Half of GDP Is Labor. Figure Wants Humanoids to Do the Work.

Sarah Guo opens No Priors with the scale of the round: Figure AI, 21 months old, had just raised $675 million from Microsoft, OpenAI, Nvidia, Intel, and Jeff Bezos. The guest is founder and CEO Brett Adcock, whose company is building humanoid general-purpose robots for jobs that are unsafe or that nobody wants. Elad Gil, the other host, is on the episode too. Guo's first question is the résumé: how does anyone get from a farm to software to electric vertical takeoff to humanoids? Adcock calls it a pretty normal path. Gil, dryly, says that is what he did too.

Adcock grew up on a third-generation farm in Illinois, started coding young, and has spent about 20 years building companies — a little over ten in software, a little under ten in hardware. After selling the recruiting platform Vettery in 2017, he immediately started Archer Aviation, which builds fully electric vertical-takeoff-and-landing aircraft. The planes are meant to use helicopter pads inside a city and, in his example, get someone from the studio back to San Francisco in under 20 minutes instead of a two-hour drive. Archer is now a public company.

Guo presses on the exotic part: most people do not sell software and then start an aircraft company. Adcock says he grew up around hardware and had wanted to work in deep tech for a long time; the only way he saw to do it was to self-fund. The motives were traffic — half the world lives in cities, and congestion keeps getting worse with no real solution — and a belief that all transport besides rockets will go electric in our lifetime. He started Archer out of the University of Florida's engineering school, built a lab and aircraft there for the first two years, then moved the company to California. Figure, he says, was less of a leap from software into Archer than Archer was into humanoids.

The thesis, Adcock says, starts with a single assumption: that it is possible to build a humanoid, meaning legs, arms, and hands, that can do human-like work. If that is true, he thinks it becomes the biggest business in the world by about an order of magnitude. Half of GDP is labor, which makes the opportunity much larger than the entire transportation market. The second claim is economic: robots doing real work every day would, in his view, push goods and services toward an age of abundance. The third is that, over a long enough period, the company could move the AGI timeline. That combination, he says, distilled into perhaps the most important business he could work on, which is why he left to start Figure 21 months earlier.

What Figure is actually trying to do is commercial: get humanoids into the market at significant scale, build a fleet, and build an AI data engine that trains those robots to do useful work. Guo puts the robotics-industry objection on the table. Plenty of people in the field argue that bipedal walking is unstable, that the human form is not optimal for many jobs, and that a robot could have ten arms and be stronger. Adcock splits the field into two schools. Either you rebuild thousands or millions of special-purpose robots for special use cases, or you build a humanoid.

The reason Figure takes the second path is that the whole world was built for humans. Stair slopes, door handles, shelf heights, and coffee-machine buttons were optimized around ordinary bodies. If people were ten feet taller with six arms, the built environment would look completely different. Asking whether the humanoid is the right form factor, he says, is the wrong question. Humans are a weird biological species, but that is the world that exists. Go to Mars today, he argues, and you will still want coffee, walking, and a habitat — and you will build a human world around those needs. If you want a general interface to automate work, the physical-world equivalent of the keyboard and mouse is a human form.

One hardware system, in this telling, can be amortized over millions of tasks. The alternative is today's robotics industry: a special robot for every task, which means a company, a brand, a culture, a capital raise, and a team with something like 15 skill sets no single full-stack software person has — firmware, embedded software, an operating system, rotors and stators, electromagnetics, batteries, BMS, power distribution, thermals, controls, and AI. Guo notes that specialized robots are how the industry actually works. Adcock does not dispute the description. He disputes the inference: that is not efficient, and it is the wrong analogy to reason from.

Guo asks whether this is even the right decade to start. Adcock says they have started to prove the idea at MVP level: the robot is doing pretty useful things after 21 months, of which the first year was mostly hiring. Useful run-rate, in his accounting, is closer to a year or a year and a half. Figure has shown two divergent kinds of public demo. On the industrial side, a bipedal robot moves bins from pallets onto conveyor systems, fully autonomous and end to end. On the consumer side, it does speech-to-speech reasoning and manipulation: talk to it, let it see the scene, and it can grab an apple or plates or make a coffee on a Keurig. That stack, he stresses, is neural nets only — speech and camera video in, inference, trajectories out.

