Summary

The first proven robotics market is industrial data collection, where avoided failures matter more than labor replacement. ANYbotics put ANYmal in the low hundreds of thousands of dollars, while Boston Dynamics described Spot at roughly $100,000-$300,000 depending on configuration. Customers can justify those prices when downtime costs hundreds of thousands of dollars per hour or a detected leak prevents a potential $3 million daily loss. As Péter Fankhauser put it, customers “don’t even want the robot. They want the data.”
The promised $1-an-hour worker is a scale-end-state, not a price available today. The host’s arithmetic used eventual $20,000-$40,000 robots operating about 20 hours daily for roughly 40,000 hours over five years, with that cost level possibly requiring 100,000 deployed units. Jonathan Hurst said costs are coming down and Agility offers both CapEx and robot-as-a-service models, but withheld its current hourly price; human labor remains the value reference.
1X will place only a handful of NEO humanoids into homes during 2026, despite selling its first 10,000 preorder slots within days. At $500 per month, this is explicitly an early-adopter product that will be “rough around the edges” and may fall. Bernt Børnich nevertheless said development now points toward a useful, potentially fully autonomous experience—while carefully stopping short of promising it; guidance or teleoperation may still be needed for tasks that must work immediately.
The key embodied-AI moat may be access to vastly more data, not simply ownership of more robots. 1X’s pyramid runs from small, high-quality teleoperation samples through sensor-equipped humans and egocentric footage to the “ludicrously immense” pool of general video. Its decade-long wager is that humanlike hands, proportions, skin behavior and movement make that footage transferable: “Our cross embodiment is the human.”
Chinese hardware creates two different competitive tests: cost versus trust. Fankhauser argued that a robot capable of walking and backflips remains only “a piece of hardware” without autonomy, inspection intelligence, cybersecurity and workflow integration; ANYbotics currently sources zero of its robot from China. Amanda McMaster was categorical that Chinese humanoids should not enter the US, citing reported data leaks and warning that Boston Dynamics cannot repeat the semiconductor experience. Fankhauser separately warned against putting “15 cameras in your critical infrastructure that somebody else controls.”
The speakers disagree sharply on whether physical AI approaches a discontinuity or a long compounding curve. Børnich is “extremely sure” hard takeoff is less than a decade away and currently bets on three years for robots to build robots, chip fabs, data centers and mining infrastructure. Hurst rejects a singularity framing: there is “no silver bullet,” robots still require far more training data than humans, and progress resembles “a snowball picking up steam going down a hill.”
Safety and weaponization are strategic constraints, not secondary product details. Digit V5’s scaling claim is that a balancing humanoid can leave a guarded work cell without a physical barrier. McMaster described on-robot physical control alongside reasoning that could run in the cloud, while Hurst emphasized supervisory safety systems. Fankhauser and McMaster resisted armed robots, but McMaster left open a future national-security decision under pressure: “Today we don’t have to make that decision.”

Deep dive

1. Quadrupeds won the first commercial wedge by selling data, not robots

Fankhauser’s answer to the host’s proposed four-legged “centaur” humanoid was use-case specificity: four legs offer stability, footholds and mobility across stairs, slippery floors, snow and grass; two legs fit narrower human spaces and eye-level interaction. “It really depends on the use case.”
ANYbotics packs its ANYmal inspection robots with thermal cameras, microphones, gas sensors and substantial onboard compute. The thesis is explicitly “not about labor replacement” but superhuman sensing: catching micro gas leaks or overheating equipment before assets begin losing hundreds of thousands of dollars per hour.
McMaster said Boston Dynamics has Spot deployed with more than 500 customers across 46 countries, making it, in her telling, the planet’s most-used mobile autonomous robot. The same unit can read gauges, detect vibration or acoustic anomalies during the day and patrol a security perimeter at night.
Both executives resisted reducing ROI to displaced wages. McMaster conceded labor replacement will be “an element” because it is easily measured, but emphasized work humans skip and failures robots prevent; the preferred destination is removing the “dull, dirty, dangerous” jobs that damage bodies.

