Figure AI just froze Figure 04's design and says it found a physical AI scaling law

This is huge. Figure AI just announced for a design lock for Figure 04, its next-generation humanoid. Engineering specifications are final, there will be no more changes to hardware architecture, and component manufacturing and lab testing are underway, including actuator, battery, and hand verification.

(I'm guessing that's near 100% locked down, but there's still some wiggle room for a few minor items that might pop up.)

CEO Brett Adcock called Figure 04 "our biggest leap between robot generations to date." He also cautioned: "A ton of work left... don't expect us to unveil this anytime soon." 

I'm pretty eager to see Figure 04, as I'm sure many other are ... but I guess we will have to practice patience.

Also, and maybe more importantly: a scaling law claim

Figure's AI team published results for Helix 2.5, their home robotics model. In 420 tests across 30 unfamiliar homes, it achieved a 56% task success rate. 

That's not amazing.

That's not a passing grade by conventional standards.

But Figure's argument is that 56% is not the point. The point is that the failures are a scale problem, not an architecture problem ... they are due to not having enough data, to the model size being too small, and not having enough compute. The argument is that if you fix those three things, the success rate climbs, exactly the way LLM benchmark scores climbed through GPT-2, GPT-3, and GPT-4.

If that's true, it's huge.

"Our AI model, Helix, becomes more capable the same way every learned system does: with more data and compute," says Adcock. 

He's betting this is the first measurable "human-to-robot transfer scaling law" demonstrated on a real humanoid. And the Helix 02 release added more evidence for the thesis. Figure deleted the final 109,504 lines of hand-coded C++ from its robot stack, replacing them entirely with learned behavior. 

Adcock called it a "Software 2.0" moment: the explicit logic that engineers laboriously wrote is gone, replaced by a neural net trained on data. If he's right, this will scale in performance with more data and get better and better and better with more time.

The compute bet

In September 2026, AI cloud firm Nscale committed $3.5 billion in compute to Figure under a multi-year partnership. The deal covers up to 100,000 NVIDIA Vera Rubin GPUs, with infrastructure being built in Barstow, Texas, targeting operations in the second half of 2027. 

The agreement has an intent to scale to more than $6 billion total.

This is essentially a reservation of cloud capacity, with Nscale taking a strategic equity stake in Figure in return. NVIDIA CEO Jensen Huang described the collaboration as a "physical AI flywheel": train on Vera Rubin, validate in Isaac Sim, deploy on NVIDIA GPUs inside Figure robots.

The math is ... big. 

Adcock has said he believes scaling humanoid robots will eventually require "hundreds of billions of dollars, and more data and compute than large language models." 

Committing to $3.5 billion in compute before Figure 04 has shipped a single commercial unit is a statement of conviction about how the technology works. Either the scaling law thesis is right, and this compute turns 56% into 80% or 90%, or the thesis is wrong, and Figure has a very expensive cluster and a hard story to tell investors.

Meanwhile, Figure 02 is retired

While Figure 04 is still in component testing, Figure 02 has completed its run. The robot spent roughly 11 months at BMW Group's Spartanburg, South Carolina plant, running 10-hour shifts Monday through Friday. It loaded more than 90,000 parts, logged over 1,250 hours of runtime, and contributed to production of more than 30,000 BMW X3 vehicles. 

Figure is now retiring the Figure 02 fleet entirely.

(We all saw the Terminator videos.)

Figure 03 took its place. The 1,000th Figure 03 unit was built over the summer at Figure's BotQ factory, at a production rate of one robot per hour. Cost came down more than 90% between Figure 02 and Figure 03, which is critically important, and Figure 03 is already at BMW Spartanburg, taking on a new task: sorting unsorted components into sequencing trolleys for the assembly line.

What this adds up to

Figure is making three simultaneous bets. 

1. First: the scaling law is real and measurable. 
2. Second: Figure 04 represents a hardware platform good enough to test that law against at scale. 
3. Third: $3.5 billion in compute, phased in as the software model matures, will close the gap from 56% to deployment-grade reliability.

If all three are right, Figure's $39 billion valuation starts to look conservative. If the scaling law doesn't transfer cleanly from language to physical tasks, the entire investment thesis is ... problematic.

Sources

• Humanoids Daily — Figure 04 design lock announcement
• Intelligent CIO — Nscale $3.5B compute deal
• Humanoid Guide — Figure supercomputer build
• Figure AI — Figure 03 at BMW Spartanburg
• Humanoids Daily — Adcock on data and compute