Aya Research
DEEP RESEARCH
中EN

Microsoft: It Built an Empire on Someone Else's Technology — Now Someone Else Is Doing It Back

Panoramic research report · longitudinal history + cross-sectional rivalry + synthesis Subject: Microsoft Corporation (NASDAQ: MSFT) · Fiscal year ends June 30 Report date: 2026-07-05 · Data cutoff: FY2026 Q3 results (quarter ended 2026-03-31, reported 2026-04-29) + news flow through 2026-07-05 (FY26 Q4 reports late July) Sources: company IR and SEC filings (10-Q/10-K/8-K), Microsoft and OpenAI official blogs, CNBC, Fortune, The Register, Synergy Research, Futurum, Recon Analytics, Yahoo Finance For information and research purposes only. Not investment advice.

Before we start

In 1981, Microsoft sold IBM an operating system it had bought for $75,000 — and kept the right to license it to everyone else. IBM built the market; Microsoft collected the tax. It remains the most consequential contract clause in business history, and it built a company that, forty-five years later, earned $101.8 billion in a single year — the first software company ever to clear a hundred billion dollars of annual profit.

In 2026, that company is the worst-performing stock in the Magnificent Seven: down roughly a quarter from its highs, some $850 billion of market value gone, slipped from second place to fourth, below the $3 trillion line it first crossed in January 2024. The business did not shrink — revenue grew 15% last fiscal year, Azure is reaccelerating at 40%, and an AI business the company says runs at $37 billion annualized more than doubled. The market is not punishing Microsoft's results. It is punishing a structure: the slow-motion unwinding of the most celebrated partnership in technology.

Because here is the symmetry the market has noticed. In its deal of the century, Microsoft supplied the crucial layer (DOS), kept it non-exclusive, and taxed the platform its partner built. In the OpenAI relationship — $13 billion invested for roughly 27% of the company, IP rights, revenue shares, and once-exclusive cloud rights — Microsoft cast itself as the tax collector again. Then, across eighteen months and two renegotiations, OpenAI did to Microsoft what Microsoft did to IBM: kept its technology portable, went multi-cloud (a $300 billion Oracle contract, CoreWeave, eventually anyone), capped Microsoft's revenue share at $38 billion, deleted the AGI clause, converted the IP license to non-exclusive, and reduced its contribution to Azure's revenue to under 20% — while extracting a $250 billion Azure purchase commitment that keeps Microsoft building data centers for its now-free partner. The apprentice read the old contract. The question this report exists to answer is whether the market's $850 billion fine for that reversal is a correct sentence or — as with this company once before, in a decade everyone misremembers — a category error about a business that was quietly compounding the whole time.

One more framing before the history, because it disciplines everything after: Microsoft is the only company in this series that has already lived a complete cycle of the exact situation it now occupies. It has been the underpriced supplier who kept the rights (1981). It has been the convicted monopolist who survived the sentence (2001). It has been the cash machine the market refused to pay for (2003–2013). It has been the despised laggard that converted a platform shift into leadership (2014–2021). And it has been the partner whose exclusive advantage dissolved by contract (1985–1990, when IBM and Microsoft's OS/2 partnership unwound — the forgotten first divorce, which Microsoft also survived by owning the asset that mattered). The company is, in effect, a complete reference library of its own current predicament. What the library does not contain is a case where Microsoft won a product war without a distribution advantage — which is precisely the war Copilot is losing. That asymmetry between what history predicts (survival, compounding, eventual repricing) and what it doesn't (product-led victory) is the honest boundary of the bull case.

Part One is the longitudinal history: the license, the overturned death sentence, the "lost decade" that tripled revenue, the Nadella refounding, and the OpenAI arc from $1 billion to 27% to the great unwinding. Part Two is the board in July 2026: the cloud receipts (with their one enormous asterisk), the $37 billion AI ledger and its passthrough problem, the Copilot seat-versus-usage scissors, the $190 billion capex cliff, the silicon gap, and the empire's quieter provinces. Part Three synthesizes: the snapshot, the contract read as this series reads capital-markets behavior — as confession — scenarios, and the falsifiable claims. For series readers, the through-lines connect directly: Microsoft is the second of the four hyperscaler "tenants" our Meta report frames, the $25 billion of memory-price inflation inside its new capex guide is our SK hynix report's revenue arriving on a buyer's income statement, and the OpenAI unwinding is the CoreWeave report's counterparty-quality question played at civilizational scale.

Part One · Longitudinal: the license, the sentence, the decade, the bet (1975–2026)

1. $75,000 and the clause (1975–1981)

Microsoft was founded in April 1975 in Albuquerque to sell BASIC for the Altair — software as a product, at a time when software was a giveaway that came with hardware. The founding insight was already the whole company, stated in Gates's famous open letter to hobbyists the following year: the value would migrate from the machine to the layer everyone had to license, and licenses deserve to be paid for.

Six years later came the transaction. IBM, racing to ship the PC, needed an operating system. Microsoft didn't have one; it bought 86-DOS — "Quick and Dirty Operating System" — from Seattle Computer Products for $25,000 (non-exclusive, December 1980), then $50,000 more for full rights in July 1981, weeks before the IBM PC launched. Total: $75,000. The masterstroke was not the purchase but the terms Microsoft gave IBM: a non-exclusive license. IBM assumed the hardware was the product; Microsoft understood the standard was. When Compaq and a hundred clone-makers reverse-engineered IBM's machine, every one of them needed MS-DOS, and Microsoft — not IBM — owned the tollbooth on the platform IBM had built and evangelized at its own expense. Revenue flowed to the licensor; commodity margins to the platform's patron.

Fix the structure of that trade in mind — supplier keeps portability, patron funds the buildout, tax flows to whoever owns the scarce layer — because this report will watch the same trade executed twice more: once by Microsoft against IBM, and once, forty-five years later, against Microsoft.

It is worth asking why IBM agreed to terms that look, in hindsight, like corporate malpractice. The answer is the timeless one: the dominant player priced the layer by its current economics. Operating systems in 1980 were low-margin plumbing that shipped with hardware; IBM's PC was a skunkworks project expected to sell a quarter-million units; exclusivity over plumbing wasn't worth negotiating for. IBM wasn't stupid — it was extrapolating. Every subsequent episode of this pattern, including the one in Section 6, features an incumbent making the same move: pricing a supplier's layer by what it earns today rather than what it becomes when the platform explodes. Microsoft in 2019 valued OpenAI's models as a feature supplier for Azure and Bing. The models became the platform. The clause-writers switched chairs.

