Robotics CEO: The Humanoid Robot Revolution Is Real & It Starts Now w/ Bernt Bornich & David Blundin | EP #188
Peter DiamandisBernt BornichDavid Blundin
1X’s central thesis is that the home is the fastest route to consumer scale and embodied intelligence, not merely another market for automation. Its earlier EVE robots plateaued after roughly 20–40 hours on repetitive guarding or logistics tasks, while homes have not yet shown a diversity ceiling. At 10,000 deployed robots, Bernt Bornich estimates the fleet could generate more non-duplicated useful daily data than YouTube: “The internet isn’t actually that big.”
The near-term commercial proposition could already be useful before full autonomy arrives. Peter Diamandis floated a roughly $30,000 purchase price or $300 monthly lease—$10 a day and about $0.40 an hour—and Bornich replied, “I think we could do better,” while declining to announce the actual price. The factory’s end-2026 annual run rate is planned at north of 20,000 units, although the ramp means 2026 production itself will be lower.
NEO Gamma’s industrial advantage is a deliberately simple, lightweight architecture rather than car-like complexity. The 5-foot-4-inch, 66-pound robot can reportedly lift about 150 pounds, carry roughly 50, run for four hours, and recharge from empty in about two; it contains hundreds of components versus roughly 50,000 in a car. Bornich’s manufacturing frame: “It’s closer to a refrigerator than a car.”
1X is betting that physical intelligence must be spatial, temporal, tactile, and interactive—not language-first. Internet video provides observations but not an agent’s goal, chosen action, or observed consequence; robots can instead execute the scientific-method loop of hypothesis, action, feedback, and revision. Bornich would not claim embodiment is theoretically indispensable, only that it is “a way shorter path” than text or low-fidelity simulation.
Teleoperation is part of the product and training stack, with transparency carrying much of the trust burden. Early customers will receive a mix of best-effort autonomy and scheduled, operator-assisted work; users must approve teleoperation, the robot visibly signals when a person is present, and operators see filtered scenes. Private data has a 24-hour pre-training deletion window, while human review requires approval and a user-supplied decryption key.
Safety is being bounded both physically and through action-conditioned world models. NEO Gamma is soft and intrinsically designed so an accidental strike might hurt but is unlikely to cause severe injury; cooking and other dangerous-object tasks will initially remain disabled because “once you pick up a kettle of boiling water, there’s no more guarantee that you are safe.” For model evaluation, 1X places the controller inside a simulated world where “the robot’s in the Matrix” and tests performance, red-team cases, and unsafe behavior.
The upside case is a labor-and-infrastructure flywheel, but its constraints are intensely physical. Bornich envisions the “hard takeoff moment” as robots building robots, chip fabs, data centers, energy systems, laboratories, and specialized automation; Peter framed labor as roughly half of $110 trillion in global GDP. Bornich considers 10 billion humanoids by 2040 “probably roughly correct” and possibly early, conditional on permitting, power, aluminum, rare-earth processing, magnets, fabs, and enough robots to bootstrap the required labor.
1. The home is 1X’s data engine
Bornich’s first reason for choosing homes is scale: consumer hardware can reach billions of devices, and humanoids only become compelling when volume drives reliability, cost, intelligence, and an ecosystem. For any single industrial task, he argued, “there’s always a better automation system.”
The empirical warning came from EVE deployments in guarding and logistics during 2022–2023. Learning plateaued after roughly 20–40 hours per task: moving a cup repeatedly sat near the low end, while navigating, opening doors, and guarding a facility pushed beyond 40 hours.
A factory repeating one operation offers no path to general intelligence, in Bornich’s view. At 1X’s current household scale, by contrast, the company has not found a diversity ceiling; he positioned 1X as running toward AGI rather than merely applying labor quickly in factories.
His broader equation for abundance multiplies knowledge or intelligence by labor, goods, and services. Improving only the model layer leaves the physical substrate constrained: intelligence needs machines capable of learning in society and eventually expanding the infrastructure on which intelligence runs.
2. A household robot must understand social context
Diamandis likened NEO Gamma to a toddler exploring physics, which Bornich accepted with one qualification: useful prior behavior must come from internet, simulation, and synthetic training so the robot attempts plausible actions rather than wandering randomly. Real-world interaction then begins, with no claim yet about how far that loop can scale.
The deeper household challenge is that “everything we do is social.” An empty coffee cup could mean refill it, wash it, or leave it for continued use; the correct physical action depends on the surrounding people, timing, and habits that repetitive industrial data does not capture.
That logic produced 1X’s decade-old founding requirements: robots must be safe, capable, and affordable enough to “live and learn among us.” Achieving all three meant simplifying the system while preserving human-level strength and making it affordable at scale. The design work led to tendon-driven robots and a decade of novel research.
Bornich described the machine’s social role as neither another human nor another pet, but “something kind of in between”—his Calvin and Hobbes analogy was simply “it’s the Hobbes.” A lifelong companion could remember its owner, communicate through gaze and body language, and make AI feel physically present.
3. The price target is consumer economics, not industrial capex
Diamandis tested a working range of $30,000 to buy or $300 per month to lease—about $10 per day or $0.40 per hour. Bornich did not confirm pricing but answered, “I think we could do better,” while agreeing that the proposed range was directionally plausible.
Bornich insisted 1X wants both the best product and price competitiveness. He said NEO Gamma remains competitive with Chinese alternatives when cost is considered alongside degrees of freedom and capability, arguing that 1X has driven markedly lower complexity.
Bornich said a poll found that people routinely want at least two robots, depending on the price. David Blundin later speculated that four or six might become normal because robots coordinate more precisely than human movers. Diamandis found that excessive, suggesting one NEO Gamma could invite additional units only when a team was needed.
Bornich’s response widened the frame: labor abundance would also make larger homes and more physical goods affordable, so today’s household dimensions may be the wrong basis for estimating robot count. The number of machines, in his telling, co-evolves with the built environment they can create.
4. Physical intelligence starts with actions, not words
Bornich rejected the premise that intelligence begins in language. Language is an efficient human-created compression format, but the core is “spatial and temporal”: how an agent sees, feels, predicts, and acts in the world. His preferred architecture begins in those modalities and adds text afterward.
Blundin proposed that embodiment enables intelligence while language enables it to scale. Bornich’s careful answer was that he could not rigorously prove embodiment is necessary; he claimed only “very, very strong proof” that it offers an easier engineering path to human-level—and potentially greater—intelligence.
The key data distinction is agency. YouTube records what happened but omits the actor’s goal, internal world model, selected action, and resulting feedback; a robot has the observation, its goal and action, and the observed result. Bornich mapped this directly to the scientific method: form a hypothesis, test it, observe the result, repeat, and learn.
Simulation might reproduce that loop, he conceded, so he would not declare alternative approaches impossible. His objection is practical: simulation has much lower fidelity than reality, closing the gap is extraordinarily difficult, and it consumes far more compute than gathering interactive experience in the physical world.
5. Fleet data and manufacturing scale compound together
Bornich’s napkin math says 10,000 robots operating through most of each day would collect more useful, non-duplicated data than is uploaded to YouTube daily. At mass deployment, robot-generated experience could dwarf internet data, creating a feedback loop between unit volume and model capability.
He corrected the hosts’ production premise: 1X had built more than 100 robots across several generations, but no more than 100 NEO Gammas. The planned factory is intended to exit 2026 above a 20,000-unit annual run rate, while the ramp prevents actual 2026 output from reaching that full number.
The subsequent factory is intended to approach another order-of-magnitude step, though Bornich cautioned it will not quite achieve that. He invoked the iPhone’s roughly 1.7-times growth cadence, including plateaus when scaling uncovers new constraints; asked about hundreds of thousands annually before 2030, he answered, “way more.”
Scale ultimately runs into aluminum refinement and assembly labor. Even with few parts, NEO Gamma remains more complicated to build than an iPhone; if it takes five times the labor, the labor pool fails first. Hence the required transition to robots assembling robots and expanding fabs, data centers, and energy supply.
6. NEO Gamma is engineered more like an appliance than a car
NEO Gamma stands about 5 feet 4 inches and weighs 66 pounds. Bornich said it can lift roughly 150 pounds and carry about 50, giving it “the weight-to-strength ratio of an athletic human” while supporting the company’s emphasis on a light, soft machine.
Battery life is about four hours, with a full recharge taking roughly two hours. Bornich cares less about continuous runtime than whether brief, opportunistic charging breaks keep the machine available; his unit at home generally charges during natural pauses rather than exhausting its battery.
The overlooked household specification is silence. A mechanical sound that seems acceptable on day one becomes irritating by day three, he said, making “dead quiet” operation a requirement rather than refinement. Softness and huggability likewise matter because people must remain relaxed inside the robot’s working space.
NEO Gamma has hundreds of components, while Bornich put a car near 50,000 parts and 4,000 pounds. “If you do a really good job here, it’s closer to a refrigerator”—albeit a complicated one—than an automobile, a comparison that captures the intended manufacturing economics.
7. Dexterity expands the ceiling on learnable behavior
Many humanoids stop near 26 degrees of freedom, typically covering legs, arms, and neck while omitting wrists. NEO Gamma adds three neck axes for expression, three through the spine, full arm articulation, and 22 degrees of freedom in each hand.
Bornich said the hand is functionally close to a human’s 22 degrees, though the count depends on whether small carpal motions are treated separately. That nuance matters because cupping, deformable objects, delicate handling, and in-hand manipulation generate categories of experience that a less capable hand would not encounter.
His intelligence metric is diversity, bounded by two independent variables: the environment’s variety and the robot’s physical capability. A sophisticated robot on one factory task still plateaus; an incapable robot in a rich home still cannot collect manipulation data. The strategy is therefore “max, max on both.”
8. Teleoperation is both a capability test and labeled data
When hardware and software teams dispute why a task failed, Bornich’s diagnostic is simple: have the best teleoperator attempt it. If teleoperation succeeds, “the right neural net can do it with enough data”; this is proof that the mechanism can access the behavior, though not proof of broad autonomous generalization.
The operator does not puppet every joint through a VR suit. Commands are increasingly abstract—place the hands here, grasp that object—while learned low-level control solves movement details. Bottom-up motor competence and top-down behavioral models are intended to “meet in the middle” until the operator disappears.
That debugging method is already breaking down in an encouraging way. NEO Gamma’s hands provide fast, high-fidelity tactile feedback that cannot be transmitted efficiently to a human operator; real-world reinforcement learning has begun producing manipulation the operator “could just dream of.”
Blundin defended teleoperation after initially noting that demos can mislead viewers about autonomy. Bornich’s distinction was precise: a teleoperated robot clearly can perform the physical act, but cannot yet do it autonomously. Transparent labeling matters because the demonstrations are robotics’ equivalent of expert-labeled fine-tuning data.
9. Fast control stays onboard while fleet learning remains shared
Everything translating perception into motor torque is learned end to end; Bornich said the surrounding code is only a few hundred lines because “it’s all weights.” The parameter count remains secret and comparatively small, since the controller must run quickly on-device while ingesting vision.
Compute sits in the head for engineering reasons, not visual anthropomorphism. Space elsewhere is packed, and the highest-bandwidth stream comes from the eyes; 1X uses high-resolution, high-frequency vision without LiDAR, structured light, wrist cameras, or similar sensors, making even a trip to the torso an awkward data path.
Control is hierarchical: distributed motor-level decisions run around 25 hertz, the onboard “brain” roughly 5–10 hertz with low latency, and slower language-like streaming near 1 hertz can live off-board. Cloud inference cannot close the high-frequency tactile loop required for manipulation.
A robot learning to crack an egg does not remain isolated. Validated fleet data trains shared models, and improved checkpoints can be deployed to every unit. Bornich also expects substantial on-device federated learning, so each companion could retain private personal experience while sharing a common intelligence backbone.
