Nvidia buys the middle of the stack

Nvidia is having a month. $6B to license Poolside’s Model Factory. Now, reports indicate Nvidia has agreed to buy Hugging Face for $12.9B (likely, but not officially signed yet). And earlier Stripe signed to acquire OpenRouter, reportedly north of $7B.

None of these companies build frontier models. They sit in the middle of the stack, the parts every reference architecture leans on without thinking about it. They provide the registries where open models reside and the routers that direct user requests.

The strategy isn't hidden. Jensen Huang has said it out loud: free AI is great for chips. Nvidia ships its own Nemotron open-weight models, and the Poolside engineers are joining that team. Meanwhile the most-used open models on OpenRouter have been Chinese for months, reportedly over 60% of tokens. So this looks like Nvidia assembling an American open-weight supply chain, from the factory that trains the models to the registry that hosts them. If you want open models to keep existing, that's arguably good news.

But I've spent years in enterprise software, most of it on the JVM, and I've seen what happens after the buyout. Oracle bought Sun in 2010 and Java stayed open. The weights stayed free, so to speak. The terms around them didn't, and by 2023 the Oracle JDK was priced per employee and everyone I know was migrating to OpenJDK builds. An open license guarantees a free artifact, but it cannot prevent a proprietary ecosystem from locking you in.

This is why today's AI architectures need to treat the model layer as swappable. A gateway you control, deciding which model gets each request and what data is allowed to reach it, was table stakes when the risk was just pricing. Now the router has an owner with its own P&L and the registry might belong to the company selling you the GPUs.

$12.9B for a company doing $150M in revenue. Nobody pays that for a file host. It’s about owning where ecosystem habits live.