Q&A with Sam Altman and AWS CEO Matt Garman about OpenAI's new partnership with AWS, Bedrock Managed Agents, Trainium chips, and more

As I noted yesterday, today's Stratechery Interview is early in terms of my timing — Tuesday instead of Thursday — and late in terms of delivery …

Stratechery 2026-04-28 Ben Thompson

Context & Ripple Effects

OpenAI’s earlier coverage has emphasized a unifying product vision alongside infrastructure deals and an expanding enterprise agenda. This partnership places that infrastructure discussion directly alongside AWS’s effort to extend its cloud lead through AI services.

For AWS, the interview follows coverage of its plan to add AI capabilities across its platform and of new Bedrock AgentCore controls and memory features. The OpenAI relationship therefore connects a major model provider to AWS’s agent platform and its Trainium chip strategy.

First-order effects

AWS gains a higher-profile OpenAI relationship around its cloud AI stack, including Bedrock Managed Agents and Trainium, while OpenAI gains another major infrastructure and distribution counterpart.
Enterprise buyers evaluating OpenAI workloads on AWS get a clearer connection between OpenAI, AWS-managed agent tooling, and AWS hardware choices.

Second-order effects

The arrangement raises pressure on other cloud providers to pair model access with more complete managed-agent, deployment, and infrastructure offerings rather than compete on compute alone.
Trainium’s relevance becomes more tied to whether AWS can turn an OpenAI-associated workload into sustained customer adoption, increasing the stakes of AWS’s in-house silicon strategy.

Third-order effects

If large model providers continue to spread infrastructure relationships across clouds, AI infrastructure is likely to become a more negotiated, multi-layer market in which model access, agent platforms, and specialized chips are bundled together.
The durable competitive divide may increasingly be between clouds that control both AI services and differentiated hardware and those that primarily resell third-party compute; the pace of that split remains dependent on customer deployment choices.

The trend: This is one instance of AI providers and cloud platforms turning infrastructure partnerships into integrated stacks spanning models, agent management, and purpose-built accelerators.

Discussion

@benbajarin Ben Bajarin on x
Wrote this growth thesis on $AMZN driven by @awscloud before this OpenAI deal. This only gives me far more confidence in the AWS growth story. https://thediligencestack.com/ ...
@inafried Ina Fried on x
New @axios: OpenAI to make its models available via Amazon's servers Codex and GPT-5.4 available now in preview form with GPT-5.5 coming soon, per AWS XEO @mattsgarman
@kyliebytes Kylie Robison on x
That was among the biggest sticking points between the two companies. I remember not many months ago, I was told that that this clause was critical leverage for OpenAI! Now, vamoose!
@kyliebytes Kylie Robison on x
No more AGI clause is huge: “Revenue share payments from OpenAI to Microsoft continue through 2030, independent of OpenAI's technology progress, at the same percentage but subject to a total cap.”
@aaronpholmes Aaron Holmes on x
Dropping the AGI clause is the biggest win for Microsoft in today's deal - removes a ton of uncertainty around whether they'd retain OpenAI IP rights in the coming years
@aaronpholmes Aaron Holmes on x
NEW: MSFT and OpenAI have scrapped the controversial “AGI” clause from their deal. Now MSFT has IP rights until 2032, but they aren't exclusive, meaning OAI can sell via AWS and other clouds. And Microsoft will stop sharing rev from OAI sales on Azure: https://www.theinformation.…
@reckless Nilay Patel on bluesky
If anyone thought LLMs would actually get to AGI this clause would still exist, just sayin www.theverge.com/ai-artificia...
@benbajarin Ben Bajarin on x
Given what we have been seeing/hearing on TPU/Tranium inference economics, margins are even better (inference margins are high to begin with) but OpenAI emphasizing Tranium is partially also for better margins.
@cryptopunk7213 @cryptopunk7213 on x
fantastic interview with sam altman and amazon AWS ceo matt garman on todays partnership announcement > sam thinks pre-training and post-training will converge as a singular training stack for ai models (very bullish inference test-time compute) > he also thinks the model and
@mattsgarman Matt Garman on x
Great to chat with @sama on @stratechery about what we're building together in Bedrock. It's a deep dive on how our teams are making it much easier to manage agents powered by @OpenAI models at scale. Excited to see what customers go build with this. Thanks to @benthompson [image…

Chronicles