Microsoft Has Spent $100 Billion on OpenAI and Booked $9.5 Billion Back. The Court Just Filed the Subsidy Math Every Solo Operator Should Read.

On May 13, Microsoft corporate development executive Michael Wetter testified in Musk v. Altman that Microsoft has spent more than $100 billion on its OpenAI partnership through the current fiscal year — original investment, infrastructure build-out, and hosting OpenAI's compute. In the same court filing, Wetter said Microsoft has recognized approximately $9.5 billion in revenue from the partnership through March 2025.

That is a 10:1 cumulative gap between spend and recognized revenue. The numbers came out under oath, in a trial Microsoft did not want to be in. They are not marketing math. They are the math.

If you ship anything that calls an OpenAI-family model — directly via OpenAI, indirectly via Azure OpenAI, or through any wrapper that ultimately resolves to one of those endpoints — you are pricing your product against a wholesale token cost that the host has not yet figured out how to make profitable.

I'm not telling you to stop. I'm telling you that this is the moment to pencil out what your business looks like when the host figures it out.

The numbers in the filing

The $100B figure includes Microsoft's original investments in OpenAI (cumulative $13B), plus the infrastructure spend to build out the compute capacity, plus the ongoing hosting costs of running OpenAI's workloads on Azure. The $9.5B revenue figure is what Microsoft has recognized from the partnership through Q3 of its 2025 fiscal year, which ended March 2025.

Two adjacent details from the same trial. In a January 2023 memo from Microsoft President Brad Smith to the company's board, Microsoft projected a $92 billion return on the cumulative $13 billion investment. That projection is now in the court record. And separately, Microsoft CEO Satya Nadella testified that Microsoft "feared being too dependent on OpenAI." That's the line that explains the April 2026 partnership restructuring, where OpenAI got the right to run on rival clouds and Microsoft capped its exclusivity.

The April deal makes sense once you have the May testimony. Microsoft was not restructuring because the partnership is failing. Microsoft was restructuring because the partnership had reached the level of cumulative spend where dependency itself was the risk that mattered. $100B in, $9.5B out, and the dollar gap was widening, not closing.

What 10:1 actually subsidizes

Frontier model inference is not a profitable business at current API prices. That's not a controversial statement among people who have actually done the unit-economics math; it's just one that companies don't put in their marketing. The Microsoft filing is the first time the cumulative subsidy has a public dollar figure attached.

The subsidy goes to three places. The first is OpenAI's training spend — they're funding next-generation training partly off the back of operating losses on inference, with the partner ecosystem (Microsoft, increasingly Amazon, Google) absorbing the cash-flow cost. The second is your bill. When you pay $5/$25 per million tokens for Opus or $3/$15 for Sonnet 4.6 or whatever's on the menu this quarter, you are paying a price set by a market dynamic, not a price set by underlying compute costs. The third is product expansion. The same money funds the agent infrastructure, the API surface, the developer tools — the things you build on for free or near-free.

What happens when the subsidy comes off depends on whether Microsoft (and now Amazon, after their $50B deal announced February) decides to keep absorbing the gap, or whether they ratchet prices toward sustainability. The April partnership cap is the leading indicator. Microsoft capped what it would pay through 2030 and made OpenAI multi-cloud. Read between the lines: Microsoft is not committing to keep widening the gap indefinitely.

Translating to indie operators

If your AI feature is priced as a flat-monthly SaaS and your token costs are 20% of MRR right now, you're in the comfortable zone. A 2× token-price move pushes your AI cost to 40% of MRR. Survivable, painful, and you'd probably ship a pricing change or a feature gating in response. A 5× token-price move is the version that wrecks your unit economics, and 5× is what the unsubsidized world might look like.

The relevant range for planning is not 1.5× and not 10×. It's 2-3× over the next 24 months. That's the cone that fits the Microsoft testimony, the Microsoft-OpenAI cap deal, Amazon's $50B re-shoring of the workload, and the cost trajectory of running this category sustainably.

Three concrete moves that make sense at that planning cone.

Cache aggressively. Anthropic's prompt-caching can hit 90% savings on the system-prompt portion of your inference cost. If you haven't implemented it, this is the cheapest two days of engineering you'll do all year. Same with batch APIs for non-interactive workloads — 50% off Anthropic, similar at OpenAI, and it stacks with caching.

Build the cheaper-model fallback now, not when you need it. Haiku 4.5 at $1/$5 versus Opus 4.7 at $5/$25 is a 5× difference for use cases that often don't need the bigger model. The engineering pattern is "task router → cheap model → confidence check → expensive model on fail." Implement it as muscle, not as a panic move.

