Why would Nvidia buy an open model startup?
Ok, here is a weird one. Nvidia, the company that sells the shovels in the AI gold rush, is reportedly in early talks to buy Reflection AI, or at the very least to pour a lot more money into it. Early talks, and they could still fall apart. Nvidia has already put in around $800 million. Reflection was last talked about at a $25 billion valuation, which is a lot for a company that only just showed its first model.
That model is called Beam, and it landed on October 5. It has 501 billion parameters, but only about 23 billion switch on for any given word, which is why it is cheap to run. The licence is Apache 2.0, so anyone can use it, and the weights are promised later this month. Reflection says it needs three to four times less compute than similar open models and keeps pace with a leading Chinese one on coding. Fine. But another Chinese model is still ahead on raw power, and nobody outside has had a chance to check the numbers yet.
Hold on. Why would a company that makes its money selling chips to everybody want to own a model it hands out for free?
Here is the likely logic, and to be clear, it is a reading and not a confirmed motive. Every good free model is a reason for someone to buy more GPUs. Open weights spread like wildfire, and they spread on whatever hardware people can get. Meanwhile Chinese open models keep getting better, and Nvidia's share of the Chinese AI accelerator market has reportedly slid to 34%, with a local rival at 28%. An American open model is a way to keep the world building on Nvidia's stack.
The shape of the deal is interesting too. Reports say one option is an acqui-hire: take the people, license the tech, skip the full purchase. Nvidia did something like that with Groq in December, in a deal reportedly worth about $20 billion. Why that shape? Because a straight takeover invites a very long antitrust look.
So the real question is not "why a model?" It is "who ends up owning the free stuff everyone builds on?"