2. Industrial autonomy is an uptime system, not a battery demo

ANYmal runs roughly two hours before autonomously docking, with some customers scheduling as many as 40 missions daily around specific industrial events. Real-time navigation and image-quality checks stay onboard because connectivity can fail; historical and contextual analysis can move to the cloud.
The sharpest deployments are where access itself is costly or dangerous: offshore transformer stations otherwise reached by helicopter, Norway at minus 20°C, deserts reaching 40°C to 60°C, and methane-containing facilities. ANYbotics built a special robot guaranteed not to create a spark in explosive atmospheres.
Spot’s battery lasts about 90 minutes, but paired robots and autonomous docking support near-continuous work; McMaster reported more than 3,000 hours of mean time between interventions. Spot sells for about $100,000 base to $300,000 fully configured, with customers expected to see ROI within two years.

3. Cheap Chinese hardware collides with workflow and sovereignty moats

Fankhauser said zero percent of ANYbotics’ robot is currently sourced from China, describing that as a historical consequence of local development rather than a ban. He remains open to globally commoditized components where sensible, while distinguishing passive metal from components carrying core expertise or security exposure.
His competitive answer was that Chinese robots can “walk beautifully” and “do backflips,” but customers buy autonomy, inspection intelligence, workflow integration and trusted data. ISO cybersecurity certification matters when a robot is collecting sensitive information inside critical infrastructure.
McMaster answered “No” when asked whether Chinese humanoids should be allowed into the United States. She said Boston Dynamics’ robots are entirely built outside China and Taiwan, reported that data from some imported quadrupeds is being back-channeled, and called for allied manufacturing, IP protection and a national robotics strategy.
The contrast is useful: Fankhauser framed China primarily as a hardware supplier and solution-layer competitor; McMaster framed it as a sovereignty threat. Fankhauser specifically warned against allowing another party to control cameras inside critical infrastructure.

4. Defense demand is real, but the executives draw the line before weapons

Fankhauser personally believes Europe has a responsibility to develop defensive technology, yet said ANYbotics neither serves nor presently intends to enter military markets. Commercial demand already supplies focus, and military autonomy would require different communications, millisecond response, remote control and human-in-the-loop operation—not merely four legs.
On armed robots, Fankhauser was blunt: “I personally don’t like it. I hate it.” He characterized circulating quadruped-with-gun videos as early demonstrations, contrasting them with drones matured through battlefield necessity, and recalled joining Boston Dynamics and others four years earlier in condemning robot weaponization as “dangerous and risky and stupid.”
McMaster likewise maintained an anti-weaponization stance while confirming non-weaponized government work, including explosive ordnance disposal. The host pressed that Chinese deployment might ultimately force Boston Dynamics to respond; her answer preserved the unresolved choice: the company would make “the right” decision if that time came, but industrial markets remain today’s focus.

5. NEO enters homes in 2026 as an unfinished platform

Børnich promised that 1X will ship NEO in 2026, but narrowed the commitment to “a handful of customers.” The first 10,000 preorder allocation sold out within days, and the stated subscription was $500 monthly; the product was framed as an early-adopter experience rather than polished mass-market reliability.
“Getting a home humanoid in 2026 is going to be rough around the edges,” Børnich said, agreeing that robots will fall. He now thinks NEO may ship very close to full autonomy and remain useful without intervention, but explicitly declined to promise it; for tasks that must work immediately, the host noted that guidance or teleoperation may still be involved.
NEO is also being launched as a developer platform: a customer could buy robots, fleet-management software, sensor gloves and the matching vision system, collect task data, fine-tune within 1X’s system and deploy a specialized workflow. A consumer skill store could let an outside developer sell something as specific as a salad-making skill.
Børnich also wants third parties able to run their models on NEO, while insisting 1X’s vertically integrated model should be best. His rationale is focus: if 1X spends years integrating every customer’s ERP, it cannot concentrate on the general problem of embodied AGI and “an abundance of labor.”

