Azure gets exclusive distribution of stateless OpenAI APIs. AWS gets exclusive rights to Bedrock Managed Agents, built on a co-developed stateful runtime. One is prompt-in, response-out. The other is the layer where agents maintain memory, identity, and context across long-running workflows.

Pause on that split. OpenAI could have simply listed its models on another marketplace. Instead, it built infrastructure that gives each agent its own identity within AWS's security perimeter, persists session state across steps, and logs every action against existing compliance frameworks. The runtime honors IAM roles, VPC boundaries, audit trails. It plugs into the identity architecture enterprises already trust.

When The Information reported that AWS customers greeted the arrival with something close to indifference, those customers had spent years building production systems on Claude, Llama, Mistral. They knew models are swappable. The shrug was recognition of a commodity. And the deal structure suggests OpenAI knows it too. Sam Altman, in a Stratechery interview recorded before the announcement, described the model and its "harness" as no longer entirely separable. The harness being identity, permissions, memory, evaluation. A striking admission from a company whose entire valuation story has been about the model. Even OpenAI, it seems, sees the model becoming a component inside something larger and more operational, something that touches the enterprise where it actually lives.

Google arrived at a similar recognition from a different direction. At Cloud Next, the company folded Vertex AI into the Gemini Enterprise Agent Platform. The model marketplace receded into a feature of a larger system: cryptographic agent identities, a gateway inspecting agent-to-agent traffic at the protocol level, a registry of approved tools, sandboxed runtimes launching 300 instances per second. The model platform became the agent platform. The naming change is blunt.

In value chains, profits tend to migrate away from the modular, interchangeable layers and toward the integrated ones. Both deals look like that migration happening in real time. The commodity layer gets distributed broadly. The integrated layer gets built exclusively.

Both moves converge on the same ground: state management, session persistence, permission propagation, audit logging. The infrastructure that makes intelligence persistent and accountable inside boundaries enterprises already operate within. Though "accountable" deserves a caveat here. The runtime can tell you who authorized an agent and what it did. Sound judgment remains outside its scope, and this infrastructure doesn't pretend otherwise.

None of it is glamorous. Most of it is invisible when it works. That's probably what makes it sticky. Models swap out in an afternoon. The runtime that remembers what your agents are doing, who authorized them, and what they're allowed to touch accumulates dependency quietly, in the infrastructure layer nobody's watching closely enough.