2. The double monopoly and the overturned death sentence (1986–2001)

The March 1986 IPO priced at $21 and valued the company at $777 million; ten thousand dollars subscribed that morning is worth on the order of fifty million today, a compounding record only NVIDIA's has rivaled. Windows 95 plus Office built the most profitable franchise in software history, with PC operating-system share above 90% — and, less remarked, built the distribution machine that is the true constant across every Microsoft era: the enterprise agreement, the volume license, the account team inside every Fortune 500 IT department. Products came and went (some catastrophically); the machine that could sell whatever Redmond shipped next never stopped compounding, and it is the asset actually being wagered in the Copilot bundle today. Then Microsoft became the only member of today's Mag 7 to actually receive a corporate death sentence: in June 2000, Judge Jackson ordered the company split in two — operating systems and applications — as the remedy for illegally maintaining its monopoly. A year later the D.C. Circuit upheld the liability finding but vacated the breakup (aided by the judge's own misconduct in press interviews), and the 2001 DOJ settlement imposed conduct rules instead.

Two consequences echo into 2026. First, the near-death by remedy made Microsoft institutionally cautious for a decade — slower, more lawyered, defensive precisely during the years the internet's next platforms were being founded. Second, it established the base rate this series keeps meeting (Google's remedies in 2025, Meta's FTC acquittal): US courts, faced with the actual moment of dismembering a tech giant, have blinked every single time. Investors pricing breakup tails across the Mag 7 should weight that unbroken record.

There is a third, subtler echo: the case taught Microsoft that the browser fight it won destroyed more value than the fight was worth — Netscape died, but the scorched-earth bundling that killed it produced the trial, the lost decade's caution, and a generation of talent that went elsewhere. The institutional memory shows in how differently the company fights now: Teams-versus-Slack was won with the same bundling play but settled with Brussels before trial; the Copilot bundle of 2026 arrives pre-lawyered with unbundled EU pricing from day one. Microsoft learned to pay the toll upfront. Whether regulators eventually decide the toll is too cheap for the conduct is the live question inside the FTC's slow-grinding probe — but the pattern explains why, uniquely among the Mag 7, Microsoft's 2020s regulatory record consists entirely of settlements and none of verdicts.

3. The decade everyone misremembers (2000–2014)

The Ballmer era is filed under "lost decade," and by platform scoreboard it was: search conceded to Google (Bing arrived in 2009, a decade late), mobile conceded to Apple and Android (Windows Phone's obituary was the $7.2 billion Nokia acquisition of 2013, written down for more than its purchase price — $7.6 billion — two years later), social never attempted, Vista a five-year engineering fiasco. The stock spent fourteen years flat, ending around $30.

Here is what the scoreboard leaves out: revenue tripled ($25 billion to $78 billion) and net income nearly tripled ($7.8 billion to $21.8 billion) across those "lost" years. Office and Windows compounded; the enterprise relationships deepened; Azure was founded (2008, code-name Red Dog, GA 2010) inside the era the narrative writes off entirely — as were, for completeness, the Xbox franchise, the server-and-tools business that became Intelligent Cloud's chassis, and SharePoint/Dynamics/the enterprise stack that later carried Teams and now carries Copilot. The lost decade built most of the vehicle the found decade drove. The market's fourteen-year verdict was not that Microsoft earned too little — it was that earnings without a claim on the next platform deserve no multiple. That is the exact sentence being handed down again in 2026, which makes the Ballmer decade this report's most important exhibit: the last time the market applied this specific punishment to this specific company, the punished earnings were real, the missed platforms were real, and the company was simultaneously incubating — unnoticed, unpriced — the second act that would octuple the stock. The lesson is not that the market was wrong; Windows Phone deserved its zero. The lesson is that this company's conglomerate breadth makes single-narrative pricing chronically unreliable, in both directions.

4. Nadella: the refounding (2014–2019)

Satya Nadella's February 2014 appointment — an insider, the Azure architect, the board's third choice by most tellings — began what is fairly called a refounding. The moves were cultural before they were financial: "mobile-first, cloud-first" replaced Windows-first; Carol Dweck's growth mindset ("learn-it-all over know-it-all") replaced the stack-ranked knife-fight culture; and in 2018 the Windows division was formally dissolved as a power center. Office went to iPad; Linux went from cancer (Ballmer's word) to first-class citizen; acquisitions bought communities rather than technologies — LinkedIn ($26.2 billion, 2016) and GitHub ($7.5 billion, 2018) both thrive, against the Skype counterexample ($8.5 billion, 2011; shut down entirely in May 2025). Azure, four years behind AWS at GA, closed the gap the only way second movers can: through the enterprise install base, hybrid-cloud pragmatism, and salesforce muscle no startup cloud possessed. By FY2025 Azure alone passed $75 billion of annual revenue, growing 34%.

The refounding matters to the 2026 debate for one specific reason: it is the other historical precedent. Where Section 3 shows this company can compound unpriced for a decade, Section 4 shows it can convert a despised position (2014's "irrelevant Microsoft") into index leadership within five years when the platform bet lands. The market has seen both movies. Its current problem is deciding which one it is watching.

The mechanics of the Azure catch-up deserve a paragraph, because they are the template Microsoft is trying to rerun in AI. Azure never beat AWS on product; it beat AWS's sales motion. Microsoft arrived at every enterprise negotiation already inside the building — the Office renewal, the Windows Server estate, the Active Directory identity layer — and priced Azure into agreements customers were signing anyway, tolerating years of technical inferiority while the bundle did the acquisition work and the product matured behind it. That is precisely the Copilot playbook of Section 9: ship into the suite, price into the E3 agreement, let distribution buy time for the product. The strategy has one demonstrated success (Azure, Teams) and one structural requirement — that the product eventually does mature faster than customers' patience expires. Azure closed its quality gap in roughly five years. The Copilot clock started in early 2023.

5. The bet of the century (2019–2023)

The OpenAI relationship began as a $1 billion investment in July 2019 — widely read at the time as Nadella buying research credibility for Azure. It compounded: roughly $2 billion more in 2021, then the January 2023 "multiyear, multibillion" round (reported near $10 billion, total ~$13 billion) structured as a profit-share with capped returns rather than simple equity. What Microsoft actually bought, it turned out, was the launch position for the platform shift of the age: exclusive cloud rights to the maker of ChatGPT, IP access to frontier models it embedded across Office, Bing, GitHub and Windows under the Copilot brand, and — for about two years — the aura of having pre-purchased the future at venture prices while Google slept and Amazon watched.

Score the investment on its own terms before scoring the relationship: $13 billion deployed, of which $11.8 billion had been funded by March 2026, now marked at ~$135 billion — a paper multiple north of 10× on the largest single corporate venture bet ever made, achieved inside seven years. By pure capital-allocation arithmetic it stands with Yahoo's Alibaba stake and SoftBank's own Alibaba position among the greatest investments in history, and it was made not by a fund but by an operating company that also extracted, along the way, the product positioning that drove its 2023 +58% re-rating. Whatever Part Two concludes about the strategic aftermath, the 2019 decision itself belongs in the pantheon — a point worth fixing because the current narrative, obsessed with the divorce, has quietly stopped crediting the marriage.