10. Privacy is offered on user-controlled terms, with real trade-offs
Bornich was explicit that early adopters trade some privacy for participation: “Without the data, we can’t make the product better.” Routine data can enter automated training without human access; if 1X wants to inspect a specific window, the owner receives the relevant video and decides whether to provide the decryption key.
Training also runs with a 24-hour delay, giving users time to erase an event “from existence” before it enters model weights.
Teleoperation necessarily exposes the task scene, so 1X filters people into blobs and emphasizes the manipulated object. No operator enters without approval; NEO Gamma’s lighting visibly changes, and the person must come from the user-approved roster—Bornich’s example was four selected operators servicing a household.
The product separates best-effort autonomy from scheduled completion. At home, Bornich can request white laundry, receipt and refrigeration of an Instacart delivery, and general tidying while he is away; some work is autonomous and some assisted, but “I don’t really care about the mix. The task gets done.”
11. Dangerous tools, not raw strength, define the first safety boundary
Best-effort autonomy is allowed to learn from failures: users can say “bad robot,” and Bornich said tasks progress faster after failures than successes. The harder alignment case is grandma requesting scotch, because models tend toward sycophancy and may comply with requests they should refuse.
Intrinsic safety means NEO Gamma’s low mass, softness, and compliant motion should make an accidental impact painful at worst rather than likely to cause severe injury. That guarantee disappears when it handles dangerous objects such as a kettle of boiling water.
Cooking therefore will not ship initially, despite internal work on it and Peter’s request for teriyaki salmon. Bornich expects dangerous capabilities to unlock only as behavioral confidence improves; he would not exchange early usefulness for an unbounded household hazard.
Physical safeguards are paired with model evaluation. Bornich rejected releasing a controller and waiting for a customer “vibe check,” because a robot—like an autonomous car—must demonstrate both better performance and preserved safety before deployment.
12. World models provide a “Matrix” for automated red-teaming
1X’s world model predicts consequences from chosen actions, including rendered observations and physical forces. “It’s essentially like the robot’s in the Matrix”: the controller behaves as though it is in a home, unaware that the environment is generated.
Engineers can then replay ordinary tasks, adversarial situations, and automated safety checks at scale. The model becomes both an AGI research direction and a nearer-term release-evaluation tool, testing whether a new controller improves performance without introducing dangerous behavior.
Blundin asked whether this model-and-data asset, rather than unit sales alone, explains the sector’s large valuations. Bornich said it will eventually be productized across digital and physical labor, but predicted robot revenue will “dominate forever” because the physical world carries much more value than investors assume.
Diamandis sized today’s global GDP at about $110 trillion and labor at roughly half, implying a current addressable base above $50 trillion. Bornich argued that this understates the opportunity because new capability expands the amount of work and production rather than merely replacing existing labor.
13. A humanoid wins first as the general-purpose platform
Answering Salim Ismail’s challenge—why not six arms or an octopus—Bornich conceded that humanoids are not the only viable machines. His case is that no alternative matches the human form’s generality and compatibility with a human-built world. Diamandis also argued that learnings transfer less directly to a six-armed form, while Bornich agreed it would be at least harder.
Blundin added that form establishes intuitive expectations: owners already know what tasks a person-shaped machine probably can or cannot perform. An unfamiliar six-legged design forces users to relearn its capabilities, undermining the natural interface that is supposed to eliminate technology’s adoption barrier.
An Apple device is absurdly overengineered as a typewriter—humanity mastered nanoscale chip fabrication to write a document—yet scale makes it cheaper and more reliable than a dedicated alternative, with an ecosystem no specialized machine can initially match.
Specialization returns only after the general market becomes enormous. Robotics may eventually resemble Star Wars, with repair drones, extra arms, and task-specific tools, but those niches first need the scale and knowledge created by humanoids. His definitive formulation: “Humanoid is a phase.”
14. China’s advantage is accumulated process knowledge
Bornich called China’s hardware ecosystem extraordinary: a broken board, machine, or component can be replaced across the street, enabling design and manufacturing iteration at a pace Silicon Valley cannot match. The less visible advantage is dispersed process knowledge accumulated on production lines.
His magnet example captured the moat: scientists can understand the material and follow every textbook instruction, yet lack the veteran who knows that after two hours “you have to stir to the left, not the right.” Rare-earth access matters, but repeatable high-grade production depends on that tacit craft.
Diamandis attributed China’s position partly to top-down designation of robot or magnet cities. Bornich was less certain, stressing the vitality of Chinese startups and capital; he considered free economic zones—faster permitting, lower friction, and permission to build—the more likely policy masterstroke.
Both wanted comparable US support. Bornich suggested domestic economic zones with expedited approvals, while Blundin argued that decades of software preference distorted venture portfolios. His blunt self-critique was that even his fund, despite making first checks into hard problems at the seed stage, does not finance hardware: “I’m as much a part of the problem.”
15. 1X’s moat sits in motors, tendons, and patient capital
Bornich said 1X manufactures its own motors, including winding, manufacturing, automation, and related electronics, because suitable components did not exist. He claimed NEO Gamma’s motors reach 5.5 times the world-record torque-to-weight benchmark, supplying enough force to eliminate gears.
The motors and tendon design make the robot light, drivable, and compliant while lowering manufacturing cost. They also required materials able to survive “millions and millions and millions of cycles,” alongside new motor drives, power amplification, magnetics, and manufacturing methods—research problems rather than catalog engineering.
AI entered the hardware stack more than a decade ago: Bornich programmed a neural network to learn motor design before Transformers. He described the hardware lead as measured in years, while even an excellent world-model lead might be only three months because software advantages diffuse much faster.
The company survived because an early Norwegian investor ultimately sold the farm that had housed its barn-based startup to extend the runway. 1X deliberately stayed small for seven years to develop core technology, then moved toward Palo Alto for the density of product, scaling, API, manufacturing, and robotics talent.
16. The endgame is robots building infrastructure, not doing every job by hand
Bornich’s 10-year vision starts with sustainable abundance: when energy and labor cease being scarce, society no longer needs to cut environmental corners merely to reduce cost. The next frontier is infrastructure that gives everyone a high quality of life, followed by much larger scientific systems.
Particle accelerators, biotech laboratories, chemistry campaigns, and other experiment-heavy programs require physical construction and repeated manipulation. He rejected a future where “the godlike AI in the sky” directs humans through glasses, preferring “symbiosis and co-invention between man and machine.”
Humanoids will fill gaps and build specialized automation rather than inefficiently impersonating machinery forever. They will not carry a car chassis with 30 bodies or machine every part using a Dremel; they will use existing automation such as CNC machines, build more automation systems, and help expand fabs, data centers, and energy infrastructure.
Space offers an early high-value extension: NEO Gamma’s 66-pound mass and energy efficiency reduce launch burden, though motor epoxy would need vacuum hardening and heat rejection remains difficult. For in-orbit assembly, Bornich favors low-latency teleoperation by expert humans in orbit until accumulated demonstrations enable autonomy.
17. Ten billion robots is a supply-chain and permitting call
Bornich promised an early-adopter program during 2025 but refused a specific preorder date. His expectation-setting was unusually plain: customers are buying “a ticket to be part of this transformation”—they will “adopt a Neo,” teach it, and receive something useful but imperfect, with substantial rough edges.
Asked about Elon Musk and Brett Adcock’s estimate of 10 billion humanoids by 2040, Bornich called it “probably roughly correct” and said it might happen sooner. The condition is society’s willingness to remove artificial constraints and let mines, refineries, power systems, factories, and labor scale fast enough.
Blundin sharpened the semiconductor mismatch: he estimated roughly one full GPU per robot, perhaps two, against only 20 million GPUs produced annually, with TSMC holding 66% of fabrication. Bornich pointed one layer deeper to ASML and the brittle equipment chain behind every additional fab.
The same problem recurs in aluminum, rare earths, and high-grade magnets, where material access without processing expertise is insufficient. Permitting may ultimately set the schedule; robots can help build the missing infrastructure only after an initial industrial base exists to manufacture the robots themselves.
The company’s name closes the loop on its credibility standard. Robotics videos often display “4X” or “8X” playback, Bornich said, whereas 1X shows machines at real-time speed: “All we do is real time because we build proper robots.”
Full transcript
You think about robots in the world probably more than anybody else. What's your vision 10 years from now?
Everybody, we're here at 1X Technologies in Palo Alto. Bernt Bornich is the CEO and founder. NEO Gamma One and NEO Gamma Two are over here. I imagine we're going to have the same level of AI eventually in the robot, where I feel like I'm talking to a fully intelligent being.
And one that is grounded, right? That actually understands what this existence is.
I'm shocked by that. Wow. I'm shocked by that, too.
How do we solve the remaining really hard problems in science? This isn't going to happen without humanoids. It's almost existential to us for human happiness.
Salim Ismail is constantly saying, “Have it look like an octopus and let it operate with all the elegance that an octopus can, rather than trying to constrain it into five fingers on this hand that do certain things and manipulate objects the way we're supposed to manipulate them.” So what's the definitive answer to him?
Let's just say humanoid is the face.
Now, that's a moonshot, ladies and gentlemen. That's great. I'm with David Blundin, my moonshot mate.
Hello, all.
NEO Gamma One and NEO Gamma Two are over here, and we just did a tour of the facility. It's pretty extraordinary. We saw probably dozens of NEO Gammas in different stages of development. They literally manufacture everything from head to toe. How many components are inside NEO Gamma, roughly?
Oh, top secret.
Top secret, okay. Can't say that.
It's in the hundreds, not the thousands.
I just secured my first NEO Gamma at my home by the end of the year. Is that right?
Oh, yeah.
Okay, fantastic. We're about to do a podcast, either with Bernt or with NEO Gamma, depending on what you want. Let's go ahead. We'll go over to the podcast area. Will you lead the way and maybe clear the way for me? Awesome. By the way, those bags over there—NEO Gamma can carry those, over half his body weight.
Okay, Peter. I'm not the NEO Gamma. Hey, can I give this to you to carry?
You can try. It might hit some safety limits, but it usually works.
All right, arms up.
Feel it properly. There you go. You can let it go, and it can take a few steps. There you go. It might, after a few steps, decide, “This is a bit unsafe for me.”
I mean, it's—
Thank you, Neo.
—incredibly strong. All right. It's nice to know that NEO Gamma will clean up the house around you. Well, listen, I'm not sure what number you are, but I want to say thank you so much. Thanks for cleaning up.
Of course. Thank you for your time. A pleasure. A pleasure. Have a great day, and thank you very much as well. I want to be polite. You never know when the robot overlords are going to come after us. I want you to remember that I was really polite. I was really, really polite. Okay, I'm safe. Great.
Do behave.
Everybody, welcome to Moonshots. I'm here with my moonshot mate, David Blundin. Salim Ismail is offline with his son this weekend. But I'm here in particular with the CEO and founder of 1X Technologies, Bernt Bornich. A pleasure, Bernt.
Awesome. Looking forward to this one.
Thank you. We just finished this tour, and it's pretty extraordinary. When did you move into these facilities here?
It's been 1.5 months.
Nice. There are just many levels of people building robots. No robots building robots yet.
We're getting there, but not yet.
But you're getting there.
Yeah.
1. The Home First Strategy
I'm very familiar with the robotics and humanoid robot space. While companies like Figure and Tesla are focused initially on going into factories, automotive factories in particular, you made a commitment to the home.
Yes.
Personally, I'm excited about that, but I'd like to start with why the home.
To me, there are 2 main reasons. There are a lot of reasons, but 2 main ones. The first one is kind of obvious: consumer hardware just scales at a different pace than everything else, right?
We got to more than 1 billion iPhone devices in a bit more than a decade. To me, humanoid robots don't make sense unless they're at scale, right? There's always a better automation system that you can use for 1 specific problem. You need scale so that you really get this incredible reliability, incredibly low cost, and an incredible ecosystem and intelligence.
