the confidence to underwrite the project”
— Colette Kress, Executive Vice President and CFO, NVIDIA (Q2 FY2027 earnings call) [11]
The distinction at the heart of this anecdote matters enormously. A semiconductor company normally prospers when somebody else is willing and able to finance the infrastructure needed to consume its products. Nvidia increasingly appears willing to participate in solving that financing problem itself, and to do so with every instrument available to a company that generates tens of billions of dollars of free cash flow every quarter: equity investments, credit support, residual-value guarantees, take-or-pay commitments, rent-back options, revenue participation, forward supply contracts, and partnerships with the largest allocators of private capital in the world. The sequence of events is summarized in Table 1.
Table 1. Nvidia’s Eight Weeks of Financial-Industrial Transformation (July 1 – August 27, 2026)
| Date | Action | Financial mechanism | Disclosed scale | Source |
| July 1, 2026 | AI Compute Partnership launched with neocloud partners (Firmus, Sharon AI and others) | Credit support; take-or-pay floor; rent-back of unused GPUs at fixed rate; revenue share above floor | $36 billion committed by July 26; six-year terms | [11][15][16] |
| August 10, 2026 | MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR | Independent compute financing platforms; Nvidia option to backstop 25% | >$500 billion third-party capital; up to $125 billion backstop | [3][4] |
| August 17, 2026 | Form 8-K: residual value guaranties with SB Energy for PORTS Technology Campus, Ohio (OpenAI tenant) | Guaranty of residual value under 20-year leases; OpenAI reimbursement; $1.5 billion equity in SB Energy | 4.25 GW initial; cap $105 billion; option on ~3.8 GW more | [5][6][8] |
| August 26, 2026 | Q2 FY2027 results and Q3 outlook | Operating cash generation; shareholder returns; forward guidance | $96.2B revenue; $89.0B Data Center; $108B guide; $26B returned | [1][2] |
| August 26, 2026 | Form 10-Q: supply and capacity commitments | Forward purchase commitments for memory, wafers, packaging, manufacturing facilities | $279 billion, up from $119 billion in one quarter | [12][13] |
| August 27, 2026 | WSJ/Reuters: some AI Compute Partnership deals paused amid antitrust and control concerns; Nvidia says model remains in place | Program governance; customer-allocation conditions | Undisclosed subset of deals | [9][10][11] |
Source: Author’s compilation from Nvidia disclosures and press reporting cited in the endnotes.
This is the anecdote at the center of this paper, and it can be restated as a chain of dependent clauses. Nvidia sells the scarce asset. Nvidia helps mobilize the capital used to purchase the asset. Nvidia can provide credit support to the institutions building around the asset. Nvidia can help absorb residual capacity risk when the asset sits idle. Nvidia invests in the companies that consume the asset. Nvidia makes enormous upstream commitments so that its suppliers will manufacture enough components to produce future generations of the asset. And because Nvidia’s architecture remains embedded throughout a large share of the frontier AI ecosystem—Huang’s own description of Nvidia as the only platform that runs in every cloud and powers every frontier and open-source model is not far from the truth [28]—decisions made in Santa Clara can transmit outward through AI laboratories, hyperscalers, datacenter developers, private-credit funds, insurers, utilities, construction firms, semiconductor suppliers and eventually the applications and agents running on top of the infrastructure. CNBC’s own summary of the earnings put the point bluntly: Nvidia is increasingly providing financial support through backstops and other arrangements that allow new AI data centers to get funded and built [2].
That is why this paper calls Nvidia the Central Bank of AI.
Why I Chose the Title “Central Bank of AI”
I chose the title Central Bank of AI because the economic role Nvidia is developing increasingly resembles several functions that central institutions perform when an expanding financial system requires liquidity, confidence and an ultimate backstop. A conventional central bank does not manufacture the factories, houses or machines that constitute an economy. Instead, it occupies a privileged position inside the financial architecture surrounding them. It influences liquidity, provides confidence during periods of stress, affects the cost and availability of capital and, under extraordinary circumstances, can become the institution standing behind markets when ordinary private counterparties become reluctant to do so. Nvidia obviously does not issue sovereign currency, determine statutory interest rates or possess governmental authority, and nothing in this paper should be read as suggesting otherwise. The comparison is therefore metaphorical. But within the narrower economy of artificial-intelligence capacity, Nvidia increasingly influences something analogous to liquidity: the availability, financing, allocation and utilization of accelerated compute.
The metaphor is not mine alone, and it is worth acknowledging that practitioners closest to the market reached for the same language before this paper did. In its July 2026 analysis of the GPU debt backstop, SemiAnalysis observed that smaller GPU buyers want the hardware and can pay for it, but cannot offer the credit ratings that lenders require in order to fund a buildout, and it described Nvidia’s mid-2026 posture in precisely the terms this paper adopts, writing that the company clearly
“stands ready to offer this support as the central bank”
— SemiAnalysis, “Nvidia GPU Debt Backstop Unleashes the AI Project Trinity” [17]
The title also captures an important transition in Nvidia’s corporate identity, one that CNBC crystallized in August 2026 when it reported that Nvidia’s moat was shifting from chips to capital, noting that the company held $30.2 billion of marketable equity securities as of its most recent quarter, up from $12.9 billion a year earlier, and that competitors such as AMD and Google had chipped away at Nvidia’s technology lead, pushing the company to take advantage of its other great asset [21]. Calling Nvidia merely a chip company increasingly misses the scale of its influence. The company sits near the center of a financial-industrial network connecting semiconductor fabrication, advanced memory, networking, cooling, datacenter construction, gigawatts of electricity, AI laboratories, hyperscalers and institutional capital. When Nvidia organizes hundreds of billions of dollars of infrastructure financing, makes residual-value guarantees, invests in AI companies, commits hundreds of billions to future supply and experiments with mechanisms for absorbing unused GPU capacity, it is helping create the conditions under which its own market can continue expanding. The company is therefore moving from selling scarce compute into actively shaping the market structure that finances, distributes and absorbs scarce compute. Huang himself described the August 10 financing announcement as a milestone in exactly these terms, saying that Nvidia began by building chips and was now helping create a new class of productive, investable infrastructure [3].
There is a second reason for the title, and it concerns the framework that organizes the rest of this paper. The Five-Layer AI Economy—Energy, Chips, Datacenters, Models and Applications/Agents—has become increasingly interdependent. A bottleneck in one layer transmits rapidly into the others. Insufficient electricity delays datacenters. Datacenter delays postpone GPU deployments. GPU scarcity constrains model training and inference. Expensive inference limits application economics. Weak application revenue eventually raises questions about whether enormous infrastructure commitments can earn acceptable returns, which in turn tightens the financing conditions for the next round of datacenters. Nvidia sits at an unusually powerful transmission point connecting these layers. Its financing decisions can therefore affect not merely semiconductor demand but the pace at which capital moves through the entire AI industrial system, in the same way that a central bank’s decisions affect not merely the banking system but the pace at which capital moves through the entire economy.
There is a third reason, which is that the institutions that actually monitor systemic risk have begun to describe the AI ecosystem in the vocabulary of financial stability. The Bank of England’s Financial Policy Committee, at its June 26, 2026 meeting, noted that activity across credit markets was expanding rapidly in public markets, private credit, leveraged finance and structured finance; that the pace of AI investment was unprecedented; that there was significant uncertainty about the scale of infrastructure required; and that if key assumptions proved incorrect—including the size of future compute demand, the timely availability of power, and the depreciation rate of data-centre facilities and AI chips—this could create risks to holders of debt issued by AI companies and projects [35]. The International Monetary Fund’s July 2026 World Economic Outlook Update warned that frothy equity valuations, particularly in AI-exporting economies and markets with high concentration in technology firms, could correct sharply [32]. And regional presidents of the Federal Reserve have started to ask, in public, whether the AI industry is becoming too big to fail [37]. When the stability of an industry becomes a subject for central banks, it is reasonable to ask which institution inside that industry is performing central-bank-like functions. This paper argues that the answer is Nvidia.
The phrase Central Bank of AI is thus deliberately larger than “GPU monopoly,” “chip leader” or “AI infrastructure company.” Those descriptions concern market share or technology. This paper concerns systemic function. The central question is no longer simply whether Nvidia can continue selling more GPUs. It is whether Nvidia has become economically important enough that sustaining the expansion of the broader AI-capacity market increasingly becomes part of Nvidia’s own strategic responsibility—and whether the rest of the system has begun to organize its expectations around that responsibility.
Structure of the Paper
Section 1 traces the transition from semiconductor vendor to compute market-maker, using Nvidia’s own disclosures from 2019 to 2026 to show how the scale of its forward commitments and financial participation has changed. Section 2 maps Nvidia’s central-bank-like functions onto each layer of the Five-Layer AI Economy. Section 3 identifies six specific functions—compute liquidity, credit support, capacity backstopping, compute transmission, allocation power and confidence—that justify the metaphor. Section 4 examines the risks of a central bank without a public mandate, with particular attention to circular financing, antitrust exposure, residual-value risk, supply-commitment risk, moral hazard and systemic concentration, drawing on the warnings issued during 2026 by the IMF, the Bank of England, the Federal Reserve system and leading academic economists. Section 5 looks forward to 2027–2030, the period during which compute finance will either mature into an institutional asset class or encounter its first genuine downturn. Section 6 distills seven pillars of lessons. The Conclusion returns to the title and explains why it is a framework rather than a provocation.
Section 1: From Semiconductor Vendor to Compute Market-Maker
The transformation described in this paper did not occur in a single quarter, and it cannot be understood by looking only at the events of August 2026. It is the culmination of a structural change in the economics of accelerated computing that began roughly with the arrival of large-scale generative AI in late 2022 and accelerated through every subsequent product cycle. The purpose of this section is to explain, in some depth, why the traditional chip-seller model became untenable at the scale the AI economy now demands, why compute has consequently become something closer to an investable asset than a component, and why the logical response of the dominant supplier of that asset was to move from selling it into making the entire economic environment around it financeable. Only after that explanation is in place does the concept of a compute market-maker—and, ultimately, of a Central Bank of AI—become analytically useful rather than merely rhetorical.
1.1 The End of the Traditional Chip-Seller Model
Historically, semiconductor companies manufactured components while their customers financed the infrastructure in which those components operated. The division of responsibility was clean. Intel did not guarantee the leases of the office buildings in which its processors sat; Nvidia, for most of its life, did not guarantee the leases of the gaming cafes and workstation farms that consumed its graphics cards. The AI boom disrupted this division because the cost of deploying leading accelerators expanded, within a few years, from millions of dollars to billions and increasingly to tens of billions, and because the deployment of those accelerators became inseparable from the deployment of power, land, cooling, networking and long-term customer contracts on a scale historically associated with utilities rather than with information technology.
A GPU sale today depends on an entire capital stack: land acquisition, grid interconnection, power-purchase contracts, datacenter leases, cooling infrastructure, networking equipment, servers, construction debt, equity capital and a long-duration customer willing to purchase inference or training capacity for years into the future. The Ohio campus illustrates the point at its extreme. The Nvidia guaranties do not become effective until SB Energy has satisfied ready-for-service conditions for the relevant premises, expected beginning in 2028, which means that between the August 2026 disclosure and the first dollar of guaranteed exposure lies a program of transmission construction, gas-turbine procurement, shell construction and commissioning that has to be financed in the interim by somebody [5][8]. The GPUs are, in a sense, the last item to arrive.
Nvidia therefore faces a new strategic problem that no previous semiconductor company has faced at this magnitude. Producing more GPUs does not automatically create enough financially viable places to install them. Reuters’ private-credit roundup of August 14, 2026 captured the logic with a memorable analogy, likening Huang to a car salesman hawking a popular model who sees eager customers but knows fewer of them can afford it, so that the fix is to round up deep pockets from across Wall Street to cover the gap; the column argued that the financing structure exists precisely because Nvidia’s own customers cannot fund the buildout on their own balance sheets [20]. SemiAnalysis made the same observation from the technical side: in 2025 datacenter capacity was the bottleneck, by early 2026 chip production had become the binding constraint, and by mid-2026 it had become clear that financing would be one of the most significant obstacles to ramping large-scale compute broadly [17]. That changes the corporate objective from the simple imperative to sell the accelerator into the far more complex imperative to make the entire economic environment surrounding the accelerator financeable.
1.2 Compute Becomes an Investable Asset
Nvidia’s August 10 financing initiative represents an institutional milestone because it makes explicit a proposition that had previously been implicit in a series of one-off transactions: that Nvidia-powered compute and AI factories can be transformed into assets financeable by global institutional capital. The press release describes the platforms as the first compute financing platforms of their kind at global scale, designed to create dedicated pools of capital at attractive rates for Nvidia customers, including frontier AI labs, enterprises, governments and AI clouds [3]. Goldman Sachs’ chief executive, David Solomon, characterized the moment for the financial industry.
“We’re in a pivotal moment of a historic AI investment cycle.”
— David Solomon, Chairman and CEO, Goldman Sachs [68]
This suggests the emergence of a new asset category situated somewhere between technology equipment, digital infrastructure and revenue-producing industrial machinery. An AI GPU cluster can potentially produce cash flows through model training, inference, enterprise AI services, agentic workloads, sovereign AI programs, robotics simulation, scientific computing, cloud leasing and capacity resale. If those cash flows become predictable enough—and predictability is the key word—compute can become collateral for increasingly sophisticated financing structures. The precedent already exists. CoreWeave borrowed $2.3 billion against Nvidia H100 processors in August 2023, the first time such hardware had been used as loan collateral, and by March 31, 2026 it had closed an $8.5 billion delayed-draw term loan rated A3 by Moody’s and A (low) by DBRS, the first investment-grade financing secured by high-performance computing infrastructure and an associated customer contract, priced at SOFR plus 2.25 percent on the floating tranche and roughly 5.9 percent on the fixed tranche, anchored by Blackstone Credit & Insurance and backed by contracts with Meta reportedly worth at least $19 billion [51][52]. Nvidia’s August 10 platforms are, in effect, an attempt to industrialize that transaction and to make it available across the ecosystem rather than to a single borrower with a single hyperscaler counterparty.
It is important to be precise about what Nvidia is claiming when it calls its compute investable, because the claim carries the entire financial architecture. The company argues that its compute is broadly adopted, flexible across models and workloads, fungible and transferable across customers and operators, and continuously improved through CUDA software in a way that extends its useful life and improves its economics over time [3]. Each of those adjectives is a statement about collateral quality: adoption speaks to liquidity, flexibility to redeployability, fungibility to the existence of a secondary market, and continuous software improvement to the slope of the depreciation curve. A lender underwriting a GPU-backed loan is being asked to accept those adjectives as underwriting assumptions. Whether they hold is the subject of Section 4.
1.3 The Scale of the Transition, in Nvidia’s Own Numbers
The most reliable way to see the transition is not through announcements but through the commitments Nvidia has disclosed in its own quarterly filings over seven years. In October 2019, Nvidia reported outstanding inventory purchase obligations of $980 million and other purchase obligations of $138 million [72]. By July 2022, at the height of the pandemic-era supply crunch and before ChatGPT existed, outstanding inventory purchase and long-term supply obligations had reached $9.22 billion [71]. By April 2024, they had doubled to $18.8 billion, with a further $10.6 billion of other purchase obligations including $8.8 billion of multi-year cloud service agreements [70]. By April 2026 the figure had reached $119 billion, and by July 26, 2026 it stood at $279 billion, with the company stating that the commitments cover memory and manufacturing facilities for long-term demand across current and future architectures [12]. Table 2 presents the series.
