OpenAI Didn’t Leave Azure. Anthropic Pushed Them Out.

The Microsoft-OpenAI partnership just collapsed. Not the corporate kind of “collapse” where both sides publish friendly blog posts and nothing actually changes. The real kind. The kind where OpenAI’s own Chief Revenue Officer writes an internal memo calling the partnership a barrier to selling to enterprises. The kind where a $1 billion handshake from 2019 gets renegotiated so many times that the original deal is unrecognizable.

The official story is “flexibility” and “evolution.” The actual story is simpler: Anthropic was eating OpenAI’s enterprise lunch through AWS Bedrock, and OpenAI couldn’t fight back because Microsoft had them locked in.

So they broke the lock.

The Deal That Made Sense (Until It Didn’t)

In 2019, Microsoft invested $1 billion into OpenAI. The terms were straightforward. Microsoft would provide compute and cloud infrastructure. OpenAI would share all of their pre-AGI intellectual property. Azure would be the exclusive cloud provider. Both sides would benefit.

This was pre-ChatGPT. Pre-hype. The entire OpenAI team fit in a single group photo. Microsoft was making a bet on a technology that most of the industry considered a research curiosity.

By 2023, that bet looked like the greatest corporate decision of the decade. ChatGPT exploded. Microsoft poured another $10 billion into OpenAI. Azure became the exclusive home for the most in-demand AI models on the planet. Every enterprise that wanted OpenAI’s technology had exactly two choices: the OpenAI API directly, or Azure.

Then OpenAI built something they didn’t want to share.

The Reasoning Model Fight

In September 2024, OpenAI launched o1. This wasn’t another incremental improvement. This was a different paradigm. The model would generate internal reasoning traces before answering, working through problems step by step. The results were staggering: 89th percentile on competitive programming, top 500 in the US Math Olympiad qualifier, superhuman accuracy on PhD-level science benchmarks.

And OpenAI refused to tell Microsoft how it worked.

Under the original deal, Microsoft was entitled to the IP. They were supposed to get documentation on how OpenAI’s models were built. When it came to o1 and chain-of-thought reasoning, OpenAI went quiet.

Mustafa Suleyman, Microsoft’s head of AI, was not pleased. Reports confirmed he raised his voice at OpenAI employees during a video call, telling then-CTO Mira Murati that the startup wasn’t holding up its end of the deal. The core of his frustration: Microsoft’s own researchers couldn’t replicate the reasoning breakthrough because OpenAI wouldn’t hand over the recipe.

This was October 2024. Months before DeepSeek R1 showed the world how reasoning models worked. OpenAI had a massive lead, and they intended to keep it. Even if it meant pissing off the company bankrolling them.

That was the start of the breakup.

Three Renegotiations in Two Years

The partnership has been rewritten three times since then.

October 2025: OpenAI restructured into a Public Benefit Corporation. Microsoft’s investment converted to roughly 27% equity, valued at about $135 billion. The AGI definition problem got a patch: any declaration would require verification by an independent expert panel, since Microsoft feared OpenAI would declare AGI early just to escape the IP-sharing deal. Microsoft’s IP rights were extended through 2032 and now included post-AGI models.

But here’s the thing. Satya Nadella had already been simmering for years. When Peter Lee briefed him on an early version of GPT-4 back in 2022, Nadella cut him off mid-sentence. “OpenAI built this with 250 people. Why do we have Microsoft Research at all?”

That frustration never went away. It just changed shape.

April 2026: The real gut punch. Microsoft’s license became non-exclusive. The exclusive cloud requirement was gone. OpenAI could now serve products across any cloud provider. Microsoft would no longer pay revenue share to OpenAI. And the AGI escape clause? Deleted entirely.

That last detail matters. Both sides gave something up. OpenAI lost the AGI loophole that could have ended the partnership on their terms. Microsoft lost exclusivity over the most valuable AI technology in the world. But on balance, OpenAI walked away with far more freedom than they walked in with.

The financial restructuring tells the story best. OpenAI’s revenue share payments to Microsoft continue through 2030, but they’re now subject to a total cap. For a company burning through billions in compute costs, a capped obligation is a very different thing than an open-ended one.

Why Anthropic Is the Real Story

None of this makes sense if you only look at the Microsoft-OpenAI relationship. The catalyst was external.

Anthropic hit a $30 billion annual run rate in April 2026. Up from $9 billion at the end of 2025. Over 1,000 enterprise customers each spending more than $1 million per year. This is one of the fastest revenue ramps in the history of B2B software.

And the engine behind it wasn’t better models. It was distribution.

Anthropic’s Claude models are available on all three major clouds: AWS Bedrock, Google Vertex AI, and now Azure. But the real volume comes through Bedrock. If you’re an enterprise on AWS, which most enterprises are, and you want powerful AI for your developers, Claude is right there in your existing infrastructure. No new vendor relationship. No separate billing. No migration to a different cloud.

OpenAI models? Until April 2026, you had two options. The OpenAI API directly, or Azure. If your company runs on AWS, you were stuck doing one of two things: set up an Azure integration just for AI (which nobody wants to do), or use the direct API without the benefits of your existing cloud credits and procurement systems.

This created a bizarre dynamic. OpenAI had arguably better models for most tasks. Anthropic was growing faster in enterprise. The bottleneck wasn’t quality. It was where the models lived.

OpenAI’s CRO, Denise Dresser, said it plainly in an internal memo: “Our Microsoft partnership has been foundational to our success. But it has also limited our ability to meet enterprises where they are. For many, that is Bedrock.”

That’s not subtext. That’s a C-level executive writing to her team that the most important partnership in the company’s history was actively costing them enterprise deals.

