Physical AI: What It Means for Founders

Physical AI is AI that perceives and acts on the physical world instead of only producing text, images, or code: models that read camera and sensor data and output motor actions, running inside robots, vehicles, drones, and machines on a warehouse floor. The term covers the model layer, the simulation layer, and the data infrastructure underneath, not just the metal.

The reason it stopped being a research word and became a market word is narrower than the hype suggests. Robotics spent a decade as the field where nothing generalized: you tuned one machine for one task in one room, and the next machine started from zero. That is now changing, because robotics is deliberately running the language-model playbook. Jim Fan, who co-leads Nvidia's GEAR lab (Generalist Embodied Agent Research), laid out that plan at Sequoia's AI Ascent and gave it a name: the great parallel. His framing of the substitution is the whole idea in one line. "So, instead of simulating strings can we simulate next physical world state?"

For a founder, the useful question is not whether humanoids arrive on schedule. It is which parts of this are now cheap enough to build a company on. This piece covers what the term actually points at, why the bottleneck moved from models to data, what changes when software can act on atoms, and the vertical playbook a small team can run today without owning a fab or a research lab.

The great parallel: robotics is copying the LLM recipe

Fan's structure is worth repeating because it tells you where the value is accruing. Language models got here in three steps: pretraining to learn the shape of the domain, supervised fine-tuning to point that capability at useful work, and reinforcement learning to push past imitation. Robotics is now running the same three steps against the physical world. Pretrain a model that can predict the next state of the world, fine-tune it on actions so it does the narrow slice that matters for a real robot, and let reinforcement learning carry the last mile.

The interesting part is what the pretraining substrate turned out to be. For the last few years the dominant approach was vision-language-action models, which took a language model, bolted an action head on top, and hoped the world knowledge transferred. It partly did. Fan's critique is that most of the parameters still went to language, so the models were good at nouns and knowledge and weak on physics and verbs. As he put it, "It's kind of head-heavy in the wrong places."

The replacement substrate is video. Video generation models, trained on nothing but predicting the next frame, turn out to internalize physics as a side effect. Fan's summary is direct: "Physics emerge by predicting the next blob of pixels at scale." Gravity, buoyancy, reflection, refraction, none of it coded in. Fan shows one of them solving mazes by rolling the simulation forward in pixel space.

That gets you to a policy model that predicts the next world state and the next action jointly. Fan calls Nvidia's version Dreamer, and the diagnostic he reports is the part founders should notice, because it means the system's failures are legible: "If the video prediction works, the action works. If the video hallucinates, the action fails." A robot whose internal prediction you can watch is a robot you can debug. That is a meaningful change from a black-box controller that simply drops the object.

The bottleneck is data, and the data strategy just changed

There is no internet of robot data, and that single fact determines what a startup in physical AI can and cannot do.

Language models had an enormous head start: a decade of human text was already sitting on the internet, free. There is no equivalent corpus of physical action. Encord's co-founders, Eric Landau and Ulrik Stig Hansen, put the asymmetry cleanly on YC's Root Access while describing the data layer they are building for this market: "Now we have all the compute infrastructure, but you need the data to actually get to the scaling law." The compute overhang is real and the data is missing, which is the exact inverse of the situation that produced ChatGPT.

So the field has spent three years generating that data by teleoperation: a human in a VR rig driving the robot while the system records. Fan's objection to it is arithmetic, not aesthetics. For teleoperation, he says, "it's upper bounded by 24 hours per robot per day, the fundamental physical limit," and in practice he puts the real number at closer to 3 hours, because the robots break. You cannot scale a training corpus against a ceiling like that.

The escape route is to stop recording robots and start recording humans. Two versions of it are live:

• Wearables. Instead of driving the robot, you wear the robot's gripper on your own hand and collect the data directly, with the rest of the robot out of the loop. The universal manipulation interface started this line, and the Nvidia exoskeleton Fan calls Dex-UMI maps one-to-one onto a five-finger robot hand. Fan showed a policy trained on this data running fully autonomously with zero teleoperation data behind it.
• Egocentric video. The bigger prize is data collection that disappears. Fan's analogy is driver-assist systems, where the data upload is ambient and you never feel it. "The data collection needs to get out of the way, fade into the background," he said, so it can capture human dexterity across every kind of work. Nvidia's result here is the striking number: pretrained on "21K hours of in-the-wild egocentric human data with zero robot data whatsoever," then fine-tuned with "only 50 hours of high precision mocap data gloves and 4 hours of teleop." Four hours of robot teleoperation, less than 0.1 percent of the training mix, driving a 22-degree-of-freedom hand through tasks like sorting cards and handling a syringe.

Fan reports that this produced "this neural scaling law for dexterity," a clean log-linear relationship between pretraining hours and validation loss, six years after the original scaling law for language. That matters commercially, because a scaling law is a budgeting tool. It tells you what another order of magnitude of data buys, which is the difference between a research bet and a plan.

His forecast follows from the curve: teleoperation drops to a negligible share within a year or two, wearables cover the hundreds of thousands of hours, and egocentric video becomes the main diet at roughly 10 million hours. The second half of the strategy is environments for reinforcement learning, which is why he compresses the whole thing into an equation: "compute now equals environment now equals data."

If the shape of that argument feels familiar, it should. It is the same conclusion as data being the real bottleneck rather than compute, with the difficulty turned up: robotics is the canonical data-poor domain, hard to start in and correspondingly hard to copy once you are in.

What changes when software can act on atoms

Fan sets three remaining milestones, and each one is a different kind of market.

