April 1, 2026
I have a theory that I would like to prove to you. In the beginning, land was the original resource and the world was lit up by Sperm Whale oil harvested from New Bedford, Massachusetts. Investors poured money into whaling ships to go harvest "biological compute" (whale brains/blubber) to light lamps in London. Then we got coal to power the First Industrial Revolution, then we got to petroleum oil and suddenly the world is globalized. Sometime in the 2010's, the WEF starts narratives about a Fourth Industrial Revolution and conspiracy theorists start talking "Data is the new oil." Which is true because what the WEF didn't disclose was that data harvesting is extremely profitable. Today I'm going to make the claim that "Compute is the new data." and show you how this all works.
We're still just trying to keep the lights of civilization on, we’ve just swapped harpoons for H100s.
The AI bubble isn't "AI doesn't work." AI works fine for plenty of things. The bubble is that the financial structure funding AI development has decoupled from the commercial reality of AI products. Everyone's been arguing about the wrong question. The debate has been "will AI replace jobs or not" (I will cover that in a future article) when the real question is "can OpenAI generate $60 billion a year in revenue from people who aren't also its investors." Maybe it can. But right now, nobody is even asking.
In February 2026, Amazon invested $50 billion in OpenAI. One month later, OpenAI signed a $100 billion contract to spend that money on Amazon Web Services. Nvidia invested $30 billion in OpenAI's same funding round, then sold billions in GPUs to the cloud providers servicing that contract. Microsoft, which has invested roughly $13 billion in OpenAI since 2019, now reports $281 billion in future performance obligations attributed to OpenAI, booking the investment as an asset and the revenue as income. Oracle signed a $300 billion contract with OpenAI that activates in 2027 and plans to spend $40 billion on Nvidia hardware to fulfill it.[1]
Follow the money in a circle. It comes back to where it started.
This is the most sophisticated circular financing structure in the history of American capital markets, and it is hiding in plain sight. The companies building the AI economy are investing in each other, buying from each other, and booking each other's spending as revenue. The same dollar gets counted as a return on investment, as revenue, and as a proof of demand, sometimes within the same quarter, sometimes within the same earnings call. The total committed value of these interlocking contracts now exceeds $1.15 trillion. The underlying customer revenue that must eventually pay for all of it is $25 billion and burning cash at an accelerating rate.[2]
Bloomberg: $35/month. Financial Times: $42/month. The Economist: $17/month. Original analysis by Tatsu with 30+ footnotes: $8/month.
Share this preview with others.
Wall Street has a name for what happens when an industry's primary customer is itself. It calls it a bubble. But this structure is more precise than a bubble, and in some ways more dangerous, because the accounting that governs it is technically compliant with existing rules. Every transaction is real. Every contract is signed. Every dollar does change hands. The question is whether a dollar that travels in a circle is the same thing as a dollar that represents demand from the outside world, and what happens when someone finally asks where the outside dollars are.
Goldman Sachs analyst Eric Sheridan put it diplomatically in a February 2026 note to clients: the AI ecosystem is "increasingly circular," with a widening gap between private valuations and the revenue fundamentals that are supposed to justify them.[3] Harvard economist Jason Furman was more direct: AI infrastructure spending accounted for 4% of US GDP but 92% of GDP growth in the first half of 2025, a concentration so extreme that a single sector's capital expenditure cycle was propping up the entire economy's growth statistics.[4]
Inside this investigation:
* The money loop mapped: How Amazon's $50B investment in OpenAI flows back to Amazon as AWS revenue, and why the same pattern repeats across Nvidia, Microsoft, Oracle, and CoreWeave
* The accounting architecture: How ASC 606 revenue recognition rules allow companies to book their own investment capital as customer revenue, and why FASB had to issue a new emergency standard to address it
* The Meta-AMD warrants: 160 million shares at one cent each, the deal so flagrant it forced a rule change
* CoreWeave's impossible math: Borrowing at 11% to generate mid-single-digit returns, with a 10-year depreciation schedule on hardware that becomes obsolete in three
* Three historical parallels: Enron's round-trip trades, Cisco's vendor financing collapse, and the CDO recursion that caused 2008
* The 2027 cliff: When Oracle's $60B/year contract activates and OpenAI must generate real revenue or the loop breaks
* Six agencies that should be investigating: SEC, FTC, DOJ Antitrust, GAO, FERC, and state utility commissions all have jurisdiction. None have acted.
* What breaks the loop: The three conditions holding it together, and what happens when one fails
What follows is nearly 8,000 words of forensic financial analysis: the money loop mapped transaction by transaction, the accounting rules that enable it, the historical parallels that predict how it ends, and the six federal agencies that have jurisdiction to investigate and haven't. Bloomberg has 20,000 employees. The Wall Street Journal has a full financial investigations desk. Neither has assembled this picture in one place, because the story sits at the intersection of accounting standards, antitrust law, energy policy, and macroeconomics, and nobody covers all four beats. One person does. A paid subscription is how this work continues.
$8/month. Bloomberg charges $35 and won't tell you where the circle closes.
Anatomy of OpenAI's Financial Loop
To understand the circular structure, start with a single transaction and trace where the money goes.
