Summary

A long interview with Brett Adcock of Figure AI, recorded on site after a full campus tour, April 2026. This is the 'after' to DK-90's June 2025 Bloomberg interview, and the pairing is the reason both are worth holding: the same founder, the same claims, ten months apart. The interview is friendly rather than adversarial, and the sponsor reads are woven into the conversation, so the pushback that makes the 2025 interview useful is largely absent here.

• WHAT CHANGED, and it is worth stating plainly: in June 2025 he stood by 100,000 robots shipped within four years, on lines he said could each build 12,000 a year. In April 2026 the near-term number is 'thousands of robots' this year, with tens and hundreds of thousands and eventually a million units a year as later ambitions. Not a contradiction, but a noticeably more modest immediate figure.
• The BMW deployment, described in 2025 as ongoing daily work, is now described in the past tense: a small batch ran every day for six months, they 'learned a ton', and it caused them to refactor the whole approach to commercialising the software. That refactor produced Helix 2, their second-generation model, launched a couple of months before this interview.
• He names the bottleneck explicitly, and it is not manufacturing: 'the big gap here is getting the robots ready to do scale autonomous operations.' He says commercial demand overwhelmingly exceeds what he can ship, and that putting out a thousand robots with 'a thousand problems every hour' would be bad for everyone. On Optimus specifically: 'this is not a manufacturing problem. This is an intelligence problem.'
• The reliability trajectory, given as his own progression: Figure 1 would fault after about an hour; Figure 2 roughly once a day; Figure 3 runs all day with faults seen weekly rather than per-robot. He notes the honest corollary — as the fleet grows the absolute number of faults rises even as the rate falls.
• His stated bar for success is specific and, by his own account, unprecedented: a robot in a home doing seven to ten hours of useful work with no human intervention, every day, indefinitely. 'Nobody's ever shown that.'
• The OpenAI split, told far more bluntly than in 2025: OpenAI led the Series B and brought in Microsoft; after a year of collaboration his team was 'running circles around' them; 'so I fired them.' He adds that he suspected OpenAI was itself moving into robotics, and concedes of the original strategic rationale, 'it turned out I was kind of wrong on that.'
• The valuation arithmetic he offers when asked whether $39bn is a risk is worth quoting because of how loose it is: roughly $30-40 trillion is paid in wages annually for the work he intends to automate, tech companies trade at 10-20x revenue, therefore the business is enormous. He does not address timing, share, or margin.
• Vertical integration is his stated differentiator: Figure designs motors, rotors, stators, sensors, structure, kinematics, joints, batteries and packs. His reasoning is control of the supply chain — 'without that you're left at the mercy of some vendor'.
• Security has escalated to a degree worth recording: tinted glass and strict physical and digital controls, prompted in part by finding a drone hovering at the office window.
• He is also running two other companies he self-funds — Cover, using terahertz imaging radar licensed from NASA JPL to detect concealed weapons at school entrances, aiming at first beta schools by year end; and Hark, an AI lab out of stealth two weeks before this interview building models and hardware. He claims US school shootings are up tenfold in ten years; that figure is his and is unverified here.

Why it matters

The pair is more useful than either half. Between June 2025 and April 2026 the near-term shipping number came down, the flagship customer moved into the past tense, and the identified bottleneck moved from manufacturing to intelligence — while the total addressable market claim stayed exactly the same. That is the pattern to watch in humanoid robotics generally: the ceiling argument does not move, the floor keeps being restated. His own reliability figures are the most checkable thing here and they do show real progress. Against that, this interview had no adversarial questioning, so the claims arrive unchallenged in a way they did not in 2025.

g1ESjEGG1SM-transcript.txt