Foundation Model Competition

AI agents are the new aggregators. Just as Google aggregated web content and Facebook aggregated social connections, AI agents will aggregate services and workflows. The company that owns the agent layer -- the interface between the user and all the services they use -- captures the majority of the value. This is why every platform company is racing to be the default agent.

Act One of generative AI was about model capabilities and the infrastructure to serve them. Act Two is about applications that generate sustainable revenue. We are now seeing the first generation of AI-native applications reach $100M ARR, and they share common traits: they solve specific workflows, they improve with usage data, and they create switching costs through accumulated context.

Microsoft's 365 Copilot has generated roughly $2B in annualised revenue, well below the $12B that Wall Street analysts projected. The gap highlights a persistent challenge: enterprises are willing to pilot AI tools but slow to roll them out across entire organisations. Usage data shows that only 15-20% of licensed Copilot users are active on a weekly basis.

We are seeing inference costs drop 10x every 18 months. That changes the economics of every AI deployment. The question is no longer whether you can afford to use AI, but whether you can afford not to. The companies building on this cost curve are pulling away from those still doing manual processes.

The most important chart in AI right now: inference cost per million tokens has dropped 95% in 18 months. This is not incremental improvement. This is a phase change. Every business case that did not work at $10/M tokens works at $0.50/M tokens. Every automation that was too expensive is now cheap. 1/12