Nvidia is putting $6 billion into Poolside, a startup building large language models, to bankroll development of an open-weight model designed to compete directly with China's DeepSeek and Kimi. According to Techmeme, the deal ranks among the largest single checks written for an open-model effort this year.

The timing is not incidental. DeepSeek's and Kimi's releases spent much of the past year proving that open-weight models trained outside the US frontier labs can match proprietary systems on reasoning and coding benchmarks, often at a fraction of the reported training cost. For most cloud providers that's a pricing problem. For Nvidia, whose revenue depends on developers running workloads on its GPUs no matter which lab built the model, the bigger question is who controls the open alternative that enterprises reach for when they don't want to pay OpenAI, Anthropic, or Google list prices.

A $6 billion check answers that question by buying influence over the answer, rather than waiting to see which open model wins by default.

Why a chipmaker needs a model of its own

Nvidia's business is selling compute, not curating which labs get to use it. That's normally an advantage — it profits regardless of which model wins. But it also means Nvidia has no leverage if a handful of frontier labs decide to tighten access to their best models, negotiate hardware terms elsewhere, or push customers toward vertically integrated stacks that reduce reliance on general-purpose GPUs.

In our estimation, that's the real driver here: Nvidia is building a fallback option it fully controls, so that if OpenAI, Anthropic, or Google ever restrict who gets their top models — or if custom silicon from those same companies erodes GPU demand — there's a credible open-weight model in the market that was trained to run well on Nvidia hardware from day one.

What Poolside brings to the deal

Poolside, co-founded by former GitHub CTO Jason Warner and Eiso Kant, has built its reputation on training models inside simulated software-engineering environments rather than on static text corpora — a bet that coding capability, not general chat performance, is where enterprise budgets actually go. That focus lines up with where DeepSeek and Kimi have applied the most competitive pressure: code generation, agentic tool use, and technical reasoning, all areas where open weights let enterprises fine-tune and self-host instead of routing sensitive code through a third-party API.

What builders should watch for

If the resulting model ships as genuinely open weights, the practical implications are straightforward:

None of this ships tomorrow — training a frontier-class model from a $6 billion infusion still takes time, and Poolside has to actually deliver on benchmarks that matter to developers, not just training-cost headlines.

AiiN's takeaway

This deal is less about Nvidia entering the model business and more about Nvidia refusing to be a spectator in a market it currently just supplies. Betting on an open-weight challenger to DeepSeek and Kimi gives the company a hedge against two separate risks at once: frontier labs tightening access to their models, and open-source momentum shifting further toward labs Nvidia has no relationship with. For builders, the practical upside is a potential new open, Nvidia-tuned coding model worth evaluating once it ships — and a reminder that the GPU-versus-model power balance in AI is still actively being renegotiated by the biggest players in the stack.