Table 2. Nvidia’s Disclosed Supply and Purchase Commitments, 2019–2026
| Quarter end | Inventory / long-term supply and capacity obligations | Other purchase obligations | Primary driver disclosed | Source |
| October 27, 2019 | $0.98 billion | $0.14 billion | Ordinary component procurement | [72] |
| July 31, 2022 | $9.22 billion (incl. $0.93 billion prepaid) | n/a | Long-term supply agreements after pandemic shortages | [71] |
| April 28, 2024 | $18.8 billion | $10.6 billion (incl. $8.8 billion cloud services) | Hopper ramp; long-term supply and capacity agreements | [70] |
| April 26, 2026 | $119 billion | n/a | Blackwell / Rubin ramp | [12] |
| July 26, 2026 | $279 billion | n/a | “Primarily related to the procurement of memory”; manufacturing facilities | [12][13] |
Source: Nvidia Forms 10-Q as cited. Figures are as reported in each filing; categories were reclassified by Nvidia over time and are not perfectly comparable.
The series tells a story that no narrative could tell as clearly. Over seven years, Nvidia’s forward commitments to its own suppliers grew by a factor of roughly 285. The July 2026 figure alone is larger than the entire annual revenue of the company in its most recent completed fiscal year, which reached $215.9 billion with Data Center contributing $193.7 billion [75]. Tom’s Hardware reported that the memory component alone approached $160 billion, including a supply pact with SK hynix, and that Nvidia had historically used such commitments for wafer processing and advanced packaging at TSMC as well as HBM, but that this time the bulk was explicitly memory [74]. The Seoul Economic Daily reported that Nvidia’s sweep of HBM supply was spilling over into standard server DRAM and NAND, deepening a shortage across the memory market and forcing other large technology companies and cloud providers to offer higher prices and unusually generous terms to secure chips [14]. A company that commits $279 billion to its suppliers is not merely forecasting demand; it is, within its own supply chain, determining who gets memory and who does not. That is an allocation function, and allocation functions are among the defining characteristics of the institutions this paper compares Nvidia to.
The revenue side of the ledger is equally instructive, and Table 3 sets it out. Data Center revenue grew from $15 billion in fiscal 2023 to $193.7 billion in fiscal 2026, and in the first two quarters of fiscal 2027 alone reached $164.2 billion [28][1][75]. The company’s first-quarter CFO commentary disclosed a fact that bears directly on the central-bank thesis: hyperscaler revenue remained at approximately 50 percent of Data Center revenue, while the remaining 50 percent came from a continued diversification of customers including AI clouds, industrial, enterprise and sovereign buyers [29]. Half of Nvidia’s datacenter demand, in other words, now comes from customers who do not have the balance sheets of Microsoft, Alphabet, Amazon or Meta—which is exactly the half that needs financing help.
Table 3. Nvidia Revenue Trajectory and Forward Guidance, Fiscal 2023 – Fiscal 2027
| Period | Total revenue | Data Center revenue | Notes | Source |
| Fiscal 2023 (full year) | n/a | $15.0 billion | Pre-generative-AI baseline as cited in FY2026 reporting | [75] |
| Fiscal 2026 (full year, ended Jan 25, 2026) | $215.9 billion (+65%) | $193.7 billion | Gross margin 71%; net income $120.1 billion; $40.1 billion buybacks | [75] |
| Q1 FY2027 (ended Apr 26, 2026) | $81.6 billion (+85% y/y) | $75.2 billion (+92%) | Hyperscale ~50% of Data Center; no Hopper shipments to China; free cash flow $48.6 billion | [28][29] |
| Q2 FY2027 (ended Jul 26, 2026) | $96.2 billion (+106% y/y) | $89.0 billion (+117%) | Gross margin 75.0%; $26.0 billion returned to shareholders; supply commitments $279 billion | [1][12] |
| Q3 FY2027 (outlook) | $108.0 billion ± 2% | n/a | No China Data Center compute revenue assumed; gross margin 74.0% | [1] |
| Fiscal 2028 (outlook) | ~70% growth, supply-constrained | n/a | Huang: demand “much greater” than the guided 70% | [2][58] |
Source: Nvidia press releases, CFO commentary and filings as cited.
1.4 From GPU Sale to Capital Formation
The next phase of Nvidia’s strategy may therefore depend less on convincing companies that GPUs are technologically useful—that argument has been won—and more on reducing the cost of capital required to deploy them. This produces a new mechanism that can be written as a loop: Nvidia hardware enables institutional financing, which produces datacenter capacity, which generates AI-cloud revenue, which supports model and application demand, which requires additional Nvidia hardware. The loop is strategically attractive because financing can accelerate the very demand from which Nvidia later benefits. Forbes’ Jim Osman, writing on August 16, 2026, argued that the significance of the financing platforms lies not in any suggestion that demand for Nvidia’s chips is artificial, but in what they reveal about how much capital is now required to keep that demand growing, and he noted that Goldman Sachs was already speaking with insurers, banks and money managers about participating [19].
But the same loop creates one of the central controversies examined later in this paper: at what point does ecosystem support become circular demand creation? Bloomberg reported that when Nvidia first announced its intention to invest up to $100 billion in OpenAI in September 2025, Bernstein’s Stacy Rasgon wrote to investors that the action would clearly fuel circularity concerns, and Bloomberg noted that Nvidia had participated in more than fifty venture deals for AI companies in 2024 alone, some of whose recipients then used the capital to buy Nvidia’s products [25]. The $100 billion framework was subsequently reduced to a finalized $30 billion equity investment in OpenAI’s February 2026 round, and Huang said in March that this investment, along with Nvidia’s $10 billion investment in Anthropic, might be the last because both companies were expected to go public [24]. The pattern is instructive: Nvidia has repeatedly announced large frameworks, faced circularity criticism, and then either scaled the commitment down or restructured it into a form—such as a residual-value guaranty rather than a direct investment—that transfers less capital while preserving the strategic objective.
1.5 Revenue Participation Changes the Relationship
Revenue-sharing arrangements alter Nvidia’s relationship with its customers in a way that is easy to state and difficult to overstate. Under a traditional transaction, Nvidia receives money when hardware changes hands, and its economic interest in the hardware ends at that moment. Under the AI Compute Partnership, Nvidia potentially receives economic exposure to what happens afterward: to whether the cluster is utilized, to the price at which its capacity is rented, and to the solvency of the operator [15][16]. The distinction is enormous. The company is no longer exposed only to GPU demand; it becomes exposed to GPU utilization. One analyst described the 25 percent backstop in the August 10 platforms as the entry fee for installing a meter on hardware Nvidia has already sold once, so that every gigawatt the platforms finance becomes a chip sale today and a toll tomorrow, and noted that Nvidia’s five-year credit default swap spread had risen to roughly 80 basis points, more than double its late-May level, as the market began to price the new exposures [66].
That exposure creates incentives for Nvidia to help customers obtain financing, find users, increase utilization, deploy software, attract AI startups, distribute capacity efficiently and avoid defaults. The chip vendor gradually becomes a participant in the operating economics of the compute market. It is precisely this participation that produced the friction reported by the Wall Street Journal in late August: when a supplier has a revenue interest in a customer’s utilization, it has a reason to care which customers the cloud rents to, and the Journal reported that partners objected to restrictions on which customers they could serve and to Nvidia’s preference for spreading capacity across smaller firms [10][11]. The controversy is not incidental to the model; it is an inherent consequence of it.
1.6 The Birth of the Compute Market-Maker
A market-maker traditionally helps create liquidity between buyers and sellers by standing ready to transact on both sides. Nvidia’s emerging role produces an analogous concept in AI infrastructure. It connects capital providers with datacenter developers, datacenter developers with AI clouds, AI clouds with frontier laboratories, and frontier laboratories with enterprise users, and it does so by supplying the standardized asset around which each of those relationships is organized. If Nvidia is also willing to rent capacity that otherwise lacks a customer—which is exactly what the take-or-pay floor in the AI Compute Partnership provides [11]—its position moves one step further. It becomes not only the supplier of the underlying asset but potentially a compute buyer of last resort. Capacity Media observed that a neocloud unable to obtain a $500 million GPU loan on its own credit can obtain one once Nvidia stands behind the residual value, and that Nvidia’s willingness to serve as guarantor had unlocked capital that likely was not available on comparable terms even six months earlier, with a refinancing wall of GPU-fleet debt taken on since 2023 at loan-to-value ratios of 60 to 70 percent now maturing into a market where lenders had grown cautious about pricing residual value on their own [18]. That is the first major reason the Central Bank of AI metaphor becomes analytically useful, and Section 3 develops it in detail. Before doing so, Section 2 places the metaphor inside the Five-Layer AI Economy.
Section 2: The Central Bank of AI Across the Five-Layer AI Economy
The Five-Layer AI Economy is the organizing framework of this paper, and it deserves a full explanation before its layers are examined one at a time. The framework holds that the modern artificial-intelligence economy is a vertically dependent stack in which each layer is both a customer of the layer beneath it and a supplier to the layer above it. Energy powers Chips; Chips populate Datacenters; Datacenters run Models; Models enable Applications and Agents; and Applications and Agents generate the revenue, productivity and strategic value that ultimately justify the capital invested in every lower layer. The framework matters for this paper because Nvidia’s financial interventions do not stay in the Chips layer where its products are made. A residual-value guaranty on an Ohio campus is an intervention in the Datacenters layer that reaches down into the Energy layer through 9.2 gigawatts of natural gas and $4.2 billion of transmission [7], and reaches up into the Models layer through OpenAI’s twenty-year tenancy [5]. A $279 billion memory commitment is an intervention in the Chips layer that reaches down into the industrial capacity of Korea’s memory makers [14]. A take-or-pay floor for a neocloud is an intervention in the Datacenters layer that reaches up into the Applications layer through the AI-native startups the neocloud serves [15]. To understand Nvidia as a central bank is to understand how its decisions transmit across all five layers at once.
2.1 Layer One — Energy: Compute Finance Becomes Power Finance
Every Nvidia GPU ultimately requires electricity, and consequently financing GPU deployment indirectly means financing power demand. The magnitudes have become macroeconomic. The International Energy Agency reported in its 2026 analysis, Key Questions on Energy and AI, that electricity demand from data centres grew by 17 percent in 2025 and that of AI-focused data centres by 50 percent, both far outpacing the 3 percent growth in global electricity demand; that the capital expenditure of five large technology companies surged past $400 billion in 2025 and was set to rise a further 75 percent in 2026; that data-centre electricity consumption is set to double by 2030 and AI-focused consumption to triple; and that the pipeline of conditional offtake agreements between data-centre operators and small modular reactor projects had grown from 25 gigawatts at the end of 2024 to 45 gigawatts [55][56]. Brookings, summarizing the IEA’s base case, notes that global data-centre consumption could reach 945 terawatt-hours by 2030 and 1,200 terawatt-hours by 2035, and that if data centres were a country they would rank between Japan and Russia as the fifth-largest consumer of electricity in the world [57]. The IEA had already warned in 2024 that Nvidia held roughly 95 percent of the AI server market and that AI electricity demand could be forecast from the number of Nvidia servers sold and their rated power [76]—which is to say that Nvidia’s shipment schedule is, to a first approximation, the world’s AI electricity demand schedule.
The Portsmouth, Ohio arrangement demonstrates how far the chain has extended. Nvidia’s disclosed guaranties concern infrastructure representing approximately 4.25 gigawatts of IT load with an option on roughly 3.8 more, an electricity requirement comparable to major industrial systems rather than to conventional corporate computing [5]. The campus is to be powered by at least 10 gigawatts of new generation, most of it gas-fired, in a plant that is owned by the U.S. government and financed by Japan under a recent trade agreement, and the transmission upgrades alone exceed $4.2 billion [7]. The IEA notes that developers constrained by slow grid connections are increasingly advancing onsite natural-gas generation, largely in the United States, but that satellite tracking shows many of these projects remain at early stages with significant technical and financial hurdles, and that the rapid and large swings in AI data-centre demand can stretch the capabilities of onsite gas plants [55]. AI finance is therefore becoming energy finance. Future Nvidia ecosystem arrangements will increasingly depend on nuclear generation, natural gas, renewables, battery storage, transmission, substations, grid interconnection, behind-the-meter power and long-term power-purchase contracts. The Central Bank of AI cannot ultimately stabilize compute without indirectly confronting electricity, and it cannot confront electricity without confronting the state regulators, utilities and ratepayers who govern it—a subject Section 5 returns to.
2.2 Layer Two — Chips: Securing the Monetary Base of AI
Within the metaphor developed in this paper, the GPU operates almost like the reserve asset of the modern AI infrastructure economy. Datacenters are financed around expected access to accelerators. AI-cloud companies advertise GPU inventories the way banks once advertised gold reserves. Frontier laboratories negotiate for clusters measured in hundreds of thousands or millions of accelerators. Countries announce national GPU capacity as an element of sovereignty. Investors increasingly evaluate AI infrastructure according to the quality, quantity and generation of the installed silicon, and rating agencies underwrite GPU-backed loans on the strength of the offtaker contract with the hardware as the recovery floor [52]. Nvidia’s huge supplier commitments therefore resemble a form of forward reserve accumulation: the company is securing fabrication, HBM, advanced packaging and other inputs needed to ensure that future compute supply exists when expected demand materializes [12][74]. The analogy should not be taken literally. But strategically, control over accelerator supply gives Nvidia extraordinary influence over who can build and when, and the memory market in the second half of 2026—in which Samsung’s and SK hynix’s operating profits were forecast to rise by 804 percent and 464 percent respectively while other buyers scrambled for allocation [14]—shows what it looks like when the central institution of a system decides to expand its reserves.
There is an important qualification. A reserve asset is only useful if it retains value, and the GPU’s claim to that status depends on the durability of its economic life. Nvidia itself argues that CUDA extends useful life and improves economics over time [3], and its supporters point out that an older chip that still earns rent still pays Nvidia. Critics, examined in Section 4, argue that a three-year product cycle in which each generation delivers two to three times the performance per watt makes the older generation uneconomic for frontier training far faster than five- or six-year depreciation schedules imply [49][50]. The monetary base of AI, in other words, may be subject to a rate of technological depreciation that no fiat reserve has ever faced.
2.3 Layer Three — Datacenters: From Equipment Supplier to Capacity Backstop
Datacenters are where the central-bank metaphor becomes strongest, because they are where financing risk concentrates. A GPU without a powered rack is inventory. A powered rack without a customer is stranded capacity. A heavily leveraged datacenter without sufficient utilization can become a credit problem, and a credit problem in a sector that has borrowed as heavily as this one can become a financial-stability problem. The Bank of England’s July 2026 Financial Stability Report observed that AI-related companies’ use of credit markets had accelerated rapidly across public markets, private credit, leveraged and structured finance, and was set to increase further as financing needs continued to expand [34]. The Quinn Emanuel client alert of March 2026 catalogued the structures: hyperscalers issued approximately $121 billion in bonds in 2025 alone, more than four times the five-year average; Meta financed its Hyperion campus through a $30 billion special-purpose vehicle; Google provided a $3.2 billion credit backstop for a Fluidstack/TeraWulf transaction; and CoreWeave’s original $7.5 billion GPU-collateralized facility carried a variable rate averaging roughly 11 percent with repayments beginning in January 2026 just as the collateral’s market value was declining [54]. By August 2026, Moody’s was warning that hyperscaler AI spending projected at $785 billion in 2026 and roughly $1 trillion in 2027 was eroding free cash flow and threatening credit quality across six major technology firms, and loan investors had begun demanding covenants after CoreWeave’s spreads widened by 125 basis points [53].
