An accounting rule introduced by the Financial Accounting Standards Board in 2016 was designed to address a flaw in the previous regime that contributed to the 2008 Financial Crisis. However, this same rule is enabling the circuit of investments that flows from Big Tech companies to artificial intelligence startups, whose increased valuation from these investments increases the value of the Big Tech companies, which they can then reinvest in the AI startups. The risk is an AI bubble that, if it pops, will also blow up Americans’ savings, writes Hera Hyeonseo Lee.
When Amazon reported first-quarter earnings on April 29, the headline number was extraordinary: $30.3 billion in net income, amounting to a 77% increase over the prior year. Revenue for Amazon Web Services, its cloud computing division, grew 28%. Advertising surged. Analysts celebrated what appeared to be the strongest evidence yet that its investment—and investment in general—in artificial intelligence was paying off.
However, the figure that should have led the coverage was the $16.8 billion of that net income that came from a mark-to-market revaluation of its stake in Anthropic. Mark-to-market is an accounting method that records the fair value of an asset or liability based on current market price versus its historical cost. In this case, reevaluation produced a pre-tax gain for Amazon that exceeded the company’s entire operating income growth for the quarter based solely on Anthropic’s most recent funding round. Amazon’s shares rose more than 4% after it reported earnings.
Amazon was not alone. Alphabet disclosed roughly $28.7 billion in similar unrealized gains for the first quarter, driven primarily by its portfolio of non-marketable equity securities, which Fortune reported is dominated by its Anthropic stake. Nearly half of Alphabet’s record $62.6 billion profit came from updating the value of equity it owns in private companies. Two of the three largest cloud providers reported their best quarters in years, and in both cases, the single largest contributor to the bottom line was a paper gain on a private company that has never paid them a dividend.
Fortune reported the numbers. Bloomberg mapped the circular deals that show the Big Tech companies recycling funding between one another, boosting the market capitalization of all. But neither outlet connected the disclosure to the financial architecture that produced it. The mark-to-market gain organizes one of the two self-referential circuits that has organized AI infrastructure investment since at least 2023—the other being what I have called the cloud credit circuit. The consequence of the two circuits is the inflation of Big Tech and AI startup valuations, which, if they stop growing or underperform, will reverse the circuits and send valuations, and Americans’ savings, spiraling.
The cloud credit circuit, briefly
The Federal Trade Commission’s January 2025 report on AI partnerships traced how the large cloud providers invest billions in AI startups: Microsoft has investments in OpenAI, Amazon and Google in Anthropic. A substantial portion of that investment takes the form of credits for the investor’s own cloud platform. The startup spends the credits training models on the investor’s infrastructure. Anthropic spent $1.35 billion on AWS in 2024 and $2.66 billion from January to September 2025 alone. The investor books the resulting usage as revenue. That revenue growth supports the stock price that justifies the next round of investment.
This circuit of the established Big Tech companies plowing funds into the AI startups and the AI startups sending that money back for cloud services has four nodes: invest, spend, book, justify. Q1 2026 reveals a second parallel loop running on the same investment.
The mark-to-market circuit
The latest accounting standards (ASU 2016-01) issued by the Financial Accounting Standards Board, a nonprofit empowered by the Securities and Exchange Commission to set accounting and reporting guidelines for public companies in the United States, requires companies holding equity stakes in other firms to update those stakes to fair value each quarter. Any changes flow directly through net income. Before the rule went into effect in 2018, unrealized gains on equity holdings could sit on the balance sheet as “accumulated other comprehensive income,” which is excluded from net income. After the rule, every upward revaluation becomes reported profit.
The rule had a sound rationale. Under the prior regime, financial institutions could hold appreciated assets at historical cost indefinitely, obscuring their true economic position from investors and regulators. This treatment of an assets value delayed the ability of accountants and investors to recognize how much assets had devalued before and during the 2008 Financial Crisis. By requiring fair-value measurement, the new accounting regime aimed to give markets a more accurate and timely picture of a company’s financial exposure.
However, in the context of these two circuits running between Big Tech investments and AI startups, the rule produces a specific and unintended outcome.
When Amazon invests $8 billion in Anthropic, and that investment is structured partly as cloud credits for AWS, the money flows back to Amazon as revenue. But the investment also purchases equity. When Anthropic raises its next funding round at a higher valuation—as it did in February with its Series G—Amazon’s stake is revalued upward. The gain flows directly into Amazon’s net income. Amazon’s $8 billion investment in Anthropic is now worth more than $70 billion according to Amazon.
