Pichai and Jassy are marking up their AI venture stakes and calling the paper profit.
It is true, in the narrow sense in which accounting standards usually mean things, that the rules required Alphabet and Amazon to do it. Under U.S. generally accepted accounting principles — specifically Accounting Standards Codification 321 — companies mark equity investments they do not consolidate, jointly control, or exercise significant influence over to fair value each reporting period. Fair value, absent a public market price, is whatever the most recent financing round said it was. Anthropic’s last round said something large. SpaceX’s, separately, said something larger. The accounting gets you there. Data centers, to put it in the same register, do not in fact materialize from the ether, and the cloudscape is in fact quite expensive to build.
The trouble is what the accounting is being asked to do, and what the analysts covering these companies are allowing it to do. Alphabet reported last quarter that 71 percent of its quarterly profit came from “other income,” almost entirely the markups. Amazon’s number was 66 percent. The Wall Street Journal’s Jonathan Weil did the arithmetic: strip the gains out, and the S&P 500’s earnings growth over the trailing four quarters drops from 31 percent to 24 percent — still spectacular, materially different. These are not operating earnings. They are paper gains on bets the parent companies made on AI labs whose entire valuation premise is the same AI buildout that Alphabet and Amazon are also financing through capex on their own data centers. The bet, the gain, and the buildout are the same trade viewed from three different ledgers. The company that posted the $112 billion Q2 profit the market read as a beat — and the cloud-revenue surge that gave Wall Street its answer on AI returns the week before — is now telling the same market that three-quarters of the profit number is paper.
The consistency Wall Street applies in deciding what is real and what is paper is — to put it in the polite Canadian register — selective. Nvidia, last quarter, reported $58.3 billion of GAAP net income and pointed analysts to $45.5 billion of adjusted earnings after excluding unrealized investment gains. Analysts, sensibly, took the adjustment. Broadcom reported $9.3 billion of GAAP net income and $12.1 billion of non-GAAP earnings, the difference being that management had stripped out stock-based compensation and amortization of intangibles, both of which are recurring operating costs that recur every quarter and are paid in real money to real engineers. Analysts, dutifully, ignored the real money and counted the higher number. FactSet counts at least 65 S&P 500 companies for which Street earnings exclude stock-based pay. Nvidia did the conservative thing — same accounting rule applies — and explicitly flagged investment gains as not part of a recurring earnings stream. Alphabet and Amazon, holding comparable stakes in comparable companies, did not. The analyst practice is not a rulebook. It is a series of negotiations between management and the sell-side desks whose business depends on the company. The companies that want the gains included get them included. The companies that don’t, don’t.
The asymmetry is the column. Real labor costs — the people who write the code and train the models — are subtracted as one-time. Paper gains on private-company bets whose valuations exist because other people have agreed to mark them up are counted as recurring. If a market multiple gets applied to the result, the equity gets valued at twenty-seven times earnings that the company itself would not, by its own reasoning, defend as a recurring earnings stream. The whole apparatus depends on the AI buildout narrative holding up long enough that the marks stop going down. They will eventually stop going up.
What the marks are saying right now is that the AI capex cycle is being financed, in part, by the paper wealth the AI capex cycle itself is supposed to create. Alphabet, Amazon, Microsoft, and Meta are on track to spend closer to $700 billion on AI infrastructure and data centers this year, by most industry tallies. The earnings they book on their AI venture stakes are, in aggregate, large enough to cover roughly one-sixth of what the same companies are spending. The accounting is, in a strict and slightly obscene sense, the capex finding its own revenue stream. The capex is also the thing that is supposed to justify the markup. The cycle is closed.
What makes it a bezzle, and not merely aggressive accounting, is the recognition event. The term is Galbraith’s, from The Great Crash, 1929 — the interval in which the gains are taken and the corresponding losses have not yet been recognized, because the recognition mechanism is the next financing round, which has not yet been printed at a lower number. The bezzle ends when the trigger arrives. The trigger is most likely to be a public-market rerating of the AI capex cycle itself — because the marks presuppose continued capex appetite, and a rerating therefore arrives before the cycle breaks, not after. At that point Anthropic’s next round comes in flat or down, or an exit event — an IPO, a strategic sale, a forced liquidation — values the stake below the carrying value. The paper gain reverses, the “other income” goes negative, the consensus adjusts, and the analyst class will discover, suddenly and with great conviction, that these were never recurring earnings. The discovery will be expensive. The S&P 500 at 27 times earnings is the index that pension funds and retail investors actually own. They will own it on the day the recognition event arrives.
The sell-side analyst class is where the standard gets enforced, and the standard is enforced by negotiation, not by rule. Analysts work for brokers whose revenue comes from the companies they cover — from underwriting, from trading volume, from the accretion of relationships that produce the next deal. The analyst who dissents on the Street earnings calculation is the analyst who loses the mandate. The first three of the four forces Cory Doctorow uses to describe enshittification — competition, regulation, and self-help — are basically absent on the question of whether to include unrealized investment gains in recurring earnings. The fourth, labor, is structurally weak because the analysts who would dissent do not have the protection of a union, and the value of their labor is computed against the value of getting the next mandate. The net result is that the figure most investors see is the figure the company wants them to see, dressed in the syntax of professional consensus.
In Canada, this is a familiar mechanism dressed in different clothes. Calgary and Vancouver resource companies have been marking up oil-and-gas and mineral properties to fair value for decades — paper gains during commodity booms treated as recurring, real labor and capital costs treated as exceptional. National Instrument 43-101, brought in after the Bre-X scandal, added disclosures and changed nothing material. The discipline of an independent audit is the only check that has ever worked on mark-to-narrative accounting, and the current consensus metric does not allow it, because the consensus metric is negotiated between the analyst class and the company class, and an independent audit would produce a number neither side wants.
It is worth being precise about who is harmed when the marks go the other way. The shareholders, eventually, when the gains reverse and the consensus catches up. The workers, immediately, because the labor cost is what got stripped first and is what gets cut first when the cycle turns. The customers and end-users of the AI products the buildout is financing, in the form of higher prices for service, lower quality, and the predictable decay that follows when a market consolidates around capital structures that depend on continued markups. The pension funds and index investors who own the S&P 500 by definition and who are being told, by the 27-times-earnings number on the screen, that the broad market is reasonably priced. The market is not reasonably priced. The market is being priced on a denominator that excludes labor costs and includes paper bets.
The fix is older than this column. Strip stock-based compensation back into earnings. Strip the investment markups out of earnings. Use GAAP. Tell investors what the company made, before and after the financial engineering. The Securities and Exchange Commission could require a non-GAAP-to-GAAP reconciliation that disaggregates investment-markup gains from operating items. The Financial Accounting Standards Board could amend ASC 321 to require disclosure of carrying value versus the most recent financing round’s implied valuation, with vintage dating. The auditor of each company could add a critical-audit-matter paragraph on mark-to-fair-value estimates for Level 3 private holdings. None of them will, because the marks are technically correct and the technical correctness is the shield.
Until then, the next time a tech executive stands in front of a microphone and tells the market that the company is on track to deliver double-digit earnings growth, the listener is entitled to ask which earnings, and which double-digit, and whether the 71 percent at Alphabet or the 66 percent at Amazon is the number that actually moves when the company is run. As the trade union my father belonged to would have put it: that is the bookkeeping the bookkeeping committee decided to keep. The departures from truth are not in the next quarter’s numbers. They are in the present.