We have three dominant AI labs consuming the majority of compute demand nowadays:

1. OpenAI

2. Google (Gemini)

3. Anthropic

Google runs Gemini almost entirely on TPUs, while using Nvidia GPUs in the cloud mainly for enterprise clients and inference when needed.

Anthropic runs almost entirely on Google TPUs and Amazon Trainium, though in recent months they’ve also been tweaking their stack to work on Nvidia GPUs. Google and Anthropic inference workloads mostly run on TPUs, not Nvidia. Anthropic is currently leading compute demand.

OpenAI is, in fact, the biggest customer of Nvidia GPUs, which is why Microsoft, Oracle, CoreWeave, and others are all providing Nvidia GPU compute to OpenAI.

But there is a problem: OpenAI is losing market share, in enterprise, in coding, and among normal users. Google is aggressively pushing Gemini across all of its services and has been very successful at making Gemini feel native to almost every Google product. The same is happening across Android, and soon likely with a new OS replacing ChromeOS.

OpenAI has no real moat in any of the above. OpenAI is losing market share. Which means Nvidia’s biggest customer may not end up consuming all the demand it already committed to hyperscalers and neoclouds.