A developer opening the AWS Bedrock model picker today scrolls past Claude, Llama, Mistral, Cohere — and finds ChatGPT in the list. That sentence would have been incomprehensible eighteen months ago. The picker normalized OpenAI overnight.

It happened in 48 hours. On Monday, Microsoft and OpenAI removed the AGI clause from their seven-year-old contract, capping Microsoft's revenue share, and confirmed Microsoft no longer had exclusive rights to OpenAI's models. On Tuesday, Amazon and OpenAI announced the expanded deal that put those models on AWS, with Sam Altman and AWS CEO Matt Garman sitting together for a Stratechery interview to explain it. The same Tuesday, the Wall Street Journal reported that OpenAI missed its internal goal of reaching 1B weekly active ChatGPT users by the end of 2025 and missed multiple monthly revenue targets in early 2026.

Three announcements. Read together, they describe one event: the moat that defined AI infrastructure for a generation just expired.

The Founding Deal

The Microsoft–OpenAI exclusivity wasn't a partnership clause. It was the strategic premise of both companies for most of a decade.

The pact began in November 2016: a small partnership announcement, OpenAI's first cloud deal. By May 2020, Microsoft had built OpenAI a 285,000-processor supercomputer. In September 2020, Microsoft secured an exclusive license to GPT-3. The framing then was unambiguous: OpenAI's compute would only run on Azure; OpenAI's models would only be sold by Microsoft. The total investment reached roughly $13 billion. The exclusivity was the consideration.

That structure built the modern industry. ChatGPT launched on Azure. GPT-4 trained on Azure. Microsoft's Copilot, the entire enterprise distribution channel, the developer story, the Office integration — all of it ran through one company's infrastructure because no other infrastructure had access. For Microsoft, this was the AI moat: not better models, but the only way to get the best models. For OpenAI, it was the capital that made frontier training possible. Both companies built their valuations on the assumption that this arrangement would hold.

The Mechanism

The dissolution accelerated in 2025. In July, OpenAI disclosed it would use Google Cloud to run ChatGPT and its API — the first public crack in the Azure-only architecture. Stargate, the $500B Oracle/SoftBank build-out, made Oracle a primary infrastructure partner. By autumn, Microsoft was pushing to strip the AGI clause — the contract provision that would have let OpenAI cut off Microsoft's access once it declared general intelligence achieved. In November, Microsoft AI CEO Mustafa Suleyman announced Microsoft would build its own frontier models. The same month, Anthropic committed $30B to Azure capacity — Microsoft was now selling compute to OpenAI's primary rival. The seller was hedging. The buyer was hedging. The exclusivity became a thing both parties wanted out of.

The economic explanation runs upstream of the contract. AI infrastructure is being commoditized at the marketplace layer, and Bedrock is the marketplace winning. Once the model is reachable through one cloud's API, the customer's switching cost collapses to a string change. AWS, Azure, and Google Cloud each compete to be the place where every model lives — because the customer wants every model in one bill, one IAM, one VPC. Multi-model marketplaces are the equilibrium toward which databases, CDNs, identity providers, and every other infrastructure layer eventually converged. AI was supposed to be different because the model was the asset. It turns out the model is the asset, and the marketplace is the channel, and both can be true simultaneously.

Exclusivity was the moat. Distribution turned out to be the asset.

The Confession

The other Tuesday story explains why now. The CFO and board, the WSJ reported, have begun questioning the wisdom of the company's data-center spending given the slowing growth. Bloomberg, Reuters, and The Information corroborated.

April 28, 2026
OpenAI missed an internal goal of reaching 1B weekly active ChatGPT users by 2025's end, and missed multiple monthly revenue targets earlier in 2026
Wall Street Journal

Both announcements were timed to land together. They are the same announcement.

The premium structure of exclusivity made sense when ChatGPT subscriptions were doubling. It stopped making sense when the next billion users wouldn't come from direct sign-ups but from API calls embedded in someone else's product. Bedrock is where the API calls happen. Companies don't trade exclusivity for distribution unless distribution matters more than premium positioning. OpenAI just made that trade. The data-center scrutiny, the slipped IPO timeline, the missed goal — those are the same signal as the Bedrock listing. The strategy that worked when OpenAI was capturing all of the top of the funnel can't work when the top of the funnel is multi-vendor.

The Inversion

The structural picture is cleaner than the noise. In 2019, OpenAI was Azure's exclusive AI tenant; Anthropic didn't yet exist. By 2024, OpenAI was Microsoft-only and Anthropic was Amazon's flagship Bedrock model after $4B and then $8B in investments. By 2026, both labs have inverted onto each other's home cloud. Anthropic runs $30B in Azure capacity. OpenAI runs in AWS Bedrock. Both are also on Google Cloud. Both are also on Oracle. The two frontier labs have, between them, dissolved the concept of cloud exclusivity entirely.

The same week — last Tuesday, April 21 — Amazon agreed to invest up to $25B more in Anthropic, on top of the existing $8B. That commitment held even as OpenAI was added to Bedrock. The clouds aren't picking favorites. They are stocking shelves.

The Bedrock model picker is the visible artifact of this. A developer building an agent can call Claude for one task and GPT-5 for another and Llama for the cheap path, billed through one AWS account. That is the AI infrastructure layer that won. It is not the layer Microsoft built its $13B bet on.

However

Microsoft did not lose. The amended deal preserves Microsoft's access to OpenAI technology through 2032 and removes the AGI clause that could have cut that access off early. Microsoft no longer pays a revenue share to OpenAI, which improves Microsoft's economics on Copilot. The investment, viewed as a financial position, has appreciated dramatically — Microsoft's stake is reportedly worth multiples of the cash invested. Microsoft also kept the customer relationship: Office, GitHub, and the enterprise distribution channel still run on OpenAI models, just not exclusively.

What Microsoft lost was the strategic premise. The reason to invest $13B in 2019 wasn't financial return. It was the conviction that AI infrastructure would be a winner-take-most market and Azure would be the winner because OpenAI was the only model worth running. That conviction is what expired this week. The AGI clause was the last clause that operationalized it. Once it was gone, the rest of the partnership was just a supplier relationship between a large customer and a model vendor — which is also what AWS now has, and Google now has, and Oracle now has. The exclusivity premium evaporated. The vendor relationship remains.

What Survived

The bedrock metaphor isn't accidental. AI exclusivity was supposed to be the foundation rock — durable, immovable, the strategic substrate beneath which nothing changes. It turned out to be sediment. The same forces that turned databases into multi-cloud commodities, that turned CDNs into a vendor list, that turned every previous layer of infrastructure into a marketplace, turned the most expensive partnership in software history into a SKU on AWS Bedrock within seven years.

What survives is what always survives at the infrastructure layer: the thing that gets distributed. Microsoft kept distribution. AWS bought distribution. Google built distribution. OpenAI traded the premium of being exclusively distributed for the volume of being everywhere distributed — and that trade was forced by growth deceleration, not chosen from strength. Anthropic made the same trade earlier and from a different position. Both labs now sit in every picker. Neither owns the channel.

The layer below the model — the one Microsoft thought it had locked — turned out to be a marketplace. The model is just the inventory.