November 2023 stress-tested the position. When OpenAI's board fired Sam Altman without warning Microsoft, Nadella spent a weekend arranging, in effect, a lifeboat reportedly budgeted at $25 billion — a Microsoft subsidiary ready to hire Altman, Brockman and any researcher who followed. Altman was reinstated in five days; Microsoft got a board observer seat. The episode was scored at the time as Nadella's masterclass. It reads differently with three years' hindsight: the weekend proved that Microsoft's most strategic asset was governed by a nonprofit board it could not control, staffed by people it did not employ, building technology it merely licensed. Everything in Section 6 follows from the players absorbing that same lesson — OpenAI concluding it needed independence from its patron, and Microsoft concluding it needed insurance against its partner.

The insurance policies, for the record, were taken out promptly and in every direction at once: Mustafa Suleyman hired (March 2024) to build Microsoft AI's own models — a program that by mid-2026 had shipped seven MAI foundation models including a trillion-parameter flagship, with Suleyman publicly describing the new OpenAI terms as Microsoft being "set free"; the Anthropic triangle (Section 11) signed; Grok, Llama and open-weights families onboarded to Foundry; and Maia silicon accelerated. Each hedge was individually sensible. Collectively they guaranteed the outcome they insured against: a partner watching its patron systematically de-risk the partnership concludes the partnership is ending, and negotiates accordingly. By the time the April 2026 amendment was signed, both sides were merely documenting a separation both had spent two years engineering. The divorce metaphor is imprecise in one respect — divorces end cohabitation. This one didn't: $250 billion of committed rent says the exes still share the house.

6. The unwinding (2024–2026)

The reversal ran in four dated steps. January 2025: OpenAI, with Oracle and SoftBank, announced the $500 billion Stargate program — and Microsoft's cloud exclusivity died in the press release, downgraded to a right of first refusal. October 2025: the grand restructuring — OpenAI became a public benefit corporation; Microsoft's accumulated position converted to ~27% (book value ~$135 billion); model and product IP rights extended to 2032; the AGI declaration (which once let OpenAI's board unilaterally terminate Microsoft's rights) moved to independent expert verification; Microsoft gave up the right of first refusal on compute; and OpenAI committed to purchase an incremental $250 billion of Azure — the number that let both sides declare victory. April 2026: the fine print turned decisively — OpenAI's ~20% revenue share to Microsoft capped at $38 billion through 2030; Microsoft's reciprocal payments to OpenAI ended; the IP license became non-exclusive; OpenAI became free to sell everything, everywhere, on any cloud; the AGI clause was deleted entirely. May–June 2026: the partners became open competitors — OpenAI launched a $4 billion-funded enterprise deployment company; Microsoft answered with a $2.5 billion "Frontier" services unit; Fortune reported Microsoft insiders admitting the company had "lost its way" in the assistant war; and Nadella handed the Copilot product to a 33-year-old ex-Snap executive with a mandate summarized by the press as save the AI strategy.

Overlay the stock chart on those dates and the market's fine tracks the unwinding almost tick for tick: +58% in 2023 (the exclusive-partner premium), +13% and +16% in 2024–25 (doubt creeping), roughly -24% year-to-date in 2026 (the premium extracted, with interest). Meanwhile Microsoft's own operating results, quarter by quarter through the same stretch, beat expectations nearly every time. The market is not grading the quarters. It is grading the clause — and it believes Microsoft signed away the platform tax it spent 1981 teaching the world to fear.

Before Part Two, one fairness note on the unwinding, because hindsight flattens it into a blunder and it wasn't one. Every clause Microsoft surrendered, it surrendered for consideration: the ROFR went in exchange for $250 billion of committed purchases; exclusivity's end coincided with the freedom to host Anthropic (Section 11) and build MAI; the revenue-share cap traded uncertain upside for $38 billion of certain cash and the deletion of the AGI clause — which had been, quietly, the most dangerous sentence in Microsoft's entire contractual universe, a provision under which a nonprofit board (the same one that fired Altman on a Friday) could have unilaterally terminated Microsoft's rights to its most strategic technology. Microsoft's negotiators traded a deteriorating exclusive position for cash, certainty, diversification freedom, and the removal of a doomsday clause. The alternative — litigating to hold an unwilling partner exclusive — is how the IBM-Microsoft relationship ended, and Microsoft knows exactly who won that one. The market's fine prices what was lost. Section 14 will price what was received.

Part Two · Cross-sectional: the board in July 2026

7. Azure: real acceleration, and a receipt with one enormous asterisk

Start with what is unambiguously working. Azure's constant-currency growth over eight quarters reads 30 → 34 → 31 → 35 → 39 → 39 → 38 → 39–40%: a genuine reacceleration at a scale (third-party estimates put the run-rate above $105 billion) where reacceleration is supposed to be impossible, guided to hold 39–40% in the quarter reporting this month. Microsoft Cloud in aggregate runs $54.5 billion a quarter growing 29%; commercial RPO — the receipts drawer this series audits at every company — stands at $627 billion, roughly doubled in a year, larger than Google's $462 billion and second to no one.

The competitive texture behind the headline: Azure's 40% at a ~$105 billion run-rate is arithmetically more impressive than Google Cloud's 63% at $80 billion (more absolute dollars added per quarter), yet the market narrates GCP as the winner — partly momentum bias, partly because Google's growth is visibly merchant (TPUs, Gemini, named external logos) while Azure's is partially related-party (the asterisk below). Against AWS the story inverts: Amazon's 19% at a $117 billion run-rate makes Azure the share-taker in the classic enterprise segment, where Microsoft's identity-and-compliance bundle keeps converting datacenter migrations that never considered anyone else. Azure's structural position — second in share, first in enterprise gravity, uniquely dual-frontier in models — remains the strongest in cloud. Its narrative position is the weakest of the three, and in this market the narrative sets the multiple.

Now the asterisk, which the company itself disclosed: management attributes ~45% of that RPO to OpenAI's commitments — substantially the $250 billion purchase agreement extracted in the October restructuring. Apply the audit standard we applied to Google's Anthropic concentration, and it bites harder here: nearly half the backlog is one counterparty; that counterparty is unprofitable at historic scale (Microsoft's own equity-method line implies OpenAI lost $11.5 billion in a single quarter); its commitment was negotiated as consideration in a divorce rather than won competitively; and — the circularity our Google report flagged, squared — Microsoft is a 27% owner of the customer whose future purchases constitute its own largest receipt. None of this makes the revenue fake: OpenAI genuinely consumes gigawatts, and its sub-20%-of-Azure current contribution means the reported growth is mostly everyone else. But it means Azure's headline backlog and its true diversification are different numbers, and the gap between them is precisely the kind of thing this series exists to say out loud. Strip the OpenAI component and the ex-OpenAI RPO ($345 billion, still growing fast) remains a first-class receipt — it is simply not the $627 billion on the slide.

