On Tuesday July 29, 2026, Microsoft closed its fiscal 2026 by reporting a $90.0 billion quarter, $35.8 billion in net income, and a specific pair of numbers investors are still digesting. The Anthropic investment produced a $3.2 billion gain in the three months ending June 30 — worth $0.33 in diluted earnings per share. The OpenAI investment, on a stake Microsoft has been building since 2019 and which sits at roughly 27 per cent, contributed $0.07 to diluted EPS for the same quarter. Eight months of Anthropic, marked to Microsoft's own valuation, delivered roughly five times the Q4 net-income impact of the OpenAI relationship — and closed most of the way to what the OpenAI position contributed to the full year ($0.67, or about $5 billion).

That is a sentence that costs $5 billion in cash and $30 billion in prepaid Azure services.

What Microsoft actually said

The Anthropic gain is on page one of the press release: results "exceeded forward guidance and included a $3.2 billion gain on the investment in Anthropic that was made during the quarter reflecting fair value adjustments." CFO Amy Hood surfaced the same figure on the earnings call as one of the reasons the print beat guide. The mechanical description matters. Microsoft took its Anthropic stake — $5 billion of cash, put in during November 2025, alongside a $30 billion Azure services purchase commitment that Anthropic will draw down over the life of the relationship — and reset the carrying value based on Anthropic's most recent private-market round. The $3.2 billion is a mark-to-market, not a cash flow.

The OpenAI line is quieter and stranger. Microsoft's release says Q4 net income "included a $480 million gain from OpenAI investments, versus a $1.575 billion loss in Q4 FY2025" — a $2 billion year-over-year swing on the equity-method flow, which is what net income captures. Separately, multiple trade press readouts describe the same quarter as containing a "roughly $600 million write-down" on the OpenAI carrying value. Both can be true at the same time: Microsoft can be recognising its share of OpenAI's improving operating results through equity method while separately adjusting the balance-sheet carrying value of the stake if the internal valuation model has moved. The 10-K, when it files later this year, will show the reconciliation. What is not in dispute: whatever OpenAI contributed to Q4 was much smaller than Anthropic did, and Microsoft's public framing treated OpenAI as, per the headline that ran everywhere, "a mixed bag."

The Anthropic mark did not appear from nowhere

Some context on the mark. Microsoft's $5 billion cash investment in Anthropic in November 2025 was made at a headline valuation the company has never disclosed on the record, but which trade press pegged in the $150–170 billion range at the time. Anthropic ran a further raise this spring — the Series H at $65 billion new capital and a $965 billion post — which is the primary mark Microsoft is using here. If the November 2025 stake was struck at $160 billion post and the mark is at $965 billion post, the arithmetic on a $5 billion cheque produces something like a six-fold paper gain; the $3.2 billion Microsoft actually took to income likely represents a partial recognition of that with tax and hold-period discounting applied.

The OpenAI equivalent looks very different. Microsoft's stake in OpenAI — worked up in roughly $13 billion of tranches since 2019 and restructured through the pre-IPO conversion in fall 2025 — is now carried at whatever internal model Microsoft applies to a company whose valuation has moved through several regimes in the last twelve months. The $600 million write-down (if the trade press reading is right) is a small percentage haircut on a very large carrying value; the $480 million equity-method gain is Microsoft's share of OpenAI's actual operating result flowing through GAAP. Neither number is bad. Both are dwarfed, this quarter, by the Anthropic mark.

The shape of the strategy Nadella keeps naming

Nadella spent the two weekend days before the print pre-positioning the story. On Sunday's Fareed Zakaria GPS — his July 27 CNN appearance — the CEO told enterprise buyers that "any firm that doesn't have this control, I will claim will not remain a firm, because you've essentially outsourced your thinking." On the earnings call itself he doubled it: "You can't depend on any one model." He named the number to back it up: multi-provider customers on Azure grew fivefold since the start of the year. He described the architecture Microsoft is selling: "Keep your harness separate from the model — any model at any given time is swappable." The Azure catalog now lists over 11,000 models from OpenAI, Anthropic, Mistral, xAI, DeepSeek, and Microsoft's own MAI family, plus the long tail.