The question is how independent that change in valuation really is. Amazon is both an investor in Anthropic and one of its principal business partners. When Amazon commits additional capital, or when it deepens its cloud partnership in ways that make Anthropic’s business more viable, it contributes to the conditions that push the valuation—and thus its reported net income—higher. As Robert Willens, a tax and accounting consultant who has served as an adjunct at Columbia Business School, told Fortune: the companies “are able to control or influence the value of one of their own assets” through “business transactions with that entity.”
To be precise about the causal claim: new funding rounds by Anthropic and other AI startups are priced by incoming investors—venture capital and private equity firms—who conduct their own due diligence. The valuation is not set unilaterally by Amazon or Alphabet. But these incoming investors price the round in a context shaped by the commercial commitments of the existing strategic investors. Anthropic’s revenue trajectory, its access to compute infrastructure, and its competitive position all depend substantially on the cloud partnerships that Amazon and Google provide. The valuation is not manufactured by a single party, but neither is it determined independently of the strategic investors’ own actions. The circularity is partial, but it is real—and it is precisely the kind of entanglement that ASU 2016-01 was not designed to address.
The mark-to-market revaluation reveals a second loop running parallel to the first. The cloud credit circuit converts investment into revenue. The mark-to-market loop converts the same investment directly into profit—bypassing the revenue step entirely. A single dollar invested in Anthropic does double duty: it returns as cloud revenue and simultaneously appreciates as equity. The more Amazon invests, the more revenue it books and the more profit it reports, without Anthropic ever returning a dollar on the equity itself. In a sense, Amazon, Google, and Microsoft have created twin engines from a single investment.
Accounting for a bubble
Amazon and Alphabet followed the Generally Accepted Accounting Principles (GAAP), which is likewise overseen by the FASB, in their SEC filing for the first quarter. They clearly disclosed the gains, according to the 2016 FASB rule. No fraud has been committed. The cloud credit circuit, as well as the booked profits from investments that raise the value of Anthropic and OpenAI, are, as Bloomgberg pointed out, perfectly legal.
However, the question is not whether these deals violate existing rules. It is whether existing rules have become infrastructure for a financial circuit that inflates reported earnings without corresponding cash generation. ASU 2016-01 was created to address the risk of the previous accounting regime revealed in the 2008 financial crisis. What we have to ask now is if it created a different risk and is contributing to a potential bubble in the AI and tech markets.
Some readers will note that this is not the first time ASU 2016-01 has distorted earnings. Berkshire Hathaway’s quarterly net income has swung wildly since the rule took effect, driven by mark-to-market movements in its public equity portfolio, leading Warren Buffett to caution investors to focus on operating earnings instead. But the Berkshire case differs in a critical respect. Berkshire holds shares of publicly traded companies—Apple, Coca-Cola, Bank of America—whose valuations are set by liquid markets and whose business outcomes are independent of Berkshire’s own commercial relationships with them. Berkshire does not sell cloud services to Apple, book Apple’s spending as revenue, and then record a gain when Apple’s stock rises. The cloud credit circuit is distinctive because the investor has a material commercial relationship with the entity whose revaluation generates the gain—making the investor both beneficiary of, and contributor to, the conditions that produce the revaluation. It is this entanglement, not fair-value accounting per se, that creates the self-referential loop.
The cash flow data bears this out. In the same quarter that Amazon reported $30.3 billion in net income, its trailing twelve-month free cash flow—the cash a company generates after covering its operating expenses and capital expenditures—fell 95% to $1.2 billion. Capital expenditures hit $44.2 billion. The company is spending unprecedented sums on AI infrastructure while reporting record profits—and the profits are substantially driven by the revaluation of an asset whose value depends on the continuation of that spending. If Anthropic’s next valuation round is flat rather than up, Amazon’s own valuation would tumble. That could in turn impact Anthropic’s value if Amazon has to pull back its investments in the startup.
What happens if the bubble pops
The mark-to-market gain does not stay on Amazon’s income statement. Its inflation flows into every investment vehicle it touches, and these include many Americans’ household savings.
When Amazon’s earnings per share come in at $2.78 against a consensus estimate of $1.64, the stock responds. Market capitalization rises. Because the S&P 500 is weighted by market capitalization, Amazon’s share of the index increases. Every target-date fund, every passively managed 401(k), every index-tracking ETF adjusts its allocation accordingly. More of the money that flows automatically into these funds on each payday is routed toward Amazon.
Since the Pension Protection Act of 2006, employers have been encouraged to auto-enroll workers into 401(k) plans and designate target-date funds as the default. By 2025, 69% of participants in Vanguard-administered plans were invested in a professionally managed allocation—the vast majority in a single target-date fund. The vast majority of retirement savers passively channel contributions into whatever the market-cap-weighted index contains. As I argued in a recent piece for ProMarket, this architecture creates a dedicated constituency for AI capital: participants did not choose to be exposed to it; the default architecture chose for them.
