The Microsoft exclusivity unwind, in two parts

OpenAI and Microsoft amended their partnership again last week, the second amendment in six months, and the same week OpenAI announced that its models will be available on Amazon Bedrock, with a co-developed Codex agent product alongside. The two announcements are the same announcement. I want to walk through why.

Start with what the amendment actually does. Microsoft's exclusive IP licence to OpenAI's models is gone. The revenue-share arrangement, which previously ran without an explicit ceiling tied to a profit cap that everyone had stopped believing in, is now capped at 20% and sunsets in 2030. Azure remains the "primary cloud partner" through 2032, note the word "primary", which is doing a lot of work where "exclusive" used to sit, and retains preferential rights on new compute commitments without holding a veto over them.¹

This is the third structural form the partnership has taken. The original 2019/2023 shape was: Microsoft funds OpenAI, gets exclusive IP, takes a share of profits up to a multiple of investment, hosts everything on Azure. The October 2025 restructuring around OpenAI's PBC conversion loosened the profit cap and softened exclusivity language. This amendment finishes the job. Exclusivity is replaced by primacy; the share is capped and time-boxed; the IP licence is non-exclusive.

If you read the amendment alongside the AWS announcement, the sequence is legible. OpenAI cannot run a frontier-model business on a single cloud at the scale its compute commitments now imply. The reported Stargate-related capex, the Oracle compute deal from last year, and now Bedrock distribution are not three separate stories about OpenAI hedging Microsoft. They are one story about what inference at OpenAI's scale costs, and where the GPUs actually are.

Apply inference economics. The binding constraint at the frontier is no longer training compute, it is serving compute, the marginal cost of each token generated against an enterprise contract that was sold on a per-seat or per-call basis priced before the reasoning-model era pushed average tokens-per-query up by an order of magnitude. OpenAI's gross margins on inference, to the extent we can triangulate them from the disclosed Azure revenue-share figures and the reported compute spend, are not the SaaS margins the equity story implied two years ago.² In that environment, the question "can we get the same H100-hour cheaper somewhere that isn't Azure" stops being a procurement question and becomes an existence question.

AWS solves part of it. Bedrock distribution solves a different part. Enterprise customers who standardised on AWS, which is most enterprise customers, by revenue, could previously get OpenAI models only by routing through Azure or through OpenAI's direct API, which meant a second vendor relationship, a second data-residency conversation, and a second procurement cycle. Putting GPT-class models on Bedrock collapses that friction. It also, and this is the part Microsoft will have negotiated hard over, makes OpenAI a peer to Anthropic inside the Bedrock catalogue, where Anthropic has been the default frontier choice for eighteen months.

So what did Microsoft give up and what did it get?

Microsoft gave up exclusivity on the IP, the right to be the sole hyperscaler reselling OpenAI models, and the open-ended revenue share. It got: a 20% cap (which is high, higher than I'd have guessed Microsoft would settle for if the relationship were unwinding from weakness), a sunset in 2030 (which is, let's say, four years of guaranteed cashflow on every OpenAI dollar), primary-cloud status through 2032, and, crucially, release from the implicit obligation to fund OpenAI's compute build-out single-handedly. Microsoft's own capex commentary over the last two quarters has hinted at this: the company would like to spend on AI infrastructure that serves Microsoft's own products, not infrastructure that serves OpenAI's third-party API revenue.³

This is the bit I find most interesting structurally. The amendment is being read in the press as OpenAI breaking free of Microsoft. Read the cashflow shape and it is at least as much Microsoft breaking free of OpenAI. A 20% cap on a sunsetting revenue share, in exchange for not having to underwrite the next $50bn of GPU buildout, is a trade Microsoft would make on a Friday afternoon.

What this is a case of: the de-coupling of frontier-lab capital structure from hyperscaler capital structure. The 2023-2024 model, lab raises from hyperscaler, lab spends raise on hyperscaler's compute, hyperscaler books revenue, lab books loss, was always a closed loop with awkward accounting properties. It worked while there was one frontier lab that mattered to each hyperscaler. It stops working when the lab's compute needs exceed any single hyperscaler's willingness to allocate, and when the hyperscaler's customers want optionality across labs. Anthropic-on-AWS-and-GCP was the first instance. OpenAI-on-Azure-and-AWS is the second. The third, somebody at Google having a conversation about whether Gemini ought to ship on Azure, is the one to watch.

The AI-safety-as-market-position frame applies here too, though more quietly. Anthropic's Bedrock primacy was sold partly on safety posture: enterprise buyers who needed a defensible AI procurement story bought Claude. OpenAI joining Bedrock with a co-developed Codex agent, pitched at the same enterprise developer buyer, contests that positioning directly. I'd watch whether AWS's marketing language around Bedrock shifts, whether "responsible AI" stays as the umbrella or whether "model choice" quietly moves to the front of the page.

A few things I notice that aren't getting written about:

The Codex co-development language in the AWS release is unusual. Frontier labs don't usually co-develop products with distribution partners; they ship models and let the partner build on top. Co-developed agent product implies shared roadmap, shared revenue treatment, and, I suspect, shared liability allocation on agent actions, which is the bit nobody has solved. The contract terms on that, if they ever surface, will be worth reading.

The amendment caps Microsoft's share at 20% but does not, on the available reporting, change the floor. If OpenAI's revenue grows faster than its Azure consumption, which is what happens if AWS picks up a meaningful share of inference traffic, Microsoft still gets 20% of the AWS-served revenue, because the share is on OpenAI revenue, not on Azure revenue. This is good for Microsoft and slightly strange for OpenAI, and I'd want to see the actual contractual language on revenue definition before being sure.⁴

The 2030 sunset is the date that matters. Everything OpenAI is doing, the AWS deal, the Oracle compute, the Stargate posture, the consumer product push, reads as preparation for a 2030 in which the Microsoft revenue share is gone and OpenAI is a standalone capital structure. That's four budget cycles away. The compute commitments being signed now are sized for the post-2030 business, not the current one.

What to watch: the Bedrock pricing page when OpenAI models go live, specifically whether per-token pricing matches Azure or undercuts it; Microsoft's next capex disclosure and whether the AI-infrastructure line decelerates; any disclosure of the Codex agent revenue-share arrangement; and whether Google does anything that suggests Gemini is about to become available somewhere that isn't GCP. The exclusivity era of frontier-lab distribution is ending. The amendment last week was the document that made it official.

Footnotes

¶