8. The $37 billion question: AI revenue or AI passthrough?

Microsoft's AI business runs at $37 billion annualized, up 123% — the largest disclosed AI revenue line in the world, nearly triple the $13 billion disclosed eighteen months earlier. The bear case compresses to one sentence: how much of that is OpenAI renting Azure to serve ChatGPT — revenue Microsoft books as its own AI triumph but which is really its partner's business passing through its data centers, at margins compressed by the very partnership terms just renegotiated? The company does not disclose the split. What it has disclosed, across two years of calls: AI services contributed 8, then 12, 13, 16 points of Azure's growth — and then, from FY25 Q4 onward, Microsoft stopped disclosing the metric. In this series' experience (Meta's buyback line, hynix's price caps), the metrics companies stop publishing are usually the ones that stopped flattering.

The honest decomposition, from available fragments: GitHub Copilot (4.7 million paid seats, ~75% growth) is real, chosen, sticky product revenue. M365 Copilot (Section 9) is real but discounted and thinly used. Foundry hosting of third-party models — now including Anthropic's Claude, making Azure the only hyperscaler serving both frontier labs — is genuine merchant AI revenue with a future. And some large, undisclosed share is OpenAI consumption, which is simultaneously real cloud revenue and the least durable component, since OpenAI's multi-cloud freedom (April 2026) means every incremental ChatGPT token can now be served from Oracle, CoreWeave, AWS or Google — and OpenAI's Stargate buildout exists specifically to internalize it. The $37 billion is true. Its quality distribution is the debate, and Microsoft's disclosure choices suggest management knows which way the distribution leans.

Contrast the disclosure architecture with the two peers this series has already audited, because the three companies have chosen three different transparency postures toward the same question. Google discloses backlog and lets analysts find the Anthropic concentration; Meta discloses nothing and takes the unreceipted-capex discount in full; Microsoft discloses a headline ($37B, $627B RPO) engineered to impress while retiring every sub-metric (AI points of Azure growth, OpenAI's revenue share trajectory) that would let outsiders decompose it. Markets eventually charge for opacity at the rate of the worst plausible interpretation — that is the deep reason a company growing 18% with a $627 billion backlog trades at 20× forward. The single cheapest multiple-repair available to Microsoft's CFO is not operational at all: it is a disclosure — the OpenAI share of Azure, published quarterly, declining. If the diversification story is true, showing it costs nothing and reprices everything. That the company hasn't is either conservatism or a tell; the FY26 Q4 call this month is the standing opportunity to prove which.

9. Copilot: the seat/usage scissors

The enterprise AI assistant was supposed to be Microsoft's birthright — the Office monopoly converted into an AI annuity at $30 a seat. The scoreboard is genuinely double-sided. Official: M365 Copilot passed 20 million paid seats in the March quarter, adding 5 million in a single quarter, the fastest growth since launch, +250% year over year; per-user engagement up ~20% sequentially; from July 2026 Copilot folds into the standard E3 bundle (with a 13% price rise) — the classic Microsoft endgame of converting a product into a line item. Twenty million seats at list price would imply ~$7 billion of annualized revenue; the third-party estimates of $2–4 billion actual, implying realized prices around half of list, tell you the sales machine has been buying the adoption curve with discounts — normal for enterprise software land-grabs, but it means the real Copilot P&L test is not the current seat count, it is what happens to those seats at first renewal, at bundle pricing, after the pilot budgets expire. That test begins this month, which is why it is Signal 2. Independent: Recon Analytics' 150,000-respondent survey puts Copilot's share of paid-assistant usage at 11.5% and falling (from 18.8% six months earlier) against ChatGPT's 55%; where employees can use both, 76% choose ChatGPT; the have-access-to-weekly-use conversion runs 36% for Copilot versus 83% for ChatGPT; analysts estimate 20–30% weekly activity on sold seats, and third parties put real M365 Copilot revenue at $2–4 billion against list-price implications several times higher, thanks to pervasive 40–60% discounting.

Why does the product lose the head-to-head it was born to win? The reporting converges on three answers. Latency and model lag: Copilot historically ran a model generation behind ChatGPT because Microsoft's deployment pipeline added quarters, and knowledge workers notice a worse answer instantly. Context poverty, ironically: ChatGPT knows everything the user pastes into it; Copilot was supposed to win on tenant data (email, files, meetings) but enterprise permissioning kept its retrieval shallow — the moat data was fenced off from the moat product by the compliance apparatus that sells E5. And the interface tax: Copilot lives inside apps built for direct manipulation, while the assistant paradigm wants a blank page. None of these is unfixable; all are the kind of product debt that a company whose center of gravity is selling — not building — accumulates by default. Hence the June reorganization handing the product to an outsider, over the heads of the org that shipped it.

This is the same scissors we documented at Meta (1.2 billion reached, 40 million engaged) — distribution manufacturing adoption statistics that usage data refuses to ratify — except Microsoft is charging for its side of the scissors, which makes the eventual reconciliation a renewal-cycle event rather than an abstraction. The bundling move is the honest strategic response: if the product cannot win the choice, remove the choice — make Copilot part of the suite's price the way Teams became part of the suite (a move that killed Slack's growth and drew, note, an EU case Microsoft settled). It may well work; suite economics usually do. But investors should be clear-eyed that the 20-million-seat line measures Microsoft's selling power, the 11.5% line measures the product, and the company's future pricing power in AI depends on the second number converging up toward the first — under competition from the best-funded product company in history, which happens to know everything about Microsoft's stack, because until recently it powered it.

10. The capex cliff — and the memory bill arrives

Microsoft guided calendar-2026 capex to roughly $190 billion including finance leases — up ~61%, some $35 billion above street consensus — with two details that connect this report to the rest of the series. First, about $25 billion of the increase is component price inflation, principally memory: the DRAM and HBM repricing our SK hynix and Micron reports documented from the sellers' side is now a disclosed, guided line item on the largest buyer's budget. The memory supercycle has crossed the aisle from thesis to invoice; when Microsoft's CFO cites memory prices as a capex driver, the "single organism" now includes the customers. Run the loop once to see the machine whole: Microsoft guides capex up $25 billion for memory prices → that guide is the demand signal hynix and Micron cite when negotiating 2027 HBM contracts up → those contracts justify the memory makers' own record capex → which orders the equipment and wafers that eventually break the shortage → in 2028, on schedule, at whatever demand level then prevails. Every party is behaving rationally; the loop as a whole is the cycle. Our memory dashboard watches it from the sellers' seats; this line item is the first time a buyer has printed the loop's cost in its own guidance, and it will not be the last — Meta's next guide and Alphabet's will carry the same inflation, disclosed or not. Second, roughly two-thirds of the spend goes to short-lived assets (GPUs, CPUs) on 4–6 year depreciation schedules — the same depreciation wave we quantified at Meta, arriving here against a gross margin already guided down to ~64% from a historical 68–70%.