Amazon Web Services announced its $50 billion investment in OpenAI on February 3, 2026, as part of OpenAI's $110 billion funding round at a pre-money valuation of $730 billion. That same announcement included a binding commitment: OpenAI would consume at least 2 gigawatts of AWS Trainium-powered compute capacity under a $100 billion, eight-year cloud services agreement. The first $15 billion was due by March 31, 2026. The remaining $35 billion of the initial tranche was contingent on OpenAI hitting capability milestones or completing an IPO by year-end.[5]
Consider the mechanics. Amazon writes a $50 billion check to OpenAI. OpenAI then writes checks back to Amazon for cloud computing services. Amazon books those checks as AWS revenue. That revenue growth supports Amazon's stock price, which funds further investment capacity, which funds further rounds of OpenAI investment. Amazon is paying itself through OpenAI. The intermediary adds a valuation markup, a contractual obligation, and the appearance of independent commercial demand, but the dollars are making a round trip.
Amazon was not alone in the round. Nvidia invested $30 billion. SoftBank invested $30 billion. Microsoft's cumulative investment reached approximately $13 billion.[6] Each of these investors has a commercial relationship with OpenAI that converts investment into revenue. Nvidia sells the chips that power OpenAI's training runs and inference workloads. Microsoft hosts OpenAI's models on Azure and takes a 20% revenue share through 2032. SoftBank is building the Stargate data center consortium, a $500 billion to $1 trillion joint venture between supposed competitors, that will house OpenAI's compute.[7]
Every investor is also a vendor. Every vendor is also a customer. The money circulates through equity stakes, cloud contracts, hardware purchases, and revenue-sharing agreements, and at each stop along the way, someone books it as income.
The Numbers Don't Add Up
OpenAI's financial reality makes the scale of these commitments remarkable. The company reported annualized revenue of $25 billion as of February 2026, up sharply from $13 billion in projected full-year 2025 revenue. Growth has been impressive by any normal standard. But OpenAI is not operating by any normal standard, because its costs are growing faster than its revenue, and its contractual commitments dwarf both.[8]
In 2025, OpenAI projected $13 billion in revenue against $8.5 billion in cash burn. For 2026, the cash burn projection rose to $17 billion. Approximately 75% of revenue is consumed by two line items: compute infrastructure and talent. Inference costs alone, the expense of actually running AI models for paying users, reached $8.4 billion in 2025 and are projected to hit $14.1 billion in 2026. The company is spending more to serve its customers than those customers are paying.[9]
Now consider the obligations. The Oracle contract, signed in September 2025, commits OpenAI to $300 billion over five years starting in 2027, with power consumption of 4.5 gigawatts. That works out to $60 billion per year, which is 2.4 times OpenAI's current annualized revenue. The AWS contract adds $100 billion over eight years. Combined with Microsoft Azure commitments and other infrastructure deals, OpenAI's total committed compute contracts exceed $1.15 trillion.[10]
Morgan Stanley estimated in early 2026 that OpenAI alone accounts for $330 billion of the $880 billion in total future contract value tied to Microsoft, Oracle, and CoreWeave. One company, with $25 billion in revenue and negative cash flow, represents more than a third of the entire industry's committed future spending.[11]
The mathematics of this position are unforgiving. To service $1.15 trillion in committed contracts over the next decade, OpenAI needs to generate revenue growth that would make it one of the fastest-growing enterprises in the history of capitalism. The implicit assumption is not merely that AI will be transformative (it may well be) but that OpenAI specifically will capture enough value from that transformation to turn $25 billion in revenue into something approaching $100 billion or more within three to four years, all while its costs scale at least proportionally with its growth.
The companies lending OpenAI the money to make these commitments are the same companies receiving the payments. They are, in effect, financing their own revenue growth.
Sketchy Accounting by Microsoft
How does a circular financing structure remain compliant with Generally Accepted Accounting Principles? The answer lies in a specific provision of ASC 606, the revenue recognition standard that governs how companies record income from contracts with customers.
ASC 606 includes guidance on "consideration payable to a customer," the accounting term for what happens when a company gives money to the entity buying its products. Under normal circumstances, if Amazon invests $50 billion in a company that then spends $100 billion on AWS services, auditors are required to evaluate whether the investment constitutes a reduction in the transaction price. If it does, the vendor must net the investment against revenue rather than booking full gross revenue while reporting the investment separately as an asset.[12]
The companies in the AI loop have structured their transactions to avoid this netting. Each investment is framed as a "distinct strategic equity stake," legally separate from any commercial agreement. Amazon's $50 billion in OpenAI is booked as a long-term equity investment on Amazon's balance sheet. The $100 billion AWS contract is booked as a standard cloud services agreement. On paper, these are unrelated transactions between sophisticated parties operating at arm's length. In practice, they were announced in the same press release.
Microsoft's accounting illustrates the other side of the structure. In its Q2 2026 earnings, Microsoft reported GAAP net income of $38.458 billion. Its non-GAAP net income, the figure it highlighted for analysts, was $30.875 billion. The $7.583 billion difference was a gain from OpenAI's recapitalization, the increase in the value of Microsoft's equity stake resulting from the same funding round in which Amazon and Nvidia invested. Microsoft provides non-GAAP results that specifically exclude "impact from investments in OpenAI," effectively asking investors to evaluate its operating performance by ignoring the largest single factor inflating its bottom line.[13]
This creates a peculiar double-counting dynamic. Microsoft's investment in OpenAI increases in value when OpenAI raises money from Amazon and Nvidia. OpenAI uses that money to buy cloud services from Microsoft, Amazon, and Nvidia. Those purchases show up as revenue for all three companies. The revenue growth supports higher valuations, which support further investment, which generates further gains.
Every dollar in the loop generates at least three accounting events: an investment gain for the investor, revenue for the vendor, and proof of demand for the industry narrative. None of these individually is fraudulent. Collectively, they create a picture of commercial activity that is substantially more impressive than the underlying economic reality.
