Figure commits $3.5B to Nscale for up to 100,000 GPUs
Brett Adcock says the compute goes to scaling Helix, and that Figure has entered the chapter of the humanoid race money can buy.
Figure is partnering with Nscale to deploy up to 100,000 GPUs on NVIDIA's Vera Rubin platform. Brett Adcock announced the deal, saying Figure is committing $3.5 billion initially with plans to scale beyond $6 billion. His reason for the size of it is simple. To ship a robot into every home, he says, Figure needs a massive amount of compute.
We're committing $3.5 billion initially, with plans to scale beyond $6 billion
Corey Lynch at Figure spelled out what the GPUs are for. The deal scales Helix, Figure's robot AI model, to up to 100,000 Vera Rubin GPUs. Lynch framed it the same way Adcock did, that a general purpose robot in every home needs scale in both data and compute.
What is actually new here
Nothing was shown on a robot. This is a purchase of compute and a hiring pitch, not a demo. Adcock says he is watching Helix do things you would not believe and that Figure is going all in, but nobody outside the company has seen that. He also says buyers will see perfection in F.04, the next generation, after Figure chose to iterate across generations rather than chase perfection on v1.
Adcock's own framing is the clearest read on why the money is being spent now. He says there are chapters of the humanoid race you cannot buy with unlimited money, the deep engineering of the hardware and designing the right AI recipe. Then there are chapters money can buy, and Figure says it has entered the one about scaling intelligence through data and compute.
The data half
Adrian Macneil of Foxglove put the announcement next to Index, Figure's creator network for real world task video, which landed a week earlier. His point is that Figure just stacked the two halves of the physical AI bottleneck, Index as the data bet and Nscale as the compute bet. He notes Figure's own writeup says data alone cannot solve this. His caution is that capture is not the hard part, and that finding the small slice of data that actually teaches the model is what compounds.
Not everyone was impressed by the number. Ryan Julian's entire reply to the $3.5 billion figure was a comparison to buying a house on a mortgage in San Francisco.
I’m watching Helix doing things you wouldn’t believe
buying a house on mortgage in San Francisco