2. Diversity Drives Intelligence
The slightly deeper reason is also that intelligence comes from diversity. This has been very clear from the beginning, in all kinds of AI research and also in more practical applications of AI across all different domains, whether it's a language model, an image model, a video model, or, in this case, a robotics model.
You don't really need data of the same thing over and over. If you think about it, it's very logical, right?
So if you're in an automotive factory, you're basically doing the same thing over and over again. You're not learning new stuff.
Yeah, and we actually have some data on this. We have some real data because our previous-generation humanoid, EVE, was deployed in both guarding and logistics back in 2022 and 2023.
Mm-hmm.
After about 20 to 40 hours, our robots plateau and stop learning for that specific task. It depends on how complex it is. If you're guarding a facility and driving around—because EVE had wheels but was also humanoid—and opening doors, there's some diversity to that, so you're more in the 40-plus-hour range. If you're just moving this cup from here to over here all day, then you're at the lower end of 20 hours.
Yeah.
There's just no path from there to general intelligence. We may be a bit different from the rest of the humanoid space in this, but I see us more as a company really running toward AGI—
Yeah.
—and asking how we can come there as fast as possible, versus how we can apply labor in industrial or similar settings as quickly as possible.
So it's robotics in service of building true AGI models and getting enough new, rich data to train up these models.
Yeah, and what would you want to— You said 20 to 40 hours for a security guard robot. What's the equivalent for all the variety of things you can do in the home? How many hours of—
We don't know yet.
—do you need? So, tens of thousands of these?
No, we don't know. At our current scale, we don't really see any kind of cap on diversity.
Yeah.
It'll get there, and we'll need to diversify. But you ask a very important question, right? We want to talk about what the goal is. To me, it's not just AI or robotics. It's a combination.
Yeah.
If you think about what this is—what abundance is—it's an abundance of knowledge or intelligence multiplied by an abundance of labor or goods and services. You need both, and they follow hand in hand. We can talk more about that, but the constraints we have in society aren't always only on the intelligence or data layer. They're also on the substrate that we're building on, right?
So when I think about it, I imagine this is why a toddler crawling around, playing, investigating the physical universe it's in, interacting with different people and different things, is learning and building a model in its neocortex.
And so, is that basically the same: your NEO Gamma is an infant learning in a diverse environment?
It is.
Yeah.
And I think just to some extent, for humans too, right? But it's more pronounced in other animals—how much of this kind of intelligence is innate and part of your instincts. You don't want your robot to just go around randomly doing anything. You want it to try to do things that might succeed. So there is room here for the more classical AI models, where we're training based on internet data, simulation data, synthetic data—everything that everyone else is doing—
Yeah.
That's useful to get you off the ground. But it doesn't fully get you there. It gets you to something that does something seemingly useful, and then you can experiment and have the robot really enter this interactive learning loop, where it's learning in the real world. That can get you somewhere—we don't know how far it can get you, right? We don't know yet.
And this whole topic of data gathering, it's amazing watching them walk around the building here and walk around the kitchen. They're so unintimidating. You walk right up to it intuitively. You don't feel like it's ever going to do anything awkward, hit you, or anything like that. So that's got to be incredibly important—
You say that—
The data gathering.
It's cozy. It's cozy.
It's cozy, and it doesn't seem to break the glasses or anything. So that's got to be really core to the data-gathering mission, right? Because you have to, as you said, let it experiment; otherwise, how's it going to learn?
3. Designing For Home
So, along those lines, what design elements did you build into NEO Gamma to make it suited for the home?
Sure. This actually goes all the way back to the founding of the company a decade ago. Really, I've been in the field for a long time.
How long?
Since I was a kid.
You were building robots at age what?
I was 11 when I decided that I was going to do humanoids and that sort of thing.
And what was the humanoid robot that you modeled? Was it Star Wars? Was it Star Trek? What was it? Lost in Space?
Honda ASIMO.
What's that?
Honda. Honda's ASIMO.
ASIMO?
ASIMO.
Yeah.
It's a beautiful robot, right? They started very early, and you can check out the Honda ASIMO piece. There are more modern ones, but the Honda Asimo P6 was like the end of the '90s.
Yeah.
Mm-hmm.
And that was walking up stairs—
Yeah.
—running around a stage, giving someone a bottle.
Yeah.
It greeted President Obama, I think, at one point. Yeah.
Yeah. That was a bit later, but yes.
Okay.
It was so ahead of its time, right?
Yeah.
But I built a lot of stuff through the years. Importantly, when I started the company, I sat down and thought really deeply about this: There are all these amazing robots that we worked on, and it didn't really work. Why didn't it work, right?
Mm.
Why didn't it work? It comes down to these fundamental principles. First of all, if you actually want to make something that's scalable with respect to intelligence, it needs to be able to live and learn among us. There are just so many nuances to this through everyday life, right? Everything we do is social. Work is social. Every task is social, and we navigate these social situations all the time while we do the things we do.
Most of the world's labor also happens in a social context, in that there are other people around you when you do it. Objects have social context, right? The coffee cup is empty. Do you need a new one? Is it dirty? Or do you want a refill, or do you keep your cup out through the day? There's this trove of diversity that you want to access.
So if you're a big believer in that, it boils down to this: The robot needs to be safe from a first-principles point of view, not able to harm people. It still needs to be very capable and needs to be as strong as a human. And then it just needs to be incredibly affordable. You need to find this beautiful combination where you can simplify, simplify, simplify, and still get a very capable system so that you can manufacture this at scale and really drive quality up and cost down, right?
That's really what we set out to do, and that's also why it took a decade, right? There's so much novel research that's been done in the company to get to where we have these tendon-driven robots that have—
So what's the vision there relative to the car, say? One in every household, two? You mentioned the iPhone—you know, go direct-to-consumer with iPhone sales, get to a billion, but it's exactly one per person. It's pretty obvious, right? But robots could be 2, could be 4, could be—
I've done that poll, and everybody routinely says, “I would have at least 2”—
Mm-hmm.
—depending on the price point, right?
So, price-point-wise, when I think about this, what I've heard is 30K, 20K. We've seen Chinese robots at much cheaper price points, but not as capable as NEO Gamma. Do you have a price point that you're thinking about?
Yeah, you're not far off. It's cheaper than what people think.
Okay.
It's quite interesting because I think this is very important. I want to make sure that we're not only making the best product; we want it to be price-competitive. I think that's going to be incredibly important. And we are actually still price-competitive with the Chinese ones.
Okay.
But you have to count, as you said, that it's not the same, right? So if you think about the number of degrees of freedom that the robot has—how much capability, basically how many joints—
Mm-hmm.
—then we actually have a significantly lower cost. So I think we've done a really good job reducing complexity to get there.
So the numbers, Bernt, that I keep in my mind are 30K to purchase, or 300 bucks a month to lease, 10 bucks a day, 40 cents an hour. Am I in the right range there?
Yeah, I think we could do better, but yes.
Okay. Even better. That's fantastic. And do you need to do better? No. Forty cents per hour—
I mean, in a heartbeat—
I'll pay that in a heartbeat. That's good enough.
Yeah, yeah. But in that case, I think people could imagine owning a couple of those robots. So I think it really depends on the lens you see this through. Clearly, everyone's going to want a robot, and I think there's this beautiful thing about the companion aspect of this, which is so underrated, right?
The humanoid is just such a beautiful interface for AI. When you talk to it and see the body language, it can look at you, it sees who's talking to it, directional audio—all these things. All my 11-year-old daughter can do if she has the robot is just sit next to it on the couch and talk about things, right?
Yeah.
And that is clearly going to be such a big aspect of it. I see it as—not another pet, but it's not another human either. It's something kind of in between.
Yeah.
And, like I said, it's kind of like Hobbes. If you ever read Calvin and Hobbes, it's Hobbes. I think it's going to be incredibly exciting to see how these relationships develop, because it's the thing that will be around you all your life, right? It will remember everything about you.
There are 2 things that really jump out at me if you compare C-3PO and that vision of an assistant robot and compare it to what you've actually built. One, it's soft. It's not a metal outside. And 2, the voice is perfect. When you're speaking to it, you immediately are disarmed and just talk to it, because it doesn't have a C-3PO robotic voice. It has just a perfectly soothing, normal voice, and it's very responsive to anything you say, any gesture, or anything.
Oh, thank you.
I imagine that these robots will all have advanced AIs at the level of GPT-5 or Gemini 3. And in so being, those robots will be hyperintelligent and able to fully understand and answer what you need.
And once they've learned the physics models fully, they can do whatever you need. You've made a decision to build your AI systems in-house, and I find that fascinating. In fact, a number of the other humanoid robot companies—not going to put you into a comparison mode here—have made that same decision versus partnering with the large hyperscalers. Can you speak to that?
Well, we're not doing the same thing. To me, intelligence does not begin with language. Language is this generative, artificial construct that we have come up with, and it's such an efficient, compressed way of conveying meaning and instruction. So language is very useful, but it's not the core of your intelligence.
The core of your intelligence is spatial and temporal, and it has to do with how you perceive the world around you. Both with respect to how you see the world, but also how you feel the world, right? And we're getting to where we're seeing that models that are native to that modality, and then you add text—
Mm-hmm.
—will be more intelligent and more powerful than language-first models.
I've read about intelligence, and the belief is that you needed embodiment for intelligence to exist and language for intelligence to scale.
I don't have rigorous proof that embodiment is needed. I do have very, very strong proof that, from an engineering perspective, it's just a way easier path, right? So if you think about the information in the world and whether you can access it, you could train a world model that can predict video and tell you, like, “Hey, here's a new video frame. Render this for you.”
In theory, you could probably train that only on text. If you have enough text descriptions of things, maybe at some point you could get a high enough signal-to-noise ratio that you actually can get something useful out, at least if you have some feedback loop with RLHF or something where you're like, “Am I happy with this frame?”
Mm.
But why would you do that?
Yeah.
That's just such an inefficient way of doing that. You, of course, train on video because you're going to output video, right?
Right.
So from that perspective, I think it's just obvious that you need all the modalities that we experience if you want to get to, first and foremost, human-level intelligence, and hopefully past that again.
But then I think there are 2 other things that are quite important when it comes to learning. The first one is quite obvious, and I think we all identify with this: robots can do interactive learning, right? You interact with the world, and therefore you can learn.
But if you think about it more from an academic point of view of how intelligence evolves, how do you get reasoning and all these things? What we generally do is have an observation of the world. We kind of know how the world works. So I know that if I do this, I know what is going to happen, right?
Yeah. Of course.
I've seen this before. So I actually start with that, and now I have a goal. I want to pick up the cup. So now I have a model of the world. I have a goal of picking up the cup. I take an action. I know which action I took. I know the action I took was to reach for and grasp the cup.
Yeah.
And then I observe the result. If you look at the internet, or in general, you can look at YouTube, right? All you have are the observations.
Yeah, right.
You don't have any of the mental model of the person in that video. You don't know which actions they took. You don't know what they tried to achieve. You only have the observation.
Yeah.
This is not how we learn. You can actually bring it all the way back to the scientific method. You should have a theory, come up with a hypothesis, test your hypothesis, observe the result, and then do it again and learn.
Yeah.
And that is just not possible with the internet data.
Definitely impossible with next-token prediction, raw internet scrapes, and all the video scrapes. So then, in these limited domains like coding and physics experiments, you can actually have that same experience, but it's only within that domain.
Coding is a good example. “Let me try writing it this way.” It didn't work. “Let me try writing it that way.” It didn't work. So you get very, very good at that narrow domain, but you still have no intuition about how the world works.
No.
You know?
You can do simulation, no?
Yeah.
So again, back to, it's hard to prove that this won't work, right? Sure, if you have a really good simulator and really scale simulation-based learning and simulation with agents, maybe you can get something similar.
But the fidelity of your simulator is nowhere near the real world. It's incredibly hard to get there and close that gap. It's also so compute-inefficient compared to just being in the real world.
