Why Nvidia and CoreWeave Want to Buy AI Inference Startups

Yesterday, Kevin and I broke the news that Nvidia was in advanced talks to buy Lepton AI, a startup that rents out servers powered by Nvidia’s chips and helps AI developers train and run open-source models on those servers, otherwise known as an inference provider.

A deal would put Nvidia in direct competition with a slew of other inference providers we’ve written a lot about, namely Together AI. Lepton, Together and others in the field don’t operate data centers. They rent graphics processing units from other cloud providers and resell them. But there are other reasons they’re quickly being targeted for acquisition by AI cloud providers that do run their own data centers—such as CoreWeave.

That AI cloud provider’s M&A team met with Toronto-based inference startup CentML in recent months, according to a person with direct knowledge of the discussions. The M&A team also met with Together, the largest startup in the inference provider field with more than $150 million in annualized revenue, in the last six months, said another person briefed on those discussions. (Those discussions were early and are unlikely to lead to deals.)