The cash-flow arithmetic is the bear case's hardest exhibit: FY2025 free cash flow fell 3% to $71.6 billion; in the December quarter, FCF compressed to $5.9 billion — a company generating $35.8 billion of quarterly operating cash flow spending essentially all of it on infrastructure, with capex/revenue touching 47%. Fund a $190 billion program from a ~$140 billion operating cash flow and the difference comes from somewhere: finance leases (total debt including leases now ~$125 billion, the leases being the fastest-growing "borrowing"), slower buybacks, or the bond market. Unlike Meta, Microsoft still pays a growing dividend and repurchases stock ($10.2 billion combined last quarter) — but the direction of travel is identical across every tenant this series covers, and October's synchronized capex disclosures (the fourth signal in every dashboard we run) will show whether the whole cohort keeps escalating into the memory-inflated cost curve or blinks together.

One structural mitigant separates Microsoft's capex from Meta's, and fairness requires it: attachment to receipts. Meta's spending amortizes against ad margins on faith; roughly half of Microsoft's maps to the RPO drawer — contracted, scheduled, enforceable demand (asterisk and all). The right mental model is a spectrum this series can now draw completely: CoreWeave (100% contracted, leveraged), Microsoft and Google (~half receipted), Meta (unreceipted, annuity-collateralized), with the memory complex selling into all four. Where a company sits on that spectrum determines which macro event hurts it: contract-quality shocks hit the left end, sentiment shocks hit the right, and only a genuine AI demand break hits everyone — which is, once more, why every dashboard in this series converges on the same October.

11. The silicon gap and the Anthropic triangle

Among the four companies bidding to escape the NVIDIA tax, Microsoft is furthest behind. Maia 200 — announced January 2026 on TSMC 3nm with 216GB of HBM3e, pitched as "30% cheaper than anything on the market" for inference — runs GPT-5.2 and Copilot workloads in two US regions but remains unavailable to Azure customers at mid-2026; capacity lead times run 18–24 months against Google's 2–3 months for committed TPU capacity. Google sells TPUs externally at gigawatt scale (Anthropic, possibly Meta); Amazon deploys Trainium for its anchor tenant; Microsoft rents NVIDIA at record scale (first cloud to deploy GB300 NVL72 at volume) and hopes Maia matures. The strategic consequence is margin, not capability: every AI dollar Microsoft serves carries NVIDIA's markup that Google increasingly doesn't pay, which compounds through exactly the gross-margin guide-down of Section 10.

Why is Microsoft — the company with the deepest systems-engineering bench in software — behind on silicon? The record suggests a sequencing error rather than incapacity: for the critical 2019–2024 window, Microsoft's chip urgency was blunted by exactly the OpenAI exclusivity it enjoyed. When your differentiation is contractual (exclusive frontier models), silicon economics are someone else's problem — you pass NVIDIA's margin through to customers who have nowhere else to get GPT. Google, lacking any such contract, had to make its differentiation physical, and its decade head start in silicon is the compounding of that necessity. The 2026 irony writes itself: the contractual moat is gone, the physical moat takes five years to pour, and Microsoft must now rent competitiveness from NVIDIA at record scale precisely when its two cloud rivals are vertically integrating past it. The Maia program's importance is therefore strategic rather than financial in this report's horizon — its first shipping gigawatt matters less than the demonstration that Microsoft can execute the transition from contractual to physical advantage, the same transition every episode of its history required and the one the OpenAI unwinding has now forced.

The November 2025 Anthropic triangle was the cleverest move on this board and deserves its geometry spelled out: Anthropic committed to buy $30 billion of Azure compute (scalable to a gigawatt); NVIDIA and Microsoft invest up to $10 billion and $5 billion respectively in Anthropic; Claude joins Azure Foundry — making Microsoft the only cloud offering both OpenAI and Anthropic frontier models, and (reportedly) opening talks for Claude inference on Maia silicon. In one stroke Microsoft diversified its model dependency, backfilled future OpenAI-shaped holes in Azure demand with OpenAI's chief rival, and acquired an equity call on the lab that wins the enterprise API market (~40% share of enterprise LLM spend). It is also, of course, another entry in the AI economy's circular-receipts ledger — investor funds tenant, tenant commits spend — and the third such triangle this series has now mapped (Google-Anthropic, Microsoft-OpenAI, Microsoft-Anthropic). The machine's revenue certificates keep cross-referencing; we keep a running count.

Anthropic's own position in this geometry deserves a sentence, because it has quietly become the AI economy's designated diversifier: it now trains on Google TPUs, serves from Azure and GCP, banks investment from Google, Microsoft and NVIDIA simultaneously, and anchors receipt drawers at two rival clouds at once. Every hyperscaler's backlog quality now partially depends on the same private company's trajectory — a concentration the individual disclosures each mention and the aggregate picture, which only emerges across reports like these, makes rather starker. It is the memory complex's single-organism thesis, reconstituted one layer up the stack.

12. The quieter provinces: security, gaming, and the rest

Three facts round out the empire. Security is the business the market forgot: roughly $37 billion of annual revenue — triple CrowdStrike and Palo Alto combined — sold through E5 bundling, growing quietly, and structurally advantaged by exactly the platform gravity that makes regulators circle. It is also the sleeper AI beneficiary nobody models: security is the one enterprise category where agentic AI has an unambiguous, already-paying use case (triage, response, threat hunting at machine speed), where the customer's alternative to automation is an unfillable analyst shortage rather than an existing workflow, and where Microsoft's telemetry position — signals from a billion endpoints, Entra identities and Azure workloads — is a data moat no pure-play can assemble. If the "agents" narrative of the next two years produces one clean revenue proof at Microsoft, the odds favor it appearing here rather than in Office. Gaming is a $23.5 billion province in visible disorder: the $68.7 billion Activision purchase (the largest in gaming history) fed Game Pass to ~$5 billion of annual revenue, but a botched 2025 price rise to $29.99 shed subscribers by the millions, forcing a public walk-back to $22.99, a 33% hardware revenue collapse, July layoffs, and a strategy retreat from console exclusivity toward multiplatform publishing — Xbox is becoming a software brand wearing a console costume, which is probably rational and definitely deflationary. The generous read: Microsoft is executing in gaming the same hardware-to-software conversion it executed with Windows Phone's corpse (exit the device, keep the services), trading a $500 million console revenue hole for a publishing business with Call of Duty, Minecraft and Candy Crush that would rank among the world's largest standalone game companies. The ungenerous read: it paid $69 billion for content whose subscription flywheel it then broke with its own pricing, in a division that has consumed three strategy resets in four years. Either way the province is too small to move the empire's valuation — 7% of revenue — and exists in this report mainly as a case study in what happens to Microsoft businesses that lack the enterprise bundle's gravity: they compete on product alone, and the results speak. LinkedIn and GitHub, the two community acquisitions, deserve their single sentence of credit: ~$17 billion and the developer ecosystem respectively, both compounding, both strategically load-bearing (LinkedIn as the B2B data asset no rival owns; GitHub as the distribution channel that made Copilot-for-code the one AI product Microsoft unambiguously leads). Regulation, at Microsoft's usual simmer: the FTC's broad probe (cloud licensing, bundling, AI) grinds on without a complaint; the EU Teams case settled without fine; CISPE settled for ~€20 million plus licensing reform; and the Microsoft-OpenAI relationship itself sits under FTC 6(b) scrutiny as a lock-in structure — a probe whose premise, ironically, the 2026 amendments have largely dissolved. After the 2000 near-breakup, Microsoft's regulatory craft — settle early, concede visibly, preserve the bundle — is the industry's best; it is also a recurring small tax on the very bundling economics (Teams, Copilot, security) that power Sections 9 and 12.

