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.
-
Visa's agent payment rail: Visa's Intelligent Commerce Connect is now in pilot, enabling agents to initiate payments across multiple protocols and card networks, which adds a transactional layer to the runtime dependency story.
-
Agent identity on the open web: Cloudflare and GoDaddy are building authentication standards for AI agents using DNS and cryptographic signatures, extending the identity question beyond cloud platforms to the web itself.
-
Stanford's jagged frontier data: The 2026 AI Index shows agent accuracy on computer tasks jumping from 12% to 66% in a year while transparency scores dropped from 58 to 40, sharpening the tension between capability and accountability that runtime governance is supposed to address.
-
The control plane thesis: a16z's Big Ideas 2026 argues that enterprise backends built for human-speed, 1:1 interactions aren't architected for a single agent goal to trigger thousands of recursive sub-tasks, which is the architectural pressure driving the runtime investments described here.