But I think, for me, it boils down not to this academic exercise of proving who's right and wrong. It's more about what's the engineering approach that makes sense here.
Yeah.
And it's just a way shorter path.
You mentioned before in our conversation the amount of data that's being collected relative to Google, YouTube, or Tesla. Can you speak to that? Your mission is to get as much data as possible during the day from interactions with these robots in the home.
Yeah. You can do some napkin math, right? Of course, we don't know exactly what is the most useful data from which modalities yet.
But if you think about it, if you have 10,000 robots out there and they gather data most of the day, then that is more data than the non-duplicated, useful data that gets uploaded to YouTube each day.
Yeah.
So already at that scale, you have your fleet of robots generating more useful data than YouTube.
Yeah.
That's just at 10,000. And then if you think about how we scale manufacturing here as this starts deploying into society, you very quickly come to the conclusion that the internet isn't actually that big. You're going to have way more data from robots than you're going to have from the internet.
4. Scaling Robot Production
So I want to hit some numbers here just to set them as foundations. You built roughly hundreds of the NEO Gammas, but you're about to open a new manufacturing plant. Can you give us a sense of that? And then there's another one that's in the plans, right? Without disclosing anything you're not willing to, can you give me a sense of, by the end of 2026, how many you're manufacturing at an annual run rate, and then in 2027 and 2028? What's the growth path you imagine?
Yeah. First of all, just a small correction.
Okay.
We haven't built more than 100 of the NEO Gammas.
Oh, the Gammas. Okay.
But we've built more than 100 of the robots.
Right.
There have been multiple versions. The factory run rate at the end of 2026 is north of 20,000.
20,000—
Yeah.
—annualized.
Annual, yep. Of course, there's a ramp to get there, so you don't reach quite that number in 2026.
So a couple thousand a month.
Now, after that, we're trying to follow an order of magnitude, right? We're not going to quite be able to do that. I think the iPhone ramp is a very good comparison here, where you see that they almost double, but you have a few plateaus as you reach certain scales, where you run into problems.
There are some quite interesting problems if you're going to scale the manufacturing of humanoids to the iPhone level, right? You run out of some basic stuff like aluminum, for example. You don't use all the aluminum on the planet—that's not what I mean—but there's a certain percentage of the current aluminum-refinement capacity you can use before you start to really struggle sourcing aluminum.
Yeah.
That might be a challenge, I think.
Wait, was the iPhone ramp about doubling? That's an interesting statistic I hadn't even thought of. You get to 1 billion in the end.
It's more like 1.7, but—
1.7 annualized—
Yeah.
—over. Wow, that's not as fast as I remember.
So you can imagine 100,000—
Well, exponentials are quite powerful.
Yeah. No, I know.
We've heard the physics, actually. So you can imagine a run rate, before the end of this decade, of hundreds of thousands per year.
End of this decade, way more.
Way more at that rate.
Yeah.
Yeah.
At that point, you need to really think about what the things are that will slow you down, right? It comes down to refining—mining and refinement, of course.
Mm-hmm.
But increasingly, it actually comes down to labor. You’re not going to get there without really using robots for labor. If you think about the iPhone ramp, Apple displaced a large part of the Chinese population across the country for labor—
To parts of China.
—and they still ran out of labor and had to expand into neighboring countries.
That’s wild.
Now—
Yeah.
I think we’ve done an incredible job in the design, so there are very few parts. It’s very simple to assemble.
Yeah, I was looking for a new iPhone.
But it’s still more complicated to assemble than an iPhone. So let’s say it takes 5 times as long, which means we need 5 times as much labor as the iPhone.
I was going to say, it looks more complicated. I mean—
Yeah, it is more complicated than an iPhone, right?
Yeah.
Then you’re in trouble.
That’s a good metric.
Then you’re in trouble.
So it’s got to be solved.
You have to automate, right? And, of course, that’s the goal anyway. We want to get as quickly as possible to what I call this hard takeoff moment—
Yeah.
—where you have robots building robots, robots building out the data centers, the chip fabs, and the energy infrastructure.
So what can we learn from the car, actually? Here, you’ve got the iPhone: fewer parts, one-fifth the labor per unit. Then over here, you have a car. How does the part count compare to a car?
We have a few hundred parts.
Oh.
A car has roughly 50,000.
50,000. So it’s much simpler—
And, I mean—
—to get scale.
A car weighs 4,000 pounds.
Yeah, a lot of material.
Our robot weighs 66 pounds.
Okay.
I don’t think it’s really comparable to a car. I’ve seen a lot of people in the space compare humanoids to cars, but I think then you should go back to the drawing board, to be honest. It’s not a car. If you do a really good job here, it’s closer to a refrigerator.
All right.
It’s a very complicated refrigerator, but it’s closer to a refrigerator than a car.
Okay, so let’s dive in a little bit and shape our viewers’ and listeners’ understanding of the robot. It’s 66 pounds.
Mm-hmm.
Let’s talk about battery life and its abilities. Describe it from a specific stats point of view, if you would.
Oh, yeah, sure. I think the most important stat is that it’s huggable.
It’s huggable.
Yeah.
Yes, it is huggable. I have hugged a robot.
Yeah. Just the safety and how it feels to be safe in its space—soft. But from a pure stats point of view, it’s 66 pounds. It can lift about 150 pounds—
Which is amazing. In terms of the weight-to-strength ratio, it’s huge.
It is the weight-to-strength ratio of an athletic human. And then it can carry about 50 pounds around. That’s what you hopefully saw earlier here. Battery life is about 4 hours.
Rechargeable in—
Half an hour.
Half an hour? Or 2 hours?
Half of it, so about 2 hours if you use the full battery.
2 hours.
Yeah. Interestingly enough, I have one in my house, right? So I’m starting to get some data on this now—
Of course.
—and—
It’s 5'4", 5'5". What is it?
Yeah, 5'4", I think.
5'4", okay. That’s a perfect height, by the way, just in case you were wondering.
Yeah, it’s also the height of my wife, so I agree with you.
It’s mine, so that’s good.
Yeah. What’s very interesting is that once you start actually using the product, you notice a lot of things that don’t usually show up on a spec sheet. The robot is completely quiet.
Uh-huh.
That’s not a coincidence. That’s something we worked so hard on. The first time you put this in your home, you think, “The robot’s very quiet. It’s fine.” You put it in your home, and on the first day it’s fine. The second day it’s a bit annoying. The third day you’re like, “Oh, man, is it going to leave my living room soon?” Because of this sound.
Mm-hmm.
It’s such a requirement for it to be dead quiet if you’re going to have this in your space.
Interesting.
Charging-wise, I don’t really run into the problem because the robot just takes these micro-breaks every now and then when it’s not doing something.
Yeah, that’s good.
I actually don’t care that much about how many hours it can run. I care that it charges fast enough that it can always do whatever I want it to do.
Yeah.
Nice.
Well, I want to talk quickly about specifications, since you said that—
Yeah.
—the number of degrees of freedom, right?
Yeah, basically—
Which is basically how many joints the robot has, right?
Yes.
Humans have about 6 joints in each leg. That’s 12. If you have 7 in each arm, that’s 14 more. So now you’re at 12 plus 14—that’s 26. You see a lot of robots today that have 26. That’s quite common. Usually, they don’t have the wrists; they actually have the neck instead. So, 2 here, and then you’re at 26.
Okay.
We have 3 here, so you have proper expression with your head.
Oh.
That’s quite important.
Oh.
We have all 7 here. We have 3 in the spine, and then, of course, we have 22 in each hand—
I mean, what I saw in the arm design was incredible. Yeah.
Yeah.
How many does a human have in their hand?
22.
So you matched it.
Well, depending on how you count your carpal bones—the small bones that you have here that allow you to cup your hand—you could, to some extent, see that that’s more like 4 or 5 degrees of freedom, not really 2. So humans have a bit more. Functionally, it’s quite similar. This is incredibly important to be able to do all those tasks in a home.
But also, from an AI perspective, we talk about diversity initially, right? It is the one metric for intelligence.
Mm-hmm.
And the diversity of your data—
Diversity of environment and data—
Well, diversity of your data. Your diversity comes from 2 things, or the limit to the diversity you can achieve comes from 2 things. It comes from the environment you’re deploying in. If you’re in a factory doing the same thing every day, it doesn’t matter how good your robot is; it’s not going to be diverse.
Mm-hmm.
And then, how capable is your robot? How many things can it do, right? Because if it cannot do any kind of in-hand manipulation or handle soft deformables, all these kinds of things, or delicate objects or whatever, then you get no data from that. So you really have to go max, max on both if you want to maximize your diversity.
A geeky question for you, but I’m really curious to know: when you build something physical and then attach a neural net to it, it’s very hard to tell whether the constraint in what it can and can’t do is in the neural net or in the physical construction of the hardware. Is there any way to decouple that and debug the 2 different sides? Or is it incredibly impossible? Once they’re meshed together, can you just not?
Well, we have a pretty good neural net here.
Yeah.
Usually, the way I approach this is: can we do it in teleoperation? If we can, the right neural net can do it with enough data.
Interesting.
That’s generally been proven to be true. If we manage to do something in teleoperation, it’s like, “Okay, now we need a lot of diverse data of similar tasks, so we get some transfer learning, and we need a lot of data of that specific task.” Almost irrespective of how complicated that task is, you can get it to work.
Yeah.
Now, of course, that doesn’t mean you can get everything to work with generalization across tasks. We’re not there yet.
Yeah.
But you can see that, okay, you can get the neural network to do this. Now we need to scale it so we get this beautiful transfer of knowledge between tasks, out-of-distribution generalization, and all these things that we currently see in large language models that we don’t see that much in robotics yet. We have some pretty cool stuff internally where we see some signs of life.
What I’m picturing, though, is you ask it to make crêpes Suzette, or you ask it to do microsurgery, and it can’t quite do it. Then you say, “Well, look, the hardware guy is claiming the hardware is good enough. It must be the software guy.”
Mm-hmm.
And then the software guy is saying, “No, no, no, the software—the neural net—is fine.”
Mm-hmm.
“The hardware just can’t do it.” And then they fight it out.
Then we bring in our best teleoperator and say, “He can do it.”
Mm-hmm.
Then the hardware can do it, clearly.
Clearly.
It’s proof of access.
Yeah. Okay, so that’s where I was going. You have a remote operator option—
Yep.
—who can control the hardware. That’s really interesting.
Yeah. Then you’re again like, “Well, but we’re getting to where this gets hard, where we can no longer do this.”
Because?
Because the hands are just so good.
Yeah.
And they have very high-fidelity tactile feedback.
The human hands are so good.
No, the robot hands.
Okay.
The human hands are still even better, but the problem is that the robot hands are really, really good, and they have really fast, highly detailed tactile feedback.
Yeah.
We can’t really transfer this efficiently enough from the human.
Yeah, because the teleoperator—
So, yeah—
—is using some kind of clumsy—
Back to the XPRIZE, right?
Yeah.
The Avatar Challenge.
Yeah.
It’s a really hard problem to transfer that fast enough.
Yeah, you got that right.
Now we’re starting to see that the robot actually learns how to do manipulation much better from reinforcement learning in the real world. You actually have the robot interactively learn in real time how to handle objects, and it can do things that the operator could just dream of.
Ah, amazing.
Now we’re kind of screwed. No, we can’t do that anymore.
5. Autonomy Meets Teleoperation
All right. I want to talk about 3 things in sequence: teleoperation versus full automation, safety in the home, and privacy in the home.
Mm-hmm.
Those have got to be critically important as you’re entering the home. We saw the NEO Gamma out here operating in teleoperation mode, but also in full AI mode.
Mm-hmm.
Right? It was able to do both, and its AI systems are going to increasingly get better and more capable. Again, as I’m talking to Gemini 3, Grok 4, or GPT-5 soon, I’m talking to a highly intelligent human and getting a feeling that it understands what I want, and it’s able to take action on my requests. I imagine we’re eventually going to have that same level of AI in the robot, where I feel like I’m talking to a fully intelligent being, in one sense.