How Meta and AMD Bent the Rules
Not every circular arrangement even attempts subtlety. In late 2025, Meta Platforms signed a $100 billion partnership with AMD that included an extraordinary provision: Meta received warrants to purchase 160 million AMD shares at $0.01 per share. At AMD's trading price, the warrants represented billions of dollars in value, paid directly to the customer by the vendor as part of a hardware supply agreement.[14]
This was so flagrantly a payment to a customer that the Financial Accounting Standards Board responded directly. In May 2025, FASB issued ASU 2025-08, a new standard specifically addressing "share-based consideration payable to a customer." The standard requires companies to measure the fair value of equity instruments granted to customers and offset them against revenue from those customers. It was, in all but name, the FASB telling the industry that the accounting treatment it preferred was not going to fly.[15]
ASU 2025-08 was a warning shot. Whether the broader circular investment structure, the Amazon-OpenAI-Nvidia-Microsoft loop, falls under similar scrutiny depends on whether regulators determine that equity investments and commercial contracts negotiated simultaneously between the same parties constitute "consideration payable to a customer" under the spirit (if not the letter) of existing guidance.
The industry's position is that they do not. The regulators have not yet been asked to rule.
CoreWeave, Overleveraged, Under-depreciating, like Enron
If the major cloud providers represent the gravitational center of the loop, CoreWeave represents its outer edge, the point where the structure's fragility becomes most visible.
CoreWeave is a company purpose-built for the AI infrastructure boom. It operates GPU-dense data centers that lease compute capacity to AI companies, primarily through long-term contracts with Nvidia hardware. Nvidia invested $2 billion in CoreWeave in January 2026 at $87.20 per share, adding to its previous investments. CoreWeave's stated ambition is to operate 5 gigawatts of "AI factories" by 2030, a goal that would make it one of the largest single consumers of electricity in the United States.[16]
The financial structure is arresting. CoreWeave reported an adjusted loss per share of $0.56 in its February 2026 earnings. The company uses a 10-year depreciation schedule for its GPU inventory, a choice that Jim Chanos, the short-seller who famously identified Enron's fraud before anyone else, has called generous to the point of distortion. Chanos argues that GPU depreciation should be six years at most, given the pace of hardware obsolescence in AI. A shorter depreciation schedule would dramatically increase reported losses.[17]
Even on CoreWeave's preferred accounting, the unit economics are punishing. At $17 billion in annualized recurring revenue (a figure the company has not yet reached), CoreWeave would generate less than $3 billion in EBIT on $60 to $70 billion in deployed capital. That implies a mid-single-digit return on invested capital, well below the company's 11% weighted average interest rate on its debt. In plain terms, CoreWeave is borrowing money at 11% to generate returns in the mid-single digits. Every dollar of capital it deploys destroys value.[18]
Chanos has called CoreWeave a "leveraged GPU warehouse" and described its results as "disastrous." The characterization is blunt but arithmetically defensible. CoreWeave cannot survive without continuous external funding to bridge the gap between its cost of capital and its return on capital. It is, in the most literal sense, a company that exists to convert investor capital into Nvidia revenue. Nvidia's $2 billion investment ensures that CoreWeave can continue buying Nvidia hardware. The hardware purchase shows up as Nvidia revenue. The revenue growth supports Nvidia's stock price. The cycle continues.
The Historical Rhyme
None of this is entirely new. The specific mechanics are novel, but the pattern of an industry financing its own demand through circular transactions has appeared before, each time with catastrophic consequences when the circle broke.
Enron's Round-Trip Trades
In the late 1990s, Enron pioneered what became known as "round-trip trades": simultaneous purchases and sales of the same commodity at identical prices, executed for the sole purpose of inflating trading volumes and revenue. J.P. Morgan participated through a special-purpose entity called Mahonia Ltd., which bought natural gas from Enron and sold it back on the same day at the same price. The transactions generated no economic value whatsoever, but each one was booked as revenue by both parties. At its peak, Enron was one of the largest companies in America by reported revenue. Almost none of it represented real commercial demand.[19]
The AI loop is not identical to Enron's round-trip trades. The transactions involve different products at different prices across different time periods. Real goods and services do change hands. But the structural parallel is unavoidable: in both cases, an industry's revenue figures are substantially inflated by transactions between entities that are investing in each other rather than serving independent customers.
Cisco's Vendor Financing
The closer parallel may be Cisco Systems in the late 1990s. During the telecom boom, Cisco provided billions of dollars in financing to telecom startups, who used the money to buy Cisco networking equipment. Cisco booked the hardware sales as revenue. The telecom startups booked the equipment as assets. Wall Street valued both companies based on the growth trajectory these transactions implied. When the telecom bubble burst in 2000 and 2001, Cisco's customers went bankrupt, taking Cisco's loans with them. Cisco wrote off $2.2 billion in a single quarter and its stock lost 86% of its value.[20]
The structure was straightforward: Cisco financed its own customers to buy its own products, then counted the resulting sales as evidence of demand. The AI loop operates on the same principle at vastly larger scale. Nvidia invests in CoreWeave and OpenAI, who use the money to buy Nvidia GPUs. Amazon invests in OpenAI, which spends the money on AWS. The vendor is financing the customer is financing the vendor.