This is not the language of the largest OpenAI shareholder describing a monogamous partnership. It is the language of that same shareholder selling every customer a way out of any single-vendor dependency — and pointing to its own books for evidence that the strategy is working. The tell is what Microsoft chose to highlight, and in what order. On the call, Hood led with the $3.2 billion Anthropic mark. She did not lead with the OpenAI numbers. That is a real editorial decision, made by a CFO who does not make editorial decisions by accident.

The Copilot line, and why it lands here

Buried in the same release: Microsoft 365 Copilot crossed 30 million paid seats. Copilot has been the single most watched product metric in the industry — the leading indicator on whether enterprise AI ever comes back on the return curve. Thirty million paid seats matters not because it is huge (thirty million out of roughly 400 million M365 commercial seats is still under 10 per cent penetration), but because it is finally moving. Two years of squinting-at-the-slide has ended.

Copilot is also the product that most benefits from Nadella's harness-not-model architecture. Microsoft made Claude models selectable inside Copilot CLI in Q3 — Rajesh Jha, EVP for Experiences + Devices, called Copilot CLI the "default and preferred coding assistant" while confirming Claude stays available as a selectable model — and the same routing logic sits inside enterprise Copilot for text tasks Anthropic wins on. The $30 billion Anthropic Azure-services commitment is, among other things, the settlement mechanism for that traffic: Anthropic gets the revenue for the tokens, Microsoft gets the margin on the harness, the memory, the M365 integration, the enterprise contract. Any enterprise architect who told their board in 2024 that Copilot was a single-vendor decision has now been given two consecutive quarters of financial evidence that Microsoft's answer is no.

The circular-financing critique, with a receipt

The pattern where a hyperscaler invests cash into a lab, the lab commits to buy cloud services back, and the hyperscaler books the appreciation on its stake has been the deciding critique in the AI-business trade press for a year. Q4 FY26 is the first quarter where Microsoft's own books put a specific dollar figure ($3.2 billion) on the Anthropic version of that pattern. Whether that is a healthy accounting recognition of value creation, or a self-referential mark that will look silly in a downturn, is now an empirical question with a Microsoft-authored answer. It is also a question every other hyperscaler with an AI-lab stake is quietly staring at: Amazon marks Anthropic on its balance sheet through a different mechanism, Google marks its own DeepMind and A24-adjacent bets through yet another, and none of the rest have yet booked a comparable single-quarter mark from a private-lab position. Microsoft went first.

What this means

Three takeaways.

1. OpenAI is no longer the AI story Microsoft leads on. For most of 2024 and 2025, the OpenAI stake was the AI narrative on Microsoft earnings calls. This quarter, Anthropic was the AI story, and OpenAI was — per Hood's own framing — recovering. That framing shift is doing more work than the numbers themselves. It signals, to every customer procurement committee, that Microsoft's own investor communication no longer requires OpenAI to be the hero of the paragraph.

2. The "harness separate from the model" sentence is now a strategy with financials. Nadella has said the sentence in four venues in two weeks — CNN Sunday, the Investor call, the analyst Q&A, and a follow-up in Fortune's Azure milestone piece. Customers who ignored it in 2024 (when it read as hedging) or in early 2025 (when it read as posturing) now have to reckon with it as accounting reality. Microsoft is telling its enterprise base that model-swap-ability is the load-bearing architectural decision, and its own books are showing why.

3. The circular-financing story got a specific number. Not a thesis, not a spread, not a "some analysts warn." Three-point-two billion dollars, ledger-line, EPS-attributable, in fair-value adjustment on a stake that closed eight months ago. Every subsequent hyperscaler earnings call — Alphabet next, then AWS at the end of October — is now going to be read against this line item. The industry has a benchmark. Microsoft set it.

What to watch: the Q1 FY27 print on October 22 or 23. If the Anthropic mark holds or moves up on the next round, and OpenAI's carrying-value question resolves without another write-down, the story stops being one quarter and starts being the shape of the pipeline. If Anthropic marks down — because a follow-on round prints below the last, or because the S-1 process changes the valuation basis — Microsoft has just given the market a very specific number to short against. Either way, this is the quarter where Microsoft chose to make the Anthropic column readable to anyone with a calculator, and the OpenAI column readable as "a mixed bag."

That is a sentence that costs $30 billion in Azure services.