The critical point is not that passive indexation transmits inflated valuations—it does this for any overvalued stock. The critical point is that the cloud credit circuit uniquely generates the inputs that passive investors rely on to assess value. Earnings per share, the denominator of the price-to-earnings ratio, is inflated by mark-to-market gains that originate within the circuit itself. This makes the standard valuation heuristics that retail investors and automated rebalancing systems use—headline P/E, earnings surprises, consensus beats—less reliable as signals of underlying economic performance. The Magnificent Seven trade at roughly 28 to 29 times forward earnings—elevated, but not in obvious bubble territory by headline metrics. Strip out unrealized investment gains from entities with which the reporting company has commercial relationships, and the multiple looks different. But the headline number is what drives index composition and what appears on the brokerage statement that a 401(k) participant glances at once a quarter.
Americans do not need to understand the cloud credit circuit or mark-to-market accounting. It needs only to see that its retirement account went up.
What regulators are not seeing
Current regulatory attention focuses on the competitive structure of AI markets. The FTC has documented the cloud-AI partnerships. Senators Elizabeth Warren and Ron Wyden have argued that these partnerships function as de facto mergers. The Department of Justice and FTC’s joint inquiry on competitor collaboration guidance (Docket No. ATR-2026-0001), to which I submitted public comment, is examining non-traditional transaction structures including cloud credit investments.
These are important interventions. But they share a common analytical horizon: market power and competitive structure. They do not address the accounting mechanism that converts the circuit’s self-referential dynamics into reported earnings.
Three measures would help.
First, the SEC should require cloud providers to separately disclose the portion of their reported earnings derived from mark-to-market revaluations of entities with which they have commercial cloud relationships. The SEC already has authority under Regulation S-K to mandate segment disclosures. When the entity whose revaluation generates the gain is also a major customer whose spending constitutes revenue for the investor, the standard disclosure framework is inadequate. A targeted interpretive release or comment-letter guidance requiring disaggregation of such gains—similar to the enhanced disclosure requirements the SEC adopted for stock-based compensation in recent years—would give investors the information they need to distinguish operating performance from circuit-generated earnings.
Second, the FASB should revisit whether ASU 2016-01’s requirement to flow unrealized gains through net income is appropriate for equity holdings in entities with which the holder has a commercial relationship. The original rule addressed a real problem: historical-cost accounting allowed companies to obscure their true economic exposure, and the 2008 Financial Crisis demonstrated the consequences. The goal should not be to return to historical cost, but to ensure that the transparency gains of fair-value accounting are not undermined by the blending of commercially entangled revaluations with operating income. At minimum, requiring separate presentation—below the operating income line, clearly labeled as gains from commercially related entities—would preserve the information value of fair-value measurement while preventing mark-to-market gains from blending into operating performance.
Third, the Financial Stability Oversight Council should examine whether the interaction between cloud credit investments, mark-to-market accounting, and passive index flows constitutes a systemic risk. Each element is individually. Their combination is not. The FSOC’s mandate is to identify risks that emerge from the interaction of individually regulated components. Specifically, FSOC should assess what happens if the circuit reverses: a flat or down valuation round for a major AI startup would trigger mark-to-market losses that flow through the same income statements, depress the same earnings-per-share figures, reduce the same market capitalizations, and cascade through the same index funds that currently transmit the gains. The concentration of passive retirement assets in a handful of companies whose earnings depend on a self-referential investment circuit creates the kind of correlated exposure that FSOC was established—after the 2008 crisis—to identify before it materializes. A formal assessment under its systemic risk authority, analogous to its reviews of asset management and non-bank financial intermediation, would be a proportionate first step.
Conclusion
The cloud credit circuit already converts investment into revenue. The mark-to-market gain converts investment directly into earnings. Household retirement savings absorb the inflated valuations through passive indexation. Each step is legal. Each step is individually rational. The aggregate result is a self-referential system in which investment, revenue, earnings, valuation, and household wealth reinforce each other without requiring the underlying technology to generate commensurate returns. The risk is a bubble that, once popped, likewise blows up Americans’ savings.
The accounting standards that enable this circuit were designed for legitimate purposes—transparency, timely disclosure, accurate representation of economic exposure. The argument here is not that fair-value accounting is wrong, but that its application in the context of commercially entangled strategic investments produces outcomes its architects did not anticipate: a loop in which a company’s own commercial commitments generate the valuation gains that it then reports as profit.
The question facing regulators is whether the financial architecture underwriting the AI industry has become sophisticated enough to generate its own evidence of success, and whether the accounting standards, securities rules, and retirement structures we rely on can distinguish manufactured performance from real economic value.
Author Disclosure: The author reports no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