Nvidia’s AI Compute Partnership addresses precisely this problem. Under the program, Nvidia provides a take-or-pay commitment on a portion of a facility’s capacity, which functions as a minimum revenue guarantee that lenders can underwrite, and in exchange Nvidia shares in a portion of the neocloud’s revenue earned above that floor [11]. If demand is weak and a facility rents only a fraction of its capacity, Nvidia’s obligation requires it to cover the difference between actual revenue and the contracted minimum, or alternatively to rent back the unused compute for its own needs [11][16]. Economically, that mechanism reduces perceived residual demand risk. Lower residual risk improves financing. Improved financing produces more datacenters. More datacenters support more GPU purchases. The circularity is both the strategic genius and the central financial-policy question surrounding the model, and it is exactly the circularity that a central bank’s lender-of-last-resort function creates in a banking system: the existence of the backstop changes the behavior of everyone who might someday rely on it.
2.4 Layer Four — Models: Financing the Buyers of Intelligence Factories
Frontier laboratories sit in an unusual economic position. They can generate extraordinary revenues and valuations while simultaneously requiring capital expenditure on a scale historically associated with telecommunications networks, utilities, semiconductor fabs or national infrastructure programs. The 2026 figures make the point. OpenAI closed a $122 billion funding round in March 2026 at a post-money valuation of $852 billion, its annualized revenue run rate passed $40 billion in July, and its chief financial officer told employees on August 19 that the company would be public in 2027 or sooner [62]. Yet its second-quarter revenue was $6.7 billion and its operating loss widened from $9.3 billion in the first quarter to $12.3 billion in the second, driven by datacenter and computing costs [63]. Sacra’s reconstruction of the renegotiated Microsoft agreement projects OpenAI’s cash burn rising to roughly $27 billion in 2026 and $63 billion in 2027, with cash-flow positivity not arriving until 2030 [73]. Anthropic, meanwhile, told investors its run rate had reached $65 billion at the end of July, up roughly sevenfold in a year [21]. Sarah Friar’s own framing of the public listing is the clearest statement available of how the frontier laboratories understand their financial condition.
“The IPO is not a finish line, it is a milestone, another fundraise.”
— Sarah Friar, Chief Financial Officer, OpenAI [62]
OpenAI, Anthropic, xAI, Meta, Google and others therefore depend increasingly on access not merely to chips but to complete industrial systems, and on capital markets willing to fund those systems years ahead of the revenue they will produce. Nvidia benefits whenever more laboratories can obtain that capacity. This helps explain why Nvidia’s strategic interest extends beyond transactional hardware sales toward investment, infrastructure support, cloud relationships and capital formation. Huang defended the pattern explicitly in the days around the August earnings, arguing that AI-linked startups have capital requirements unlike any previous generation of companies [10].
“There has never been a startup that needed billions of dollars to get started”
— Jensen Huang, Founder and CEO, NVIDIA, on the capital needs of frontier AI laboratories [23]
Layer Two, in short, increasingly has an incentive to ensure that Layer Four remains sufficiently financed to continue consuming Layer Two. That incentive is rational, and it is also the seed of the moral-hazard problem examined in Section 4.
2.5 Layer Five — Applications and Agents: Where the System Must Ultimately Earn Its Return
No financing structure can permanently substitute for economic demand. The hundreds of billions invested in AI infrastructure must eventually be justified by applications capable of producing enough revenue, productivity or strategic value to service the debt, remunerate the equity and replace the hardware when it wears out or becomes obsolete. That means the ultimate foundation beneath Nvidia’s central role is not Blackwell or Rubin. It is useful AI. Agents must perform work. Models must produce valuable inference. Robots must improve industrial productivity. Enterprises must pay for AI services. Consumers must find AI sufficiently valuable to sustain subscriptions, transactions, advertising or other business models.
The evidence on this question in 2026 is genuinely divided, and honesty requires presenting both sides. On the optimistic side, Stanford’s Erik Brynjolfsson argued in February 2026 that the AI productivity take-off had finally become visible in aggregate data, pointing to a downward revision of 2025 U.S. job gains to just 181,000 alongside fourth-quarter GDP growth near 3.7 percent, which implied productivity growth of roughly 2.7 percent, nearly double the decade average, and he framed the moment as a transition along the productivity J-curve from an investment phase to a harvest phase [45].
“We are transitioning from an era of AI experimentation to one of structural utility.”
— Erik Brynjolfsson, Director, Stanford Digital Economy Lab, writing in the Financial Times [45]
Brynjolfsson’s earlier field research with Danielle Li and Lindsey Raymond, published in the Quarterly Journal of Economics in 2025, found that access to a generative-AI assistant raised the productivity of customer-support agents by 14 percent on average, with gains of more than 30 percent for the least experienced workers [46]. The Federal Reserve Board’s July 2026 FEDS Note found that AI-related components of investment had contributed meaningfully to quarterly GDP growth from 2025 through the first quarter of 2026, with software and computer equipment the largest positive contributors [38]. On the skeptical side, MIT’s Daron Acemoglu, a 2024 Nobel laureate, continued in 2026 to estimate that AI would deliver total-factor-productivity gains of roughly 0.55 percent over the coming decade, that only about 5 percent of tasks would be profitably automated in the near term, and that the resulting GDP boost would be on the order of 1 to 1.5 percent—a fraction of Wall Street’s projections [43]. He told MIT Technology Review in May 2026 that the missing signal for real economic impact was a layer of broadly usable consumer applications, and he expressed doubt that agents could yet handle the orchestration between tasks that humans do naturally [44]. His assessment of the quality of the surrounding debate was blunt.
“I find all of this discussion of capitalism so brainless”
— Daron Acemoglu, Institute Professor, MIT; 2024 Nobel Laureate in Economic Sciences [43]
Between those poles sits NYU’s Aswath Damodaran, whose August 20, 2026 essay argued that AI had reached what he called its bar mitzvah moment, the point at which the business must move from hype and hope to hard questions about revenues, profits and moats. He estimated that even generous calculations put the current AI products and services market at roughly $250 billion, that infrastructure providers such as Nvidia, electrical-equipment makers and utilities had captured much of the economic value so far, and that the valuations being attached to the frontier laboratories would require revenues on the order of a trillion dollars within a decade to justify [40][41]. The true balance sheet behind the Central Bank of AI is therefore the future productivity of artificial intelligence itself, and as of August 2026 that balance sheet is being audited in real time by the most serious economists in the field, without agreement.
Table 4. Nvidia’s Interventions Mapped Across the Five-Layer AI Economy
| Layer | What the layer does | Nvidia’s traditional relationship | Nvidia’s central-bank-like intervention (2025–2026) | Illustrative evidence |
| 1. Energy | Generation, transmission, grid interconnection | None (indirect consumer through customers) | Guaranties on a campus requiring ~10 GW of new generation and $4.2B of transmission; $1.5B equity in SB Energy | PORTS-Pike, Ohio [5][7] |
| 2. Chips | Accelerators, memory, packaging, networking | Designer and seller | Forward reserve accumulation: $279B supply commitments, mostly memory; allocation influence across HBM/DRAM market | 10-Q; memory-market reporting [12][14] |
| 3. Datacenters | Land, shells, power, cooling, racks, leases | Equipment supplier | Take-or-pay floors; rent-back of unused GPUs; residual-value guaranties; $500B financing platforms with 25% backstop option | AI Compute Partnership; Aug 10 MOUs [3][4][11][16] |
| 4. Models | Frontier labs training and serving models | Customer | Equity: $30B OpenAI, $10B Anthropic; $30.2B marketable equity portfolio; credit support behind OpenAI’s tenancy | CNBC; 8-K [21][24][5] |
| 5. Applications & Agents | Enterprise, consumer, agentic, robotics workloads | Indirect beneficiary | Revenue participation in neocloud usage; steering capacity toward AI-native startups (Baseten, Fireworks, Together) | AI Compute Partnership launch [15] |
Source: Author’s analysis.
Section 3: What Makes Nvidia “Central-Bank-Like”?
A metaphor earns its place in analysis only if it can be decomposed into specific, observable functions, each of which can be tested against evidence. This section therefore identifies six functions that central institutions perform in a financial system—the supply of liquidity, credit enhancement, backstopping, transmission, allocation and the anchoring of confidence—and asks, for each, whether Nvidia now performs an analogous function in the AI-capacity economy. The exercise is not intended to prove that Nvidia is a central bank, which it is not. It is intended to show that the resemblance is structural rather than superficial, and that it holds across enough dimensions that the metaphor illuminates more than it obscures. It concludes with a consolidated view of Nvidia’s financial exposure stack as disclosed through August 2026, because a central bank is defined as much by its balance sheet as by its functions.
3.1 Compute Liquidity
The first central-bank-like characteristic is the ability to increase the availability of compute by helping capital reach infrastructure projects that otherwise could struggle to secure financing. This is compute liquidity. Compute liquidity describes not simply how many GPUs exist, but how easily capital, hardware, power and customers can be assembled into functioning AI capacity. The AI Compute Partnership and the August 10 financing platforms are liquidity operations in this sense: they do not create new GPUs, but they lower the cost and raise the availability of the capital that turns GPUs into revenue-producing clusters [3][15]. Huang described the purpose of the platforms precisely in these terms, saying they would help customers access scarce compute at scale [4]. The analogy to a central bank’s open-market operations is imperfect, because Nvidia is also the seller of the asset whose liquidity it is enhancing, but the functional effect on the system—more capacity financed at a lower spread than the market would otherwise offer—is the same.
3.2 Credit Support
The second function is credit enhancement. Nvidia’s residual-value guaranties and financing partnerships help shift risk between developers, tenants, lenders and investors. The structure of the Ohio guaranty is instructive because it is deliberately engineered as a guaranty of an object rather than of a tenant: if OpenAI defaults or becomes insolvent, SB Energy must first attempt to relet the premises at the same price and then to sell them, and only the remaining shortfall against the guaranteed minimum value is Nvidia’s, subject to the $105 billion cumulative cap, with Nvidia choosing among remedies that include assuming the lease, reletting, selling, terminating or deferring for up to a year [7][8]. Global Data Center Hub noted that the guaranteed minimum value that determines what Nvidia owes is not disclosed in the public documents, that the cap binds only where recovery approaches zero, and that OpenAI’s reimbursement obligation is triggered by the same insolvency that would make reimbursement uncollectible [8]. The company is effectively using the strength of its balance sheet and ecosystem position to reduce uncertainty surrounding infrastructure deployment. That is fundamentally different from cutting GPU prices. It addresses the financing architecture around the GPU, and it does so in the way that a central bank’s collateral framework addresses the financing architecture around a bank’s assets: by declaring, in advance, what will be accepted and at what haircut.
3.3 Capacity Backstopping
The third function is potentially the most important. If Nvidia helps absorb computing capacity that cannot immediately find outside customers, it effectively places a floor beneath part of the utilization risk. This resembles a private-sector version of a lender of last resort, except that what Nvidia lends is not money but demand. The logic runs as follows: unused compute triggers Nvidia absorption, which reduces stranding risk, which raises financing confidence, which produces more infrastructure construction. Data Center Dynamics reported that the backstop sees Nvidia agreeing to rent back unused GPUs at a fixed rate, and that the model formalizes and extends earlier demand guarantees under which Nvidia had helped neoclouds such as CoreWeave and Lambda raise billions in debt [16]. Capacity Media noted that CoreWeave’s 2025 IPO filing disclosed Nvidia simultaneously holding an equity stake, serving as a major customer and acting as a capacity backstop, and that the new partnership formalizes that same pattern across a wider set of smaller operators [18]. This mechanism deserves close scrutiny because it can both stabilize a young market and obscure whether underlying demand is sufficiently independent; the same property is true of every lender of last resort ever created, which is why such institutions are ordinarily placed under public governance.
3.4 Compute Transmission
Central banks influence economies through transmission mechanisms. Nvidia has its own increasingly powerful transmission mechanism, which runs from a Nvidia financing decision to datacenter financing, to GPU deployment, to electricity demand, to model capacity, to inference supply, and finally to application economics. A decision that begins as a financing arrangement in Santa Clara can eventually affect electrical infrastructure in Ohio, Texas, Virginia, Pennsylvania, Indiana or Arizona. The Ohio transaction is the cleanest illustration: an 8-K filed by a semiconductor company describes a twenty-year lease, a gas plant, four substations, and 765-kilovolt transmission lines in Pike County [5][7]. The IMF’s July 2026 press briefing described the same transmission mechanism at the level of nations, noting that the strength of AI investment and its boost to some economies had been a surprise, that Korea was uniquely placed to benefit from the upturn and correspondingly exposed in a downturn, and that a market correction driven by a reassessment of AI profitability was a key downside risk [33]. When the Seoul Economic Daily reports that Nvidia’s memory commitments will lift the operating profits of Samsung and SK hynix by several hundred percent [14], it is describing the compute transmission mechanism of the Five-Layer AI Economy reaching a sovereign balance of payments.
3.5 Allocation Power
Nvidia’s importance also gives it influence over scarce capacity. When accelerators are constrained, decisions about allocation determine which AI clouds expand, which laboratories receive sufficient compute and which countries build sovereign AI infrastructure fastest. Huang has said that Nvidia prioritizes customers that can use its products now [65], which is a rationing rule, and the August revenue-sharing controversy is particularly instructive because reports indicated sensitivities about Nvidia’s preferred conditions regarding who could use participating capacity: the Journal reported that Nvidia sought to restrict which customers partner clouds could rent to and preferred that capacity be spread across smaller firms rather than concentrated with hyperscalers [10][11]. The University of Chicago’s Randal Picker, commenting on the earlier DOJ inquiry, offered the classical antitrust framing.
“You are allowed to get market power, you’re allowed to have it.”
— Randal Picker, Professor, University of Chicago Law School [65]
The difficulty, as Picker’s formulation implies, lies in what one does with market power once obtained. When the supplier of the scarce asset also influences financing and potentially downstream allocation, market power becomes more complicated than conventional semiconductor market share, because the supplier is no longer merely choosing whom to sell to but shaping the customer base of its customers. That is an allocation function of a kind that, in monetary systems, is ordinarily reserved to public institutions and hedged with rules.
3.6 Confidence
The final central-bank-like function is psychological. Infrastructure markets depend on confidence. Lenders must believe customers will pay. Datacenter operators must believe tenants will remain solvent. Utilities must believe projects will actually materialize. Suppliers must believe orders will not disappear. Investors must believe AI workloads will grow fast enough to justify today’s enormous capital commitments. Nvidia’s willingness to commit its own capital acts as a confidence signal across that network, and the August 10 announcement was explicitly designed to be read that way: six of the largest allocators in the world declaring, alongside Nvidia, that compute is an investable asset class [3][67]. Huang told CNBC after the earnings that the risk of Nvidia’s financial support was low and described the investments as a once-in-a-generation opportunity [22].
