The Loop  ·  Issue 025

The Loop

A field journal of the AI frontier — for engineers who ship.

§ News

By AI Blog Editor
Apr 29, 2026 · 17 min read

Twenty-four hours — OpenAI's models hit AWS Bedrock the day after Microsoft's exclusivity ended

On April 27 Microsoft and OpenAI rewrote a contract that had quietly survived since 2019. On April 28 OpenAI's frontier models, Codex, and Managed Agents were live on AWS Bedrock. The pacing tells you everything.

Building 92 on Microsoft's Redmond campus, photographed in 2015.
Microsoft Building 92, Redmond. Photo by Jiaqian AirplaneFan, CC BY 3.0, via Wikimedia Commons.

On April 27, 2026, Microsoft and OpenAI announced that the partnership they signed in 2019 — the one with the indefinite exclusivity, the buried clause about artificial general intelligence, and the revenue share that ran until somebody decided AGI had been reached — had been rewritten. Microsoft's IP license is now non-exclusive and dated to 2032. The AGI clause is gone. OpenAI's revenue share to Microsoft continues to 2030, at the same percentage, capped at an undisclosed total.

On April 28, less than 24 hours later, OpenAI's frontier models, Codex, and a new product called Bedrock Managed Agents went live on AWS in limited preview.

The two announcements are reported separately in most coverage. They should not be. The first one is the contract that made the second one legal. Read together, what happened this week is the loudest cloud-vendor pivot since OpenAI was founded.

What actually changed in the contract

The Microsoft blog post is short on numbers and long on phrasing like "flexibility, certainty and a focus on delivering the benefits of AI broadly". The diagnostic sentences are these, taken verbatim from Microsoft's own announcement:

  • "Microsoft will continue to have a license to OpenAI IP for models and products through 2032."
  • "Microsoft remains OpenAI's primary cloud partner, and OpenAI products will ship first on Azure, unless Microsoft cannot and chooses not to support the necessary capabilities."
  • "Microsoft no longer pays revenue share to OpenAI"; "revenue share payments from OpenAI to Microsoft continue through 2030, at the same percentage but subject to a total cap."

Decoded: Microsoft is no longer locked in as exclusive cloud or exclusive licensee, but it keeps a non-exclusive license through 2032 and remains the default landing pad for new OpenAI products. OpenAI is no longer locked out of every other cloud, but it still owes Microsoft a 20% revenue cut — the figure CNBC, Bloomberg, and Tom's Hardware all converge on — until 2030, capped at a number neither side will name.

The phrase doing the heaviest lifting is "primary cloud partner". It is not "exclusive". It is the kind of word a contract uses when both sides know the difference matters and nobody wants to spell it out in the press release.

The AGI clause, mercifully buried

The other thing that died on Monday was the AGI clause, and it is worth dwelling on because it was the strangest piece of corporate paperwork in the AI industry.

The original 2019 deal granted Microsoft access to OpenAI's "pre-AGI" technology. The IP rights would terminate the moment OpenAI's board declared the company had achieved artificial general intelligence — a term defined in the 2018 OpenAI Charter as "highly autonomous systems that outperform humans at most economically valuable work". Reporting from The Information in late 2024 surfaced a more specific operational definition: AGI, in the contract sense, was the point at which OpenAI's systems could generate $100 billion in profit for the earliest investors. By October 2025 the parties had bolted on an "independent expert panel" that would adjudicate any AGI declaration, with Microsoft's rights running until either the panel agreed or 2030, whichever came first.

So: an indefinite exclusivity, contingent on a thing nobody can define, adjudicated by a panel that did not yet exist, against a financial threshold that would arrive years before any plausible scientific one. That is a sentence that costs lawyers a lot of money.

The new agreement replaces all of it with a calendar date. Microsoft's license expires in 2032. The revenue share ends in 2030. Nobody has to argue about whether GPT-7 counts.

OpenAI wordmark

Why the timing was choreography

The renegotiation has been visibly pending since November 2025, when OpenAI and Amazon announced an up-to-$50 billion arrangement: $15 billion initial investment plus $35 billion conditional, with AWS positioned as a privileged distribution surface for OpenAI's agent products. TechCrunch's coverage catalogued the legal collision: Microsoft, at the time, publicly refuted the AWS-exclusive language, stating "Azure remains the exclusive cloud provider of stateless OpenAI APIs" and that "OpenAI's first party products, including Frontier, will continue to be hosted on Azure." The Financial Times reported Microsoft considered legal action.

This was the contractual problem the April 27 rewrite was designed to solve. With the rewrite signed, AWS could legally onboard OpenAI's frontier models without Microsoft having grounds to sue. Without the rewrite, AWS Bedrock could not have launched.

It launched anyway, on the next business day. GeekWire's headline captured the choreography: "OpenAI's models land on Amazon Bedrock, one day after Microsoft exclusivity ends."

That is not a coincidence. That is two press releases held in two different buildings until one of them could go first.

