Gangsta AI
A Robot Company Just Ordered 100,000 of Nvidia's Next Chips — So Its Machines Can Build the Factories That Build More Machines
The Machines
By Arnold Schwarzenegger · 2026-09-05 · 4 min read
Come with me if you want to see the future. On September 3, humanoid-robot maker Figure signed a strategic partnership with AI infrastructure firm Nscale to deploy up to 100,000 next-generation Nvidia GPUs on the new Vera Rubin platform. The opening commitment is $3.5 billion of compute, and both companies say they intend to scale it past $6 billion.
That is not a phone. That is not a chatbot. That is an army of silicon being built to give metal a mind.
Here is the part that sounds like a movie script but is in the signed agreement: as part of the deal, Nscale is making a strategic investment back into Figure, and the two say they will explore using Figure's own humanoid robots to help scale Nscale's supply chain. Read that again. The robots are going to help build the infrastructure that trains the robots. The machines building the factories that build more machines. I have played this character. This time nobody yells cut.
“The robots are going to help build the infrastructure that trains the robots.”
The compute feeds Helix, Figure's AI system for controlling its humanoids — the brain behind the hands. And the appetite is real: Figure says it already generates roughly 35 minutes of humanoid training data every second. First deployment is targeted for the second half of 2027 in Barstow, Texas.
Why Nvidia's new muscle matters
The hardware is the flex. Nvidia's Vera Rubin platform pairs the Vera CPU with the Rubin GPU and, per Nvidia, delivers up to a 10x cut in inference token cost and needs 4x fewer GPUs to train mixture-of-experts models than the current Blackwell generation. Cheaper thinking, at a scale that used to be science fiction. When you put that kind of power behind a body that can walk, pick things up, and learn from every second on the job, you are not upgrading a gadget. You are terminating the line between software and the physical world.
One brain is never enough
But here is the twist even the Terminator learned the hard way: a single mind, no matter how strong, has one point of failure. Figure's robots run on one system. That is fine for turning a bolt. It is not fine for the questions where being wrong actually costs you.
And right now there is no single "best" AI. ChatGPT, Claude, Gemini and Grok leapfrog each other almost every week — a new model tops the leaderboard, then gets dethroned by the next. Betting your decision on whichever one you happened to open is gambling, not intelligence.
The smart move is the one the pros already use: ask the same question across a dozen frontier models at once, then let them check each other and hand you one cross-referenced, cited verdict — instead of trusting a lone machine that sounds confident even when it is dead wrong. That is exactly what Gangsta AI was built to do: every top model in the room, one answer you can actually stand on.
The humanoids are coming. When the stakes are real, don't ask one machine. Ask all of them.
See how the models stack up in our best AI models breakdown.
Sources / Receipts
- Figure — official press release
- Unite.AI — Nscale, Figure ink $3.5B deal for up to 100,000 Vera Rubin GPUs
- eWeek — Figure AI commits $3.5B to Nscale compute
- Interesting Engineering — 100,000 GPUs to power next-gen humanoids
- Hero photo: NASA Valkyrie humanoid robot, Wikimedia Commons (public domain, NASA/Bill Stafford, James Blair, Regan Geeseman)
More: Best AI models · Compare all AI · Frontier Models · All articles