Part Three · Synthesis: the contract as confession

13. Where it stands: the July 2026 snapshot

The operating engine, five quarters:

Quarter (FY) Revenue Op. profit Op. margin Azure growth (cc) GAAP EPS
FY25 Q3 $70.1B $32.0B 45.7% +35% $3.46
FY25 Q4 $76.4B $34.3B 44.9% +39% $3.65
FY26 Q1 $77.7B $38.0B 48.9% +39% $3.72
FY26 Q2 $81.3B $38.3B 47.1% +38% $5.16*
FY26 Q3 $82.9B (+18%) $38.4B (+20%) 46.3% +39% $4.27 (+23%)

(*FY26 Q2 GAAP EPS includes a $7.6 billion one-time gain from the OpenAI restructuring — +$1.02 of EPS; the clean comparative is non-GAAP $4.14. The OpenAI equity-method line has whipsawed EPS by more than a dollar in both directions across two quarters: -$0.41 in Q1, +$1.02 in Q2, near zero in Q3 — the accounting shadow of Section 6's drama.)

Read the table's quality before its levels: operating margin holding at 46–49% through the capex wave's early depreciation is the number the bears keep failing to dent — Amazon would take Microsoft's worst margin quarter as its best ever, and the margin is why $190 billion of capex is survivable here in a way it would not be almost anywhere else. The growth mix underneath: Intelligent Cloud +30%, Productivity +17%, More Personal Computing -1% — the old Windows-and-devices business is now a rounding error on the growth math, which means the "PC company" mental model (still embedded in some retail flows) prices a segment that no longer moves the needle in either direction.

FY2025 closed at $281.7 billion of revenue (+15%) and $101.8 billion of net income; FY2026 consensus sits near $325–329 billion and ~$17 of EPS, with the June quarter guided to $86.7–87.8 billion and Azure to 39–40% constant currency. Balance sheet: $78 billion of cash and short-term investments against ~$40 billion of conventional debt and ~$125 billion including the finance leases that increasingly are the capex program. The valuation: $2.90 trillion — fourth in the Mag 7, having been passed by Alphabet during the round trip — at 23.3× trailing and 20.2× forward, below its own five-year average, below every peer except Meta on forward earnings, with a 0.9% dividend yield and 55 analysts averaging $561 against a ~$390 price. As at Meta and SK hynix, the street believes a scenario the tape refuses to fund; the spread between $561 and $390 is, almost to the dollar, the market's valuation of the OpenAI divorce plus the capex cliff.

Against the family (July 3–4 data): NVIDIA $4.72T (29.8×/15.3×), Apple $4.53T (37.3×/32.1×), Alphabet $4.39T (27.5×/24.7×), Microsoft $2.90T (23.3×/20.2×), Amazon $2.61T (31.6×/24.5×), Meta $1.48T (21.2×/15.8×), Tesla $1.48T (357.7×/154.5×). The 2026 sort order is itself the story: the two companies the market has decided are AI's relative losers — Microsoft and Meta — are the two cheapest, and they are cheap for mirror-image reasons. Meta is discounted for spending without receipts; Microsoft is discounted for receipts of contested quality. Meta's discount is an allocation discount (removable by disclosure and discipline); Microsoft's is a relationship discount (removable by demonstrated independence from OpenAI). Both discounts are, in principle, self-repairable — which distinguishes them from Apple's growth problem and Tesla's arithmetic problem, and is why this series keeps finding its interesting questions at the cheap end of the index.

14. The contract as confession: reading the unwinding the way we read balance sheets

This series' method is to treat companies' capital-markets behavior as their most honest disclosure — hynix's ADR, Meta's vanished buyback. Microsoft's equivalent document is the April 2026 amendment, and it rewards close reading, because each clause is a confession about how the two most informed parties in AI price its future.

The $38 billion revenue-share cap. OpenAI bought out its open-ended ~20% obligation to Microsoft for a fixed ceiling through 2030. A party that expects modest revenue does not spend negotiating capital capping a percentage; OpenAI's own conduct says it expects revenue so large that 20% of it dwarfs $38 billion. Microsoft accepting the cap says the reciprocal: it preferred certain, bounded cash to uncertain, unbounded participation — the exact opposite of the 1981 instinct, where Microsoft chose unbounded participation over bounded cash. The non-exclusive IP conversion says Microsoft now values freedom to use other models (MAI's seven-model family, Anthropic's Claude) above the exclusivity premium of OpenAI's — a rational hedge that is also an admission the crown-jewel license stopped being a moat. The deleted AGI clause says both sides concluded the science-fiction contingency was distorting a commercial relationship — and its deletion quietly removed the single wildest tail in Microsoft's risk profile, an underappreciated positive. The $250 billion Azure commitment says OpenAI still cannot build capacity as fast as it consumes it — the tenant's confession, familiar from our CoreWeave report, that time-to-power is the binding constraint even for the best-funded startup alive.

Net the confessions: both parties priced OpenAI's future as enormous, Microsoft's claim on that future as shrinking, and compute as the scarce layer. Which yields the reframe the market has only half-absorbed: Microsoft's OpenAI position is transitioning from platform tax (dead, by contract) to landlord plus 27% shareholder (alive, and worth $135 billion at the last private mark before any IPO). The 1981 analogy the bears deploy — Microsoft as IBM, funding a partner who keeps the standard — is apt for the strategic layer and wrong for the financial one: IBM owned none of Microsoft. Microsoft owns 27% of its usurper, holds IP rights to 2032, and collects rent on $250 billion of its compute. IBM should have had it so good. The fine of $850 billion prices the death of the tax; it gives little credit to the estate.