But confidence works in both directions. The same interconnectedness that stabilizes an expanding market can transmit losses rapidly if assumptions fail, and the market’s response to Nvidia’s August commitments—a widening in its own credit default swap spread [66], a 3 percent decline in the stock on the day the backstop figure was explained, and a subsequent recovery on the earnings—shows that investors are pricing both the stabilizing and the destabilizing possibilities simultaneously. Table 5 summarizes the six functions, and Table 6 consolidates the exposure stack that stands behind them.
Table 5. Six Central-Bank-Like Functions and Their Nvidia Analogues
| Central-bank function | What it does in a monetary system | Nvidia analogue in the AI-capacity economy | Evidence (2026) |
| Liquidity provision | Supplies reserves so that credit can expand | Compute financing platforms; credit support that lets capital reach projects | $500B platforms; AI Compute Partnership [3][15] |
| Credit enhancement / collateral framework | Defines eligible collateral and haircuts | Residual-value guaranties; declaring compute an investable asset with defined characteristics | Ohio 8-K; $125B backstop option [4][5] |
| Lender (buyer) of last resort | Stands behind solvent institutions in stress | Take-or-pay floors; rent-back of unused GPUs at fixed rate | Kress remarks; DCD reporting [11][16] |
| Transmission mechanism | Policy rate moves through banks to the real economy | Financing decisions move through datacenters to power, memory makers, sovereign economies | IMF on Korea; memory profits [14][33] |
| Allocation / rationing | Allocates reserves among counterparties in scarcity | Decides who gets HBM, who gets Blackwell/Rubin, which tenants clouds may serve | WSJ reporting; Huang on prioritization [10][65] |
| Anchoring confidence | Signals commitment to stability | Own-capital commitments and joint declarations with Wall Street | Aug 10 MOUs; Huang: “risk is low” [3][22] |
Source: Author’s analysis.
Table 6. Nvidia’s Disclosed Financial Exposure Stack as of August 27, 2026
| Instrument | Counterparty / structure | Disclosed amount | Nature of exposure | Source |
| Supply and capacity commitments | Memory makers (incl. SK hynix), TSMC, manufacturing facilities | $279 billion | Forward purchase obligations; partly cancelable or reschedulable with costs | [12][74] |
| Residual value guaranties | SB Energy (lessor); OpenAI (tenant); PORTS-Pike, Ohio | Capped at $105 billion; option on ~3.8 GW more | Contingent; triggered by tenant insolvency/default, net of recoveries; reimbursable by OpenAI | [5][8] |
| AI Compute Partnership commitments | Neoclouds (Firmus, Sharon AI and others) | $36 billion (six-year terms) | Take-or-pay / minimum revenue; rent-back option; revenue share above floor | [11][12] |
| Financing-platform backstop option | Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR | Option up to $125 billion (25% of >$500 billion) | Residual-value support on chips pledged as collateral; MOUs not yet final agreements | [3][4][66] |
| Equity investments | OpenAI ($30B), Anthropic ($10B), SB Energy ($1.5B), CoreWeave, others | $30.2 billion marketable equity securities (latest quarter) plus private stakes | Mark-to-market and illiquidity risk; concentration in own customers | [21][24] |
| Prior demand guarantees | CoreWeave, Lambda and other neoclouds | Historical: e.g., $6.3B CoreWeave capacity commitment | Precedent structures now formalized in the Partnership | [16][18] |
Source: Author’s compilation from Nvidia filings and reporting cited. Amounts are caps, options or commitments as disclosed and are not additive measures of expected loss.
Two observations follow from Table 6. First, the headline figures—$279 billion, $105 billion, $125 billion, $36 billion—are of different kinds and must not be summed; a supply commitment is a promise to pay for goods Nvidia intends to sell at a 75 percent gross margin, while a residual-value guaranty is a contingent liability that pays only on a trigger event and net of recoveries. Second, and more importantly for the thesis of this paper, every one of these instruments has the same directional exposure. Each pays off, or avoids paying out, in a world where AI demand keeps growing, and each becomes more expensive in a world where it does not. One analyst described this as wrong-way risk: Nvidia’s obligations grow as AI demand weakens, at exactly the moment its revenue falls [66]. A central bank whose balance sheet is entirely long the economy it stabilizes is a central bank with a pro-cyclical balance sheet, and that is the structural feature that Section 4 examines.
Section 4: The Risks of Having a Central Bank Without a Public Mandate
Every advantage described in the preceding sections has a corresponding hazard, and the purpose of this section is to examine those hazards with the same seriousness that the preceding sections applied to the advantages. The essential difficulty can be stated at the outset. A public central bank is placed under statute, supervised by legislatures, constrained by mandates, staffed by officials who do not profit from the institutions they stabilize, and granted the one power that makes a backstop credible in any state of the world: the ability to create the reserve asset without limit. Nvidia possesses none of these features. It is a shareholder-owned corporation whose executives are compensated in its equity, whose backstops are financed from cash flows that depend on the health of the very system it is backstopping, and whose reserve asset—the GPU—depreciates according to a product roadmap that Nvidia itself controls. A central bank without a public mandate is therefore not merely a private institution performing public functions; it is a private institution performing public functions with a pro-cyclical balance sheet and an unavoidable conflict of interest. The six risks examined below all flow from that structural fact, and the institutional warnings issued during 2026 by the IMF, the Bank of England, the Federal Reserve system and academic economists are best understood as attempts to describe it.
4.1 Circular Financing
The most obvious concern is circularity. Suppose Nvidia invests in an AI company; helps that company obtain financing; sells GPUs to infrastructure supporting that company; guarantees part of the infrastructure; receives cloud revenue participation; and rents unused capacity if outside customers do not materialize. Each individual transaction may have economic logic. Collectively, however, analysts must determine how much final demand originates from truly independent customers. That determination has become the central analytical problem of the AI investment cycle. CNN’s August 13, 2026 analysis defined circular financing as one company paying money to another—in the form of a loan, investment, lease or other support—in exchange for that second company buying the first’s products, and observed that during booms such arrangements create a virtuous circle that works until it does not; it also reported Apollo’s chief economist, Torsten Slok, warning that the math in the AI space did not yet add up [27]. Bloomberg had made the same point a year earlier, when the original $100 billion OpenAI framework was announced.
“The action will clearly fuel ‘circular’ concerns”
— Stacy Rasgon, Senior Analyst, Bernstein Research (investor note reported by Bloomberg) [25]
Nvidia’s response has been consistent and deserves to be stated fairly. Kress told analysts that Nvidia’s investment decisions rest on independent commercial judgment and that portfolio companies purchase Nvidia products on the basis of their own technical requirements without tie-in or mandatory-procurement clauses [10]. Huang argued on CNBC that critics were missing the essential point, which is that frontier laboratories are the first generation of companies requiring tens of billions of dollars of capital simply to get started, that the risk is manageable because the core assets of portfolio companies are GPU clusters, and that his only regret was not having invested earlier and more [22][23]. There is force in this defense. The distinction that matters analytically is not whether Nvidia participates financially—vendors have financed customers in every capital-intensive industry—but whether the participation is accelerating demand that would eventually exist or creating demand that cannot survive without continued support. This distinction will become increasingly important for investors and regulators, and it is the subject of the next subsection.
4.2 Artificial Demand Versus Accelerated Demand
Circular financing does not automatically mean false demand. Young infrastructure systems frequently require vendor finance. Aircraft manufacturers support customers. Equipment makers provide leasing. Telecommunications vendors extended credit throughout the 1990s buildout. Automakers finance buyers through captive finance arms that are, in some years, more profitable than the manufacturing operations they support. The analytical question is therefore not whether Nvidia participates financially, but whether financial support is accelerating demand that would eventually exist or creating demand that cannot survive without continued support. That distinction may determine whether the 2020s AI infrastructure boom resembles the early development of railroads, electricity and cloud computing—or an overbuilt speculative cycle of the kind that consumed telecommunications equipment vendors after 2000.
The Vanderbilt Policy Accelerator’s March 2026 paper, After the AI Crash, argues that the scale of the current cycle makes the second outcome a systemic rather than a sectoral risk. Its author, Asad Ramzanali, a former chief of staff at the White House Office of Science and Technology Policy, calculated that the hundreds of billions of dollars of planned 2026 hyperscaler capital expenditure was on a path to represent a larger share of U.S. GDP than peak investment in the Manhattan Project, the expansion of electricity, the Apollo program, the interstate highway system or the dot-com broadband buildout, and he argued that the combination of overreliance on AI investment with opaque financial engineering means a correction could resemble 2008 rather than 2000 [47][48]. In a subsequent discussion of the paper he emphasized how quickly formerly asset-light companies had become heavy debt issuers, pointing to Alphabet’s issuance of a hundred-year bond as the emblem of the shift [77], and the paper drew bipartisan attention when Senators Elizabeth Warren and Marsha Blackburn headlined a Vanderbilt event on the financial foundations of AI in April 2026 [78].
“lawmakers should prepare for this anticipated crisis now”
— Asad Ramzanali, Director of AI and Technology Policy, Vanderbilt Policy Accelerator [48]
Jim Cramer, reacting to the initial reports of a $250 billion Nvidia backstop for the Ohio campus in late July, reached for the same historical comparison and framed the question in terms of the buyer’s ability to pay: if OpenAI can afford the chips, perhaps because it goes public, Nvidia is in excellent shape, and if it cannot, the story is very different; he warned that if the market stops funding datacenters and the companies themselves lack the money, the outcome would be a return to 2000 [26].
“I don’t want the sequel.”
— Jim Cramer, CNBC, on the parallels between AI vendor financing and the dot-com bubble [26]
What distinguishes the present cycle from its predecessors is the quality of the revenue evidence. OpenAI’s run rate has risen from $3.7 billion in late 2024 to more than $40 billion in August 2026 [62]; Anthropic reports $65 billion [21]; Google Cloud’s backlog reached roughly $460 billion, double the prior year, and AWS grew 37 percent in the second quarter of 2026 [30]. Those are not the numbers of a demand-free bubble. But they are also not yet the numbers that justify the capital: Damodaran’s estimate of a $250 billion AI products-and-services market sits against more than $725 billion of 2026 hyperscaler capital expenditure [30][41], and his earlier warning about what he calls the Big Market Delusion—the tendency of too many investors to assume that each of their companies will capture the majority of a vast new market—was accompanied by an estimate of what the sector would collectively need to earn.
“two, three, four trillion in revenues eventually”
— Aswath Damodaran, Professor of Finance, NYU Stern School of Business, on the revenue required to justify AI capital [42]
The honest conclusion is that the demand is real and the financing is running ahead of it, and that Nvidia’s backstops are the mechanism by which the gap between the two is being bridged. Whether that bridge holds depends on the rate at which Layer Five revenue arrives, which no financing structure can control.
4.3 Antitrust and Control
Reuters’ August 27 report makes competition policy impossible to ignore, because it reveals that the concern originated inside Nvidia. Some Nvidia employees expressed to current and potential customers that the AI Compute Partnership could draw antitrust scrutiny, and the Journal described sensitivities around the extent to which Nvidia can dictate how its customers do business [9][10]. If Nvidia simultaneously supplies the dominant accelerator architecture, supports financing, participates in revenues, influences customers and helps determine where capacity is placed, regulators may eventually ask whether financial integration reinforces technological market power. The question is not new. The Department of Justice sent subpoenas to Nvidia in 2024 amid concerns that the company made it harder for customers to switch suppliers and penalized buyers that did not use its chips exclusively [65]. When the OpenAI framework was announced in 2025, Vanderbilt Law School’s Rebecca Haw Allensworth warned that Nvidia’s financial interest in OpenAI could distort market behavior, and the DOJ’s antitrust chief emphasized the need to prevent exclusionary conduct over the resources needed to build competitive AI systems [64].
Future antitrust questions could therefore involve far more than GPU market share. Regulators may examine tying arrangements, preferential access, customer restrictions, cloud allocation, financing conditions, interoperability, software lock-in, investment relationships and vertical influence across AI infrastructure. The central-bank metaphor sharpens the problem rather than softening it. A public central bank is permitted to discriminate among counterparties—to lend to some banks and not others, to accept some collateral and not other collateral—precisely because it is a public institution acting under a mandate and subject to review. A private supplier that discriminates among its customers’ customers as a condition of credit support is doing something that looks structurally similar and is treated by law entirely differently. The Vanderbilt paper goes so far as to propose a Glass-Steagall for AI, a structural separation between the provision of AI infrastructure and financial engineering around it, alongside utility-style regulation of critical layers of the stack [47]. Whether or not such proposals advance, the fact that Nvidia paused deals within seven weeks of launching the program suggests that its own lawyers had reached a similar view of the exposure.
4.4 Residual-Value Risk
AI chips depreciate unusually quickly because technological generations advance rapidly. A datacenter financed around today’s premium accelerator may contain equipment that has lost significant economic value several years later. This creates a fundamental question: what is the residual value of yesterday’s intelligence machine when tomorrow’s accelerator delivers dramatically more inference per watt? The question moved from accounting seminars to the center of the market in November 2025, when Michael Burry argued that hyperscalers were understating depreciation by extending the useful lives of Nvidia-based hardware to five or six years while the product cycle ran two to three, and estimated that the industry would understate depreciation by roughly $176 billion between 2026 and 2028 [49][50].
“Understating depreciation by extending useful life of assets artificially boosts earnings”
— Michael Burry, Scion Asset Management, writing on X [50]
Nvidia pushed back, arguing that customers consistently observe four-to-six-year lives on the basis of utilization and longevity, and the hyperscalers themselves diverged: Amazon shortened the useful life of a subset of its servers in 2025 while Meta extended its estimate further [49]. Microsoft’s chief executive offered the most candid statement of the underlying dynamic, acknowledging that Nvidia’s migration pace had accelerated and that the biggest competitor to any new Nvidia chip is its predecessor.
“stuck with four or five years of depreciation on one generation”
— Satya Nadella, Chairman and CEO, Microsoft, on why Microsoft spaces its AI chip purchases [49]
The relevance to the Central Bank of AI is direct. Every residual-value guaranty, every rent-back at a fixed rate, and every backstop of chips pledged as collateral is a bet on the slope of the depreciation curve. Quartz reported in May 2026 that CoreWeave’s GPU rental rates had fallen 50 to 70 percent from their peaks, and that GPU-collateralized debt structures depend on customer contracts holding, so that their credit quality is only as strong as the counterparty’s willingness and ability to keep paying [51]. The Bank of England’s FPC listed the depreciation rate of data-centre facilities and AI chips among the key assumptions whose failure could create risks to holders of AI debt [35]. Guaranteeing infrastructure whose economics depend on rapidly depreciating technology could eventually expose Nvidia to risks very different from ordinary semiconductor manufacturing, and it is worth noticing that Nvidia is the one party in the system that controls the variable in question: the faster it ships the next generation, the faster it depreciates the collateral behind its own guaranties.
4.5 Supply-Commitment Risk
Nvidia is simultaneously making enormous upstream commitments. That strategy protects the company against shortage during accelerating demand; Kress said the memory bottleneck would constrain growth through fiscal 2028, and Huang said demand was far in excess of what supply allowed the company to forecast [14][58].