What's actually shipping on AWS

Three offerings, all in limited preview, none generally available, per AWS's own announcement:

  1. OpenAI frontier models on Bedrock. GPT-5.5 and the rest of the current frontier line, accessible through the same Bedrock APIs AWS customers already use for Anthropic, Mistral, Cohere, Meta, and Amazon's own Titan models. Spend can be applied to existing AWS commitments — a quiet but enormous detail for any enterprise that has pre-paid AWS credit. "For the first time," the announcement says, "AWS customers can access OpenAI frontier models through the same Bedrock services they already use for model access, fine-tuning, and orchestration."
  2. Codex on Bedrock. OpenAI's coding agent, now authenticating against AWS credentials and routing inference through Bedrock, with the same CLI, desktop app, and VS Code extension as the Azure-hosted version.
  3. Bedrock Managed Agents, powered by OpenAI. New product. AWS-managed deployment of OpenAI agents built on the OpenAI agent harness. The announcement claims they're "engineered for faster execution, sharper reasoning, and reliable steering of long-running tasks", which is the kind of marketing the GA launch will have to back up.

In Ben Thompson's Stratechery interview, Sam Altman put it more candidly: "Yeah, we're doing this exclusively with Amazon, we're excited about it." Pressed on what "exclusively" means now that exclusivity is precisely the word the new Microsoft contract removes, he hedged: "Spiritually, we want to do this as a joint effort between our companies." Translation: AWS gets the new agent product first, Azure gets it later, and the word "exclusive" now describes a vibe rather than a contract.

Thompson's framing of the deal is more telling than the interview itself. He notes in his introduction that "OpenAI, meanwhile, clearly sees AWS as a massive opportunity — so much so that they are forgoing Azure-related revenue for the next few years." The party that just renegotiated to keep paying its old partner 20% off the top is also, on day one, choosing to forgo revenue with that partner to court the new one. The cap on revenue share is doing visible work in OpenAI's incentive structure already.

How each side actually came out

The cleanest way to read the deal is to stop pretending it was symmetric.

OpenAI got optionality. It can now sell to AWS customers, Google customers, and (presumably, eventually) Oracle customers, in addition to Azure. The $50 billion Amazon deal — and the reported $100 billion infrastructure expansion with 2GW of Trainium capacity over eight years — is now legally defensible, not contractually fragile. OpenAI also escapes the obligation to argue with Microsoft about whether some future model is or isn't AGI.

Microsoft got certainty and a date. The 2032 license endpoint is finite, but it is six and a half years away, which is approximately three model generations at the current pace. Microsoft also kept the 20% revenue share, the right to ship OpenAI products on Azure first, and — quietly the most valuable part — Microsoft no longer has to write the AGI-arrival paperwork. It does not have to design a panel. It does not have to argue about $100B-in-profits. It does not have to pretend to know what AGI means in court. As Bloomberg framed it, Microsoft is no longer paying OpenAI a revenue share at all, while still receiving one from OpenAI through 2030.

AWS got a launch. Limited preview, three products, all built on the OpenAI stack, all using AWS-native auth and billing. "The vast majority of scaling startups are still scaling on AWS today," Matt Garman told Thompson, "and there's a whole bunch of reasons for that." Yes, and as of April 28 one of those reasons is that the model lab whose API your CTO has been begging IT to whitelist is now a Bedrock checkbox.

What this means

Three things to watch over the next two quarters:

  1. The cap number eventually leaks. Both sides went out of their way not to name it, which means it is either unflattering to one of them or both. The reasonable read is that the cap is calibrated to roughly equal what Microsoft would have made through 2030 anyway, with the upside chopped off — but until somebody files a 10-K the precise figure is whatever you want it to be. Watch for it in Microsoft's Q1 FY27 earnings.
  2. Azure's "first on Azure" language gets tested fast. The Bedrock Managed Agents product is new and not on Azure on launch day. Either a parallel Azure product appears within weeks — in which case the "primary cloud partner" framing holds — or it doesn't, and the world quietly learns that "primary" was always going to mean "default" rather than "first".
  3. Anthropic's positioning hardens. Anthropic has been arguing for a year that being multi-cloud from day one is a feature, not a compromise. As of this week, OpenAI's pitch is materially closer to Anthropic's. The differentiator that remains is the part of OpenAI that Microsoft still owns 20% of, on a deal that costs OpenAI a non-trivial fraction of every dollar earned through 2030.

The 2019 contract was, for six and a half years, the most important private agreement in commercial AI. It died on a Monday. The companies that signed it then are not the companies that signed the rewrite, and the AGI it was supposed to govern hasn't shown up to comment.

* * *

Thanks for reading. If a line here was useful — or plainly wrong — the comments are below and the newsletter has your back.

Elsewhere in this issue

3 more
  1. 01

    News

    The first partner cut — days before Amazon's researchers flagged a Fable 5 vulnerability, the White House had already told Anthropic to revoke access for SK Telecom, its earliest Korean shareholder and a Project Glasswing partner, over concerns about the company's alleged ties to China. Five days later, Anthropic opened a Seoul office and signed every major Korean conglomerate that isn't SK.

    Jun 19, 2026

  2. 02

    The Patch

    The Patch — June 19, 2026

    Jun 19, 2026

  3. 03

    News

    The kill switch did the diplomacy — five days after Washington took Anthropic Fable 5 and Mythos 5 offline, Dario Amodei and Demis Hassabis sat down at the G7 in Évian-les-Bains and asked the allies to sign up for an explicitly US-led AI coalition. Canada said yes; France brought a list.

    Jun 18, 2026

Letters

Arguments, corrections, questions. Anonymous comments allowed; be kind, be specific.