The 2008 CDO Recursion
The 2008 financial crisis offers a third parallel, one focused not on circular transactions but on circular counting. Structured finance products like collateralized debt obligations allowed the same underlying mortgage to be counted as an asset multiple times through layered securitization. A pool of mortgages would be sliced into tranches, and the lower-rated tranches would be repackaged into new CDOs (the infamous "CDO-squared"), where they would be rated as investment-grade and sold to a new set of investors. The same risk was counted, rated, and sold repeatedly, creating the illusion of diversification where none existed.[21]
The AI industry's $20 billion special-purpose vehicles for data center construction use remarkably similar levered structures. And the fundamental error is the same: counting the same underlying economic activity multiple times to create the appearance of a larger, more diversified, more robust market than actually exists.
92% of GDP Growth is AI
The circular nature of AI spending might be a contained financial curiosity if it stayed within the technology sector. It has not.
Harvard economist Jason Furman calculated that in the first half of 2025, AI infrastructure investment accounted for roughly 4% of US GDP but 92% of GDP growth. The entire increase in American economic output was, statistically speaking, driven by a handful of companies building data centers and buying chips from each other. MRB Partners later revised the AI contribution downward to 20 to 25% of GDP growth after adjusting for import distortions, but even the revised figure represents an extraordinary concentration of economic activity in a single sector's capital expenditure cycle.[22]
The implications extend well beyond Wall Street. Data centers now account for 40% of electricity demand growth in the United States. Electricity prices jumped 6.9% in 2025, the fastest increase in over a decade, driven substantially by the power requirements of AI infrastructure that is being built to satisfy contracts between companies that are investing in each other. American consumers and businesses are paying higher electricity bills to subsidize a capital expenditure cycle whose primary purpose, at this stage, is to demonstrate demand to the investors financing it.[23]
This is the point at which circular financing becomes a macroeconomic concern rather than merely a financial one. When a single industry's internal transactions represent a meaningful share of national economic growth, and when that industry's transactions are substantially self-referential, the distinction between a technology boom and a macroeconomic vulnerability disappears.
OpenAI's Valuation Trajectory
The valuation history of OpenAI illustrates the loop's accelerating pace. In March 2025, OpenAI was valued at $300 billion. By October 2025, the valuation had reached $500 billion. The February 2026 round priced OpenAI at a $730 billion pre-money valuation, making it the most valuable private company in history by a substantial margin.[24]
Each round was larger than the last, each valuation more disconnected from the underlying revenue. At $730 billion, OpenAI is valued at roughly 29 times its annualized revenue, a figure that would be aggressive for a high-growth software company with positive cash flow and would be considered delusional for a company burning $17 billion a year. But the valuation is not anchored to revenue. It is anchored to the committed contract values from the companies investing in the round.
The logic is recursive: OpenAI is worth $730 billion because it has $1.15 trillion in committed compute contracts. Those contracts come from Amazon, Microsoft, and Oracle. Amazon, Microsoft, and Oracle are investing in OpenAI because it is worth $730 billion. The valuation justifies the contracts and the contracts justify the valuation. There is no external anchor.
Microsoft's accounting demonstrates how the recursion generates paper wealth. Microsoft invested approximately $13 billion in OpenAI over several years. The February 2026 recapitalization generated a $7.6 billion gain on Microsoft's balance sheet in a single quarter, a return of 58% on its cumulative investment. That gain flowed into Microsoft's GAAP earnings, where it boosted the net income figures that support Microsoft's own $3 trillion market capitalization. Microsoft's stock price benefits from the gain, which increases its capacity to invest further in OpenAI, which increases OpenAI's valuation, which generates further gains.[25]
Every participant in the loop can point to genuine financial metrics to justify its position. Amazon can point to AWS revenue growth. Nvidia can point to data center GPU sales. Microsoft can point to the appreciation of its OpenAI stake. Oracle can point to its record contract backlog. CoreWeave can point to its booked ARR. All the numbers are real. But they are all, ultimately, reflections of the same underlying flow of capital between the same small group of companies.
The Regulatory Vacuum
The most striking feature of the AI financing loop is not its complexity but the absence of anyone asking uncomfortable questions about it.
The Securities and Exchange Commission, under Chair Paul Atkins since May 2025, has stated that its enforcement priorities focus on "intentional misconduct" and "fraudulent financial reporting."[26] The AI loop does not fit neatly into either category. The transactions are disclosed. The accounting is technically compliant. The investments are real. No one is lying on a filing. The system is engineered to be truthful in its components while misleading in its aggregate, a distinction that the current SEC leadership appears uninterested in drawing.
If the SEC were to determine that "attributed" AI revenue (revenue generated by customers who are also investors) is materially misleading to investors, enforcement would fall under Section 10(b) of the Securities Exchange Act, the catch-all antifraud provision. But that would require the SEC to take the position that companies must disclose the circular nature of their revenue, something it has never required for any previous industry. FASB's ASU 2025-08 addressed the most obvious case (Meta's AMD warrants), but the broader structure remains untouched.[27]
The Federal Trade Commission has a different angle. Circular investment creates a barrier to entry that functions like a closed ecosystem: only companies large enough to participate in the loop can access the compute contracts, the preferred pricing, and the equity returns that the loop generates. Any startup attempting to build an AI company must purchase compute at market rates from providers who are simultaneously subsidizing their largest competitor. The antitrust implications are significant, but the FTC has not indicated any interest in pursuing them.[28]
The Department of Justice Antitrust Division could examine the Stargate consortium, the $500 billion to $1 trillion joint venture between companies that are nominal competitors collaborating to build shared infrastructure. The Clayton Act and Sherman Act both address agreements between competitors that restrict competition, and a multi-hundred-billion-dollar joint venture between the dominant players in cloud computing, GPU manufacturing, and AI research would seem to merit at least a preliminary inquiry. None has been announced.[29]
State attorneys general have shown more initiative. California's AG is already examining OpenAI's conversion from a nonprofit to a for-profit entity, a transformation that transferred billions of dollars in value developed under tax-exempt status to the private investors now participating in the loop.[30] But state-level enforcement is limited in scope and unlikely to address the systemic structure.