“much greater than 70 per cent”
— Jensen Huang, Founder and CEO, NVIDIA, on unconstrained demand relative to guided fiscal-2028 growth [58]
But if demand slows, upstream commitments can become liabilities. The same mechanism operates in two directions: in a boom, supply commitments protect growth; in a slowdown, supply commitments amplify excess capacity. Nvidia’s filing notes that some agreements are cancelable, reschedulable or adjustable before firm orders are placed, but that changes may result in additional costs [12], and the company has already absorbed margin pressure from rising memory prices, guiding gross margin to 74 percent for the third quarter and, according to analysts, toward 71 to 72 percent by the fourth [13]. The larger Nvidia becomes, the more its forecasting errors could propagate across HBM manufacturers, foundries, packaging suppliers, datacenter developers and electricity providers. A central bank that has pre-committed to buy $279 billion of the reserve asset from its suppliers has, in effect, adopted a forward guidance so aggressive that reversing it would itself be a systemic event.
4.6 Moral Hazard and the Backstop
Stabilization is never free. A credible backstop encourages investment because lenders and developers perceive less downside risk; that is its purpose. But backstops can also encourage excessive construction if participants believe another institution will absorb losses. AI infrastructure could therefore encounter a form of moral hazard familiar from banking: the more Nvidia protects its ecosystem from failure, the more aggressively the ecosystem may expand, and the more the ecosystem expands on the assumption of protection, the more expensive protection becomes to withdraw. Capacity Media put the point precisely, observing that unlocking financing and removing risk from the system are not the same thing, and asking whether the backstop model resolves the neocloud refinancing wall or simply relocates the same risk onto a single, highly concentrated balance sheet that also controls chip supply, pricing and the pace of the next hardware generation [18].
The regulators have noticed. In January 2026, Senator Elizabeth Warren and colleagues pressed the Financial Stability Oversight Council to investigate the more than $1 trillion of debt projected to flow into AI infrastructure, warning that AI companies rely increasingly on complex and opaque debt including private credit, securitizations and off-balance-sheet financing, and that some AI executives were already laying the groundwork for a taxpayer-funded bailout if the bubble bursts [39].
“could cause destabilizing losses for an interconnected set of financial institutions”
— Senator Elizabeth Warren et al., letter to the Financial Stability Oversight Council [39]
The moral-hazard concern is also the reason transparency matters so much. When Nvidia’s guaranteed minimum value in the Ohio leases is undisclosed [8], when the revenue-share percentage in the AI Compute Partnership is undisclosed [16], and when the financing platforms are memoranda of understanding whose final terms have not been agreed [3], the market is being asked to extend confidence on the basis of headline caps rather than expected losses. A public central bank publishes its balance sheet weekly. The Central Bank of AI publishes it quarterly, in footnotes.
4.7 Systemic Concentration
The deepest question is systemic. If one company becomes simultaneously the dominant accelerator supplier, a major software-platform provider, a networking supplier, an AI infrastructure investor, a financing facilitator, a residual-value guarantor, a capacity backstop, an AI-cloud customer, a model investor and a major purchaser of future semiconductor supply, then Nvidia becomes more than a successful corporation. It becomes a systemically important institution within the AI economy. That does not imply imminent instability. It does mean that Nvidia’s financial health, technological roadmap and strategic decisions increasingly matter to counterparties far beyond Nvidia shareholders, and the institutions charged with financial stability have begun to say so in their own vocabulary.
The Bank of England’s Financial Policy Committee judged in July 2026 that AI-related equity valuations rested on expectations that depended on the successful build-out of infrastructure, continued access to financing and the pace of adoption, that these assumptions were uncertain, and that any fall in AI equity prices would be amplified by high index concentration, momentum-driven positioning and increasing leverage [34].
“could trigger sharp adjustments in asset prices”
— Bank of England Financial Policy Committee, Financial Stability Report, July 2026 [34]
The IMF’s July 2026 World Economic Outlook Update identified a correction in AI-linked valuations as a key downside risk and warned that a repricing could be amplified by risk sensitivity among AI-exposed investors and would weigh on private investment [32]. The Federal Reserve Bank of Chicago found in February 2026 that large banks’ direct exposure to AI-adjacent industries averaged only 0.8 percent of assets, but cautioned that indirect exposure through lending to funds and other channels was most likely larger, and noted that 45 percent of institutional investors in a Bank of America survey identified an AI bubble as their top tail risk [36]. And in August 2026, three regional Federal Reserve presidents spoke publicly about the question. New York’s John Williams said he did not see a bubble-like situation, because AI-related borrowing was being handled by companies with strong earnings; San Francisco’s Mary Daly said the Fed was building a forward-looking risk dashboard; and Kansas City’s Jeff Schmid asked the question that this paper’s title implies [37].
“too big to fail”
— Jeff Schmid, President, Federal Reserve Bank of Kansas City, on the web of leveraged commitments linking data centers, energy providers and communities [37]
A systemically important institution in a banking system is designated as such and subjected to enhanced supervision, capital requirements, resolution planning and stress testing. Nothing of the kind applies to Nvidia. The Bank of England’s April 2026 record noted that valuations remained particularly stretched for U.S. AI-focused technology companies and that debt-financing needs and doubts about returns had already produced selling pressure before the Middle East conflict [80]. If the FPC and the IMF are correct that the AI ecosystem’s interconnectedness with the broader financial system is rising, then the central institution of that ecosystem is, for the first time, a company whose systemic importance has been recognized by regulators before it has been recognized by regulation. Table 7 consolidates the risk assessment.
Table 7. Risk Matrix for the Central Bank of AI
| Risk | Mechanism | Who bears it first | Amplifier | Institutional warning (2025–2026) |
| Circular financing | Nvidia capital returns to Nvidia as chip purchases; independence of final demand obscured | Nvidia shareholders; lenders relying on demand signals | Headline frameworks far larger than signed contracts ($100B → $30B; $250B → $105B) | Bernstein; CNN/Slok; Vanderbilt [25][27][47] |
| Antitrust / control | Credit support conditioned on customer allocation; vertical integration of supply and finance | Neoclouds; competitors; Nvidia (litigation) | Internal employee concerns; prior DOJ subpoenas | WSJ/Reuters; Allensworth; Picker [9][64][65] |
| Residual-value | Guaranties and rent-backs assume durable GPU value; product cycle accelerates depreciation | Nvidia (guarantor); GPU-backed lenders | Nvidia controls the roadmap that depreciates its own collateral | Burry; Nadella; BoE FPC [35][49][50] |
| Supply-commitment | $279B forward purchases become excess if demand slows | Nvidia; memory makers; TSMC | Concentration of commitments in ~2 years; margin compression already visible | Nvidia 10-Q; Seoul Economic Daily [12][14] |
| Moral hazard | Backstops encourage overbuilding; risk relocated to one balance sheet | Whole ecosystem; ultimately taxpayers if bailout sought | Undisclosed floors, shares and minimum values | Warren/FSOC letter; Capacity [18][39] |
| Systemic concentration | One firm is supplier, financier, guarantor, customer and investor | Financial system via banks, insurers, private credit, index concentration | Leverage in equity markets; momentum positioning | BoE FSR; IMF WEO; Chicago Fed; Schmid [32][34][36][37] |
Source: Author’s analysis.
Section 5: The Central Bank of AI in 2027–2030
The period from 2027 to 2030 is the one in which the thesis of this paper will be tested, and this section attempts to describe what that test will look like. It begins with the innovation that is most likely to occur regardless of the cycle’s outcome—the emergence of a mature market for compute finance—and then examines the conditions under which compute can become collateral, the strategic contest between Nvidia and the hyperscalers that are simultaneously its customers and its rivals, the specific challenge of custom silicon, the geopolitical constraint of China and export controls, the political economy of a backstop that will increasingly intersect with public infrastructure, and finally the scenario that matters most: the first AI capacity recession. The section does not predict which scenario will occur. It argues that the Central Bank of AI will be revealed as either a metaphor or a description depending on how Nvidia behaves when the scenario arrives.
5.1 The Emergence of Compute Finance
The next major financial innovation surrounding AI is likely to be the creation of a mature market for financing compute, and the building blocks are already visible. CoreWeave’s March 2026 facility demonstrated that a GPU-backed, non-recourse, delayed-draw term loan could achieve an investment-grade rating when anchored by a hyperscaler take-or-pay contract [52]. Meta’s Hyperion SPV demonstrated that a hyperscaler could move $30 billion of datacenter financing off its own balance sheet [54]. Nvidia’s August 10 platforms are designed to standardize the asset around which such structures form [3]. Reuters reported that Bank of America analysts expected chip-funding programs to raise debt at a pace similar to or faster than the hyperscalers themselves, estimating that Broadcom’s parallel chip-financing vehicle could reach $370 billion of senior debt by mid-2029 to fund 20 gigawatts of compute [20]. Table 8 lists the instruments that a mature compute-finance market would contain, together with their current status.
Table 8. Instruments of a Mature Compute-Finance Market: Status as of August 2026
| Instrument | Function | Status in 2026 | Example / evidence |
| GPU-backed loans | Debt secured by accelerators and customer contracts | Established; first investment-grade rating achieved | CoreWeave $8.5B DDTL 4.0, A3/A(low) [52] |
| Compute leases and take-or-pay contracts | Long-duration offtake that lenders can underwrite | Established at hyperscaler scale; extended to neoclouds by Nvidia floors | Meta–CoreWeave $19B; Nvidia AI Compute Partnership [11][52] |
| Datacenter asset-backed securities / SPVs | Off-balance-sheet project finance | Established | Meta Hyperion $30B SPV [54] |
| Residual-value guarantees | Third-party promise about asset value at lease end | New at scale | Nvidia–SB Energy $105B cap [5] |
| Vendor rent-back / demand backstop | Buyer of last resort for unused capacity | New; $36B committed | AI Compute Partnership [11][16] |
| Compute financing platforms | Dedicated institutional capital pools with vendor backstop | MOU stage; >$500B target | Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR [3] |
| Inference-revenue bonds; capacity futures; depreciation hedges; residual-value insurance | Price and transfer utilization and obsolescence risk | Not yet observed at scale | Author’s projection |
| Sovereign compute funds | National capacity as strategic reserve | Emerging | EU €30B gigafactory tender; sovereign share of Nvidia Data Center revenue [29][79] |
Source: Author’s compilation.
Nvidia does not need to become a bank for this market to emerge. It merely needs to help standardize the asset around which the market forms, and that is precisely what the language of the August 10 announcement—broadly adopted, fungible, transferable, continuously improved—was intended to do [3]. In this respect, Nvidia’s role resembles that of the institutions that standardized the residential mortgage into a securitizable instrument in the second half of the twentieth century, an analogy Blackstone’s president drew explicitly.
“financeable asset class”
— Jon Gray, President and COO, Blackstone, on CNBC, comparing AI compute to mortgage-lender assessment of homes [69]
The analogy is illuminating and should also be sobering, because the history of standardized mortgage finance is a history of both extraordinary capital formation and, in 2008, of a systemic crisis produced when the standardized asset turned out to be more heterogeneous than its rating implied.
5.2 Compute as Collateral
The crucial development will occur if lenders increasingly treat accelerator clusters as financeable collateral in their own right rather than as the recovery floor beneath a hyperscaler contract. Traditional collateral has observable resale markets and reasonably predictable depreciation schedules. GPUs present more complicated characteristics: rapid technological obsolescence, extraordinary current scarcity, high secondary-market value that may not persist, export restrictions that limit where the asset can be redeployed, geographic constraints imposed by power availability, software dependencies that tie value to a particular ecosystem, and enormous differences between generations. The rating of CoreWeave’s facility was explicitly a rating of the Meta contract with the hardware as recovery floor; one analysis of the transaction noted that rating agencies were underwriting the hyperscaler customer, not the GPU, and that the correct framework was project finance in which the contract is the primary credit instrument [52]. If financial markets learn to price the risks of the hardware itself, compute could become a major institutional asset class. If they do not, compute finance will remain a derivative of hyperscaler and frontier-laboratory credit, and the Central Bank of AI will be backstopping not an asset class but a handful of counterparties.
5.3 Nvidia Versus the Hyperscalers
Nvidia’s expanding financial role creates a strategic tension with Amazon, Microsoft, Google and Meta. The hyperscalers are simultaneously Nvidia customers, Nvidia infrastructure partners, AI-cloud competitors, custom-chip designers and potential alternatives to Nvidia-centered financing ecosystems. Their spending remains the largest single force in the system: the four companies planned roughly $725 billion of capital expenditure in 2026, up about 77 percent from roughly $410 billion in 2025, with Amazon raising its guidance to $220 billion, Alphabet to $195–205 billion and Meta to $130–145 billion, and analysts projecting more than $1 trillion in 2027 [30]. CNBC reported in February 2026 that reaching those numbers would mean a large drop in free cash flow, with Amazon’s projected to turn negative for the year [31]. Table 9 sets out the trajectory.
Table 9. Hyperscaler Capital Expenditure, 2024–2027
| Company | 2025 (approx.) | 2026 guidance (latest) | Notes | Source |
| Amazon | ~$125B | $220B (raised from $200B, July 30) | AWS +37% y/y in Q2 2026; capex $54.2B in the quarter; capacity expected to trail demand through 2027 | [30] |
| Alphabet | ~$91B | $195–205B (raised twice) | Google Cloud backlog ~$460B, roughly double prior year | [30] |
| Microsoft | ~$88B (FY) | ~$190B (FY basis) | Azure growth ~31–39% cc in H1 2026 | [30] |
| Meta | ~$72B | $130–145B (raised twice) | Higher component pricing and additional capacity | [30] |
| Four combined | ~$410B | ~$725B (+77%) | Analysts project >$1 trillion in 2027; Moody’s: $785B in 2026 | [30][53] |
Source: Company guidance as compiled in the cited reporting; 2025 company-level figures are approximate and drawn from the same compilations.
Nvidia’s own disclosure that hyperscalers account for only about half of its Data Center revenue [29] reveals the strategic logic of the Central Bank of AI with unusual clarity. The other half—AI clouds, sovereigns, enterprises, industrial customers—is the half that cannot self-finance, and it is the half whose growth Nvidia’s backstops are designed to unlock. By financing the non-hyperscaler half of the market, Nvidia reduces its dependence on the four companies most capable of replacing it. AWS Trainium, Google TPU and Meta’s internal silicon can therefore be understood partly as efforts to prevent Nvidia from controlling too much of the economics beneath AI, and Nvidia’s financing platforms can be understood partly as an effort to build a customer base that the hyperscalers do not control. The competition is no longer GPU versus GPU. It is increasingly AI industrial system versus AI industrial system.
