OpenAI and AWS Close the Loop: Codex Is Now an AWS Service

OpenAI and AWS closed a loop that has been conspicuously open for two years. As of April 28, Codex, GPT-5.5, and OpenAI Managed Agents are available in limited preview on Amazon Bedrock. AWS CEO Matt Garman put the logic plainly: "Their production applications run in AWS. Their data is in AWS." For enterprises already running on AWS infrastructure, this is the last meaningful barrier to a fully sanctioned Codex deployment — with existing security policies, billing visibility, and compliance controls — removed in a single announcement.

The timing is impossible to ignore. This dropped one day after OpenAI renegotiated its Microsoft partnership to end Azure exclusivity. OpenAI can now serve across any cloud provider, not just Microsoft's. The very next day, AWS became the first non-Azure cloud to offer OpenAI's flagship coding agent. That sequence is not coincidental. OpenAI spent years bound to Microsoft Azure as its exclusive cloud path, which created a real adoption friction for any enterprise not already deep in Microsoft's stack. The Azure exclusivity is over. AWS is the immediate beneficiary.

The practical implications are more structural than they first appear. When Codex runs on Bedrock, it inherits AWS's enterprise control plane — IAM, PrivateLink, guardrails, encryption, CloudTrail logging. You authenticate with AWS credentials. You route inference through Bedrock. You charge usage against existing AWS cloud commitments. There is no separate OpenAI account required, no separate procurement process, no separate security review for an external AI vendor. For IT organizations that have already approved AWS as their compute and data platform, that is a dramatically lower activation barrier than anything that existed before this week.

The billing angle deserves more attention than it has gotten in early coverage. AWS cloud commitments are a line item that many enterprises have already negotiated and approved. When Codex usage can flow through that commitment instead of creating a new budget line for an external AI vendor, procurement becomes a non-issue for a large class of potential customers. CFOs who have already approved AWS spend do not need to have a second AI tooling conversation. They just need to know Codex is on Bedrock. That is a meaningful wedge against Anthropic, which competes on Bedrock alongside OpenAI's models — but only Anthropic's models, not OpenAI's — and has to be evaluated, approved, and budgeted separately from the cloud infrastructure that everything else runs on.

There is also the $100 billion number. Amazon invested $50 billion in OpenAI, and OpenAI committed to $100 billion in AWS spend over eight years while agreeing to use 2 gigawatts of Amazon Trainium chips. That is not a vendor relationship. That is a structural interdependence that puts OpenAI's infrastructure roadmap partially inside AWS's chip and compute roadmap. For developers building on OpenAI's APIs today, that kind of deep alignment tends to produce more stability than a pure commercial licensing arrangement — though it also means OpenAI's roadmap is now partly shaped by AWS's commercial incentives rather than just OpenAI's own product vision.

Ben Thompson's Stratechery interview with both Altman and Garman produced the most useful frame: this is "Codex in AWS," and the key insight was that a lot of Codex's operational complexity comes from running it securely. Hosting on Bedrock offloads that security and infrastructure burden to AWS, which is the actual pitch for regulated and legacy-heavy enterprises. Thompson framed it as "the infrastructure company doing what it always does — taking software complexity and turning it into managed service simplicity." That is exactly right, and it is exactly the wedge that OpenAI needs to drive adoption beyond the developer-forward companies that already figured out how to run Codex on their own.

The Managed Agents piece is the longer-term bet and should not get lost in the Codex announcement. AWS is positioning this as "Codex in AWS" today, but the real target is enterprise agents that go beyond coding — workflow automation, document processing, research tasks, cross-system orchestration. If that product matures and enterprises start deploying OpenAI-powered agents as managed AWS services, OpenAI gets a second distribution layer for agentic work that does not depend on a developer deciding to use Codex. That is a meaningful strategic hedge for OpenAI, and a meaningful new capability for AWS to sell against Microsoft and Google in the enterprise AI race.

For practitioners, the near-term question is simpler: if your company runs on AWS and has been waiting for a sanctioned path to Codex that maps to your existing security review and procurement workflow, that path exists now. The limited preview means it is not production-ready for everyone yet, but the direction is clear. You can begin evaluating what it looks like to run Codex inside your existing AWS tenancy, bill it against your cloud commitment, and govern it with your existing IAM policies. That is not nothing. It is the difference between "we're evaluating AI coding tools" and "we're deploying AI coding tools at scale."

What is less clear is what this means for the Microsoft relationship. OpenAI still ships on Azure first — that remains the primary cloud path — and Microsoft remains a major shareholder. But the non-exclusivity arrangement means AWS gets to compete for OpenAI enterprise revenue on equal footing with Azure for the first time. Enterprises that have been locked into Azure because that was the only game in town now have a choice. That is probably good for OpenAI's negotiating position with both cloud vendors. It is probably good for enterprise customers who want leverage. It is probably more complicated for Microsoft, which spent years building Azure's OpenAI practice as a competitive moat.

The headline is straightforward: OpenAI just turned AWS into its largest potential distribution channel without needing a separate sales relationship. The sub-headline is more interesting: OpenAI just made multi-cloud a real option for enterprise AI tooling, which means the coding agent race is now running through cloud commitments rather than cloud exclusivity. That changes how the competitive dynamics work, who wins the next tranche of enterprise deals, and what it means for developers who have been building on OpenAI's APIs with one cloud provider in mind. The answer is no longer just Azure.

Sources: OpenAI Blog, AWS About Amazon, Reuters, Stratechery