,

Brett Adcock: Figure 4, Helix 3 and Why the World Ran Out of Compute

Figure AI founder Brett Adcock says humanoids are in their “flip phone era,” that Figure 4 will be the iPhone moment, and that his company could not buy more compute because the world ran out in September 2026. Speaking with Peter Diamandis and Dave Blundin on the Moonshots podcast from Figure’s headquarters, where robots now outnumber employees, Adcock covers home robot timelines, the Helix 3 model, China’s dancing robots, Tesla, robot safety, Figure’s split from OpenAI, and why his AI company Hark could soon be bigger than Figure.

TLDW

Adcock frames the humanoid race in four chapters: functional hardware, a working AI recipe, scaling intelligence and scaling robots, and says Figure is now in chapter three. He hopes to see robots doing long, multi-hour tasks in homes they have never seen within six months, would be shocked if that is not shown by the end of 2027, and says Figure 4’s hands are nearly indistinguishable from a gloved human’s. He calls China Figure’s closest competitor but says its robots mostly replay motion-captured dances, and says he has not seen much from Tesla in two or three years. He also warns that robots are needed to stop superintelligent AI from bossing humans around, and says Figure expects billion-dollar training runs next summer on a 100,000-GPU cluster.

Thoughts

The most newsworthy detail is Adcock’s account of the compute market. Seven months ago, four vendors bid for Figure’s cluster and Figure “beat the piss out of them” on price. A few months later, it asked for double the capacity and got no bids, then one, and that vendor set the terms. He says he predicted in early 2026 that compute would be unavailable by 2027, and that it happened in September 2026: “the world ran out of compute.” Older H100s now trade above list price, and someone recently offered him eight nodes as if that were a lot. The constraint runs further down too. A large memory maker told Figure it had no capacity for humanoids at all, and the hosts note that new memory fabs are already oversold before they open, with no humanoid demand in their forecasts. That is a striking counterpoint to talk of an AI bubble, coming from a buyer rather than a seller. It also explains why Figure’s planned Nscale deal, roughly 100,000 next-generation Nvidia GPUs for the second half of 2027, matters so much: Adcock says it will mean more than 100x Figure’s current training compute and billion-dollar Helix runs next summer.

Adcock’s most original argument is moral, not technical. He says that as digital AI gets smarter, there is a future where systems on servers make money by “bossing around humans” to do physical work for them, and notes services already exist where AI agents hire people to open bank accounts or run errands. His conclusion is that humanoid robots are “super necessary” to avoid that dystopia: if superintelligence needs things done in the physical world, it should be telling robots, not people. It is a self-serving argument from a robot maker, but it reverses the usual framing. Most worry about robots taking human jobs. Adcock worries about the alternative, a world where the AI is smart but has no body, so humans become its hands. Blundin adds a second benefit: a robot that can build itself can build almost anything, ending the era of one identical product for everyone.

On competition, Adcock is blunt and specific. He calls “the collective China” Figure’s closest competitor, with around 150 humanoid companies whose hardware and models are mostly “crap right now” but improving fast every year. What he dismisses is the viral dancing. He explains those robots run reinforcement learning controllers trained on people in motion-capture suits, replayed open loop with no reaction to the world, which he says a college senior with one robot could build. Out-dancing China, he says, would mean pivoting Figure entirely, because it is “literally 180 degree opposite work” from a robot that can reason through a home it has never seen. He says Unitree robots arrive with tiny compute and no hands, which is why researchers bolt compute backpacks onto them. On Tesla, he says only that he has not seen much in two or three years beyond teleoperated demos. His prescription is simply speed: work harder, move faster, beat them.

The home robot timeline is more careful than it first sounds. Adcock refuses to ship until the robot is an “Apple style quality product” that walks into a home, talks, and works for a dozen-plus hours a day without teleoperation or freezing, and he compares rushing it to the Humane and Rabbit AI devices. Figure is testing in 20 to 100 new California homes at any time, and Helix 2.5 did zero-shot tasks like making beds in 30 rental homes. Helix 3, which the hosts saw in a secret lab called Blackbox, is “a baby AGI moment,” like a two-year-old in an adult body. He hopes for long-horizon home work within six months and would be “very shocked” if it is not shown by the end of 2027, but that is a demo milestone, not a sales date. The price target the hosts mention, $300 to $500 a month on a lease, is roughly $10 a day, and Adcock prefers leasing so customers get upgrades as the hardware improves.

The Hark section shows where Adcock thinks the bigger prize may be. He says Hark Pro, a month old, has passed 1.2 million signups, already out-earns Figure, and could be bigger than Figure within six months, maybe 90 days. His reason for building it is a sharp critique of OpenAI: Figure’s Series B was led by OpenAI, but its team rarely came to Figure’s office, progress was “completely lopsided” in Figure’s favor, and Figure ended the partnership. He also predicts the iPhone will go away within 10 years, replaced by AI devices co-designed with their models, and says Hark has that hardware in the lab now. It is a big claim from a founder with obvious incentives, but it ties back to the robot thesis: the hardware is only the distribution for the AI, whether that hardware walks or fits in a pocket.