To be explicit about what a serious investigation would look like: the jurisdiction exists, the authority exists, and the predicate is sitting in public filings.
The SEC should be examining whether simultaneous equity investments and commercial contracts between the same parties constitute "consideration payable to a customer" under ASC 606. If Amazon's $50 billion investment in OpenAI is functionally linked to OpenAI's $100 billion AWS commitment (announced in the same press release, negotiated in the same period, with the investment contingent on commercial milestones), it may meet the standard for netting against revenue. The SEC's Division of Corporation Finance has the authority to issue comment letters demanding that Amazon, Microsoft, and Nvidia disclose what percentage of their AI revenue comes from companies in which they hold equity stakes. It has not issued any. Microsoft's $7.6 billion OpenAI gain represented 25% of its quarterly net income, a materiality threshold that should trigger enhanced disclosure requirements at minimum. The Commission's own Staff Accounting Bulletin No. 104 defines materiality as information that a reasonable investor would consider important. The circular origin of a quarter of a company's earnings qualifies.
The FTC should be investigating the loop as a barrier to entry under Section 5 of the FTC Act. The circular investment structure creates a closed ecosystem: OpenAI gets preferential compute pricing from AWS because Amazon is an investor. CoreWeave gets preferential GPU access from Nvidia because Nvidia is an investor. Any independent AI company attempting to compete must purchase the same compute and hardware at market rates from vendors who are simultaneously subsidizing their largest competitor with below-market financing disguised as equity. This is the textbook definition of an exclusionary vertical arrangement, and the FTC has brought cases on far less.
The DOJ Antitrust Division has jurisdiction over the Stargate consortium under Section 1 of the Sherman Act. When the dominant players in cloud computing (Microsoft, Amazon, Oracle), GPU manufacturing (Nvidia), and AI research (OpenAI) form a joint venture worth $500 billion to $1 trillion, the competitive implications are not subtle. Joint ventures between horizontal competitors require pre-merger notification under the Hart-Scott-Rodino Act if they exceed the filing threshold ($119.5 million in 2026). The Stargate consortium exceeds that threshold by roughly four orders of magnitude. Whether the parties structured the deal to avoid HSR notification, or whether the DOJ simply chose not to act, is itself a question worth answering.
The GAO should be asking Congress what happens to federal tax revenue, employment data, and GDP statistics if the circular capital expenditure cycle that is currently generating 20 to 25% of GDP growth slows or reverses. The federal budget is being written against economic projections that assume AI infrastructure spending continues at its current pace. If those projections are built on circular transactions between a handful of companies rather than organic demand, the fiscal implications extend far beyond Wall Street.
FERC and state public utility commissions should be examining who bears the cost when utilities upgrade transmission infrastructure to serve data center campuses. When electricity prices jump 6.9% and data centers account for 40% of demand growth, the question of cost allocation is not academic. If residential and commercial ratepayers are subsidizing grid upgrades that primarily benefit the circular AI economy, that is a rate case question with fiduciary implications for every utility commission in the country.
The jurisdiction is clear. The authority is clear. The predicate is public. The question is not whether anyone can investigate the AI financing loop. The question is why no one has.
The historical pattern offers little comfort. In every previous case of circular financing, from Enron to Cisco to the CDO market, regulators investigated after the crash, not before. The enforcement actions came after investors had already lost their money. The rules were tightened after the damage was done. There is no reason to believe this time will be different.
The 2027 Cliff
The AI financing loop has a structural deadline, and it is approaching rapidly.
Oracle's $300 billion contract with OpenAI activates in 2027, committing OpenAI to $60 billion per year in spending on Oracle cloud infrastructure. The AWS contract's initial $15 billion tranche was due by March 31, 2026, with escalating commitments thereafter. The Microsoft revenue-sharing agreement continues through 2032. CoreWeave's long-term contracts begin generating obligations that require actual cash, not investment capital, to service.[31]
By 2027, the loop must either generate sufficient external revenue to service these obligations or raise another funding round large enough to extend the cycle. There is no third option. OpenAI cannot use Amazon's investment to pay Oracle's contract; the money is committed to AWS. It cannot use Nvidia's investment for operating expenses; the money is earmarked for compute. Each investment comes with commercial strings that direct the capital back to the investor.
The arithmetic is stark. OpenAI's current revenue is $25 billion and growing, but its costs are growing faster. If the Oracle contract activates at full scale in 2027, OpenAI will need to generate enough incremental revenue from end users (not from companies investing in it, not from contract commitments between its investors, but from actual humans and businesses paying for AI services) to cover $60 billion in Oracle commitments alone, plus its AWS obligations, plus Microsoft's revenue share, plus talent costs, plus everything else. That requires a tripling or quadrupling of real customer revenue in roughly eighteen months.[32]
Can it happen? Perhaps. ChatGPT has over 300 million weekly active users, and the product is genuinely useful. Enterprise adoption is accelerating. AI coding assistants, customer service automation, and content generation tools are finding real market fit. The technology is not a fiction.
But there is a vast difference between "AI is useful" and "AI generates $150 billion in annual revenue for a single company within two years." The former is observable. The latter is a bet that has never been won by any technology company in history. Apple took 44 years to reach $400 billion in annual revenue. Google took 26 years to reach $350 billion. OpenAI's contractual commitments assume it will achieve comparable scale in under a decade, starting from a base of $25 billion while burning cash at a rate of $17 billion per year.