Industrial first is a sequencing choice, not a lack of interest in the home. Shipping into manufacturing, supply chain, and logistics, he argues, builds the AI data engine faster because robots go out the door sooner, and it builds manufacturing volume that pulls cost down. The home is the market they are extremely excited about, and also a mess: unstructured, every house different, a higher variance of failure. Drop a bin in a warehouse and nobody cares. Drop "the number one dad cup" at home and it is a different story. Safety, pricing, and consumer expectations all get harder. So the company is using factories as the on-ramp.

What entrepreneurs fear, Adcock says, is a technical wall — an upper bound you cannot push through. They do not even know where that bound is right now, which is why the next 24 months excite him. They are still searching for it. Guo then asks the timing question directly: why now, and why not five years ago? He gives three reasons, plus a fourth. First, the powertrain is batteries and motors. Energy in cells and torque density in motors improved enough over the last two decades that the robot can actually run. Ten or twenty years ago, lithium-ion packs and electric actuators could not deliver useful runtime or force. His analogy is a Tesla: about 100 miles of range a decade ago, 300-plus now, on the back of roughly 7 percent compound growth in watt-hours per kilogram.

Second is locomotion. A decade ago, bipedal humanoids were clumsy, slow, and falling over; the DARPA Robotics Challenge is his example of why you could not look at one and imagine it in a home. Figure walked a humanoid in under a year from the C-corp filing, work he thinks is among the best in the world on bipedal controllers. Third is AI: the compute and the algorithms were not feasible ten years ago. The fourth, which he calls arguable, is speech as the default user interface, including the industrial unboxing and initialization. Humans already work with gestures and language. Five years ago, talking to the robot to configure it was not possible either. That is the OpenAI partnership's opening.

Guo asks what walls they cannot imagine hitting. Adcock's answer is that the whole problem is walls. If you are afraid of them, this is the nightmare scenario — "a funhouse of walls." One of Figure's five corporate values is being aggressively optimistic. Guo had looked at an internal risk slide with him: for a component risk, the second mitigating factor was sprint harder, as in, we will figure it out. He owns that. Work harder is, at this point, sometimes the only risk mitigation on the page.

The risks are not rhetorical. Nobody, he says, has delivered a commercial humanoid into a market in human history. The bar is not a demo. It is an equivalent of human-like work, done reliably and continuously over months and years, while humans remain very productive. The robot has more than 30 degrees of freedom, so the action space is enormous. You cannot write a C++ script for every orientation in the world. The short answer is robot learning, which has not worked extremely well in human history either — the closest live experiment is self-driving cars. Hardware that has never shipped, plus reliability, safety, and performance, plus learning, plus unknown unknowns in the shadows: that is the stack.

The reason they do not feel an upper bound now is more mundane. They can see the roadmap for the next 12 to 24 months, and it is optimization: make the robot more reliable and faster. They are not trying to get an end-to-end application working for the first time. They have already done that, including for a first client's use cases. It just has to be more robust. Guo then asks how you assemble a team that multi-domain in a year. Adcock's method is the same one he used at Vettery and Archer, and it is not glamorous.

Before Figure, he wrote mission, vision, and values, a ten-year master plan, and a culture document that went online: move fast, aggressively optimistic, work in the office five to seven days a week. Then he drew an org chart in detail — controls, AI, actuation, battery systems, kinematics, integration and test, industrial design — and the skill sets under each, down to rotor, stator, transmission, sensors, thermals, and motor controls. He spent the first year hyper-focused on building what he calls the world's best engineering team. The execution was brute force: find the people he thought were the best in the world, then 300 phone calls over six months, all starting from cold email. The first 30 to 50 people, he identified the role, found the person, closed them, wrote the offer, wrote the 30-60-90, sat on the floor next to them, and attended every engineering meeting.

On deployment, Adcock says companies are already paying for robots as Figure delivers them this year. Over the next 12 to 18 months, he hopes those robots are in clients' real facilities, in real work cells, getting paid for the work. A work cell, for listeners, is a job: at BMW, moving a bin or a box, or moving bins onto a conveyor, in a defined location. After that come the known-hard problems: extreme reliability and safety as the fleet goes from hundreds to thousands on a factory floor, fleet management, AI training at scale, over-the-air updates, and manufacturing at scale. If the humanoid can do useful stuff reliably, he thinks the rest is extremely doable because those problems have been solved before. The failure mode he does not want is waking up in ten years with a robot that works and an inability to make enough of them.

