Inference. 99% of the customers that aren't buying GPUs to train on are either using it for inference or putting it in a datacenter where inference is the intended use-case.
For some reason inference seems to be overlooked. A lot of “ink” has been spilled over GPUs for training tasks but at the end of the day if you can’t do inference you can’t serve users and you can’t make money.
The real root of the problem is that GPU compute is not a competitive market. The demand is less for GPUs and more for Nvidia hardware, because nobody else is shipping CUDA or CUDA-equivalent software. Thus the demand is artificially raised beyond whatever is reasonable since buyers aren't shopping in a reactive market. Basically the same story as what happened to Nvidia's hardware during the crypto mining rush.
This is absolutely not true. The gap is narrowing as providers (Google, Anthropic, Deepseek) introduce cross request KV caching but it’s definitely not true for OAI (yet).