OpenAI on Bedrock: the distribution side of the Microsoft amendment

The announcement, jointly published yesterday by OpenAI and AWS,¹ is short enough that it can be summarised in one sentence: OpenAI's latest GPT models, the Codex coding agent, and a new "Bedrock Managed Agents" service built on OpenAI are available in limited preview on Amazon Bedrock, and enterprise customers can draw down their existing AWS commitments to pay for it.

That last clause is the one that matters. The models being on Bedrock is the headline; the AWS commitment drawdown is the deal.

I want to walk through why.

What was actually announced

The AWS post lists three things going live in limited preview:

1. OpenAI's current GPT family, available as Bedrock-hosted endpoints.
2. Codex, OpenAI's coding agent, available as a managed service.
3. "Bedrock Managed Agents powered by OpenAI", an agent runtime where the orchestration layer is AWS's and the model layer is OpenAI's.

The commercial wrapper, quoting the AWS post: "Enterprise customers can apply usage of OpenAI models on Bedrock toward their existing AWS enterprise agreements and committed spend." OpenAI's parallel post phrases it slightly differently, "AWS customers can use their existing committed cloud spend", but the substance is the same.²

What is not disclosed: revenue share, whether AWS is paying OpenAI a wholesale rate or a per-token rate, whether there is a minimum-commitment floor from AWS to OpenAI, or whether AWS is taking inference capacity risk (i.e. pre-buying tokens) or acting purely as a reseller.

Limited preview means a small number of named customers. Neither party has disclosed which.

The Microsoft amendment is the context

You cannot read this deal without reading the amended Microsoft partnership, announced in late 2025, which removed Azure's exclusive right to host OpenAI models for third-party API distribution and replaced it with a right of first refusal on net-new compute capacity.³ In the original 2023 arrangement, OpenAI's API was a Microsoft product everywhere except on openai.com. After the amendment, OpenAI can sell its API through any cloud, subject to specific structural constraints around new training compute.

So this is the first deal that uses the freedom the amendment created. Google Cloud will be next; the question is weeks, not quarters.

The structural read: OpenAI has spent two years being a single-channel API business and is now becoming a multi-channel one. That is a meaningful change in distribution shape, and the AWS deal is the proof.

Inference economics: who is bearing what cost

The frame I want to test this against is inference economics, specifically, who pays for the GPUs that serve these tokens.

There are two plausible structures and the announcement doesn't tell us which:

Structure A: AWS resells OpenAI-hosted inference. AWS routes Bedrock calls to OpenAI's own infrastructure (Azure, CoreWeave, Oracle, the various clusters OpenAI has been assembling). AWS takes a margin on the pass-through. AWS bears no GPU capacity risk. This is operationally simple and economically thin, AWS is essentially a billing relationship.

Structure B: AWS hosts the weights on its own Trainium and Nvidia capacity. OpenAI licenses weights to AWS for hosted serving. AWS bears capacity cost and earns gross margin on inference. This is the Anthropic-on-Bedrock model, more or less.

Structure B is operationally hard, it requires OpenAI to ship weights to AWS, which is a thing OpenAI has not historically done, but is the only structure that meaningfully advantages AWS, because under Structure A AWS is just a payment processor with a logo.

The presence of "Bedrock Managed Agents powered by OpenAI", an AWS-built orchestration layer, is the small tell that suggests we are closer to B than A, at least for the agent product. AWS does not generally build managed services on top of pure-resale APIs; the operational dependency is too tight. But for the raw model endpoints, A is more likely. Mixed structures are normal in these deals.

Either way, the binding constraint for OpenAI here is not revenue per token. It is distribution. Which brings us to the AWS commitment drawdown.

The commitment drawdown is the actual product

Large enterprises do not buy AI by signing a new vendor contract for each model. They buy AI by spending money they have already committed to a hyperscaler. AWS, Azure, and Google Cloud sit on hundreds of billions of dollars of customer commitments,⁴ which are essentially budget the customer has pre-allocated and is now looking for ways to consume.

When AWS says OpenAI usage counts against your AWS commitment, what they are saying to the enterprise buyer is: you do not need to convince procurement to open a new vendor. You do not need a new MSA. You do not need a new security review of the billing entity. The model is just another line item on the AWS invoice you were going to pay anyway.

This is a much bigger deal than it sounds. The single largest friction in enterprise AI procurement, by some distance, is vendor onboarding. Removing it for a frontier model is a meaningful unlock. It is also the precise reason Anthropic's Bedrock distribution has been so commercially effective.

The structural implication: OpenAI is now competing with Anthropic on Anthropic's distribution turf. Anthropic's enterprise lead through Bedrock has been one of the more durable commercial moats in the model layer. It is, as of yesterday, a less durable one.

What this is a case of

This is a case of the model layer accepting that distribution is the binding constraint, not capability. OpenAI's product advantage at the API tier has not translated into the enterprise share its consumer brand would predict, and the diagnosis, for some time, has been channel. Microsoft was a strong channel into Microsoft customers and a weak channel everywhere else.

It is also a case of hyperscalers becoming agnostic. AWS now hosts Anthropic, OpenAI, Meta's Llama, Mistral, and its own Nova family on the same procurement rail. The Bedrock thesis, be the place where you buy frontier models, regardless of which one, is closer to delivered than it was last week.

What to watch

• Pricing parity. Whether OpenAI's per-token rates on Bedrock match openai.com or carry an AWS markup. Anthropic's Bedrock pricing is at parity with its direct API; OpenAI may or may not follow.
• General availability date. Limited preview is a soft launch. The shape of the deal becomes clearer at GA, when the customer list and pricing are public.
• The Google announcement. If OpenAI on Vertex AI follows in the next quarter, the multi-channel pivot is complete and Microsoft's distribution premium evaporates entirely.
• Anthropic's response. Whether Anthropic announces deeper Bedrock integration, exclusive Bedrock features, or a counter-move into Azure. The Bedrock relationship was Anthropic's moat; it now has a tenant problem.
• OpenAI's disclosed enterprise ARR mix. If the next disclosed figure shows a step-change in non-Microsoft-channel enterprise revenue, the AWS deal is doing what it was built to do.

Footnotes

¶