5.4 Nvidia Versus Custom Silicon
As Google, Amazon, Meta, Microsoft, OpenAI and others develop custom accelerators, Nvidia’s financial architecture may become one of its strongest strategic defenses. The custom-silicon challenge is real and growing. Tom’s Hardware’s May 2026 survey reported that Nvidia still held roughly 70 percent of the AI chip market but that ASIC-based server shipments were projected to reach 27.8 percent of the market in 2026, growing 44.6 percent year over year against 16.1 percent for merchant GPUs; that Anthropic had committed to up to one million Google TPUs in the largest deal in Google Cloud history and then expanded the arrangement with Google and Broadcom to multiple additional gigawatts from 2027; that AWS had deployed more than one million Trainium processors and confirmed a 2-gigawatt Trainium deal with OpenAI; and that Meta had entered talks for multi-billion-dollar TPU deployments [60]. Anthropic explicitly pursues a tri-platform strategy across TPU, Trainium and Nvidia GPUs, and its April 2026 expansion with Google and Broadcom secured up to 5 gigawatts of next-generation TPU capacity starting in 2027 [61]. Table 10 summarizes the landscape.
Table 10. Custom Silicon Versus the Nvidia Financial-Industrial System, Mid-2026
| Program | Owner / partner | Scale disclosed (2025–2026) | Availability to third parties | Financing ecosystem | Source |
| TPU (Ironwood, 7th gen) | Google / Broadcom | Up to 1M TPUs and >1 GW for Anthropic (2026); up to 5 GW from 2027; Meta talks | Google Cloud only | Google balance sheet; Google credit backstops (e.g., Fluidstack) | [54][60][61] |
| Trainium 3 / 4 | AWS (Annapurna) | >1M Trainium deployed; Project Rainier ~500k Trainium2; 2 GW OpenAI deal; Anthropic up to 5 GW | AWS only | Amazon balance sheet ($220B capex) | [30][60] |
| MTIA | Meta | Four generations in two years announced | Captive | Hyperion SPV; Blue Owl-style structures | [54][60] |
| Maia 200 | Microsoft | 140B transistors; inference for Copilot/Azure | Azure only | Microsoft balance sheet | [60] |
| OpenAI custom chip | OpenAI / Broadcom | Limited deployment late 2026; broader 2027 | Captive | OpenAI fundraising; Nvidia/SoftBank/Amazon equity | [60][62] |
| Nvidia GPU (Blackwell / Vera Rubin) | Nvidia | $89B Data Center revenue in one quarter; ~70% chip share; 90%+ of training | Every cloud; every neocloud; on-premises | AI Compute Partnership; $500B platforms; residual-value guaranties; $279B supply commitments | [1][3][12][60] |
Source: Reporting as cited. Shares and volumes are as reported by the cited sources and vary by methodology.
The table exposes the asymmetry that underlies Nvidia’s strategy. Competitors can design chips, and several have designed very good ones. But custom silicon is captive: TPUs exist only inside Google Cloud, Trainium only inside AWS, MTIA only inside Meta. Reproducing CUDA, networking, systems engineering, supplier relationships, cloud distribution, institutional financing and residual-value support is considerably harder than designing an accelerator, and none of the custom programs has attempted to reproduce the financial layer at all. Nvidia’s moat could therefore migrate from best accelerator toward lowest-friction financial-industrial ecosystem for deploying accelerated compute. That would represent an extraordinary evolution in competitive strategy, and it is the one that CNBC identified when it wrote that Nvidia’s moat was shifting from chips to capital [21]. It is also the strategy that the AI Compute Partnership pursues most directly, by placing Nvidia’s balance sheet behind the neoclouds and AI-native startups that custom silicon cannot reach.
5.5 China and the Limits of the Central Bank
No analysis of Nvidia can ignore geopolitics, and the 2026 record shows a central institution operating with a large part of the world excluded from its jurisdiction. U.S. export controls have restricted Nvidia’s ability to treat the global AI market as one integrated system since 2022, and each control prompted a compliant chip that in turn prompted a new control. In January 2026, the U.S. approved H200 sales to roughly ten Chinese companies under an arrangement in which advanced chips bound for China route through the United States for testing and the U.S. government receives 25 percent of sale revenue; Beijing, however, initially restricted purchases while promoting domestic alternatives, and a Commerce Department official told Congress in July that shipments remained very few [59]. Nvidia halted production of China-configured H200 units in March and redirected the freed TSMC capacity to Vera Rubin [58]. By the second quarter of fiscal 2027 the company had shipped its first H200s to China, accounting for less than 1 percent of Data Center revenue, and it still assumed no China Data Center compute revenue in its $108 billion outlook [58][79]. Bernstein has projected that Nvidia’s share of China’s AI chip market could fall to about 8 percent by the end of 2026, with Huawei rising toward 50 percent [59].
A Central Bank of AI operating under American export law therefore faces a fundamental contradiction: global economics encourage universal distribution; national-security policy requires selective exclusion. The Next Web observed that a single supplier could write the world’s second-largest economy out of its forecast and still guide revenue up by roughly a fifth [79], which is a remarkable statement about demand elsewhere and an equally remarkable statement about the fragmentation of the compute economy into blocs. Nvidia’s future growth will consequently depend not only on technology and finance but on Washington’s decisions about which countries, companies and intermediaries may receive advanced compute, and on Beijing’s decisions about whether to accept it. The monetary analogy holds here too: a reserve currency whose issuer restricts its circulation to allies invites the creation of alternative reserves, and Huawei’s rising share is the compute-economy equivalent.
5.6 The Political Economy of the Backstop
If Nvidia-backed projects eventually require tens or hundreds of gigawatts, the company’s financial decisions will increasingly intersect with public infrastructure, and the Ohio campus is the first case in which that intersection is fully documented. The generation plant behind PORTS-Pike is owned by the U.S. government and financed by Japan under a trade agreement; the transmission is being built by a regulated utility; the campus sits partly on federal land leased from the Department of Energy; and power, as one report noted, is the part local residents will feel [7]. Governors and state regulators will ask who pays for transmission, who finances new generation, who bears stranded-asset risk, who pays if a datacenter project is canceled, whether households should subsidize grid upgrades, whether AI companies should supply their own generation, whether large loads should be required to curtail, how tax incentives should be structured, how water consumption should be treated, and what happens if guaranteed AI capacity becomes uneconomic. The IEA notes that supply chains for gas turbines and transformers have tightened alongside those for chips, and that onsite battery storage is becoming critical for AI datacenters whose demand swings can stretch onsite generation [55]. Thus the Central Bank of AI eventually encounters actual governments—not as a metaphor but as counterparties, landlords, regulators and, in the Ohio case, co-financiers.
5.7 What Happens During the First AI Capacity Recession?
The strongest test of this entire thesis will not occur during a boom. It will occur during a downturn. Imagine a future period in which model revenues disappoint; inference prices collapse under competition among open and closed models; enterprises slow AI adoption; a major AI laboratory fails or is forced to shrink its compute commitments; custom silicon reduces Nvidia demand from the hyperscalers; electricity becomes more expensive; datacenter utilization falls; or capital markets stop financing speculative capacity. Several of these conditions have already been rehearsed in miniature. The IMF’s baseline already assumes the AI-driven technology cycle moderates with no exogenous productivity boost [33]. OpenAI’s own chief financial officer was reported in April 2026 to have questioned whether the company’s revenue growth would support its spending commitments [73]. CoreWeave’s loan spreads widened 125 basis points in a single episode and lenders demanded covenants [53]. Nvidia’s own supporters concede the wrong-way structure of its exposures [66].
In that scenario, the questions become concrete. Who absorbs the unused GPUs? Under the AI Compute Partnership, Nvidia has contractually agreed to rent back a portion of them at a fixed rate [16]. Who restructures the debt? The residual-value guaranties give Nvidia the right to assume leases, relet, sell or defer remedies for up to a year [8], which is to say that Nvidia has already written itself the role of workout agent for the Ohio campus. Who purchases distressed datacenters? The financing platforms’ backstop option places Nvidia in the first-loss position on chips pledged as collateral [66]. Who protects suppliers? Nvidia’s $279 billion of commitments already do, at Nvidia’s expense. Who determines which AI clouds survive? Whoever controls the allocation of scarce hardware and the terms of credit support, which is Nvidia. If Nvidia becomes deeply enough interconnected with those institutions, market participants may expect it to intervene—and the Warren letter’s warning that some executives were already laying the groundwork for public support suggests that at least some participants expect intervention from an institution larger than Nvidia [39].
At that moment, Central Bank of AI would stop being merely a metaphor for influence. It would become a description of market expectations, and the decisive question would be whether Nvidia was willing—and financially able—to meet them. A public central bank meets such expectations by creating reserves. Nvidia can meet them only by spending cash flows that will, in exactly that scenario, be falling. That asymmetry is the single most important reason this paper’s title should be read as a warning as much as a description.
Section 6: What Have We Learned? Seven Pillars
The preceding sections have moved from anecdote to framework to function to risk to forecast. This section consolidates what that journey teaches into seven pillars, each of which is both a conclusion drawn from the evidence of 2025–2026 and a proposition that the events of 2027–2030 will confirm or refute. The pillars are ordered deliberately, from the most descriptive to the most fundamental. The first three describe what has changed in the economics of compute and in Nvidia’s role; the fourth and fifth describe the paradoxes that change creates; the sixth describes the institutional gap that the change exposes; and the seventh identifies the only reserve that can ultimately back the entire structure. Together they constitute the analytical core of the Central Bank of AI thesis.
Pillar 1 — Compute Has Become a Financial Asset, Not Merely a Technology Product
The first lesson is that accelerated computing has crossed an economic threshold. A GPU was once primarily a semiconductor product, purchased from operating budgets and depreciated quietly. At AI-factory scale, compute becomes a long-duration capital asset surrounded by financing, leases, guarantees, insurance, power contracts and utilization risk. The evidence is now institutional rather than anecdotal: an investment-grade rating on GPU-backed debt [52], six of the world’s largest allocators declaring compute an investable asset class [3], a residual-value guaranty structure borrowed from automobile leasing and applied to a 4.25-gigawatt campus [5][7], and a Bank of England financial-stability framework built specifically to monitor the macrofinancial implications of AI [34]. The financial system is therefore becoming an additional layer surrounding the Five-Layer AI Economy. One might describe it as an invisible connective tissue rather than a sixth layer: capital finances Energy; Energy powers Chips; Chips populate Datacenters; Datacenters run Models; Models enable Applications and Agents. Without financing, the physical chain cannot expand, and the institution that most shapes the financing of the chain is the institution this paper has been describing.
“we are helping create a new class of productive, investable infrastructure”
— Jensen Huang, Founder and CEO, NVIDIA, announcing the compute financing platforms [3]
Pillar 2 — Nvidia Is Moving From Supplier to System Stabilizer
The second lesson is that Nvidia’s strategic role has changed in kind and not merely in degree. Its competitive advantage increasingly comes not merely from designing powerful accelerators but from making the surrounding ecosystem easier to finance and operate. Financing platforms, guarantees, investments, supplier commitments and capacity backstops all point in the same direction, and Table 6 showed that they all carry the same directional exposure. Nvidia increasingly has an economic interest in ensuring that customers can continue purchasing, financing and utilizing Nvidia infrastructure, and it has begun to act on that interest with instruments that were, until 2025, foreign to the semiconductor industry. That is the foundation of the Central Bank of AI thesis. It is also, importantly, a choice rather than an inevitability: Nvidia could have remained a pure vendor and allowed the financing problem to constrain its addressable market. It chose instead to expand the market by underwriting it, and Huang has been explicit that the company regrets only not having done so sooner [23].
Pillar 3 — The Half of the Market That Cannot Self-Finance Is the Half Nvidia Is Financing
The third lesson is quantitative and is easy to miss. Nvidia disclosed that hyperscalers account for roughly half of its Data Center revenue and that the other half comes from AI clouds, industrial, enterprise and sovereign customers [29]. The hyperscalers have balance sheets capable of funding $725 billion of annual capital expenditure [30], and they are simultaneously building the custom silicon that most threatens Nvidia’s share [60]. The non-hyperscaler half has neither the balance sheets nor the alternative silicon. Every financial instrument Nvidia has introduced since July 2026—the take-or-pay floors, the rent-backs, the financing platforms, the residual-value guaranties—is aimed at that half. The Central Bank of AI is therefore not a general-purpose stabilizer of the AI economy. It is a targeted intervention designed to grow the segment of demand that is least able to finance itself and least able to defect to a competitor. That is sound strategy. It is also the segment in which credit risk is highest, which is why the intervention required a balance sheet of Nvidia’s size to attempt.
Pillar 4 — Backstopping Creates Both Stability and Moral Hazard
The fourth lesson is that stabilization is never free. A credible backstop encourages investment because lenders and developers perceive less downside risk; Capacity Media documented that Nvidia’s guaranty unlocked capital that was not available on comparable terms six months earlier [18]. But backstops also encourage excessive construction if participants believe another institution will absorb losses, and the AI infrastructure economy could therefore encounter a form of moral hazard that the banking system has spent a century learning to manage. The more Nvidia protects its ecosystem from failure, the more aggressively the ecosystem may expand, and the larger the eventual claim on the protector. This makes transparency around guarantees, capacity commitments, related-party investments and utilization increasingly important, and it makes the undisclosed elements of Nvidia’s arrangements—the guaranteed minimum values, the revenue-share percentages, the final terms of the platforms—the most important numbers in the AI economy that nobody outside Nvidia can see.
Pillar 5 — Nvidia’s Greatest Risk May Come From the Success of Its Own Ecosystem
The fifth lesson appears paradoxical. The larger Nvidia becomes, the more capital the ecosystem requires. The more capital it requires, the more Nvidia may become involved in helping mobilize that capital. The more financially involved Nvidia becomes, the more exposure it acquires to customer performance and infrastructure economics. Success therefore produces interconnectedness, and interconnectedness produces systemic responsibility. Nvidia may eventually discover that dominating the AI accelerator market means inheriting some responsibility for maintaining the economic system that dominance created. The Federal Reserve’s regional presidents have already begun to describe that responsibility in the language of systemic importance [37], and the IMF has already identified a reassessment of AI profitability as a key downside risk to the global economy [32][33].
“could correct sharply”
— International Monetary Fund, World Economic Outlook Update, July 2026, on frothy AI-linked equity valuations [32]
Huang’s own assessment, offered to CNBC on the day of the August earnings, is the counterpoint that the next several years will test.
“the risk is low”
— Jensen Huang, Founder and CEO, NVIDIA, on Nvidia’s financial support for the AI ecosystem [22]
Pillar 6 — A Systemically Important Institution Has Emerged Before the Institutions That Govern It
The sixth lesson concerns governance. Public central banks are the product of crises: the Federal Reserve of 1907, the modern lender-of-last-resort doctrine of the 1930s, the post-2008 architecture of designation, stress testing and resolution planning. Each was built after a private institution or a private market had been discovered to be performing a systemic function without the mandate, the transparency or the resources to perform it safely. The evidence assembled in this paper suggests that the AI economy has reached the stage that precedes such institution-building. A private company is supplying liquidity, enhancing credit, backstopping capacity, transmitting its decisions across sovereign economies, allocating scarce inputs and anchoring confidence. Regulators have noticed: the Bank of England has published a monitoring framework [34], the Federal Reserve is building a risk dashboard [37], the IMF has flagged the risk in two consecutive outlooks [32], the Chicago Fed has measured bank exposure [36], senators have petitioned the Financial Stability Oversight Council [39], and a Vanderbilt paper has proposed a Glass-Steagall for AI [47]. What has not happened is any formal recognition that the central institution of the AI-capacity economy is a corporation, or any framework for what its obligations and constraints should be. The gap between the systemic function and the governance of that function is, in the author’s view, the most consequential finding of this paper.
