71% of Alphabet's Pretax Profit Was Paper Gains • The Weighted Average

Alphabet reported $97.98 billion of non-operating income in the quarter ended June 30 against $138.75 billion of pretax profit, per the company’s Form 10-Q as tagged in SEC XBRL data. That is 71% of pretax profit arriving from outside operations. A year earlier the same line was $2.66 billion — a 37-fold increase, and a number that should change how any operator reads a cloud vendor’s balance sheet before committing to a multi-year contract.

Amazon’s filing tells a parallel story. Its June-quarter 10-Q reports $53.4 billion of non-operating income against $80.86 billion of pretax profit — 66% — up from $1.69 billion in the year-ago quarter. Neither company earned that money selling anything. As reporting on the Financial Times’ $160 billion tally documents, Alphabet’s gain traces mostly to SpaceX and one unnamed private company widely understood to be Anthropic, while Amazon’s filing attributes its own to roughly $50.5 billion of upward adjustments on Anthropic preferred stock after observable price changes in Anthropic’s funding rounds.

Why a buyer should care about someone else’s accounting

The obvious reading is a market-quality story: record profits that are not cash. The operator reading is narrower and more useful. When two of the largest AI infrastructure providers derive the majority of their reported pretax profit from marks on private AI companies, the earnings figures a procurement team uses to judge vendor stability have stopped measuring vendor stability.

Work the ratio the other way. Strip non-operating income and Alphabet’s pretax profit falls to about $40.8 billion, Amazon’s to about $27.5 billion — both still enormous, both a fraction of the headline. That is the number to use when assessing whether a vendor can absorb a downturn in the business you actually buy from, because a lockup-restricted stake in a private company cannot fund a data center or honor a price commitment. This paper’s tracking of the three-trillion-dollar data-center buildout assumed capex would be funded from operations and debt; a growing slice of the reported profit backing those plans is a mark that reverses if the next funding round prices lower.

Nvidia shows the same volatility from the opposite direction, which is the point. Its own non-operating line ran $16.37 billion in the April quarter and then $7.77 billion in the quarter ended July 26, per the figures in its Form 10-Q — a swing of more than half in three months on a line that has nothing to do with selling accelerators. A number that can halve in a quarter without any change in the underlying business is not a measure of the business.

The circularity is the part with teeth. Cloud providers invest in model companies, model companies spend the proceeds on cloud and chips, and each subsequent round lets the investor mark up the earlier stake. Every leg is legitimate under the accounting rules and every leg depends on the next round clearing at a higher price. The paper found the same loop in physical form when it priced OpenAI’s Ohio warrants at $688 million per gigawatt: the tenant, the landlord, the chip vendor, and the parent all hold positions in the same asset.

What would break the read

The strongest counterpoint is that unrealized does not mean unreal. Alphabet’s SpaceX position followed an actual June IPO, which is a public price, not a model. Amazon’s Anthropic markup followed observable transactions in real funding rounds. Under US GAAP these are the correct entries, and calling them fictional is lazy.

Nor is the exposure evenly distributed. Microsoft’s fiscal-year non-operating income was $10.7 billion against $165.93 billion of pretax profit in its Form 10-K for the year ended June 30 — about 6% — which is what a conventional quarter looks like and a reminder that this is a company-specific risk rather than a sector-wide one. Buyers can therefore price it: two of the largest cloud vendors carry a large equity-mark dependency and one does not, and that distinction belongs in a vendor scorecard alongside uptime and regional coverage.

The second counterpoint is duration. A mark only reverses if the underlying reprices, and both Anthropic and OpenAI have moved toward listings that could validate current valuations rather than puncture them. If they list at or above their private marks, the gains harden into something closer to the real thing.

What would change the verdict is a single data point: the first major AI lab to list below its last private round. That event converts the same balance sheets from write-ups to write-downs, and the quarter in which it happens will reveal how much of the sector’s reported profitability was reflexive. Until then, operators should do three things. Read the operating income line, not net income, when scoring vendor health. Ask cloud vendors whether committed discounts are underwritten by segment operating margin or by group profit. And treat any vendor whose pretax profit is majority non-operating as carrying an equity-market dependency in addition to its business risk — the same discipline today’s lead applies to pricing structures where the visible price and the borne cost have separated, and that the paper applied when only 6% of firms could show a real AI earnings impact.