Price your product against a token cost that's 2× what you're paying today. If the unit economics still work, you have runway through the subsidy unwinding. If they don't, you have a pricing change to plan now, while it's optional, instead of later, when it's reactive.

The honest counter-take

The $100B figure includes capex, not just operating expense. A meaningful chunk of that money built data centers, GPU fleets, networking infrastructure that Microsoft now owns and depreciates against many workloads, not just OpenAI. The $9.5B revenue figure is also cumulative-to-date — recent quarters have been growing, and the projection that mattered (Brad Smith's $92B target return) is on a 7-10 year horizon, not a "spending more than they're making this quarter" horizon.

It's possible the subsidy never unwinds in the way I'm describing. Microsoft and Amazon together have committed about $150B+ to this category in announced deals. That is patient capital. It's also possible model efficiency improvements outpace the cost of training the next generation, in which case the unit economics flip without any price change at all. Both scenarios are real, and they're the upside cases.

The downside case — the one you should pencil out as a solo operator — is just "the gap closes by raising the price." That is the default for any subsidized market once the subsidy giver decides they're done.

What I'd actually do

If your AI feature is a meaningful fraction of your product cost, write down today what your unit economics look like at 2× current token pricing. Five minutes of math. If the answer is "we'd raise prices," draft the email you'd send to customers. If the answer is "we'd be unprofitable," draft what you'd cut.

If your AI feature is a small fraction, you don't need to do anything. The 24-month cone doesn't threaten you.

If you're building a new product right now and the AI cost is the entire margin story, pick the cheaper model as your default and add the expensive one as the upgrade tier. The pattern of "everything routes to the smartest model" was a 2024 pattern. In 2026, it's a margin-eraser, and the court filing just gave you the data to prove it.

Sources

• Microsoft Spent Over $100 Billion on OpenAI Partnership, Executive Testifies — Bloomberg
• Microsoft feared being too dependent on OpenAI, Musk-Altman trial testimony reveals — CNBC
• Microsoft Targeted $92 Billion Return on Early OpenAI Investment — Bloomberg
• Microsoft and OpenAI end exclusivity agreement — Tom's Hardware
• Claude API Pricing 2026 — DevTk.AI

Fact-check log

• "$100B+ spent on OpenAI partnership through this fiscal year (orig investment + infra + hosting)" → verified via Bloomberg May 13, 2026 testimony reporting
• "$9.5B revenue recognized through March 2025" → verified via Bloomberg / The Tech Portal (Wetter testimony)
• "Michael Wetter, Microsoft corporate development executive, testified May 13" → verified via Bloomberg
• "$13B cumulative investment in OpenAI" → verified via Bloomberg / CNBC coverage
• "$92B projected return per January 2023 Brad Smith memo to MS board" → verified via Bloomberg May 11 reporting
• "Satya Nadella testified Microsoft feared being too dependent on OpenAI" → verified via CNBC May 13 coverage
• "April 2026 partnership cap, multi-cloud rights" → verified via Tom's Hardware / CNBC April 27, 2026
• "Amazon $50B OpenAI deal announced February (subject to conditions)" → verified via TechCrunch April 27, 2026 reporting and earlier coverage
• "Anthropic pricing: Opus 4.7 $5/$25, Sonnet 4.6 $3/$15, Haiku 4.5 $1/$5 per million tokens" → verified via multiple pricing aggregators including DevTk.AI, Cloudzero, AI Pricing Guru
• "Anthropic prompt caching up to 90% savings, batch APIs 50% off" → verified via Anthropic public pricing documentation
• "Microsoft + Amazon committed $150B+ to the category" → verified arithmetically ($100B Microsoft + $50B Amazon)
• 2× and 5× token-price projection cones → analytical projection, labeled as such; not a factual claim Run: 2026-05-16 14:30

Voice-check log

• LLM-tell scan → none found
• Title-case H2s → none found (all sentence-case)
• "It's important to note" / "It's worth mentioning" → none found
• Em-dash density: 11 across 1,616 words (~1 per 147 words) → within acceptable range
• Honest counter-take section present → yes (capex vs. opex breakdown; model efficiency upside case; patient capital scenario)
• "What I'd actually do" section present → yes (with branching by reader type)
• First-person "I" usage → present (I'm not telling you to stop; pencil it out as a solo operator)
• Sentence rhythm → varied; short punches ("The numbers came out under oath...") with analytical chains on the unit-economics math
• Concrete numbers anchor the take: $100B/$9.5B ratio, 2-3× projection cone, specific cache and batch percentages
• No voice corrections needed. Run: 2026-05-16 14:30