Post

A few thoughts on yesterday’s twin AI model releases: The pitch from the big labs is changing from "this thing is smarter than humans" to "it can do the same work faster and less annoying for less money," implying that there are, in fact, limits to what AI is capable of.

Compute is the main factor, and the moat. Opus 5.5 can be better and faster because Anthropic has made a lot of progress toward delivering inference-as-a-service more cheaply. It's not like their unit costs make sense, but they aren't having to do so much to ration access to this model.

More on compute — Google has the biggest compute advantage (they've been building/rolling out efficient AI processors for years, TPUs) but their models are mid. OpenAI and Anthropic have the leading models, but have struggled to find enough compute at a non-bonkers price.

If/when open-weight models fully pass GPT/Claude in capability, AI compute and user-facing services may start to be better businesses than frontier AI labs. Like, AWS/Azure selling high-performance instances for running models, and products using those models.