The Cloud Credit Game Nobody Talks About

There’s a layer to this that the official announcements don’t cover.

Cloud providers hand out massive credits to startups. AWS offers up to $100,000. Google goes to $350,000. Azure is known for giving $500,000 or more, especially to Y Combinator companies.

These credits are supposed to be flexible. Use them for compute, storage, whatever you need. And for the most part, they are. You can use Google credits to run Gemini. You can use Azure credits to run OpenAI models. No restrictions.

Except for Anthropic models. Try using your cloud credits for Claude on Bedrock or Vertex AI, and the answer is consistently no. The revenue share deals Anthropic struck with cloud providers appear to be so aggressive that the providers won’t absorb the cost. If hypothetically 50% of every dollar spent on Claude goes back to Anthropic, then giving someone a million dollars in credits to run Claude means the cloud provider just wrote a $500,000 check to Anthropic for free.

The result: enterprises can burn through their cloud credits on literally any other AI provider except Anthropic. And Anthropic is still growing faster.

That’s how strong the Bedrock distribution advantage is. Anthropic can charge a premium, deny credit usage, and still win because they’re embedded in the infrastructure enterprises already use.

Azure’s Infrastructure Problem

There’s another reason OpenAI wanted out, and it has nothing to do with business strategy. Azure’s inference infrastructure was broken.

Not “slightly slow” broken. The data was damning. GPT-5.5 on Azure was 2.2x slower on average than the same model on OpenAI’s own endpoints. The worst-case performance was 15x worse. Token generation would randomly crater from 70+ tokens per second down to 2. Time to first token would spike to 100 seconds or more.

This wasn’t a one-time outage. It persisted for over a year. Detailed complaints to Microsoft’s head of AI infrastructure went nowhere. The performance charts that were shared with Microsoft’s team were apparently so impressive that Microsoft’s engineers cloned them for their internal documentation. Then didn’t fix the underlying problem.

One developer who received a million dollars in Azure credits used about $5,000 of it, hated the experience, and went back to paying OpenAI directly because it was cheaper than dealing with Azure’s inconsistency.

To Microsoft’s credit, when someone finally built a public benchmark (azure.t3.gg) exposing the severity of the problem, the fix came within 24 hours. The speed of the resolution suggests this was a solvable bug, not a fundamental infrastructure limitation. OpenRouter data confirmed that Azure subsequently matched or outperformed OpenAI’s own endpoints.

But the damage was done. The lesson wasn’t that Azure is permanently broken. The lesson was that Azure required public shaming to fix a problem that had been reported privately for over a year. That’s not the kind of infrastructure reliability you build a business on.

The AWS Deal

In February 2026, OpenAI and Amazon announced a strategic partnership. The numbers are staggering.

Amazon invested $50 billion in OpenAI. AWS became the exclusive third-party cloud distribution provider for OpenAI’s Frontier platform. OpenAI committed to consuming 2 gigawatts of Trainium capacity through AWS. An existing $38 billion compute agreement was extended by another $100 billion over eight years.

The technical angle is interesting too. OpenAI and AWS are co-creating a “stateful runtime environment” on Bedrock. If you’ve followed how the OpenAI API has been shifting toward persistent, websocket-based connections and aggressive caching, you understand why statefulness matters. These are the kinds of capabilities that make AI agents viable at production scale. Building this on top of Azure’s infrastructure would have been, to put it diplomatically, challenging.

The Trainium Question

There’s a genuine risk buried in this deal that nobody’s talking about enough.

OpenAI has never run models on anything other than NVIDIA GPUs. Trainium is Amazon’s custom silicon. It’s a completely different architecture with different compilers and optimization paths. The failure modes are different too.

Anthropic went through this transition already. And it wasn’t smooth. In August 2025, Claude users reported significant quality degradation. The community suspected Trainium was the cause. Anthropic’s post-mortem pinned it on software bugs and compiler errors on TPU hardware. Not the silicon itself.

But the community debate never fully resolved. When Claude quality dipped again in early 2026, the first suspicion was always “are they running this on cheaper hardware?” Anthropic traced that incident to unintentional system prompt changes and a long-session memory bug, not hardware. The fixes went in and quality recovered.

OpenAI is about to walk the same path. The commitment spans both Trainium 3 and the next-generation Trainium 4 chips (expected 2027). These chips offer more memory bandwidth and larger HBM capacity, which could actually benefit larger models and longer context windows. But getting inference quality and consistency right on new silicon is a hard engineering problem. One that Anthropic has already stumbled on publicly.

Where This Ends Up

The AI infrastructure war is shifting. For the last three years, it’s been about model quality: OpenAI vs. Anthropic vs. Google. That competition isn’t going away, but a second front is opening.

The fight for where models run is becoming as important as the fight for which models are best. NVIDIA vs. Trainium vs. Google’s TPUs. AWS vs. Azure vs. GCP. The companies that control the silicon and the cloud infrastructure will have enormous influence over which AI models enterprises actually use.

Google and NVIDIA seem best positioned to maintain relevance across both fronts. Google has both the models (Gemini) and the hardware (TPUs). NVIDIA has the chips everyone depends on today. AWS has the distribution network that apparently matters more than model quality.

Microsoft? They still own 27% of OpenAI. They still have IP rights through 2032. They’re not losing money on this deal.

But they lost the thing that mattered most: the guarantee that every enterprise wanting OpenAI had to go through them. That competitive moat is filled in. The wall is down. And the company that filled it wasn’t even in the room.

It was Anthropic, quietly signing Bedrock deals while Microsoft and OpenAI fought over chain-of-thought documentation.