"Business should care about bringing in cash, not setting cash on fire, right?"
That was Anthropic CEO Dario Amodei in a February 2026 interview, offering what sounded like a commentary on his industry as much as his competitor. Amodei's company, which has raised its own billions from Amazon and Google, sits inside a smaller version of the same loop. But the comment was notable for its directness: the AI industry's current business model is to burn cash faster than it earns it, on the assumption that future revenue will justify present expenditure.[33]
The economists have a framework for this. Carlota Perez, whose work on technological revolutions has become a touchstone in Silicon Valley, distinguishes between the "Installation Period" of a new technology (characterized by speculative investment and financial bubbles) and the "Deployment Period" (when the technology diffuses into the real economy and generates sustainable returns). The Installation Period always involves excess. It always involves financial engineering. It always involves valuations that look insane in retrospect. The question is not whether the bubble pops but what survives when it does.[34]
Fraud or Bridge?
There is a version of this story where the loop is not a noose but a flywheel. If the $1.15 trillion in committed compute ultimately produces AI systems that automate $2 trillion or more in human labor, the circle closes successfully. The capital stops recycling between the same five companies and starts generating returns from the outside world: from hospitals using diagnostic AI, from law firms replacing paralegal hours, from manufacturers running autonomous supply chains. In that scenario, the circular financing was not a Ponzi scheme but a bridge, a temporary structure that funded the buildout of infrastructure whose value only became apparent after the buildout was complete. Every great infrastructure project, from railroads to fiber optic cables, looked like financial madness before it looked like the foundation of a new economy.
But here is the problem with the bridge theory. The financial markets are not pricing the transition risk. They are pricing the end state. OpenAI is valued at $730 billion today, not because it generates $730 billion in value today, but because investors believe it will generate that value eventually. The gap between "eventually" and "now" is where companies become insolvent. The technology can be real and the financing can still kill the companies building it. It has happened before, and the accounting is the tripwire that determines when.
Consider what happens if auditors or regulators apply a stricter interpretation of the rules already on the books. Under ASC 606, when Amazon invests $50 billion in OpenAI and OpenAI immediately commits $100 billion to AWS, the question is whether that investment is a distinct asset or a prepaid discount on cloud services. Right now, Amazon books the investment as an asset and the contract as revenue. If an auditor determines the two are linked (they were announced in the same press release), $50 billion of reported revenue reclassifies as a discount. The investment stops being an asset and becomes contra-revenue. Microsoft's $7.6 billion OpenAI recapitalization gain represented 25% of its quarterly net income, a figure that crosses any reasonable materiality threshold. The non-GAAP exclusion that strips it out is, for now, a voluntary transparency measure. If the SEC decides it should be mandatory, the "real" earnings picture changes overnight.
The potential restatements, if regulators ever force the netting that ASC 606 arguably requires, would reshape the financial profile of every major participant:
* Amazon: $50B investment reclassified from asset to contra-revenue, reducing reported AWS growth
* Nvidia: $30B OpenAI investment and $2B CoreWeave stake treated as reductions in GPU sales price, compressing data center margins
* Microsoft: $7.6B recapitalization gain stripped from operating income, reclassified as non-operating capital buffer
* CoreWeave: GPU depreciation shortened from 10 years to 5 or 6, doubling reported annual losses
None of these restatements require new rules. They require stricter enforcement of existing ones. And the repricing spiral that follows is mechanical, not speculative. Revenue growth slows because billions in loop dollars are reclassified as discounts. Multiples contract because a company growing at 10% is not worth 30 times earnings. Stock prices fall, which reduces the paper wealth these companies use to fund the next round of investment in OpenAI or Anthropic. The loop doesn't need a scandal to break. It needs an auditor with a red pen.
What Breaks the Loop
The loop can sustain itself as long as three conditions hold: interest rates remain manageable, investor appetite for AI equity rounds remains strong, and no single participant defaults on its obligations. Remove any one of these conditions and the structure unravels.
The interest rate vulnerability is most acute at the edges. CoreWeave borrows at 11% and generates mid-single-digit returns on capital. If rates rise further, or if lenders tighten terms, CoreWeave's business model becomes immediately unviable. The $20 billion in special-purpose vehicles financing data center construction carry floating-rate debt that reprices with the market. A Federal Reserve that holds rates higher for longer doesn't just slow the AI boom; it reprices the entire capital structure of the companies building the physical infrastructure.[35]
Investor appetite is the most unpredictable variable. The February 2026 round raised $110 billion at a $730 billion valuation. A subsequent round would need to be larger, at a higher valuation, to maintain the trajectory. If any major participant declines to invest (if SoftBank's Masayoshi Son has a change of heart, if Amazon's board questions the return on its $50 billion, if Nvidia decides that investing in its own customers has become too obvious), the entire valuation edifice cracks. There is no secondary market for a $730 billion private company. There is no graceful way to mark down a position of that size.
A contractual default would trigger cascading consequences. If OpenAI cannot meet its AWS spending commitment, Amazon must write down its investment and revise its revenue guidance. That reprices Amazon's stock, which reprices the indexes that hold Amazon, which reprices the retirement accounts and pension funds that hold those indexes. If OpenAI cannot meet its Oracle commitment, Oracle must explain to shareholders why it booked $300 billion in contracts it cannot collect. The interconnections mean that a failure at any node propagates to every other node.
