The AGI clause is gone, and OpenAI is on Bedrock by Tuesday

The thing to notice about the restructured OpenAI–Microsoft partnership is the timing. The amended agreement was announced; within 24 hours, OpenAI's frontier models, Codex, and the Managed Agents runtime were live on AWS Bedrock. Twenty-four hours is not a procurement timeline. It is the timeline of something that was wired up weeks ago and waiting for the legal language to clear.

So the structural story is not the announcement. It is the pre-positioning.

What actually changed

Three substantive amendments, per the joint release and the supporting materials Microsoft and OpenAI each posted to their newsrooms.¹

First, the AGI clause is removed. The original 2019 agreement, as reported by The Information and confirmed in fragments across years of OpenAI board statements, contained a provision under which Microsoft's commercial rights would terminate when OpenAI's board declared AGI achieved. That clause is gone. Microsoft's IP licence now runs to 2032 on a fixed term, with no AGI-trigger termination.

Second, the IP licence is non-exclusive. Microsoft retains access to model weights and research through 2032, but OpenAI is free to license the same IP to other cloud providers. AWS is the first counterparty; the agreement does not preclude GCP, Oracle, or others.

Third, OpenAI's revenue-share obligations to Microsoft are capped. The original 20% gross revenue share is replaced with a structure that the parties describe as "tiered with a defined ceiling", the ceiling itself is not disclosed in the public release, though the 10-Q filing Microsoft is expected to make in early May should clarify the accounting treatment.

In exchange, Microsoft's equity position is converted from the previous capped-profit interest into a more conventional preferred stake, and Microsoft's $13bn cumulative investment is restructured with revised liquidation preferences. The press release uses the phrase "aligned with standard venture preferences", which is doing some work.

Reading the AWS launch

Bedrock now offers gpt-5.2, gpt-5.2-mini, the Codex code-generation endpoint, and the Managed Agents runtime, with pricing roughly 8–12% below Azure OpenAI Service equivalents at the input-token tier and identical at output.² AWS is also offering a 90-day Bedrock credit for enterprise migrations from Azure OpenAI, capped at $500k per account.

A few things are worth noting about how this was structured.

The Managed Agents runtime on Bedrock is the interesting piece. Codex and the chat models are commodity inference; any cloud can host them. Managed Agents, OpenAI's hosted agent execution environment, with tool use, memory, and the orchestration layer, is the thing OpenAI has been positioning as its enterprise moat. Putting it on Bedrock day one signals that OpenAI does not believe the moat sits in distribution. It sits in the runtime itself.

The pricing delta is small but pointed. AWS undercut Azure on input tokens by roughly a tenth. This is not a cost-pass-through move, input tokens cost roughly the same to serve regardless of which hyperscaler is doing the serving, and the underlying compute economics are similar. It is a strategic discount, funded out of AWS's margin, designed to give enterprise buyers a reason to switch the default. The frame here is inference economics: at the cloud layer, inference is now competitive enough that hyperscalers will eat margin to capture the workload.

What this is a case of

This is a case of an exclusive distribution arrangement converting into a multi-cloud one, in the standard pattern. The pattern is: exclusive partner provides early capital and credibility; partnership matures; exclusive partner's strategic interest diverges from the IP holder's; the arrangement is restructured to non-exclusive; the IP holder lights up alternative distribution within days because the alternative distribution was already built.

Salesforce on Azure in 2021 followed this shape. VMware's relationship with AWS in 2017 followed this shape. The OpenAI–Microsoft restructure is the AI version, with the additional wrinkle that the original deal contained the AGI clause, a provision that became, as the capability questions hardened, less of a theoretical termination and more of a live commercial uncertainty. Removing it is, among other things, a clarification: neither party wants the partnership's commercial terms hostage to a board vote on a definition.

The AI safety as market position frame applies here in an inverted form. The AGI clause was, originally, a safety provision, the idea being that if AGI was achieved, the commercial arrangement should terminate so that AGI deployment wasn't governed by a vanilla cloud-services contract. Removing it does not mean OpenAI has abandoned the safety posture; the RSP is still in place and was updated in March. But it does mean the safety posture is no longer doing structural work in the commercial agreement. Safety is now a product feature and a regulatory positioning lever, not a contractual termination trigger. That is a meaningful shift in how OpenAI thinks safety relates to its enterprise business.

The Microsoft side

Microsoft gives up exclusivity and a permanent revenue share, and gains: a fixed-term licence through 2032, capped revenue obligations (Microsoft's payments to OpenAI for compute are presumably also restructured, though this isn't disclosed), and freedom to invest in and ship competing models without the awkwardness of the existing arrangement. Microsoft has been visibly building out its in-house model capability, the MAI-1 family, the Suleyman-led Microsoft AI division, and the previous structure made the internal politics of that build-out fraught. Now it doesn't.

What Microsoft loses is the implicit promise that Azure was the only place to get OpenAI at scale. That promise was already eroding, Azure OpenAI Service capacity constraints have been a running enterprise complaint through 2025, and several large accounts had been quietly multi-sourcing through the OpenAI API direct anyway. Making it official lets Microsoft stop pretending the constraint wasn't there.

What to watch

A few specific follow-ons.

Google Cloud's response. GCP does not currently host OpenAI models. The non-exclusive licence makes that possible. If a GCP–OpenAI announcement arrives within the next quarter, the read is that OpenAI is treating cloud distribution as fully commoditised and will license to whoever pays the bandwidth. If it doesn't arrive, AWS may have a near-term exclusive on the multi-cloud expansion, which would be its own interesting deal structure.

The revenue-share ceiling. Microsoft's next 10-Q should disclose enough about the restructured economics to back out the cap. If the cap is low, say, total cumulative payments not far above what Microsoft has already received, then the restructure is closer to a buyout-in-instalments than an ongoing partnership. If the cap is high, the partnership remains economically meaningful.

Azure OpenAI Service pricing. Microsoft has two options: hold pricing and accept share loss, or match AWS and compress its own margin. The choice will tell you how Microsoft is now thinking about Azure OpenAI as a business, defensive moat or active growth product.

Managed Agents adoption telemetry. OpenAI does not disclose runtime usage by cloud, but the developer ecosystem signals, SDK download splits, integration partner announcements, the cloud affiliations of the largest agent deployments, will be visible within a quarter. The FDE market structure frame predicts that enterprises will follow their existing cloud relationships rather than chase the 8–12% input-token discount; switching costs at the cloud layer dwarf model-pricing deltas. If Bedrock OpenAI usage grows faster than that frame predicts, the read is that enterprise AI buying is decoupling from incumbent cloud relationships faster than the rest of the cloud market did.

The last point is the one I'd weight most. Decoupling model choice from cloud choice is the headline of the announcement. Whether enterprise buyers actually decouple, or whether they keep buying OpenAI from whichever hyperscaler their VPC is already in, is the structural question the next two quarters will answer.

Footnotes

¶