Guo has not seen many orgs this vertically integrated: Figure makes its own actuators and wrote its own operating system. Why do that when the project is already huge? Adcock rejects the idea that build-versus-buy is a religion. They run the trade at every component. The philosophy is that they would rather not build. Building means job listings, hiring, managing people, performance, integration, repairs, and a supply chain — "it's a mess." They ended up building almost everything anyway, GPU, CPU, and battery cell excepted, because there is no mature supply chain for a humanoid. Nothing they could buy satisfied the need. They were forced to design, and in a lot of cases to manufacture. A sensor vendor became a dangerous dependency last year; they had to stand up teams on the sensor and software sides to fix it.

Product development is iterative design, not long stretches of research and analysis. They want to test, learn, and fold the learning into a continuum of hardware and software that, he says, will never be good enough to stop updating. The process starts with customers and requirements: how many pounds the robot must lift, how long it must run, where it charges, safety so the battery cannot burn down a building, IP ratings on actuators. Then come three company-wide gates Adcock brought over from Archer — conceptual, preliminary, and critical design review. Compared with software sprints, story points, and A/B tests, the method is the same scientific loop on a much longer clock. Hardware timelines are an order of magnitude longer, sometimes a hundred times longer, which is why they are more methodical about trade studies before cutting metal.

Actuator design is the example. At the top: hydraulic, pneumatic, or electromechanical? Then linear versus rotary drive — Figure uses a lot of rotary drives. Then the cross-section: how the joint is actually built, against the requirements of what it has to do. All of that is top-down through trade studies, then iterative design: pick, build as fast as you can, ask whether it did what you wanted. The trick is compressing the cycle. A friend who runs the drone-delivery company Zipline told Guo that beyond raw technical ability, the interview screen is intuition for rapid prototyping. Speed is one of Figure's five values for the same reason: if you find the mistake in three months you have time to fix it; if it takes a year or two and you are wrong, you die. They still run a 9 a.m. standup every morning, rain or shine, because bringing up a new robot is a mess — software and hardware talking for the first time, everything breaking.

BMW is Figure's first announced commercial customer, buying robots for manufacturing. The first site is Spartanburg, South Carolina, BMW's largest plant by vehicle production globally, about 1,400 cars a day, and conveniently in the U.S. Over the next 12 to 18 months, Figure hopes robots there are doing real work. Five use cases are already picked and ranked; the first is in end-to-end reliability work now. The OpenAI partnership is the other public pillar. OpenAI started in robotics, Adcock notes, and some of the people on this project date to that period. At the highest level they are building new models for commercial robots. OpenAI brings what he calls the best vision-language model and the team to implement it for language reasoning: you say fold the laundry, the centralized brain breaks that into grab the hamper, walk there, pick it up.

OpenAI does not know how to grab the hamper or command the motors. That is Figure's stack: neural nets that drive hands and joints. High-level language reasoning plus low-level motor control, combined. The more time they spend on it, the more they feel you cannot do this at real scale without an intelligent VLM sitting on top. Thirteen days after the announcement they put out a speech-to-speech video: the robot could talk, ask what was in the scene, and use memory of what had happened. Custom models with OpenAI are the next six to 12 months of work.

At home, Adcock's personal task list is ordinary: laundry, dinner, and the kids' toys on the floor every evening. Over a long enough period he thinks everybody will own a humanoid to do work, and labor will be optional — you will choose to work or not. The robots they have shown are expensive; he does not think they stay that way. Cost comes down from design-for-manufacturing plus the experience curve of volume, "Ford Model T school." Ship millions into industrial settings first, and use that as the data-collection and manufacturing-volume path into the home. Social acceptance, he says, will not arrive on a Friday morning. It has to be earned through a safety record in places like BMW, the way Archer had to think about aircraft over cities. Sci-fi has not helped: almost every movie ends poorly for humanoid robots. Trust will be gradual, and every video and every commercial launch has to be designed with that in mind.

Gil's last question is the sharp one: do humanoids make AGI come faster, and is the missing actuator the real safety barrier? Adcock's view is that the humanoid has to be figured out pre-AGI. A dystopian version is AGI that wants something done and pays you, or forces you, to go do the physical work. Walk into a warehouse today, he says, and people are already actuators for a model: a computer, a barcode scanner, a warehouse management system telling them the next step. "They're literally a cyborg." His hope is that humanoids take that work before superintelligence arrives. The team also thinks large language models still struggle to reason about the physical world and plan actions, and that data coming off the robot may be the last big piece of the AGI timeline. Figure hopes to matter to that timeline in five to ten years. Jury's still out, he says. They will see.