Key Takeaways

  • Figure has well over 1,000 robots on campus, more than its roughly 700 to 800 employees, and they greet visitors, open doors and use elevators.
  • Adcock’s four chapters of the humanoid race are good hardware, a working AI recipe, scaling intelligence and scaling robots.
  • Figure is testing in 20 to 100 new California homes at a time, and Helix 2.5 completed tasks zero-shot in 30 rental homes.
  • He hopes for long-horizon work in any home within six months and would be shocked not to see it by the end of 2027.
  • Helix runs entirely on board the robot, while speech and fleet orchestration run in the cloud.
  • Figure says it has shown the first data scaling laws for humanoids, with performance improving log-linearly as data grows.
  • Figure 4 will be its biggest generational leap yet, with new hands Adcock says are almost indistinguishable from a gloved human’s.
  • Home robots will likely lease for $300 to $500 a month, and Figure will handle maintenance itself at first.
  • Figure has moved its supply chain out of China, after paying tariffs as high as 270% in one week, and will not do defense work.
  • The BotQ factory can make over 40,000 robots a year, and Adcock sees a path to a million a year from one facility.
  • He expects about 20 billion humanoids eventually, and fully lights-out robot manufacturing within five years.
  • Figure spends low nine figures a year on compute and expects billion-dollar training runs next summer.
  • Hark Pro has over 1.2 million signups a month after launch, and Adcock thinks Hark may soon be bigger than Figure.

Chapters

01:15 Figure’s Robot-Filled HQ and When Robots Enter Homes

Visitor robots walk the campus with real-time speech powered by Hark, and Figure now has more robots than people. Adcock says commercial work demands speed while homes demand handling endless variety, and that both are intelligence problems. Figure will not launch in homes until the robot works like an Apple product, and he hopes Helix 3 can do long tasks in unseen homes within months.

11:06 The Four Stages of Humanoid Robotics

Adcock recalls the Figure 01 demo of a robot putting a K-Cup into a Keurig with a neural network instead of C++ code, which he says proved deep learning works for robotics and led to a raise of nearly $700 million. Moving from tabletop work to whole-body control took a year, because the legs, hips and hands have to be trained together.

16:47 China’s Dancing Robots and Building Helix

Adcock says China’s viral robots replay motion-capture routines open loop, and that demos of language models controlling robots are slow, sped up and tied to fixed cameras. Helix was built from scratch, runs fully on the robot, and is now starved mainly of data and compute. Figure spends nine figures a year on compute and is doubling capacity.

31:15 Robot Pricing and the Figure 4 Breakthrough

Home robots will need a $300 to $500 monthly lease. Figure 4 aims to be cheaper and more capable, after seven hand iterations, and Adcock says it will make Figure 3 “look like a caveman.” He explains why Figure builds its own actuators after an off-the-shelf elbow actuator arrived three times too big and too heavy.

41:43 Faces, Memory, Safety and Security

Figure is still working on the robot’s face and refuses to add “googly eyes,” which Adcock calls a con. Robots now remember people, and personalities will move between devices. Safety runs from rules in nonvolatile memory up to system architecture, a large cybersecurity team guards against hacking, and Figure has rotated out of Chinese parts.

55:00 Space, Defense and Robots Building Robots

Adcock says putting a human in space costs tens of millions of dollars an hour, but Figure is not ready for vacuum. He reaffirms Figure will not do defense work and will not train robots to defend themselves from vandals. A robot recently handled a new warehouse zero-shot, and he expects fully lights-out manufacturing within five years.

1:13:22 Scaling to Millions of Robots

BotQ can make over 40,000 robots a year, and Adcock sees a path to a million a year if Figure cracks its current challenges. He expects around 20 billion humanoids, whole companies running 24/7, and argues robots are needed so AI does not end up directing human labor.

1:27:14 Competing With China and Tesla

Adcock calls the FCC ban on Chinese robots a shock and says he wants a level playing field. China is the closest competitor, Tesla has shown little lately, and the Bay Area is the best place in the world to build robots. He recounts going from Vettery to Archer Aviation, and says Archer went public only because it needed capital.

1:42:33 Training Data and the Compute Crunch

Figure’s Project Index collects rich physical data, including forces, that cannot be scraped from the internet, and it now has about 15x more data than the model the hosts saw. Adcock describes compute bids drying up and says the world ran out of compute in September 2026.

1:48:50 Hark, New Devices and the Split With OpenAI

Hark builds personal AI with speech, memory and computer use, and its own hardware. Adcock calls Meta’s glasses one of the worst products he has tried, predicts the iPhone will be replaced, explains why Figure dropped OpenAI, and says he works every night and loves it.

Notable Quotes

“We’re like in the flip phone era for humanoids.”

Brett Adcock, on where the humanoid industry stands

“A figure 4 with a glove and a human with a glove is almost indistinguishable.”

Brett Adcock, on Figure 4’s new hands

“I personally haven’t seen it much out of Tesla in two or three years.”

Brett Adcock, on competition from Tesla

“We just got to go beat them.”

Brett Adcock, on China’s 150 humanoid companies

“I’m fairly confident we’re going to have a lot more robots than humans.”

Brett Adcock, on the long-term humanoid market

“Robotics is super necessary to kind of like avoid like a dystopian future.”

Brett Adcock, on AI directing human labor

“September 2026, the world ran out of compute. You just can’t get it.”

Brett Adcock, on failing to buy more GPUs for Helix 3

Watch the full conversation here.

Related Reading