Could Nvidia soon own the AI lab it already funds?
A few weeks ago, hardly anyone outside the AI bubble had heard of Reflection. Now a chipmaker is reportedly thinking about buying it. And the way we got here is a better story than the headline makes it sound.
Two researchers and a very cheeky pitch
Reflection was started in 2024 by two researchers who used to work at Google DeepMind. The pitch was simple: America needs its own open frontier lab. A DeepSeek of the West. In October 2025 a round of about 2 billion dollars, led by Nvidia, put its value at 8 billion. By spring it was reportedly raising again at 25 billion, which is roughly 19 billion pounds, the figure British papers have been quoting.
Then came Beam
On 5 October the lab finally showed what it had been cooking. Beam is a model with 501 billion parameters, but it only switches on 23 billion of them for each token, which is what keeps it cheap to run. Reflection says it needs three to four times less inference compute than comparable open models, and promises the weights under the Apache 2.0 licence later this month. Refreshingly, it also admits that a Chinese model, Kimi K3, is still ahead on raw capability.
Five days later, the plot thickens
On 10 October the Financial Times reported that Nvidia is in talks to buy Reflection outright, or to put in a lot more money, or to do an acqui-hire: take the people, license the technology, skip the purchase and quite possibly a long regulatory review. It is early, it could collapse, and a deal could land within weeks.
Why would Nvidia want it? It already has around 800 million dollars in there. It makes its own open models, called Nemotron, but reportedly has nothing at truly frontier scale. And it has done this trick before: last December it struck a licence and hiring deal with the chip startup Groq that was valued at around 20 billion dollars.
Hold on. Nvidia sells the chips, funds the lab, and the lab spends big money renting Nvidia chips. Now Nvidia might buy the lab. Isn't that a loop?
It does look like one. The models may be open, but the machines are rented: a deal with SpaceX reportedly worth around 150 million dollars a month through 2029, plus a Nebius contract worth more than a billion. The open part is the weights. The expensive part is the hardware underneath, and that is exactly where Nvidia sits.
The detail worth watching
If the Apache 2.0 weights ship before any deal closes, they stay open for good. If a deal closes first, the release plan is whatever the new owner decides. So the calendar matters more than the price tag. Add the reported hope in Washington that Reflection becomes an American answer to cheap Chinese models, and this stops being a pure business story.
One honest caveat: the talks come from press reports, Reflection declined to comment, and the benchmark numbers are the lab's own. Nothing is signed.