6. Teleoperation is both fallback labor and a permanent expert interface

Børnich laid out a continuum rather than a binary choice: an operator can intervene occasionally, improve the resulting dataset and reduce future intervention, while exceptionally difficult workflows may remain fully teleoperated. The host’s operating model was a small remote team monitoring 30 or 40 intermittently active NEOs.
Some teleoperation will never disappear, Børnich argued. While traveling, he wants to inhabit a NEO in Norway—“Put the hat on NEO. I am NEO”—to inspect parts and join meetings; another enduring use is “expert in place” at remote power stations where rare failures do not justify automating every repair.
Yet teleoperation no longer exploits NEO’s full hardware fidelity: a remote operator cannot feel everything the robot’s new hands sense, and comprehensive haptics become slow and clumsy. For baseline dexterity, 1X increasingly prefers humans performing tasks naturally while wearing unobtrusive sensors matching the robot.

7. Human likeness is 1X’s bridge to internet-scale training

Børnich’s data pyramid starts with small, high-quality teleoperation samples selected from successful attempts, then expands through humans wearing robot-equivalent sensors, egocentric human footage and finally general video. Each layer is less robot-specific but dramatically larger; the top helps align the model, while the bottom supplies scale.
General video is “absolutely ludicrously immense” relative to any robotics dataset likely to be collected soon. Making it usable requires obsessive embodiment matching—from humanlike hand geometry and arm span to nonlinear skin deformation, friction and impact energy—because the robot must be as close to a human as possible.
The catch-22 is deployment: robot data is ultimately better because it includes actions, forces and touch, but useful data at societal scale requires a capable base model first. 1X therefore uses human video to make robots productive enough that customers will pay to deploy them, generating the richer robotics data needed for the next cycle.
Børnich defines “hard takeoff” as a self-sufficient labor system in which robots build robots, data centers and chip fabs while performing mining and refining. He is “extremely sure” it is under a decade away, currently bets on three years, and agreed with the host’s point that digital intelligence cannot create its own physical substrate.

8. Foundation models solved context before they solved motion

Hurst’s 20-year retrospective separated today’s breakthrough from earlier humanoid spectacle: it is easy to build something person-shaped, but hard to make it useful in human spaces. Perception is now “all but solved,” giving robots broad recognition and semantic context that would previously have required narrow programming.
Internet data do not contain the required record of every motor’s torque command against every sensor input, so robotics lacks the direct equivalent of the training corpus that propelled large language models.
Learning from demonstration, teleoperation, animation and motion capture all contribute, but human control caps what a robot can discover about its own optimal behavior. Simulation and world models enable immense practice, yet the host noted that condensation, fluid dynamics, surface friction and imperfect robot dynamics preserve a “massive sim-to-real gap” requiring physical experience.
Hurst therefore sees “all of the tools,” not a silver bullet or singularity. Humans evolved to learn from very little data; robots still need many more examples. Their long-run advantage arrives afterward: “Robots have Wi-Fi,” so one machine’s learned skill can be loaded onto every compatible machine.

9. Digit V5 makes safety the gate to warehouse scale

Digit currently handles scoped, multi-purpose workflows such as lifting bins and totes, carrying them through narrow spaces and placing them on tall shelves. Two arms and whole-body balance justify the humanoid form; Hurst sees too little added utility from a third arm to offset its coordination and mechanical complexity.
Digit V5, due later in the year discussed, is intended to be the first balancing humanoid that can leave a work cell without a physical barrier from people. Amazon’s earlier response—successful task performance still failed its safety requirements—led to a two-to-three-year, bottom-up redesign touching every robot system.
The host translated five years into roughly 40,000 work hours and an eventual dollar-an-hour machine if hardware falls to $20,000-$40,000; Hurst said costs are coming down, agreed with the 20-hours-per-day operating premise, and declined to quote today’s hourly rate. The host suggested that reaching those hardware costs may require 100,000 deployed robots.
Fully lights-out facilities may favor conveyors, fixed arms and specialty automation over humanoids. Digit’s larger opportunity is legacy warehouses, retail, hospitals, construction and front-door delivery—the latter demonstrated with Ford about seven years earlier—while new work grows around robot assembly, operation, deployment and maintenance. The hoped-for legacy is that children regard today’s dull, dirty and dangerous jobs as people now regard dangerous historical labor.

The full-length interview with Elon Musk Gary Gallagher: American Civil War, Slavery, Lincoln, Grant & Lee | Lex Fridman Podcast #499