Oh, yeah. Clearly.
Yeah.
And one that is—
We’re very close already.
It is. And one that is grounded, right? That actually understands, to some extent, what this existence is. Today’s large language models have this abstract notion of it, but it’s a facade that quickly falls away if you start to probe at it. That will get there, I think.
In teleoperation mode, you’ve got humans wearing VR headsets and using haptic controls?
No.
What are the humans doing?
They’re giving slightly more high-level commands, just guiding it: “Hey, put your hands over here. Grasp this thing.” You don’t want to overconstrain a system. You want to give it some opportunity to solve how to do the task.
Okay.
We have the learning coming up from the bottom, enabling a more and more abstract interface for the operator. Then we have the learning from all the large amounts of data we have coming from the top, getting more and more of the general behavior that you want the robot to do. They meet in the middle, where—
So you’re using—
Gradually, the operator goes away.
—you’re using full automation and teleoperation always together in that regard.
Yes.
And learning.
Everything that enables the robot to do anything that the teleoperator does is fully learned end to end. The network outputs torques to the motors.
That’s very similar to what Tesla and Elon Musk were saying, where the self-driving car was originally all C++ code with a little bit of neural net—maybe 80% C++, 20% neural net. Then every year that went by, it became more neural net. And now there is no—
300,000 lines of C++ were eliminated. They’re gone.
Just a few guardrails left—
Yeah.
—and the rest is just one neural net. So, same thing here, I guess.
Yeah, it’s all weights. The code is just a few hundred lines.
Is it really?
Yeah.
That’s crazy.
Just all parameters. What’s the parameter count, or is that all super-secret?
It’s kind of secret, but it would be small if you compared it to today’s neural networks because it’s running on the robot very fast. It’s kind of like your muscle neural system. But it does take in vision, so it’s not very small.
Well, that begs a question I’m dying to ask. You’ve seen Ex Machina, right?
Mm-hmm.
When I saw that movie, I’m like, why—
Don’t go dystopian on us here, okay?
Okay. But why is the brain the blue blob in the head?
Yeah.
Why isn’t it in the server room?
Yeah, so learning is shared between all robots.
Well, yeah, and it can be much bigger. If half the power of the robot is going into the thinking, you could save energy. You could run twice as long on a battery charge if you moved it over to the server room and had it just communicate remotely. So why did you choose to put it in a head, aside from being anthropomorphic and cool?
No, no, it has nothing to do with that.
Okay.
There are some simple answers to that. The head is where nothing else is unless you put the brain there.
The room is not, yeah.
Everything else is pretty freaking full. Building a humanoid with this kind of power level in such a miniaturized form, while still having enough space to make it completely soft and all this, is a really hard engineering problem.
So where are we going to put this if we don't put it in the head, if we don't put it on the physical robot?
There are smaller arguments. The very high-bandwidth thing that happens in your brain is vision.
Yeah.
And to some extent, audio, smell, and tactile, but vision just dominates.
Yeah.
You just want to minimize the distance between your eyes and the compute.
For real? So the bandwidth between the sensors—the eye, mostly—
It's very high.
Wouldn't make it over the home Wi-Fi password?
Well, it wouldn't even make it down to the stomach of the robot.
Really?
Really?
Without getting overly complicated about which physical interfaces you would choose for this transfer, it's very high bandwidth.
I'm—
Wow.
I'm shocked by that.
I'm shocked by that too.
We're running no LiDAR, no structured light, no wrist cameras—nothing. We're running pure emulation of human vision, right?
Yeah.
We're relying so heavily on that, so it's very, very high resolution, very high bandwidth, and very high frequency.
That's funny, because that's exactly—
Yeah.
—where the human brain is very close to the eyes too.
It is. Now, that doesn't mean that you can't do things in the cloud, and we do things in the cloud, but it becomes hierarchical from an intelligence point of view.
If you think about your neuromuscular system, this runs quite fast, right? It usually runs at around 25 hertz. It doesn't necessarily go up to your brain. There are neurons distributed throughout your system that make decisions.
Right.
We have this in the robot. We have some of our stuff pushed to the power electronics that controls—
Just for latency, just for speed.
Yeah. And then you have the brain itself, which actually runs pretty fast, right? It usually runs at between 5 and 10 hertz. Even though it's 5 to 10 hertz, it has very low latency, and this runs on the robot.
Now, if you're running more like a 1-hertz streaming thing than, typically, an LLM's time-to-first-token latency, right?
Yeah.
That runs off-board, but it can't solve high-frequency tactile-feedback manipulation tasks. That's too slow.
Okay. The first time my NEO Gamma learns to crack open an egg to make an omelet, do all NEO Gammas then learn that? Is learning shared?
They do. There's shared learning in the sense that you can say, "This data goes to the cloud model that is doing this for all NEO Gammas," but there are also the distributed models.
Of course, there would be a nightly checkpoint where we say, "Hey, this model is better. We have more data, we've validated this, and we've established safety," which I'll talk about later when it comes to how you validate the models.
Mm-hmm.
Then we deploy that to all the robots.
Mm-hmm.
So even though it's distributed on the robots, they can still learn from each other, of course. You just need to do one hop through the server layer, do the training, and propagate it out.
Mm-hmm.
There is a future not so far away where I'm pretty bullish on there being a lot of federated learning happening on-device. This has to do with how we have your companion learn throughout life from all of the experiences that are dear to you, but private.
Yes.
So all robots will not be the same, but they will share an intelligence backbone.
Mm-hmm. Let's go into the conversation of privacy and safety. You're inviting these robots into your home, where there will be activities that you may not want shared with the world. Of course, you're asleep and the robot is running tasks at night. You don't want to wake up in the morning and find your safe has been opened and the robot's gone.
Or you don't want the robot to be taking care of your aging mother and find out that it's giving her shots of scotch at night when she asks for them. How do you deal with safety and privacy?
The last one is the hardest one, by the way. We can get back to that.
Okay.
Not giving Grandma scotch.
Shots of scotch.
Generally, models are always tuned to be sycophants, and they end up—
Yes.
—doing whatever you ask them to do.
Yeah.
If we start with the privacy side, I think, first of all, it's a lot about transparency.
Mm-hmm.
If you're one of the first people, like you, Peter, who will have a NEO Gamma in your house, we're trading a bit on privacy versus being an early adopter, because without the data, we can't make the product better.
Can't wait.
Of course, we're going to do everything we can to make sure this is privacy on your terms and that you're in control, but we do need your data if we're going to make the product better.
Sure.
So—
I mean, listen, I give my data to Google, Amazon, and X all the time. People don't realize that you're sitting in the home having a discussion with your spouse, and Amazon's Alexa is listening, right?
Mm-hmm.
Siri's listening.
But they're doing something very important, which we also do: no human in our company can hear or see that data.
Yes.
That is going into the training model, yes, but it doesn't go by a human.
Right.
If we want to look at that data—and sometimes you might need to, right? It might be, "Let's figure out what happens here, because something clearly is happening across multiple robots that we want to figure out"—then we'll send you a notification on your phone.
We'll say, "Hey, in this specific window, we want to review the data," and you'll get a video of what that data is.
Yeah.
If you say yes, then we get the decryption key and we can look at the data. If you say no, then we can't.
Yeah.
You're in control of that. Even with respect to going into the training data, we always run a 24-hour delay on training. If there is something that you really don't want in the training data—"This never happened. I want to erase it from existence"—you can go in and delete it before it gets into the training weights.
I just want everybody to hear that there are policies and plans that make this acceptable and are used by technology companies, and you're going to be implementing the best of those. That's a pretty bold compromise, actually.
All of them.
Yeah.
But there is something else. The mode I talked about now is what we call best-effort autonomy, which is most of the time. It's what you saw earlier today: if you talk to it and ask it to do something, hopefully it does the right thing. If it doesn't do the right thing, you can say, "Bad robot," and hopefully it's better next time.
It's actually learning in real time. This is really interactive learning. Interestingly enough, the robot progresses faster on tasks when it fails than when it succeeds.
Sure.
It learns more from failures, just as we do. In this mode, that's the privacy.
When it comes to teleoperation, of course, there's no way you can do the task without seeing the glass.
Mm-hmm.
We use some abstractions so that you don't see people. You just see blobs, and you see the object you're interacting with. We can do a lot on the filtering side to ensure privacy.
Mm-hmm.
But the most important thing we do here is that no one goes into teleoperation on your robot unless you approve it, right? It's very visible on the robot—the lighting changes, and it's like someone is in your robot.
Yeah.
It's one of the preselected operators that you have approved from a large set of operators. You choose, "Here are the 4 who service you." It's like inviting your cleaner, or whoever, into your house.
Yeah. Another human.
Another human into your house, and you just need to make sure that they're actually invited.
So, to actually take a second and spell this out in more detail: In the early days, when I have NEO Gamma in my home, it'll be basically autonomous, but there will be times when it needs to bring in a teleoperator. So you'll have teleoperators in headquarters who can step in if it needs help, is doing something complicated, or gets something wrong, and actually make the task happen.
Yeah. In the beginning, there are actually 2 different modes. So you have the mode that I call the best-effort autonomy that we just talked about.
Okay.
Then you have task scheduling, which my robot at home is doing now. I take my phone, schedule it, and say, “Hey, between these hours, here are the tasks I want you to do for me.” Today, it's: do my white laundry, and then there's a package coming from Instacart, so you can receive it at the door and unpack it in the fridge, and it's just generally tidy.
Mm-hmm.
And I've given it the hours when I'm not home: these hours, I'm at work. Just get it done, right? Now, I don't care if that happens autonomously or through a teleoperator. A lot of that happens through a teleoperator because some of these tasks are quite complicated, and we don't know how to automate them well enough yet.
Yeah.
Now, of course, the teleoperator uses autonomy to help improve the efficiency. So it's not all teleoperation, but I don't really care about the mix. The task gets done.
Yes.
So we kind of split it like that, and then there's the gray zone in the middle. If you want to have your friends over for a party and you want the robot to be the bartender, and we don't have a bartender mode yet, then you can approve a teleoperator to do that. So you can also schedule it.
Mm-hmm.
And most, if not all, of the videos we see of Optimus at Tesla's diner or at their events are teleoperation modes.
They are. But I think teleoperation has gotten this underserved, bad reputation.
Why?
Because I think people don't have enough clarity: Is this teleop or is it autonomous? But it is just labeled data. It's expert demonstrations, right?
Mm.
If you look at any of the big AI models that were trained, an enormous number of people sat down and hand-labeled data, looked at examples, wrote out question-and-answer pairs, and bootstrapped a very high-quality data set for this to work, right?
Mm.
So you pre-train on general information. We also do that—just everything that's happened with the robot. Then you have a fine-tuned data set that's very high quality.
Yeah.
And in robotics, that is teleoperation because it's the expert demonstrations. It's the hand-labeled data.
Yeah.
So there, it's no different. I think there's a lack of transparency in what's going on.
Well, I think the objection is: if you have a demo, like a video, that makes it look like it can do something, and it actually can't because you hand-coded it.
Mm.
Well, it clearly can, but it can't do it autonomously.
Yeah, it can't do it. But yeah—
But—
But I think you're dead on that if it can physically do that, if the mechanism can do that, the neural net will fill in that blind spot instantly anyway.
Yeah.
You know, once you've trained it. So I think it's perfectly legit. I want to jump into another fun subject, which is the uncanny valley and the face. Talk to us—you've probably had endless conversations internally about how do you make a face look, how human do you make it, how skin-like do you make it, how do you represent it. Can you tell us philosophically what you and Dar [?], who's on your design team, think about that? Where do you make it human enough?
It is this very delicate line where you want to make sure that body language comes across crystal clear. Because that's the magic of the device, right? Of the companion.
