Just caught the Stratechery interview with Sam Altman and Matt Garman about what's really happening between OpenAI and AWS, and honestly, it reframes everything we've been thinking about enterprise AI.

So the headline everyone's talking about is obvious: Microsoft's Azure exclusivity deal with OpenAI is done. OpenAI can now offer models through AWS and other clouds. But here's the thing—that's almost the boring part of the story.

The real shift is way more fundamental. We're moving from "companies calling APIs" to "companies running actual AI agents inside their infrastructure." Think of it like this: instead of ChatGPT being a smart chatbot you chat with, imagine an AI colleague that actually works within your company, accesses your databases, respects your permission systems, and gets stuff done.

Bedrock Managed Agents is what that looks like in practice. It's not just "OpenAI models now available on AWS." It's OpenAI's models deeply embedded into AWS's entire native stack—identity management, permissions, logging, security, deployment, the whole thing. The models run inside your VPC, your data never leaves AWS, and the agent understands your company's boundaries.

Sam made a point that stuck with me: he used to see the model and the supporting infrastructure as separate things. Now he doesn't anymore. When Codex does something amazing, you can't really tell how much is the model being smart versus the system around it being well-designed. They're becoming one thing.

Matt Garman brought up something equally important: enterprises right now are literally stitching together their own solutions. They want agents that remember context, work across their systems, understand their data, respect their security. Every company is doing this themselves. What AWS and OpenAI are doing is saying: we'll handle that integration for you, so you don't have to.

The permission and security piece is huge here. Before, it was like a "castle and moat" security model—everything's local, assume it's safe. Now you need zero-trust architecture. Every agent needs its own identity, its own permissions. They're still figuring out the mental models for this. Like, should an AI agent use your employee account but identify itself as an agent? They don't even have the basic concepts nailed down yet, but they know it matters.

What's wild is how this mirrors what happened with cloud computing 20 years ago. AWS made it so startups didn't need to buy their own servers and hire infrastructure teams. Now OpenAI and AWS are trying to do the same thing for enterprise AI: lower the barrier to actually deploying agents that work, without companies having to assemble models, permissions, data systems, and security architecture themselves.

Sam also dropped something interesting about the future of pricing. Right now they price by tokens, but that's already becoming outdated. Their new model costs more per token but needs way fewer tokens to get the same answer. What customers actually want to pay for is "intelligence delivered," not "tokens consumed." It's a subtle shift but it changes how you think about the whole business.

On the competition front, this is the moment the game shifts from "who has the best model" to "who can turn models into actual enterprise infrastructure." Google's going full vertical integration. AWS is taking a different path: strong infrastructure layer, partner with the best models, let enterprises choose what works. Sam and Matt both seem genuinely convinced they're building something differentiated here, not just "finally we can access OpenAI through AWS."

One more thing: Sam mentioned that demand for frontier models is still astronomical. People aren't price-shopping yet. Everyone wants the cutting-edge model because that's what gets work done. But as costs come down and smaller models get better, you'll probably see a mix emerge. Still feels early though.

The whole conversation reads like two companies that actually respect each other's strengths building something real. Worth reading the full Stratechery piece if you want the technical details, but the core idea is solid: we're not talking about distribution channels anymore. We're talking about a completely different computing paradigm emerging.