Size the estate against the fine, roughly, since the market won't. The 27% stake: $135 billion at the restructuring mark, plausibly $150–200+ billion at OpenAI's subsequent reported private valuations — call it 5–7% of Microsoft's market cap, carried in the public price at approximately nothing (the equity-method line actually subtracts OpenAI's losses from Microsoft's EPS, so the stake currently costs reported earnings while its value goes unmarked). The IP license through 2032: whatever six more years of frontier-model access is worth to a company shipping AI into a billion seats — not zero. The $38 billion capped revenue share: bankable. The $250 billion Azure commitment: Section 7's asterisk, but consideration nonetheless. Against this, the fine: $850 billion of market value removed. Even applying brutal haircuts — discount the stake for illiquidity and OpenAI's burn, discount the commitment for concentration — the asymmetry between what was repriced and what was received runs into the hundreds of billions. Either the market is also pricing something else (the Copilot product gap, the capex cliff — real, but Section 16 sizes them smaller), or the OpenAI divorce has been marked as a total loss when it was a forced conversion from one asset class to another. Divorce settlements are rarely triumphs. They are also rarely zeros.

15. Three scenarios for the year ahead

Anchors: ~$327B FY26 revenue consensus, ~$17 EPS, $190B calendar-26 capex, 20.2× forward at ~$390.

Scenario A — the estate revalues. Azure holds 39–40% as ex-OpenAI demand (Anthropic's $30B, enterprise agents, Foundry) backfills; Copilot's bundling converts seats to usage enough that renewal-cycle pricing sticks; Maia reaches customer GA and clips the NVIDIA tax; an OpenAI IPO or secondary marks the 27% stake publicly (at OpenAI's reported private valuations, the stake alone could exceed $200 billion — 7% of Microsoft's market cap, currently carried by the market at roughly zero). The multiple mean-reverts toward its five-year norm (~28×): a 35–40% move on flat estimates. The street's $561 average encodes most of this.

Scenario B — the grind (modal). Azure strong but its OpenAI share visibly decays quarter by quarter; Copilot seats grow while usage surveys stay embarrassing; capex guides higher again in October; FCF stays compressed; the OpenAI stake stays private and unmarked. EPS grows low-teens, the multiple stays punished, and the stock ranges $350–450 while the market waits for one of A's catalysts to fire. Two years of underperformance argue this is well underway; the Ballmer-mirror risk is that it lasts as long as the market lacks a visible second-act proof point — and "visible" is the operative word, since the first Ballmer mirror ended only when Azure got big enough to disclose. Note also what the grind pays while it grinds: at 20× forward with a growing dividend, mid-teens EPS growth, and $10 billion a quarter of capital returns still flowing, Microsoft's Scenario B is the best-compensated waiting room in the Mag 7 — unlike Meta's (no buyback) or Tesla's (no earnings). The cheapest thing about the stock is the cost of being early.

Scenario C — the tax dies faster than the estate matures. OpenAI's Stargate capacity comes online and its Azure consumption plateaus or shrinks before the $250B commitment's schedule bites; a renewal-cycle Copilot revolt (enterprises striking the line item as the E3 price rise lands) headlines the "AI shelfware" narrative; the October capex season forces a choice between the dividend-and-buyback complex and the buildout; gross margin breaks below 62%. The stock revisits $300 — roughly 17× forward — and the Mag 7's "safest" name completes its round trip to value-stock status. C's probability is the lowest of the three, but note its correlation: it is the Microsoft expression of the same 2027-capex-deceleration event that defines Scenario C at Meta, the air pocket at SK hynix, and renewal risk at CoreWeave. One wager, five tickers; October discloses.

A probability note the scenario structure imposes: unlike Meta (where the collateral either holds or it doesn't) or Alphabet (where two numbers race), Microsoft's scenarios are sequential rather than exclusive — B is almost certainly the next two quarters regardless, and the real question is which exit it takes. That shape rewards patience differently: the A-catalysts (a disclosure, a GA date, a public mark) are discrete events management largely controls, while the C-triggers are mostly exogenous (OpenAI's buildout pace, macro). A stock whose upside catalysts are endogenous and whose downside triggers are exogenous, held at a below-market multiple with paid waiting, is a specific and analyzable shape — the mirror of Tesla's, as the next report in this series will show, where the multiple is enormous, the waiting is unpaid, and every catalyst is a promise.

16. The valuation paradox, Microsoft edition — and the Ballmer test

State this company's version of the series' paradox precisely: Microsoft trades at its largest discount to its own five-year multiple in a decade, while operating the fastest-reaccelerating hyperscale cloud, the largest disclosed AI revenue line, the largest commercial backlog, and a $135 billion private-market stake the public price ignores — because the market has decided the OpenAI unwinding is a platform miss in the Ballmer class.

The four-sided-market lens sharpens the oddity. The market currently pays a premium for AI's sellers (NVIDIA at 30× trailing, the memory complex at cycle-defying trillion-won prices) and for the receipted intermediary (Alphabet, re-rated), while discounting both tenants whose spending constitutes the sellers' revenue. If the tenants' capex is value-destructive — the implicit claim in Meta's and Microsoft's multiples — then the sellers' revenue is terminal and their premiums are wrong. If it is value-creative, the tenants' discounts are wrong. The market is, in effect, charging the same dollar twice: once as risk on the buyer's books, once as quality on the seller's. This series has now written that sentence, in local dialect, five times. The trade it implies is not directional on any single name — it is convergence: someday soon, the buyers and sellers of the same gigawatt must be priced as parties to the same transaction. October, again. The test for that thesis is the Ballmer test itself: in 2000–2014, the punished company had no claim on the platforms it missed — zero search share, zero mobile share, zero social. In 2026, the allegedly-missing company owns 27% of the platform it "missed," hosts both frontier labs, holds IP rights to 2032, and books the platform's compute on its own P&L. The analogy fails on its facts. What survives of the bear case after the analogy falls is real but smaller: margin dilution from NVIDIA dependence, receipt concentration, Copilot's product gap, and the arithmetic violence of $190 billion of capex against $140 billion of operating cash flow. Those are 15–20% problems, not 40% problems. Conversely, what the bulls must concede: the 2014–2023 Microsoft premium was the certainty premium — the company that owned the enterprise default in every category it entered — and the assistant war is the first category in twenty years where Microsoft is visibly not the default and is bundling to compensate. Certainty premia, once broken, rebuild on proof, not on history. Hence the modal grind.