But at the same time, you don't want it to get to where your instincts tell you, “Hey, something's wrong. This is a human, but there's something wrong with it.” So you don't want it to be a human. And it is actually pretty surprising that there is this gap where people clearly identify this as, “Hey, this is a being I identify with. I understand its body language and everything, but it's also clearly not a human.”
Yeah, yeah.
And you want to be in that space. Where you are in that space depends a bit on who you ask, right? People have a different threshold here.
So we're trying to hit in the middle of that and ensure that, for as many people as possible, this is just an incredibly easy-to-understand product. But at the same time, it's not creepy.
Yeah.
And I think adoption here—by the way, talking about scale—is so important. Adoption of new technology usually takes some time because there's this knowledge barrier, right? There's a barrier to entry. Even using a phone, there's a barrier to entry.
This interface is just so natural. There is no barrier to entry. It's something you just talk to, like a person.
It is. What's incredibly cool to me is that there are 50 things around the house that I don't know how to do, including the fricking way to backwash the pool. All this crap. The robot can, in real time, access the information, learn how to do it, and just do it.
Yeah, from the web.
I can't do that. It would take me an hour to study, and there's no laborer who's going to come into the house and do it for under 400 bucks. So there are so many things that are in that category where I'm not trying to replace a human being; I'm doing something that there literally was no other option for because the knowledge is obscure.
There are so many of those things around the house now, like resetting the water heater when it keeps going out and the reset process. But you can look it up. The robot can look it up. And just go do it.
This is micro-units of work, essentially, right?
Hyper-specialized micro-units of work.
Yeah. And you need 5 minutes of it every now and then, and it's super high value to you. It's just really hard to get to.
Or a fricking Shop-Vac. The Shop-Vac—you can run it forward or backward. There's a manual there. You could read the manual. I just want to get this crap off the garage floor. The robot will know how the Shop-Vac works because somebody else's robot—
Right.
one of the other 10,000 has already done it.
Make my perfect teriyaki salmon on the grill.
Yeah, some obscure mixing, some food. Yeah.
Which brings us to something that you talked about earlier—
The Scotch for grandma.
That we kind of dodged: Scotch for grandma and safety.
Yeah.
So, I do hope to make you a perfect salmon teriyaki.
Thank you. Appreciate that.
But I have to do it myself. I'm not going to let the robot do it because that's one of the things we're actually not doing at launch, and that is due to safety.
Yeah.
Mm-hmm.
Because what I worked so hard on for this decade is to make robots that are safe intrinsically.
Mm.
What I mean by that is, if something goes really wrong and it accidentally hits you, that might be painful, but it's not likely to severely harm you.
Injurious, right?
Yeah. And once you pick up a kettle of boiling water, there's no more guarantee that you are safe, right?
Mm-hmm.
Mm-hmm.
So we generally avoid any kind of dangerous objects so that we can ensure safety in the beginning.
Yeah.
Of course, over time, as the AI improves and we get more and more certainty that all behaviors are safe, we will allow cooking and other things. So we're doing internal projects on this, but we're not going to be rolling it out to customers in the beginning, just due to safety concerns.
Yeah.
Um—
Yeah, cooking—
So—
Cooking and safety is a—
It's a real problem for humans.
Cooking around fire is not an easy thing.
I mean, it is a real problem for humans, too.
Yeah.
It is. But there's the notion of intrinsic safety. This is incredibly important, and that is the safety of the AI. This is the reason we have a white paper out on this that I might recommend if you guys are interested—read it.
6. World Models Make Robots Safe
But why we started very early betting extremely heavily on world models.
On role models.
World.
Oh, world models.
World models.
Yeah.
They are, of course, the currently best-known path towards AGI. But even more importantly for us in the short term, as we progress here on the data collection and model training, they give us this incredible opportunity to automate evaluation of models, including safety and red-teaming and all these things.
Mm-hmm.
So you can think about, if you train a new model and now you want to know if it's better than the previous one—
Mm—
—and you can deploy it to all your customers and get some vibe check a few days later, like, “Hey, are people happier now?” That's generally how it's done.
Yeah.
You don't want to do that with a physical system, right?
Mm-hmm.
You can't do that with an autonomous car either.
Yeah.
What a world model actually is, is a model that is able to generate what will happen if you take specific actions. So you can think about a video model where you ask it to do something, and then it actually gets not only the question of what to do, but the actions to do so. It gives you back not just a video, but how the world feels—the forces, everything for the robot. So it's essentially like the robot's in the Matrix.
Yeah.
We take the robot, we put it in a world model—
Yeah—
—and it doesn't know that it's in a world model. It thinks it's in the real world.
Yeah.
And it does its things, and we ask it to do the things we're usually doing around the house.
Yeah.
And we see what it does, and we can put in lots of automated checks to ensure that it's both performing better and not doing anything that could be deemed unsafe. So it's really—
It's wild.
—this incredibly important and powerful evaluation tool, and that's a start to solving the problem, right?
Do you think that's why you guys, Figure, and Tesla are getting monster valuations? Is the valuation purely, “Hey, we're going to sell 10,000, 20,000, then 200,000”? Or is it that the world model is such a unique asset and so valuable in thousands of different ways, and that becomes a very much self-feeding barrier to entry? That could also explain it. Do you plan to productize that core capability?
Yeah. To me, it's this: our mission is to create an abundance of artificial labor, and that goes across both the digital and the physical.
Mm-hmm.
So yes, it will be productized.
Mm.
Still a bit out, but yes, this will be productized. And—
Yeah, because it seems like no matter how much factory capacity you build, it wouldn't be until 2028 or 2029 that you could diversify into all these things, like microsurgery, warehouses, drones, all that. But that same world model could apply to those much sooner, but you'd have to somehow get it into the hands of many other companies.
Well, I actually think revenue from the robots will dominate forever. I do think the real physical world has—
I mean, it's—
I mean, just for folks to realize, we're at $110 trillion in global GDP, and labor is half of that, right? So the TAM—the total addressable market here—is $50-plus trillion.
Just—
Yeah.
Just if you keep doing what we already do—
You will achieve capability—
—but you do what we'll be capable of doing, yeah. It's—yeah.
It's going to be so much bigger.
Yeah.
And you've attracted some incredible early investors. Do you mind just sharing who has come into your cap stack?
I think we have some big classical ventures like SoftBank, Tiger Global—
NVIDIA—
—EQT, NVIDIA, OpenAI.
OpenAI. Yeah.
So there's some good names in there.
I mean, that's damn good.
A lot more.
Mm.
I think it's becoming increasingly clear that the bottleneck in society—
Mm.
—to superintelligence is not better algorithms or scraping the internet in a more thorough way.
It's better data.
It's—yeah, it's better data, and then you need robots to generate this data. But even more importantly, it's the physical parts, right? You need more data centers. You need more power. You need more labor. And to do this, it's kind of like this: it's a bootstrapping problem.
If you break down the pyramid and say superintelligence consists of this incredible amount of data and this substrate of compute and power, then you see that humanoids are a solution to both of them.
Mm-hmm.
And if you just do the math, you'll see that you're probably not going to get there without that. You're just running up against these basic constraints. I think humanoids will be surprisingly useful, surprisingly fast.
Okay. I—
Not perfect, but they're going to be surprisingly useful surprisingly early.
Yeah.
7. Humanoids Fit The World
I have a question on behalf of Salim Ismail, who's typically our third moonshot mate here. I have to ask on his behalf.
No.
So Salim is constantly saying, “Why 2 arms? Why 2 legs?”
Mm-hmm.
“Why not 6 arms? Why do we need to have a humanoid form? In the kitchen, wouldn't it be better to have an extra pair of arms?” So what's the definitive answer to him?
Well, I think first of all, he's kind of right. Humanoid isn't the only thing that will work. I do think that—and we've looked a lot—I don't know of any form factor that is as general as a human in doing any kind of labor in any kind of environment.
We've tried to simplify. We've tried to increase complexity. The human is a pretty good machine.
Mm-hmm.
If your goal is to just be as general as possible, then you need a humanoid. Now, if your goal—
I think that's the most important part of the equation there.
It's very important. And then—
You're not going to transfer to a 6-armed robot learnings from a human.
No. It's at least harder. And then—
Mm—
I mean, the world is made for humans.
Yeah.
It's what Jensen says, right? It's brownfield deployment. It's very true. And then I think lastly, it's just: do you want to live with a 6-legged robot in your kitchen?
No.
But I view humanoids as kind of the pinnacle of general technology. There is this repeat pattern through history of this happening with zero-to-one novel products. So if you think about, let’s say, the computer, it started with big mainframe computers solving very specialized tasks.
Mm.
The equivalent in robotics would be industrial robotics, right? Now comes the PC, or even before the PC, like the VIC-20s or Ataris or whatever—more general computers. This gets produced at such a scale that it just becomes generally available. Now it’s super high quality, incredibly reliable, has this huge ecosystem, and becomes the best way to solve any problem.
Yeah.
And here’s the argument against humanoids, right? It’s overly complicated for a task. Even though it’s overly complicated for the task, when you take your beautiful Apple here and write a Word document, that’s the most complicated typewriter I can think of. Humanity mastered nanoscale chip manufacturing for you to have a typewriter.
Yeah.
But it’s still actually the cheapest, most reliable typewriter because it’s made at such a scale.
Yeah.
Humanoids are exactly the same. Now, if you see what happens to computers, because the market has become so big, it starts to become segmented again. Now you can carve out niches in computing, and they’re still so large that they have scale. So now you get specialized compute for AI, specialized compute for physics, specialized for all kinds of things, right?
Yeah.
And this is because it’s become so big. The same will happen in robotics. We will get to where we have Star Wars. There will be different drones doing different tasks, and they will look more specialized, like my repair drone with six arms and scissor hands and I don’t know what else.
Yeah.
It’ll get there, but you have to go through this humanoid phase first. So let’s just say humanoid is a phase.
Mm.
Yes.
My favorite robot is still Data from Star Trek.
It’s a great robot.
Yeah. It’s kind of the closest thing I think of to what you’re building: a lovable, happy robot that you can give a hug to.
Yeah.
Do you have a favorite robot?
I wouldn’t have thought of Data, but now that you said Data, that’s top of the food chain. Everybody loves R2-D2 because, for some reason, R2-D2 has no voice even though they have voice technology everywhere.
It squeaks. Yeah, it squeaks.
All those visions, though, are built around what Hollywood could easily get on a set.
Yeah.
I think the humanoid form factor, though, has another aspect that you kind of touched on. When I bring it into my house, I have a vision of what it can do and what it can’t do based on humans.
Mm.
So I ask it to do things that are rational and not irrational because I know what a person could do. If I had a six-legged thing that Salim came up with—
Yeah.
—I’m not quite sure. Should it be able to climb on the roof and fix the shingles or not? I don’t know what this thing’s capabilities are.
Interesting. Yeah.
So it breaks the whole comfort zone of expectations.
Expectations, yeah.
The thing that does surprise me, though, about the robots is how unbelievably coordinated they are between themselves, and there are some good demos of this at MIT that are just mind-blowing. When you have 2 movers trying to take a couch up the stairs, they’re like The Three Stooges, right? They’re saying, “Oh, move it to the left. Now move a little bit.” When you see the equivalent act with 2 robots, they’re just in phase and do it seamlessly. So I think there’s a very high probability that the standard in the home is going to be 4 or 6, if you get the price point down a lot. They work so well in concert with each other, it’s almost a crime not to have that teamwork synergy.
Huh. It seems like a bit much to me. Maybe, if I need to have some movers, I can ask my NEO Gamma, and he’ll invite some friends over.
You’re not taking everything into account, Peter.
Yeah.
Because you have to remember that by the time you have this many robots in your home, everyone’s homes are really freaking big.
Yeah.
We have an abundance of labor.
Mm.
Your house is not going to be this small—
I’m trying to think.
Everything is going to be huge.
So labor is going to continue to demonetize and democratize.