Pillar 7 — The Ultimate Reserve Behind the Central Bank of AI Is AI Productivity
The seventh lesson is the most important. No financing mechanism can permanently compensate for insufficient economic value. Not guarantees. Not private credit. Not hyperscaler capex. Not GPU scarcity. Not Nvidia’s balance sheet. The entire infrastructure boom ultimately depends on whether artificial intelligence produces enough economic output to justify the capital invested in it. The ultimate reserve behind the Central Bank of AI is therefore neither dollars nor GPUs. It is productivity.
The debate over that reserve is now the most serious debate in economics, and this paper has tried to present it fairly. Brynjolfsson sees the J-curve turning and productivity growth near 2.7 percent in 2025 [45]; the Federal Reserve’s own economists see AI-related investment contributing meaningfully to GDP [38]; Acemoglu sees a decade of gains measured in fractions of a percentage point and a discourse that has lost its seriousness [43][44]; Damodaran sees a $250 billion market being asked to justify trillions of capital [41]. If AI agents, models, robotics and applications generate enormous economic value, today’s infrastructure commitments—Nvidia’s $279 billion of supply, its $105 billion of guaranties, its $500 billion of platforms—may appear conservative in hindsight, and the Central Bank of AI will be remembered as the institution that financed the electrification of intelligence. If they do not, the financial architecture constructed around AI capacity will eventually be forced to recognize the difference between compute demand and economically productive compute demand, and the institution at its center will discover what every central bank discovers in a crisis: that a backstop is only as strong as the economy it is backstopping.
Table 11. The Seven Pillars and the Evidence That Will Test Them, 2027–2030
| Pillar | Claim | What would confirm it | What would refute it |
| 1. Compute as financial asset | GPUs are now long-duration capital assets with a financial layer around them | Growth of rated GPU-backed debt; secondary markets; residual-value insurance | Impairments; failed refinancings; return to pure vendor sales |
| 2. Supplier to stabilizer | Nvidia’s role has changed in kind | Platforms move from MOU to funded vehicles; Partnership commitments grow beyond $36B | Program wound down; guaranties not extended beyond Ohio |
| 3. Financing the unfinanceable half | Non-hyperscaler demand is the target of every backstop | ACIE share of Data Center revenue rises above 50% | Neocloud consolidation; hyperscaler share rises |
| 4. Stability and moral hazard | Backstops encourage overbuilding | Utilization falls while construction continues | Utilization stays high; backstops never drawn |
| 5. Success creates systemic risk | Interconnectedness produces responsibility | Nvidia CDS and equity react to customer credit events | Exposures remain contingent and undrawn |
| 6. Governance gap | Systemic function precedes governance | Designation, disclosure or structural-separation proposals advance | Regulators conclude exposures are contained |
| 7. Productivity as reserve | Only Layer Five revenue can back the structure | AI revenue approaches multiples of $250B; productivity data confirm J-curve | Revenue stalls; Acemoglu-scale gains; write-downs across the stack |
Source: Author’s analysis.
Conclusion: Why “Central Bank of AI” Fits the Emerging Nvidia Economy
Nvidia entered the artificial-intelligence revolution as the company that happened to manufacture the processor most suited to its computational requirements. It then became the dominant supplier of AI accelerators. From there it expanded into networking, systems, software and complete AI-factory architectures, and by fiscal 2026 its Data Center business alone generated $193.7 billion of revenue on a base that had been $15 billion three years earlier [75]. The next transformation, documented in this paper, may prove even more consequential than any of those, because it changes not what Nvidia sells but what Nvidia is.
Nvidia is increasingly becoming involved in the financial architecture that determines whether AI infrastructure gets built at all. Its actions between July 1 and August 27, 2026 illustrate the transition unusually clearly. It launched a business model under which it provides credit support and take-or-pay floors to AI clouds in exchange for a share of their revenue, and committed $36 billion to it within weeks [11][15]. It partnered with six of the world’s largest financial institutions in an effort to mobilize more than $500 billion for compute infrastructure, retaining an option to backstop a quarter of it [3][4]. It disclosed residual-value guaranties capped at $105 billion behind a 4.25-gigawatt campus in Ohio where OpenAI will be the tenant, with an option on nearly four gigawatts more [5]. It contemplated renting capacity that customers could not place elsewhere [16]. It more than doubled its commitments to suppliers to $279 billion so that future generations of AI infrastructure would have enough memory and manufacturing capacity [12]. At the same time, quarterly revenue reached $96.2 billion, with $89.0 billion coming from Data Center products alone, and the company guided to $108 billion for the following quarter while assuming nothing from China [1].
Reuters’ August 27 report that Nvidia paused some of the revenue-sharing arrangements does not weaken the central thesis. In some ways, it makes the thesis more important. The reported concerns—possible antitrust scrutiny, questions about control over how customers do business, discomfort among partners about which customers they could serve—demonstrate precisely how far Nvidia has moved beyond the conventional boundaries of semiconductor manufacturing [9][10]. A chip company does not need to worry about whether it is dictating its customers’ customer lists. A central institution does. Nvidia itself said the broader model for expanding compute access remains in place and continues to evolve, and its chief financial officer described its purpose to analysts in the language of project finance rather than of product sales [9][11].
That evolution explains why I chose Central Bank of AI as the title of this paper. The phrase does not mean Nvidia controls interest rates. It does not mean Jensen Huang is a central banker. It does not mean Nvidia possesses government authority or can create sovereign money; indeed, the absence of that last power is the source of the deepest risk this paper identifies. The metaphor describes something more specific. A central institution becomes important when other participants begin organizing their decisions around its willingness and ability to supply liquidity, support markets and maintain confidence. Increasingly, AI companies organize their infrastructure around Nvidia architectures. Datacenters are designed around Nvidia systems. Financiers evaluate projects built around Nvidia compute, and the largest of them have now agreed to build dedicated vehicles for it. Suppliers expand capacity in anticipation of Nvidia orders, and their profits rise by hundreds of percent when those orders arrive [14]. AI laboratories negotiate for access to Nvidia clusters and accept Nvidia as investor, guarantor and landlord’s guarantor simultaneously. Governments consider access to Nvidia accelerators a strategic national resource and restrict or tax their movement accordingly [59]. And Nvidia itself is becoming increasingly willing to use its extraordinary financial position to keep that ecosystem expanding.
The transformation can therefore be summarized in five stages: Nvidia the Chipmaker; Nvidia the Platform; Nvidia the Infrastructure Supplier; Nvidia the Capital Mobilizer; and Nvidia the Central Bank of AI. The final stage changes the nature of the company. Once Nvidia helps finance the ecosystem, guarantees portions of its infrastructure, secures its future supply, invests in its customers and stands ready to absorb unused capacity, it acquires something that market dominance alone does not create: systemic responsibility. That responsibility will become increasingly important between 2027 and 2030. During continued expansion, Nvidia’s balance sheet can accelerate AI infrastructure construction. During scarcity, its allocations can determine who receives compute. During financial uncertainty, its guarantees can strengthen confidence. During supply shortages, its commitments can pull enormous quantities of semiconductor capacity toward AI. And during a future downturn, markets may discover whether Nvidia is willing—or financially able—to stand behind the ecosystem it helped create.
That will be the decisive test of the Central Bank of AI thesis, and it is a test that public central banks pass by virtue of a power Nvidia does not have. A central bank can always create the reserve it is asked to supply. Nvidia can supply its backstops only from the cash flows of a business that will, in the scenario in which the backstops are needed, be contracting. Its balance sheet is long the very economy it stabilizes. The regulators who have begun to describe the AI ecosystem in the vocabulary of financial stability—the IMF, the Bank of England, the Federal Reserve system, the Chicago Fed, the senators petitioning FSOC, the scholars at Vanderbilt, MIT, Stanford and NYU whose work this paper has drawn upon—are not describing a bubble in the crude sense. They are describing a system that has acquired a central institution before it has acquired the rules that such institutions require. Whether those rules are written before or after the first AI capacity recession is a question this paper cannot answer. It can only insist that the question is the right one.
Because the most consequential question surrounding Nvidia is no longer simply how many GPUs Nvidia can sell. The larger question is becoming how much of the artificial-intelligence economy now depends on Nvidia making sure that everybody else can keep buying, financing, powering and using them. When the answer becomes large enough—and the evidence of August 2026 suggests it already has—Nvidia is no longer merely selling the machinery of artificial intelligence. It is helping stabilize the monetary-equivalent asset at the center of the new industrial system: compute itself. And the reserve that ultimately stands behind that asset is not on Nvidia’s balance sheet, or on any balance sheet. It is the productivity that artificial intelligence has yet to prove it can deliver.
That is why Central Bank of AI is not simply a provocative title. It is a framework for understanding the next stage of the Five-Layer AI Economy—and a warning about what the stage after that may require.
Footnotes / Endnotes:
[1] NVIDIA Corporation (Investor Relations). “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027.” NVIDIA Newsroom / GlobeNewswire, August 26, 2026. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/default.aspx
[2] Kif Leswing et al. (CNBC). “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
[3] NVIDIA Corporation (Investor Relations). “NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital.” NVIDIA Newsroom / GlobeNewswire, August 10, 2026. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Partners-With-Apollo-BlackRock-Blackstone-Brookfield-Goldman-Sachs-and-KKR-to-Establish-AI-Compute-Infrastructure-Financing-Platforms-to-Mobilize-Over-500-Billion-of-Third-Party-Capital/default.aspx
[4] Reuters (via Yahoo Finance). “Nvidia partners with Wall Street giants to raise $500 billion for AI buildout.” Reuters, August 10, 2026. https://finance.yahoo.com/technology/ai/articles/wall-street-giants-partner-nvidia-163441421.html
[5] NVIDIA Corporation. “Form 8-K, Item 1.01: Residual value guaranties relating to the PORTS Technology Campus (Portsmouth Site), Pike County, Ohio.” U.S. Securities and Exchange Commission (EDGAR), August 17, 2026. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm
[6] Data Center Knowledge. “Nvidia Backs OpenAI’s Ohio Data Center Buildout With $105B Guarantee.” Data Center Knowledge, August 2026. https://www.datacenterknowledge.com/data-center-construction/nvidia-backs-openai-s-ohio-data-center-buildout-with-105b-guarantee
[7] The Next Web. “Nvidia’s $105bn Ohio guarantee covers buildings, not OpenAI.” TNW, August 2026. https://thenextweb.com/news/nvidia-openai-ports-pike-105bn-residual-value-guarantee
[8] Global Data Center Hub. “NVIDIA Did Not Guarantee The Rent. It Guaranteed The Resale.” Global Data Center Hub, August 2026. https://www.globaldatacenterhub.com/p/nvidia-did-not-guarantee-the-rent
[9] Reuters (via U.S. News & World Report). “Nvidia Pauses Revenue-Sharing Deals With AI Cloud Companies, WSJ Reports.” Reuters, August 27, 2026. https://money.usnews.com/investing/news/articles/2026-08-27/nvidia-pauses-revenue-sharing-deals-with-ai-cloud-companies-wsj-reports
[10] Investing.com. “Nvidia pauses revenue-sharing deals with AI cloud firms – WSJ.” Investing.com, August 27, 2026. https://www.investing.com/news/stock-market-news/nvidia-pauses-revenuesharing-deals-with-ai-cloud-firms-wsj-4880348
[11] Anton Shilov (Tom’s Hardware). “Nvidia denies pausing AI cloud commitments initiative after reported partner backlash.” Tom’s Hardware, August 28, 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-denies-pausing-ai-cloud-commitments-initiative-after-reported-partner-backlash-report-claims-company-told-cloud-providers-it-could-only-lease-its-gpus-to-nvidia-approved-customers
[12] NVIDIA Corporation. “Form 10-Q for the quarterly period ended July 26, 2026 (Supply and capacity; AI Compute Partnership commitments).” U.S. Securities and Exchange Commission (EDGAR), August 2026. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000075/nvda-20260726.htm
[13] Daniel Sparks (The Motley Fool). “Nvidia’s Supply Commitments More Than Doubled to $279 Billion. Here’s What It Does to the Stock.” The Motley Fool, August 27, 2026. https://www.fool.com/investing/2026/08/27/nvidia-s-supply-commitments-more-than-doubled-to-usd279-billion-here-s-what-it-does-to-the-stock/
[14] Seoul Economic Daily. “Nvidia Sees Memory Bottleneck Lasting Into 2028.” Seoul Economic Daily, August 27, 2026. https://en.sedaily.com/finance/2026/08/27/nvidia-sees-memory-bottleneck-lasting-into-2028
[15] 24/7 Wall St. (via Yahoo Finance). “Nvidia’s Bold New Bet on AI Neoclouds: Brilliant Platform Strategy or Latest Sign of an AI Bubble?.” Yahoo Finance, July 2, 2026. https://finance.yahoo.com/technology/ai/articles/nvidia-bold-bet-ai-neoclouds-145617772.html
[16] Data Center Dynamics. “Nvidia acts as backstop for customer GPUs in return for cut of cloud revenue.” DCD, July 27, 2026. https://www.datacenterdynamics.com/en/news/nvidia-acts-as-backstop-for-customer-gpus-in-return-for-cut-of-cloud-revenue/
[17] SemiAnalysis. “Nvidia GPU Debt Backstop Unleashes the AI Project Trinity: Capital, Offtake and Datacenters.” SemiAnalysis Newsletter, July 2026. https://newsletter.semianalysis.com/p/nvidia-gpu-debt-backstop-unleashes
[18] Capacity Media. “Nvidia now agrees to rent back unused GPU capacity from neocloud operators if customer demand falls short.” Capacity, August 2026. https://capacityglobal.com/news/nvidia-now-agrees-to-rent-back-unused-gpu-capacity-from-neocloud-operators-if-customer-demand-falls-short/