17. What would have to be true, the risk matrix, and five signals

For Scenario A to pay from ~$390, most of the following must hold:

  1. Azure ≥37% cc growth through FY2027 with OpenAI's revenue share visibly declining (diversification proven, not asserted);
  2. Copilot renewal cycles post-bundling hold seat counts with usage metrics improving toward parity with access;
  3. Capex peaks as a share of revenue within four quarters (the guide stabilizes);
  4. Maia 200 reaches external GA, or Anthropic-on-Maia lands (the NVIDIA tax clipped);
  5. A public mark on the OpenAI stake (IPO, tender, or disclosure) forces the market to price the estate.

Risk matrix (probability × severity):

And one deliberate absence from the matrix, stated for the record: MAI model risk. Microsoft's own seven-model family could plausibly matter enormously (replacing OpenAI dependency from within) or not at all (another Bing-class also-ran), and mid-2026 offers no adoption data to weight it either way. We decline to score what cannot yet be measured; the first external MAI benchmark cycle and any Copilot-on-MAI migration disclosure will move it into the matrix, in whichever direction the data points. Its absence should not be read as dismissal — a genuinely frontier MAI is Scenario A's cheapest accelerant, since it converts the entire NVIDIA-margin and OpenAI-dependency discussion simultaneously.

Signal design notes, as throughout the series: ordered by arrival, chosen for independence — the earnings print reads the receipts and their concentration, the renewal season reads the product, the Maia item reads the margin structure, the OpenAI mark reads the estate, October reads the whole chain. Three of the five (1, 2, 5) are calendar-bound within four months; the estate mark (4) is the wildcard that could fire at any OpenAI funding event. A bull thesis that survives signals 1–3 arriving hostile was thesis-shaped hope; a bear thesis that survives a public OpenAI mark above $500 billion plus a disclosed declining-OpenAI-share Azure print has stopped being about facts.

Five signals, in firing order:

  1. FY26 Q4 earnings (late July): Azure vs the 39–40% guide; any disclosure of OpenAI's declining revenue share (bullish if declining while Azure holds); the capex number against the $190B calendar frame; buyback pace. This print doubles as the fiscal-year close, so it carries FY2027 guidance — the first full-year frame constructed entirely inside the post-exclusivity world, and therefore the clearest reading yet of how management itself sizes the estate.
  2. The Copilot renewal season (from July's bundle repricing): enterprise agreements struck or renegotiated at the new E3 pricing — the seat/usage scissors resolving in public.
  3. Maia 200 GA or the Anthropic-on-Maia confirmation: the silicon gap's first closing datapoint.
  4. Any public mark on OpenAI equity: the single largest unpriced asset in the Mag 7; even a secondary transaction reprices it.
  5. October capex season: the synchronized disclosure event shared with every dashboard in this series — Meta, Alphabet, and the memory complex all read the same two weeks.

Two calibration notes for position thinking (analysis of the instrument, not advice). First, correlation: Microsoft is the least correlated of the four tenants to a pure AI-sentiment shock — Office, security, gaming and Windows are non-AI ballast that Meta (98% ads) and the pure suppliers lack; in a Scenario-C world across the complex, MSFT falls least, which is worth something in a portfolio holding several links of this chain. Second, asymmetry of information arrival: the two most likely large positive catalysts (the OpenAI-share disclosure and the OpenAI public mark) are both disclosure events rather than execution events — they reveal value that already exists rather than requiring new value to be created. Stocks discounted for opacity have a specific virtue: sunlight is a catalyst, and sunlight is cheap.

The bottom line. Microsoft built the modern world's most durable business by understanding, before anyone, that the scarce layer is never the machine — it is the license everyone must renew. For two years the market believed Microsoft had executed that insight a second time on the AI era, and priced it as the trade of the century; for the last year it has watched the license slip away clause by clause and fined the company $850 billion for the reversal. Both reactions priced the tax. Neither has seriously priced the estate: a 27% stake in the era's defining company, a $627 billion receipts drawer, the only dual-frontier-model cloud, and a $37 billion AI line that no longer depends on any single partner's loyalty. The Ballmer decade teaches that this company can compound unpriced for years when the market fixates on the platform it lost; the Nadella decade teaches what the repricing looks like when the second act becomes undeniable. The five signals above are ordered by their power to force that moment. The first arrives this month, and the fourth — whenever OpenAI finally faces a public market — will settle an argument this report can only frame: whether Redmond's second act was the tax it lost, or the estate it kept.

And hold the 1981 contract next to the 2026 amendment one last time, because the pair is this company's entire biography in two documents. In the first, a small supplier kept its rights against a giant, and the rights became the century's greatest fortune. In the second, the giant that supplier became watched a small supplier keep its rights, and paid $850 billion of market value for the tuition. The lesson the market has drawn is that Microsoft was outplayed at its own game. The lesson the documents actually support is narrower and stranger: in platform economics, the licensor always wins the rights — but IBM survived losing them for forty more years of profit, and Microsoft enters its own IBM period owning 27% of its Microsoft. History does not repeat; it invoices. The five signals will tell you the amount.

Sources

Report generated by the Aya Invest research pipeline. Every claim above traces to a public source; where figures are third-party estimates (Azure run-rate, Copilot usage shares, real Copilot revenue) or reported-but-unconfirmed (Anthropic-on-Maia talks), the text says so. Anthropic's Claude models — a commercial partner of Microsoft's Azure Foundry discussed in this report — are also the model family this research pipeline runs on; we note the fact for transparency. For information and research purposes only. Not investment advice.

FAQ

What changed in the April 2026 Microsoft-OpenAI amendment?

Three things, each loosening Microsoft's grip: the AGI clause — which once let OpenAI cut Microsoft off upon reaching AGI — was deleted, but so was exclusivity; the revenue-share was capped; and IP rights became non-exclusive. The report reads it as history inverting: Microsoft built its empire in 1980 by licensing DOS to IBM while keeping the right to sell it to everyone else. OpenAI just executed the same maneuver against Microsoft. What Microsoft kept is the estate: a ~27% stake worth perhaps $135B that appears nowhere on its balance sheet.

How concentrated is Microsoft's $627B RPO?

Remaining performance obligations reached $627B, of which roughly 45% traces to OpenAI's compute commitments. That is the strongest receipt in the series by size — and the most concentrated by counterparty: a customer that is simultaneously a partner, a competitor, and a company whose own financing depends on markets staying open. The report treats the RPO's quality, not its size, as the live question, alongside a capex guide that quietly absorbed ~$25B of memory-price inflation.

Is Copilot actually being used?

Microsoft reports about 20 million paid Copilot seats — the fastest-selling product in its history by seats. Independent usage telemetry, however, puts Copilot's share of enterprise AI assistant usage around 11.5%, far behind ChatGPT. The gap between seats sold and share used is the enterprise-AI version of this series' receipts test: procurement is a receipt for intent; usage is the receipt for value. Renewal season adjudicates between them.

© Aya Research · About · invest.aya-ai.org · For information and research purposes only. Not investment advice.