Let’s go someplace that I’d love your insight on, which is China. When I think about the robot industry, I’m tracking 50-plus well-funded humanoid robot companies in different stages around the world. The majority are in the U.S. and China. You started in Norway. There are some in Europe, some in India, parts in Japan, and Korea.
But China, by far, I think, is dominating. What I see there with the Robot Olympics and special robot villages is pretty extraordinary, where the Chinese government is really accelerating this for obvious reasons. They need access to low-cost labor to continue the manufacturing boom, and they need it to support their elderly population. How do you think about China? What do you think of the work coming out of China?
Well, first of all, I think we need the same thing here.
We do need the same kind of support.
This is not something we realize as much, maybe, but of course we need the same thing.
Yeah.
I think the Chinese ecosystem is incredible. I don’t know anywhere else in the world where you can develop hardware as fast. You need something, and you go over and get a machine on the corner here, something broken—
A Shenzhen-based—
—PCB, and you just go over the street and buy some new components. There’s someone doing a reflow over on the street corner over there. This is an incredible ecosystem.
Mm.
And I think the Bay—I say “the Bay” now, but I know it’s also, I mean, Silicon Valley. The hardware bay in the Shenzhen area is also a bay, but Silicon Valley, this bay, we have a long way to go if we want to really get to the same level of rapid iteration on hardware.
Yeah.
That’s just incredible. I think the manufacturing part is incredible. There’s so much process knowledge.
Mm-hmm.
And I think this is highly underrated. Think about magnets.
Hmm. I do.
So do I, a lot. We have great material scientists who know how magnets work and can design very good magnets. But then we lack that guy who knows that, yes, you do all of that stuff they told you in the books, but after 2 hours you have to stir to the left, not the right. There’s just so much of that.
Mm-hmm. Yeah.
And this is just so disseminated in China. There’s so much process knowledge.
Yeah. How did that evolve in China and not here? What’s the cause? Top-down incentives.
Just funding, or—
I think it’s the government saying, “You’re a robot city, you’re a neodymium magnet city, you’re…”—and just capital and people, and communist-directed, but then allowing companies to build on top of that.
Mm-hmm.
I mean, is that what you see as well or not?
I’m not sure. I think the Chinese startup community is very alive, right?
Mm.
And the capital is quite alive. I feel it runs very similarly to what we like to think of as the Bay here—
I mean, I used to take a group of investors every year to China, and we would go and visit Shenzhen, Shanghai, Hong Kong, and Beijing, and meet with Baidu, Tencent, Huawei, and the leadership of all these—
And there was a super vibrant entrepreneurial community, right?
Yes.
The mindset was 9:00 AM to 9:00 PM, 6 days a week, and that was a great lifestyle. You considered the 1.3 billion people in China your market, and the 300 million in America your market as well. But there was a fall-off after 2019, and there was a real dip in that ecosystem. I think it's beginning to reemerge, but I think the government is really pushing hard on supporting AI. Humanoid robots are an embodiment of AI. Obviously, you know this, they're closely meshed. I do think there's a lot more support that the US government needs to give to US—
100%.
—hardware companies.
But I think the most genius thing they did was the special economic zones, like the free economic zones.
Yeah. Sure.
It's not that people here don't want to build stuff. We want to build stuff.
Yeah.
It's just that it takes too long, costs too much, and it's too convoluted, right?
We do it in spite of the challenges.
Yeah. I think the US should just spin up some special economic zones. Here you have expedited permitting and—
Well, in California in particular—
Yeah.
—it'd be a no-brainer to do that. That's the simplest, best idea ever, but I don't know what it would take to get it through.
No, but this is some of what people are working on these days, right? If you look at Masa Stream, Project Crystalline, for example, it's very similar to this kind of free economic zone in the US.
Mm-hmm.
I think there's another problem you need to solve, too, though, which is that the US just did software forever. We were not only not doing hardware, we just didn't do chips. The chips—
Well, not forever. We're in Silicon Valley.
Yeah.
Yeah, where's the silicon?
There was—
You're not making silicon in Silicon Valley.
Yeah, yeah.
Yeah, yeah. What's up with that?
There was a phase in between where people lost the way.
Yeah.
They lost the plot.
Yeah.
And now we have to find the comeback.
Well, the venture community got all messed up, too, because they wouldn't fund you if you had a physical component in your business plan. They'd be like, “Well, yeah, I'm looking for the next Meta or Google.”
Yeah, hardware's hard.
They don't really care. Yeah, hardware's hard.
I've been keeping this company afloat for 10 years, so I can—
You've got scars.
—I can go on and on about—
Yeah. Do go on and on, because it needs to get fixed.
But people are so afraid of hardware.
Yeah.
Right?
Yeah.
But—
It's going to kill us if we don't find a solution, and quickly.
I think it's also going to kill VC, to be honest. If you look at returns on venture, it used to be incredible, right? If you got to be an LP in a venture, you're like, “Oh, man, I'm set, right? I'm going to make the big bucks.”
Yeah.
And now it's more like a philanthropy thing. You want to fund startup entrepreneurs because venture doesn't really make that much money. I think it has a lot to do with—
Except at Linq. You guys are doing amazing. We are doing amazing.
That is good.
Yeah.
That's good.
I know the data.
But you guys touch the hard stuff. The point is, if you don't touch the hard stuff—
We touch the ear—
—what's your moat?
—we touch the early stuff, right? So it's first checks into companies that then are scaling rapidly versus companies that are doing late stage.
But we're not. We're doing first checks into hard stuff at the seed stage, but we're not doing hardware. So I'm as much a part of the problem.
Well, you should look at the biggest companies. They all have hardware.
Yeah, I mean, listen, Elon Musk cracked the code on that. He's been able to just make hardware sexy and has generated incredible returns.
Yeah, and I think Jensen Huang says it really well, right? They want to work on the really hard problems that are super painful, that you are uniquely capable of. You know that your competitors have to go through the same pain or more, and they're not going to be willing to take as much pain as you. This is how you win.
Yeah.
And things here are actually defensible, right? The moat we have on hardware—
Yeah.
—that's years. The moat we have—I'm incredibly proud of our AI team, by the way. We've accomplished some things that are so amazing on such a budget.
Yeah.
So we're way ahead of everyone else in what we're doing on world models. Let's say we're 3 months ahead, right?
Way ahead.
Well, that is way ahead.
But you're incredibly rare. All props to Elon Musk—he's incredible—but Elon's pathway to getting to hardware was through PayPal.
Yeah, sure.
He burned a couple hundred million dollars himself, got down to near bankruptcy, and was almost dead on both of his big companies, Tesla and SpaceX. He barely pulled it out, and then made them huge, but the VCs were not touching it.
No. He had to borrow money in 2008 during a divorce, with SpaceX having its third failure and trying to borrow money for its fourth launch.
Yeah, that is not—
Yeah.
—a repeatable funding model for America.
I was very lucky. I had a very good early founding investor. The company didn't start in a garage because we're not Silicon Valley. We started in a barn because we were Norwegian.
Yeah.
At some point, 2 years later, he sold the farm, so we had to move. He sold the farm to fund the company.
Huh.
Right?
That's amazing.
Yeah.
This company in Silicon Valley wouldn't exist without—
No.
—a Norwegian investor—
No.
—believing in your vision.
And we wouldn't exist because we wouldn't have had the runway, right? Operating this in Norway was just incredibly cheap compared to operating here.
So your initial Norwegian investor, did he or she believe they were going to make a huge amount of money, or did they do it because they were passionate about your vision and your mission?
Or did they believe in you?
Yeah.
I think it's all three. It turned out pretty well.
Yeah. Well, yeah, but—
But—
—it wasn't here. That's the point I'm making. It's like—
No, but there are different phases, right? If you want to scale something, you have to come here.
Yeah.
I think you can do deep research in other parts of the world. There's talent everywhere. But really hyper-scaling that and getting it across the finish line—
Yeah.
—that's here.
Did you consider LA, Austin, Florida, versus here in Palo Alto?
Yeah. We even had manufacturing for a little time in Texas, in Dallas. There's just no comparison to the talent.
Something in the water.
It's the talent pool.
Yeah.
There's talent everywhere, but the density of talent matters. There are different types of talent because when you have a zero-to-one field like this, in the beginning you have a lot of really passionate people who have been working on this all their lives, and they're so good at, in this case, humanoid robotics, right?
Mm.
I remember back in the day, if you went to the Humanoids Conference, everyone could fit around one small table. Those people are still the ones who are at most of these companies, right? Those people don't know how to make a great product. They don't know how to scale that to a million or a billion devices.
Yeah.
They don't know how to write incredibly good APIs for the software to support the ecosystem. They know this thing and do deep research. And now your field comes of age, and it's time to actually do this because the timing is right. We purposefully stayed very small for the first 7 years, just doing core technology.
Now, suddenly, you get access to this talent pool of people who go from field to field, whichever is the hottest thing right now, and just do it again and again and again and again. That's Silicon Valley, right? But there's been an incredible inflection point in humanoid robotics.
I remember we had the ANA Avatar XPRIZE, which had teams build robotic avatars that you could telepresence. I remember the finals; we had good teams. I know some of your team members here were part of those teams, but it's come 1,000x since then, really in the last 5 years.
Is it the AI models that have made that? What's caused the inflection in the last 5 years? The AI is clearly part of it. There are things we do with AI now that we couldn't do 5 years ago. I do think we saw the breadcrumbs, and we were on the path already then, but it wasn't working yet.
I think you just hit this critical mass of accumulated innovations that have happened in hardware. I do think it's important to note, though, that it's hard to see what is real innovation and what is not in any field, especially in humanoid robotics. You can go on YouTube and find things from the early 2000s that look better than most things you see today from humanoid robotics companies.
You can't just make a beautiful robot that looks good. You have to actually make a robot that is safe, that you can manufacture at scale for a very affordable price, and that is still capable. I think that's been the main unlock and the challenge. You need to get those things right, and that just takes a lot of time.
I think the neural nets are light-years ahead of anything anyone would have predicted 5 years ago. Then the hardware—the NVIDIA chip that it runs on—is getting pushed as fast as any innovation in history because the demand is through the roof. That part is well understood.
On the physical hardware side, what's something that you use today that you couldn't have used 10 years ago? What's improving in the motors, the harnesses, the electronics, and the batteries?
Great question. I think mostly it's been on the motors and material science side. We make our own motors, including not only the IP for the motor, but also the manufacturing and automation for all of this and everything that goes into it. The supply chain is so broken.
You literally make your own motors.
We have to wind the wires. We do it kind of special; it's the 1X version of this. Motors are one of the things we really innovated in, and this is actually how I started. When I sat down a decade ago, the first thing I did was design a different kind of motor.
The motors we have now in Neo are 5.5 times the world record in torque-to-weight ratio.
The product does that, too.
And that's why we have something that's so powerful that we don't need gears. We can just pull on these tendons to loosely simulate human muscles or tendons.
That's why it's so light.
It's also why it's so drivable and compliant. It's why it's so cheap to manufacture. Everything kind of comes from this.
Now, of course, when you have these motors, you can start using tendons. But then you need to sink a lot of time into figuring out how to use tendons, and then comes all the material science to have tendons that can last millions and millions and millions of cycles.
These are really hard research problems. They're not even engineering problems; they're hard research problems. We spent so much time figuring all that out. You can't make the motors that we make without doing some pretty significant innovations in electronics, how you do power amplification, and, in general, motor drives.
There are a lot of things that come together. You couldn't have designed the motors we do today without some of the innovations that have happened in magnetics. Of course, you couldn't have done it without AI, either.
The first thing I did back in the day when I sat down was program a network to learn how to make motors.
You designed the motors via AI? How long ago was that?
It's a bit more than 10 years. It wasn't Transformers, but it doesn't matter. For that kind of use case, it was a neural net.
8. Abundance Transforms Civilization
Bernt, you think about robots in the world probably more than anybody else. What's your vision 10 years from now? What are we seeing? What does abundance in labor enable that goes beyond people's initial reaction to how I would use a robot?