[19] Jim Osman (Forbes). “Nvidia AI Financing Is The $500 Billion Risk Investors Aren’t Watching.” Forbes, August 16, 2026. https://www.forbes.com/sites/jimosman/2026/08/16/nvidia-ai-financing-is-the-500-billion-risk-investors-arent-watching/
[20] Reuters (via Investing.com). “Private credit roundup: Nvidia’s half trillion for chips financing, plus others.” Reuters, August 14, 2026. https://www.investing.com/news/stock-market-news/private-credit-roundup-nvidias-half-trillion-for-chips-financing-plus-others-4861033
[21] CNBC. “Nvidia’s AI moat is shifting from chips to capital.” CNBC, August 18, 2026. https://www.cnbc.com/2026/08/18/nvidias-ai-moat-is-shifting-from-chips-to-capital.html
[22] CNBC. “Jensen Huang defends Nvidia’s growing financial support for AI ecosystem, says ‘the risk is low’.” CNBC, August 26, 2026. https://www.cnbc.com/2026/08/26/jensen-huang-defends-nvidias-growing-financial-support-for-ai-ecosystem-says-the-risk-is-low-.html
[23] Seoul Economic Daily. “Nvidia’s Huang Rejects Circular Financing Claims Over AI Investments.” Seoul Economic Daily, August 27, 2026. https://en.sedaily.com/international/2026/08/27/nvidias-huang-rejects-circular-financing-claims-over-ai
[24] CNBC. “Nvidia CEO Huang says $30 billion OpenAI investment ‘might be the last’.” CNBC, March 4, 2026. https://www.cnbc.com/2026/03/04/nvidia-huang-openai-investment.html
[25] Bloomberg (via Business Standard). “Nvidia-OpenAI deal sparks concerns over circular financing in AI boom.” Business Standard, September 24, 2025. https://www.business-standard.com/amp/companies/news/nvidia-openai-deal-sparks-concerns-over-circular-financing-in-ai-boom-125092401589_1.html
[26] CNBC. “Jim Cramer warns AI’s circular financing frenzy echoes the dot-com bubble.” CNBC, July 27, 2026. https://www.cnbc.com/2026/07/27/jim-cramer-warns-ai-circular-financing-echoes-dot-com-bubble.html
[27] CNN Business. “AI boom or bubble? Timing is everything.” CNN, August 13, 2026. https://www.cnn.com/2026/08/13/business/ai-stock-tech-bubble
[28] NVIDIA Corporation (Investor Relations). “NVIDIA Announces Financial Results for First Quarter Fiscal 2027.” NVIDIA Newsroom / GlobeNewswire, May 20, 2026. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-First-Quarter-Fiscal-2027/default.aspx
[29] NVIDIA Corporation. “CFO Commentary on First Quarter Fiscal 2027 Results (Form 8-K exhibit).” U.S. Securities and Exchange Commission (EDGAR), May 20, 2026. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000051/q1fy27cfocommentary.htm
[30] FinanceBuzz. “Google, Amazon, Meta, and Microsoft Hiked AI Capex and It Benefits One Stock.” FinanceBuzz, August 2026. https://financebuzz.com/news/google-amazon-meta-hiked-ai-capex
[31] CNBC. “Tech AI spending approaches $700 billion in 2026, cash taking big hit.” CNBC, February 6, 2026. https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html
[32] International Monetary Fund. “World Economic Outlook Update, July 2026.” IMF, July 8, 2026. https://www.imf.org/-/media/files/publications/weo/2026/update/july/english/text.pdf
[33] International Monetary Fund (Petya Koeva Brooks et al.). “Press Briefing Transcript: World Economic Outlook Update, July 8, 2026.” IMF, July 8, 2026. https://www.imf.org/en/news/articles/2026/07/08/tr070826-weo-press-briefing-transcript-july-8-2026
[34] Bank of England (Financial Policy Committee). “Financial Stability Report – July 2026.” Bank of England, July 7, 2026. https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026
[35] Bank of England (Financial Policy Committee). “Record of the Financial Policy Committee meeting on 26 June 2026.” Bank of England, July 7, 2026. https://www.bankofengland.co.uk/-/media/boe/files/financial-policy-committee-record/2026/fpc-summary-and-record-july-2026.pdf
[36] Federal Reserve Bank of Chicago. “Tail Risk for Banks Posed by Investments in Generative Artificial Intelligence.” Chicago Fed Insights, February 2026. https://www.chicagofed.org/publications/chicago-fed-insights/2026/ai-tail-risk-for-banks
[37] BigGo Finance (reporting Reuters interviews with John Williams, Jeff Schmid and Mary Daly). “Fed Officials Begin Assessing Financial Stability Risks From Frenzied AI Investment Boom.” BigGo Finance, August 2026. https://finance.biggo.com/news/c3b5abd6-55ed-4db8-a594-8c53793715cd
[38] Board of Governors of the Federal Reserve System (FEDS Notes). “The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact.” Federal Reserve Board, July 17, 2026. https://www.federalreserve.gov/econres/notes/feds-notes/the-ai-buildout-and-the-economy-publicly-available-data-to-assess-ais-impact-20260717.html
[39] U.S. Senate Committee on Banking, Housing, and Urban Affairs (Sen. Elizabeth Warren et al.). “Warren, Colleagues Press FSOC to Launch Probe into Financial Stability Risks of AI Debt Bubble.” U.S. Senate, January 22, 2026. https://www.banking.senate.gov/newsroom/minority/warren-colleagues-press-fsoc-to-launch-probe-into-financial-stability-risks-of-ai-debt-bubble
[40] Aswath Damodaran (Professor of Finance, NYU Stern School of Business). “AI’s Bar Mitzvah Moment: From Hype & Hope to Business Questions!.” Musings on Markets, August 20, 2026. https://aswathdamodaran.blogspot.com/2026/08/ais-bar-mitzvah-moment-from-hype-hope.html
[41] Yuvraj Malik (Stocktwits). “‘Dean Of Valuation’ Aswath Damodaran Says AI Has Reached Its ‘Bar Mitzvah Moment’ — And Investors Must Start Asking Hard Questions.” Stocktwits, August 21, 2026. https://stocktwits.com/news-articles/markets/equity/dean-of-valuation-aswath-damodaran-says-ai-has-reached-its-bar-mitzvah-moment-and-investors-must-start-asking-hard-questions/cZYBp1gRJNL
[42] David Prior (Foreign Policy Journal). “NYU “Dean of Valuation” Exits Nvidia and Warns Investors to Watch for the AI Shakeout.” Foreign Policy Journal, August 4, 2026. https://www.foreignpolicyjournal.com/2026/08/04/nyu-dean-of-valuation-exits-nvidia-nasdaq-nvda-and-warns-investors-to-watch-for-the-ai-shakeout/
[43] Fortune (interview with Daron Acemoglu, Institute Professor, MIT; 2024 Nobel Laureate). “Nobel Laureate Daron Acemoglu on the ‘brainless’ AI discourse, the myth of capitalism and the Gen Z revolution risk.” Fortune, June 21, 2026. https://fortune.com/2026/06/21/nobel-laureate-daron-acemoglu-ai-productivity-capitalism-democracy/
[44] MIT Technology Review (interview with Daron Acemoglu). “Three things in AI to watch, according to a Nobel-winning economist.” MIT Technology Review, May 11, 2026. https://www.technologyreview.com/2026/05/11/1137090/three-things-in-ai-to-watch-according-to-a-nobel-winning-economist/
[45] Fortune (on Erik Brynjolfsson, Stanford Digital Economy Lab). “One of Stanford’s original AI gurus says productivity liftoff has begun after doubling in 2025 amid transition to ‘harvest phase’ along J-curve.” Fortune, February 15, 2026. https://fortune.com/2026/02/15/ai-productivity-liftoff-doubling-2025-jobs-report-transition-harvest-phase-j-curve/
[46] Erik Brynjolfsson, Danielle Li and Lindsey Raymond. “Generative AI at Work (Quarterly Journal of Economics, vol. 140(2), 2025; NBER Working Paper 31161).” NBER, 2025. https://www.nber.org/papers/w31161
[47] Asad Ramzanali (Vanderbilt Policy Accelerator, Vanderbilt University). “After the AI Crash.” SSRN / Vanderbilt Policy Accelerator, March 26, 2026. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6476159
[48] Vanderbilt Policy Accelerator. “After the AI Crash: A Proposal.” Vanderbilt Policy Accelerator (Substack), March 26, 2026. https://vanderbiltpolicyaccelerator.substack.com/p/after-the-ai-crash
[49] CNBC. “The question everyone in AI is asking: How long before a GPU depreciates?.” CNBC, November 14, 2025. https://www.cnbc.com/2025/11/14/ai-gpu-depreciation-coreweave-nvidia-michael-burry.html
[50] Forbes Business Council. “The Hidden Variable In The AI Rally: A Depreciation Reality Check.” Forbes, April 17, 2026. https://www.forbes.com/councils/forbesbusinesscouncil/2026/04/17/the-hidden-variable-in-the-ai-rally-a-depreciation-reality-check/
[51] Quartz. “GPU-collateralized debt explained: AI financing risks.” Quartz, May 8, 2026. https://qz.com/gpu-collateralized-debt-ai-neocloud-coreweave-financing-risks-050526
[52] CoreWeave, Inc. (Investor Relations). “CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment-Grade Rated GPU-backed Financing.” Business Wire, March 31, 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx
[53] TechTimes. “AI Loan Investors Demand Covenants After CoreWeave Spread Blows Out 125 Points.” TechTimes, August 3, 2026. https://www.techtimes.com/articles/322772/20260803/ai-loan-investors-demand-covenants-after-coreweave-spread-blows-out-125-points.htm
[54] Quinn Emanuel Urquhart & Sullivan, LLP. “Client Alert: Emerging Litigation Risks in Financing AI Data Centers Boom.” Quinn Emanuel, March 13, 2026. https://www.quinnemanuel.com/the-firm/publications/client-alert-emerging-litigation-risks-in-financing-ai-data-centers-boom/
[55] International Energy Agency. “Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions (Key Questions on Energy and AI).” IEA, April 2026. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
[56] International Energy Agency. “Key Questions on Energy and AI – Executive Summary.” IEA, 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
[57] Brookings Institution. “Global energy demands within the AI regulatory landscape.” Brookings, June 10, 2026. https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/
[58] South China Morning Post. “Nvidia ships first H200 chips to China, but forecasts no data-centre computing revenue.” SCMP, August 27, 2026. https://www.scmp.com/tech/big-tech/article/3365383/nvidia-ships-first-h200s-china-forecasts-no-data-centre-computing-revenue
[59] TechJournal. “US-China AI Chip War 2026: Nvidia, Tariffs & the H200.” TechJournal, July 2026. https://techjournal.org/us-imposes-25-tariff-on-nvidia-h200-ai-chips-bound-for-china
[60] Tom’s Hardware. “The custom AI ASIC state of play (May 2026) — Broadcom deals, Google TPUs, Meta MTIA & beyond.” Tom’s Hardware, May 21, 2026. https://www.tomshardware.com/tech-industry/semiconductors/custom-ai-asics-examined-from-broadcom-to-mtia
[61] Hashrate Index. “Inside the Custom AI Chip Race: Google, AWS, Microsoft, Meta, OpenAI.” Hashrate Index, May 8, 2026. https://hashrateindex.com/blog/hyperscaler-ai-asic-market-report-part-1/
[62] CNBC. “OpenAI ‘will be a public company in 2027’ or sooner, CFO Friar tells employees.” CNBC, August 19, 2026. https://www.cnbc.com/2026/08/19/open-ai-ipo-timing-2027-friar.html
[63] TheStreet. “OpenAI’s CFO says the company will go public in 2027.” TheStreet, August 2026. https://www.thestreet.com/investing/stocks/openai-cfo-says-ipo-is-in-2027
[64] Mogin Law LLP (citing Rebecca Haw Allensworth, Vanderbilt Law School). “Nvidia’s $100B Investment in OpenAI Raises Antitrust Eyebrows.” Mogin Law LLP, September 25, 2025. https://moginlawllp.com/nvidias-100b-investment-in-openai-raises-antitrust-concerns/
[65] PYMNTS (citing Randal Picker, University of Chicago Law School, via WSJ). “Report: DOJ Begins Probe Into Nvidia Contracts and Partnerships.” PYMNTS, September 8, 2024. https://www.pymnts.com/news/regulation/2024/report-doj-begins-probe-into-nvidia-contracts-and-partnerships/
[66] Business Model Analyst. “Nvidia’s 25% Backstop Bought a Toll Booth on Its Own Chips.” Business Model Analyst, August 2026. https://businessmodelanalyst.com/nvidia-usage-linked-revenue-25-percent-backstop/
[67] Apollo Global Management. “NVIDIA Partners with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms (press release with partner statements).” Apollo, August 10, 2026. https://www.apollo.com/insights-news/pressreleases/2026/08/nvidia-partners-with-apollo
[68] Global FinTech Series (reproducing partner statements incl. David Solomon, Goldman Sachs). “NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms.” Global FinTech Series, August 2026. https://globalfintechseries.com/finance/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital/
[69] Yahoo Finance (citing Jon Gray, President, Blackstone, on CNBC). “NVIDIA Corporation (NVDA) & Blackstone (BX): Nvidia Wants Wall Street to Lend Against AI Chips Like They’re Mortgages.” Yahoo Finance, August 2026. https://finance.yahoo.com/markets/stocks/articles/nvidia-corporation-nvda-blackstone-bx-174958402.html
[70] NVIDIA Corporation. “Form 10-Q for the quarterly period ended April 28, 2024 (Note 12 – Commitments and Contingencies).” U.S. Securities and Exchange Commission (EDGAR), May 2024. https://www.sec.gov/Archives/edgar/data/1045810/000104581024000124/nvda-20240428.htm
[71] NVIDIA Corporation. “Form 10-Q for the quarterly period ended July 31, 2022 (Note 13 – Commitments and Contingencies).” U.S. Securities and Exchange Commission (EDGAR), August 2022. https://www.sec.gov/Archives/edgar/data/1045810/000104581022000147/nvda-20220731.htm
[72] NVIDIA Corporation. “Form 10-Q for the quarterly period ended October 27, 2019 (Note 13 – Commitments and Contingencies).” U.S. Securities and Exchange Commission (EDGAR), November 2019. https://www.sec.gov/Archives/edgar/data/1045810/000104581019000170/nvda2020q310q.htm
[73] Sacra. “OpenAI revenue, valuation & funding (company profile).” Sacra, July 2026. https://sacra.com/c/openai/
[74] Tom’s Hardware. “Nvidia revenue tops $96 billion as memory commitments soar to $160 billion — CEO Jensen Huang says AI ‘has reached its inflection point’.” Tom’s Hardware, August 27, 2026. https://www.tomshardware.com/tech-industry/big-tech/nvidia-revenue-tops-usd96-billion-as-memory-commitments-soar-to-usd160-billion-ceo-jensen-huang-says-ai-has-reached-its-inflection-point
[75] Stocks Today. “Nvidia Shifts Focus from China as Export Impasse Halts H200 Chip (fiscal 2026 full-year figures).” Stocks Today, August 2026. https://stockstoday.com/nvidia-shifts-focus-from-china-as-export-impasse-halts-h200-chip/
[76] Data Center Frontier (reporting IEA Electricity 2024). “IEA Study Sees AI, Cryptocurrency Doubling Data Center Energy Consumption by 2026.” Data Center Frontier, January 2024. https://www.datacenterfrontier.com/energy/article/33038469/iea-study-sees-ai-cryptocurrency-doubling-data-center-energy-consumption-by-2026
[77] The Capitol Forum (Teddy Downey with Asad Ramzanali). “Transcript of Conference Call: “After the AI Crash” — Policy, Markets, and the Future of AI Infrastructure.” The Capitol Forum, May 6, 2026. https://thecapitolforum.com/resource/transcript-of-conference-call-after-the-ai-crash-policy-markets-and-the-future-of-ai-infrastructure/
[78] Vanderbilt Law School. “Senators Warren and Blackburn Headline VPA Event on Looming AI Crisis.” Vanderbilt University, April 30, 2026. https://law.vanderbilt.edu/elizabeth-warren-marsha-blackburn-vpa-event-on-looming-ai-crisis/
[79] The Next Web. “Nvidia ships first H200 chips to China, but expects no data-centre revenue from them.” TNW, August 28, 2026. https://thenextweb.com/news/nvidia-h200-china-zero-revenue-guidance-europe-gigafactories
[80] Bank of England (Financial Policy Committee). “Financial Policy Committee Record – April 2026.” Bank of England, April 2026. https://www.bankofengland.co.uk/financial-policy-committee-record/2026/april-2026