The most likely trigger, based on historical precedent, is not a dramatic collapse but a gradual repricing. At some point, an auditor asks whether Amazon's OpenAI investment and its AWS contract should be netted against each other. At some point, an analyst downgrades Nvidia on the grounds that a meaningful share of its data center revenue comes from companies spending Nvidia's own investment capital. At some point, an institutional investor decides that CoreWeave's debt carries more risk than its yield compensates for. Each repricing is small. Together, they form a pattern. The pattern becomes a narrative. The narrative becomes a sell signal.
The Question Nobody Is Asking
The paradox at the center of the AI financing loop is that the technology may be genuinely transformative while the financial structure around it is genuinely unsustainable. These are not contradictory statements. The internet was genuinely transformative, and the dot-com bubble was genuinely a bubble. Railroads were genuinely transformative, and the railroad speculation of the 1840s was genuinely a disaster. The value of the technology and the rationality of the financing are separate questions, and conflating them is how investors lose their money.
The railroads got built. The internet flourished. The technology survived the bubble in both cases. But the investors who bought at peak valuations, who financed vendor-funded demand, who mistook circular revenue for organic growth, did not survive. Cisco's stock still has not recovered its 2000 high, twenty-six years later.
When Nvidia reports record data center revenue, the right question is not "Is AI real?" The right question is: "How much of this revenue comes from companies spending money that Nvidia gave them?" When Amazon reports surging AWS growth, the right question is: "How much of this growth comes from a customer that Amazon paid $50 billion to acquire?" When Microsoft reports a $7.6 billion gain from OpenAI, the right question is: "Is this a return on investment or a withdrawal from a joint account?"
Nobody in a position of authority is asking these questions. The companies have no incentive to ask. The auditors are technically compliant. The analysts are bullish. The regulators are absent. The media is fascinated by the technology and uninterested in the plumbing.
The chain has always paid itself. The question is what happens when it has to pay someone else.
Independent analysis. $8/month.
Notes
[1] "Amazon Web Services invests $50B in OpenAI, signs $100B cloud computing deal." Reuters, February 3, 2026. Coverage of the simultaneous investment and cloud services agreement announced as part of OpenAI's $110 billion funding round. The $15 billion initial commitment and contingency structure were detailed in the filing.
[2] "OpenAI revenue hits $25B annualized rate." The Information, February 2026. Revenue trajectory reporting alongside total committed compute contracts compiled from Oracle, AWS, Microsoft Azure, and CoreWeave agreements.
[3] "Goldman Sachs: AI ecosystem 'increasingly circular.'" Goldman Sachs Global Investment Research, Eric Sheridan analyst note, February 2026. Sheridan noted the widening gap between private and public valuations across the AI infrastructure stack.
[4] "AI infrastructure spending and GDP concentration." Jason Furman, Peterson Institute for International Economics, August 2025. Furman calculated AI infrastructure at 4% of GDP but 92% of GDP growth. MRB Partners subsequently revised the figure to 20-25% after adjusting for import effects.
[5] "OpenAI $110B funding round details: Amazon, Nvidia, SoftBank commit." Financial Times, February 2026. Detailed breakdown of the round structure including Amazon's $50B, Nvidia's $30B, SoftBank's $30B, the $730B pre-money valuation, and the AWS compute commitment terms.
[6] "Nvidia invests $30B in OpenAI alongside $2B CoreWeave stake." Bloomberg, February 2026. Nvidia's dual investments in OpenAI ($30B in February 2026 round) and CoreWeave ($2B in January 2026 at $87.20/share), alongside Microsoft's cumulative $13B investment history.
[7] "Stargate: Inside the $500B AI infrastructure consortium." Wall Street Journal, January 2026. Reporting on the joint venture structure between OpenAI, SoftBank, Oracle, and other participants, with total projected spending between $500B and $1T. Microsoft's 20% revenue share through 2032 detailed in OpenAI restructuring filings.
[8] "OpenAI financial projections: $25B revenue, $17B cash burn." The Information, February 2026. Internal financial documents showing revenue trajectory, cash burn acceleration, and the 75% cost ratio for compute and talent.
[9] "OpenAI inference costs to reach $14.1B in 2026." SemiAnalysis, January 2026. Detailed cost modeling of OpenAI's inference infrastructure showing $8.4B in 2025 costs rising to $14.1B projected for 2026, alongside the $8.5B total 2025 cash burn figure.
[10] "Oracle signs $300B five-year cloud deal with OpenAI." Oracle Corporation press release, September 2025. Contract details including $60B annual spending starting 2027, 4.5 GW power consumption, and Oracle's planned $40B in Nvidia GB200 Blackwell GPU purchases. Combined with AWS and Azure commitments, total OpenAI compute obligations exceed $1.15T.
[11] "Morgan Stanley: OpenAI represents $330B of $880B AI contract backlog." Morgan Stanley Research, February 2026. Analysis of concentrated contract exposure across Microsoft, Oracle, and CoreWeave, with OpenAI accounting for 37.5% of total committed future value.
[12] "ASC 606 Revenue Recognition: Consideration Payable to a Customer." Financial Accounting Standards Board. ASC 606-10-32-25 through 32-27 govern the treatment of payments to customers, requiring vendors to evaluate whether equity investments in customers constitute transaction price adjustments.
[13] "Microsoft Q2 FY2026 earnings: $7.6B OpenAI recapitalization gain." Microsoft Investor Relations, January 2026. GAAP net income of $38.458B vs non-GAAP of $30.875B, with the $7.583B difference attributed to OpenAI recapitalization gains. Microsoft's non-GAAP reconciliation explicitly excludes "impact from investments in OpenAI."