I think, first of all, what will happen is that actual abundance means everyone can have whatever they want. But not only can you have whatever you want, you can have whatever you want in a sustainable manner. Sustainability is something we lose when we cut corners to shave costs.
If you actually have an abundance of energy and labor, why would you not do things sustainably? Then I think the next frontier that comes after building out the infrastructure across the globe that allows everyone to have an incredible quality of life is: How do we solve the remaining really hard problems in science?
I think this is not going to happen without humanoids, because you need to build particle accelerators. You need to build enormous biotech labs where you're doing experiments with really intricate networks. You need to do all the experiments.
I think it's almost existential to us for human happiness. I don't want the godlike AI in the sky to be directing all of the planet's inhabitants around with their glasses to do experiments for it to solve science. That's not the future we're aiming for. We want to have this beautiful symbiosis and co-invention between man and machine.
Yeah, that particular use case is so acute, where Dennis Aassovis is working on the full-cell simulator to try to close the loop. But you know that you're going to need people to mix a huge number of chemicals to truly unlock longevity, health, and chemistry. The humanoid robots can do the work because everything in the lab is oversized.
No, not only can they do the work, I think this is a common misconception. Humanoid robots will do a lot of the work initially, but once it gets to a certain scale, the humanoid robot will make the automation system that will do the work.
Because humanoid robots will not be machining new parts with a Dremel, right? You will use the CNC machine.
Mm-hmm.
Humanoid robots will not be moving car chassis around by carrying them with 30 humanoids. Clearly, this does not make sense, right?
Yeah.
We have existing automation systems, and we will build more. What humanoids will do for you is build all these automation systems and get them up and running, and then cover the remaining gaps that you currently can't do with humans.
Yep. Yep.
How are you going to do it—
Such an unlock.
How are you going to do it in a vacuum? I want my Neo Gamma to help me set up my space station or mine my asteroids.
Well, we can go on and on. I think, first of all, we have a huge advantage because the robot is so light.
Yes.
I guess Elon's working on this, but payload to orbit is still expensive. Secondly, most of the stuff we have actually works pretty well in space. We have to do some work with the epoxy on the motors; that's not going to be very vacuum-hard.
If you want to train in zero G, one of my companies is called Zero Gravity Corporation. They do these parabolic flights.
Oh.
Yeah, we flew Stephen Hawking in zero G. Maybe Neo Gamma should come next.
That would be great. I actually do think there are real use cases for this, and one thing is building a base on Mars or whatever, right?
Mm-hmm.
But even before we get there, just in-orbit assembly is this extremely high-value task, and I think there we will use teleop.
Yes.
The reason I'm saying that is just that the cost of mistakes is so high that you want to use the smartest, most expert humans you have. Until we get to superintelligence, that will be a human. You have people in orbit, you have robots outside, and there's very low latency—
Yeah.
—and you can teleoperate in a very natural manner, as if it were your own body, to do all of these in-orbit assembly tasks.
And they can be incredibly complex, and you can still do them with very high accuracy, and you're not endangering people.
Yeah.
And of course, when you've done this for a while, you have the data to automate all this, which is very interesting.
Yeah. That's where your weight advantage would be really amazing, too, because you can take 5, 6, 7 of these up.
And the energy efficiency.
Yeah. You guys are gonna have to somehow bleed off your heat, right?
Right.
It's really hard.
Yeah, that's right.
You must be looking to hire people.
We are.
What kind of people watching are you interested in potentially hiring?
People who are really mission-driven, who really believe in the beauty of a world where we have an abundance of labor.
Mm-hmm.
And like to solve really hard problems. People who can also demonstrate that they've solved incredibly hard problems.
Mm-hmm.
Because that's what we're doing here, right? Everything from materials science all the way at the bottom, all the way up to the foundation models at the top. And I think what we offer is just this incredible place to work.
Mm-hmm.
Not with respect to work-life balance or any of this. We're not quite Chinese, but it's a hard problem and we're in it to win. But probably the place on the planet with the most experts across all different disciplines in science.
So if you come here as a mechanical engineer, you will learn so much about AI, about electrical engineering, about batteries, about materials science, everything else. And it doesn't matter which discipline you come from, right? You will learn so much from the people around you, and I think also that's one of our biggest strengths, how we really always work in these multidisciplinary groups.
Yeah.
And we find the good solutions between the disciplines. It's like, “Hey, you don't need to do that. That's kind of costly in manufacturing. I can calibrate that away.”
Yeah.
Or, “You don't need to calibrate. This doesn't cost any more.”
I can see that, actually, when you're walking around the building here. Dean Kamen's lab in New Hampshire is very, very similar, where he was the Segway inventor, and everybody's happy.
All the MIT people that we know who work with him, they're just happy. And the reason is because when you do software, you're largely behind a workstation all day. You're sitting. When you're doing physical things, you're moving around a lot more, and you're building and making, and it energizes you all day long. It's just such a fun work environment.
It's so obviously tangible, just walking around and talking to people. So it's a good lifestyle.
And it helps when there's a lot of robots walking around with you.
Yeah, for sure. People can go to 1X Technologies' website to—
That's true.
—to go and find out what positions are open.
Yeah.
Yeah.
For sure. And one thing I'm excited about to announce is, you and Dar and the Neo Gammas are gonna be at the Abundance Summit in March.
Yeah.
Can't wait.
Yeah.
Meet a lot of great people.
Yeah, so our theme this year is the rise of digital superintelligence and the rise of humanoid robots, because the two are going together.
Sounds pretty spot on.
Yeah, I think so.
Yeah. Yeah.
I mean, it really is. And without making any promises, I'm hopeful we'll have a number of the Neo Gammas there, interacting, living, and hanging out with the Abundance members.
Yeah, how do they get there? You buy them an airplane seat? They just walk on? You sort of—
Yeah.
You don't box them up, do you?
Yeah, we're down in LA, so—
Yeah, we're probably gonna drive down to LA.
Okay.
It's easier than getting them on a plane.
Do you put them in the seats and strap them in, or—
Yeah, we do.
That's so funny.
Actually, at this point, they're starting to sit into the seat themselves. So it doesn't strap itself in yet, but that's coming.
It's a funny story, though. We put one of the first robots on a plane back in the day.
Mm-hmm.
We were rushing back home from China. It was a proper startup story. It was way back in the day. We were running out of money, and we hadn't gotten to where the product was good enough to raise more money.
So I took the entire team and we went to China, and we lived in a hotel for 5 weeks, designing and manufacturing as we went.
We designed until late into the night. In the morning, you walk down to the machine shop, you help get them some information, you get some new parts back, and we just kept iterating on this. The electronics market—everything's magical, right?
Yeah.
Then we had to go back, and we were like, “Okay, we need to rush back on the plane to meet some investors.” So we checked the robot, and we folded it up, right?
Yeah.
And we put it in a briefcase. And then when it goes through the scanner—
Oh, no.
You can see the guy just go all white. His hands are shaking as he's opening the bag. And we're like, “No, no, it's just a robot.” And he's like, “Yeah, it's a robot.”
That's hilarious.
Oh, that's awesome.
I'm really thrilled. I loved your TED Talk, and I'm excited to have both Neo Gammas there, hanging out with all our Abundance members. And hopefully, you'll be ready to sell some robots.
In the early days of getting them into the home, no promises, but you're going to have an application process to get the robots in and start to build data assets. When do you think you'll be ready to take pre-orders and orders for Neo Gamma?
I'm gonna be kind to my team and not say a specific date.
Okay.
But it is happening this year.
Okay.
It's this year, 2025.
This year, 2025?
This year, 2025.
Yeah.
Now, we're gonna talk a lot about this in the pre-order.
Yeah.
But the most important thing we do here is manage expectations.
Yes.
This is incredibly early, right?
Yeah.
And what you're buying here is a ticket to be part of this transformation.
Mm.
Adopt a Neo into your family. Help us teach it. It's gonna be a lot of fun.
I love that.
It's gonna be useful.
I love that framing. That's perfect.
It's going to be useful, but it's not gonna be perfect. It's gonna be a lot of rough edges.
Mm.
And we're gonna treat you really well, and we're gonna figure it out together, and it's gonna be an incredibly fun journey. And that's kind of the early adopter program that—
Yeah.
Well, you're gonna have a long waiting list. We need millions and millions of these, and we need to get the price point down. When you think about the constraints to human happiness globally—
Mm.
—a lot of them are gonna be solved through regular AI. But another big chunk—most of them—is related to houses and food and—
Give the jobs that are dull, dangerous, and dirty to the robots.
And then create a lot more of the things that make people happy: the parks, the homes, and all of that.
Everything.
Bigger homes and—
And—
—better things to play with. It's all constrained by our inability to manufacture because of the lack of a humanoid.
Let me ask you a numbers question. I interviewed Elon at the FII Summit. You're gonna be there in October as well, and I also interviewed Brett Adcock. They both gave a number of around 10 billion humanoid robots by 2040. Do you believe that number?
10 billion by 2040?
Yeah.
I think it's probably roughly correct. I think it might happen before. I think it really comes down to what kind of artificial constraints we put on how we scale.
Yeah.
Mm.
At that point, you have to actually really think about how you're refining rare earths, how you're mining more aluminum, and how you're ensuring that you get your labor bootstrapped really well with robots—
into labor? How do you build out a power infrastructure? We need more chip fabs, by the way.
Yeah.
We're not going to be able to build 10 billion humanoids without way more chip fabs.
Yep. Yeah.
We can help with having robots build this out, but I do think that timeline depends a lot on how permitting processes go and how much we allow ourselves to scale, and how fast. But I do hope we get there.
Yeah. For reference, there's about 1 billion automobiles on the planet.
Mm.
You would think there are more, but there are on the order of 8 billion smartphones on the planet.
I'm really glad you said what you just said, though, because the numbers are so wildly out of balance. Each one of these robots uses about a full GPU. It could probably use 2.
If you're talking about 1 billion of them by 2040, we're only making 20 million GPUs a year. And then TSMC has 66% market share in the fab market now. So they have literally one point of failure for the entire economy that we're trying to build. And so we're desperately short on the fabs.
And that's if you just go 1 layer deep. Look at ASML behind it.
Yeah. Oh, my God. Sure.
Right? So the supply chain for chip fabs—
Yep.
That's even more brittle.
Yep. I'm really surprised that we're not moving much faster, given that Elon is right in the middle of it. Elon is, or was, in Washington, and we're just letting this bottleneck fester.
Well, how long have we been talking about magnets?
How long have we been talking about magnets?
We've been talking about magnets for a long time, right? That is a problem that only China can really make high-grade magnets.
Oh, rare earths.
And it's not just the rare earths; it's the process to produce—
Mm.
Yeah.
Yeah.
Right? And I think now, finally, people are opening their eyes and saying, “Wait a minute, this is actually a real problem.”
We meet with a lot of government officials, and they're completely unaware of these bottlenecks. And it's funny—if you point them out, there's still no reaction. It's like... But it's so acute and so urgent. You're in a perfect position to actually identify those bottlenecks, so it's really great that you said it on this podcast, because then we can take that material and say, “Look—look, he would know. This is what we need. This is going to be a crisis very quickly.”
So, yeah.
Bernt, thank you for the tour today. Thank you for the work that you're doing. Super grateful. Excited to have you at the Abundance Summit with your team of robots.
And, by the way, the reason you named the company 1X—I think that's worth closing out as the story here.
Well, there are all these videos on YouTube, and there's always an 8X or 4X in the corner. All we do is real time because we build proper robots.
There, you got it.
Yeah.
What you're seeing is real 1X speed. And we had fun today with NEO Gamma.
Okay. Well, a real pleasure, my friend.
Awesome.
Thank you for today.
Awesome.
Thanks, Jeffrey.
The things I get to do because of this podcast, and just how it was.
We're having fun.
Oh, my God. Yeah.
Awesome.
So awesome.