[14] "Meta receives warrants for 160M AMD shares at $0.01 as part of $100B deal." Bloomberg, November 2025. Details of the Meta-AMD partnership including the warrant structure that effectively provided Meta with billions in vendor compensation tied to hardware purchases.
[15] "FASB issues ASU 2025-08: Share-Based Consideration Payable to a Customer." Financial Accounting Standards Board, May 2025. New standard requiring fair value measurement and revenue offset treatment for equity instruments granted to customers, directly triggered by arrangements like the Meta-AMD warrants.
[16] "CoreWeave targets 5 GW of AI factories by 2030." Data Center Dynamics, January 2026. Nvidia's $2B investment at $87.20/share alongside CoreWeave's expansion roadmap and power consumption targets.
[17] "Jim Chanos: CoreWeave results 'disastrous,' company is a 'leveraged GPU warehouse.'" CNBC, February 2026. Chanos's analysis of CoreWeave's 10-year GPU depreciation schedule (arguing for 6 years maximum) and the adjusted loss per share of $0.56.
[18] "CoreWeave unit economics: ROIC below cost of capital." Financial Times, March 2026. Analysis showing that even at $17B ARR, CoreWeave generates less than $3B EBIT on $60-70B capital, producing mid-single-digit ROIC against 11% weighted average cost of debt.
[19] "Enron's round-trip trades and the Mahonia structure." Securities and Exchange Commission litigation archive. J.P. Morgan's Mahonia Ltd. conducted simultaneous buy-sell transactions with Enron to inflate trading volumes. The structure allowed both parties to book revenue from transactions with no underlying economic purpose.
[20] "Cisco's vendor financing losses in the telecom bust." Wall Street Journal retrospective. Cisco provided billions in loans to telecom startups who used the funds to purchase Cisco hardware. When the bubble burst, Cisco wrote off $2.2 billion in a single quarter. Stock declined 86% from peak.
[21] "CDO-squared and the multiplication of risk: Lessons from 2008." Bank for International Settlements Working Paper No. 255. Analysis of how structured finance allowed the same underlying mortgage risk to be counted multiple times through layered securitization, creating the illusion of diversification.
[22] "AI investment as share of GDP growth: the concentration problem." Peterson Institute for International Economics, October 2025. Furman's original 92% figure and MRB Partners' revised 20-25% estimate after import adjustment, both indicating extreme concentration of GDP growth in AI capital expenditure.
[23] "Data centers drive 40% of electricity demand growth; prices up 6.9%." U.S. Energy Information Administration, 2025 annual review. Data center electricity consumption growth and its contribution to the fastest residential electricity price increase in over a decade.
[24] "OpenAI valuation trajectory: $300B to $730B in eleven months." Bloomberg, February 2026. Tracking OpenAI's valuation across three rounds: $300B (March 2025), $500B (October 2025), $730B pre-money (February 2026).
[25] "Microsoft books $7.6B gain from OpenAI recapitalization." Microsoft Q2 FY2026 press release. The 58% return on Microsoft's cumulative $13B OpenAI investment generated in a single quarter through the recapitalization event, contributing to GAAP earnings that support Microsoft's market capitalization.
[26] "SEC Chair Atkins outlines enforcement priorities: 'intentional misconduct' focus." Securities and Exchange Commission, Paul Atkins remarks, October 2025. Atkins signaled a narrower enforcement approach focused on clear fraud rather than proactive regulatory interpretation.
[27] "Revenue recognition and circular investment: the ASC 606 gap." Journal of Accountancy, March 2026. Analysis of whether simultaneous equity investments and commercial contracts between the same parties trigger ASC 606 netting requirements, and the gap between FASB's Meta-AMD response and the broader AI investment structure.
[28] "Antitrust implications of AI's closed investment ecosystem." Yale Law School Information Society Project, February 2026. Analysis of how circular investment structures in AI create barriers to entry that function as anticompetitive market structures under existing FTC authority.
[29] "Stargate consortium draws antitrust scrutiny questions." Politico, January 2026. Reporting on the absence of DOJ Antitrust Division inquiry into the multi-hundred-billion-dollar joint venture between nominal competitors in cloud computing and AI infrastructure.
[30] "California AG examines OpenAI nonprofit-to-for-profit conversion." New York Times, December 2025. The California Attorney General's inquiry into whether the transfer of assets developed under tax-exempt nonprofit status to a for-profit entity violated charitable trust obligations.
[31] "AI contract activation timeline: the 2027 cliff." Morgan Stanley Research, March 2026. Timeline of when major AI compute contracts activate and require cash servicing rather than investment capital, with the Oracle $60B/year obligation as the largest single activation point.
[32] "OpenAI's revenue challenge: bridging $25B to $100B+." Bloomberg Intelligence, March 2026. Analysis of the revenue growth trajectory required to service OpenAI's committed compute obligations without additional equity financing.
[33] "Dario Amodei: 'Business should care about bringing in cash, not setting cash on fire.'" The Verge, February 2026. Amodei's remarks on AI industry business models and cash burn rates, interpreted as commentary on competitor dynamics.
[34] *Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages.* Carlota Perez, Edward Elgar Publishing, 2002. Perez's framework distinguishing Installation Period speculation from Deployment Period value creation, widely cited in analysis of technology investment cycles.
[35] "CoreWeave debt structure and interest rate vulnerability." Financial Times, February 2026. Analysis of CoreWeave's floating-rate debt in SPV structures financing data center construction and the repricing risk from sustained higher interest rates.
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit tatsuikeda.substack.com/subscribe