The physical Turing test. His definition: "Across a wide range of activities, you cannot tell the difference between a human doing a task or a robot doing it." He puts it maybe 2 to 3 years out. Treat that as one researcher's estimate rather than a date to plan around, but note what it implies about sequencing. Narrow, unglamorous tasks clear that bar long before general ones do.

The physical API. This is the one that should interest software founders most, because it is the point where robotics becomes addressable by people who do not build robots. A fleet of machines, in Fan's description, where "they can be configured just like any other software using APIs and command lines." Once physical capacity has an API, everything above it is a software business: scheduling, orchestration, exception handling, observability, billing. None of that requires a hardware team.

Physical auto research, where robots design, improve, and build their next iteration. That one presumes the first two already shipped, so it is a reason to watch the field rather than a plan to fund.

Fan's timeline argument is that it took 14 years to get from AlexNet in 2012 to agentic auto research in 2026, and that "technology does not advance linearly. It advances exponentially," so he puts the end of the tree around 2040 with what he calls 95 percent certainty. Treat that as a researcher's confidence interval rather than a number to underwrite payroll against, and take the ordering instead: the API layer arrives well before the humanoid does. It is the same timing discipline in Demis Hassabis's founder playbook, which frames the useful window as being years ahead of the market rather than decades ahead of it.

The market size argument is more grounded than the timeline. As one of Encord's co-founders put it, "for as much investment that's going into physical AI, it still is actually small because 80% of the world's economy is just moving things around in the physical world or doing things in the physical world." Physical Intelligence co-founder Quan Vuong ran the same napkin math on YC's Lightcone from the other end: take US GDP, assume solving robotics contributes something like 10 percent of it, and the investment case writes itself. Both are estimates rather than measurements. What they establish is that the addressable surface is not a niche, which is the opposite of how software founders have been trained to think about robots.

Where a small team can still win in physical AI

Here is the practical answer for a founder entering physical AI, and it is not to go train a robot foundation model.

Vuong, whose company builds exactly that, spent a large part of his Lightcone appearance arguing that the vertical layer above it is now open. His framing of the change: "The equation, I think, for starting a robotic business has changed and will continue to change at an accelerating pace because the upfront cost is not that high anymore." Robotics used to demand full vertical integration, your own hardware, autonomy stack, safety certification, customer relationships, everything, which is why the barrier to entry stayed absurd. A shared foundation model removes the hardest and most capital-intensive slice of that.

His sequence, which he gave in full on the podcast, is the closest thing to a playbook this field has: "identify where the robot can fit in, focus on cheaper hardware, collect data, run evaluation, mix autonomy, break even, scale robots." Worth unpacking the parts founders get wrong:

1. Start from the workflow, not the robot. The system has to fit an existing process. The question is where in a workflow that needs a given amount of work today a machine changes the economics most. That is customer research, not robotics research.
2. Buy cheap hardware. Modern policies are reactive enough to compensate for imprecise machines, so precision hardware is no longer the price of entry. This is the single largest cost line that just moved.
3. Take mixed autonomy seriously as a product, not a fallback. A person takes over when the robot fails. Vuong's point is that plenty of real jobs tolerate robot mistakes, and those are the ones that are deployable now.
4. Get to break even per robot before you scale. He is emphatic about the ordering, and it is where a generation of robotics companies died: if the payback period does not work per unit, adding units multiplies the loss.
5. Invest in evaluation early. "Evaluation is really hard problem in robotic because it scales super linearly to model capability," Vuong said. A model that does a 20-minute task is far more than 10 times harder to evaluate than one that does a 2-minute task. Budget for it before it becomes the thing blocking every release.

There is a second business here that requires no robots at all. Vuong describes being surprised, on starting the company, that the supporting infrastructure simply did not exist: data collection tooling, annotation, data management, evaluation harnesses, operational process. Every robotics company is rebuilding all of it. That is the classic picks-and-shovels opening, and it is the one Encord is chasing with a $60 million Series C led by Wellington Management, bringing them to $110 million raised and more than 300 AI teams as customers.

One more constraint worth internalizing before you pick a wedge. Vuong notes that "in robotics, it's hard enough to get your model to work on one particular robot platform," and that in his own company's fleet, no two robot platforms are the same. Hardware heterogeneity is the tax you pay in this business the way model churn is the tax in software. Plan for it in the architecture rather than discovering it at your fifth deployment.

If you want the adjacent version of this opportunity where the AI does not yet touch the machine, AI for field service and AI for supply chain are the same physical economy attacked through documents and workflows instead of actuators. Both are shorter paths to revenue, and both build the domain knowledge a robotics wedge later requires. The broader map of where AI changes a startup's unit economics is in the AI for startups pillar.

The trust layer: physical AI has far less error tolerance

I write this as a CISSP, and this is where I would slow a founder down.

Software teams have absorbed a specific tolerance for model error: a bad answer gets a thumbs down and the user retries. Encord's co-founders drew the contrast plainly on Root Access, noting that in digital systems a bad output breaks nothing, but "the error tolerance for physical AI systems are actually much lower." A hallucinating language model produces a wrong paragraph. A hallucinating physical system moves mass. That difference propagates into every part of the company.

Four things it changes:

• Data quality stops being a quality metric and becomes a safety control. If the training data is subtly wrong, the failure lands in a room with people in it. The bar for what you accept into a training set is higher than anything you are used to from software.
• Evaluation is a release gate, not a dashboard. Combine Vuong's point about superlinear evaluation cost with a physical failure mode and you get a clear conclusion: you cannot ship on vibes here, and you should assume the evaluation harness is a permanent line item rather than a phase.
• The autonomy boundary has to be explicit and visible. Mixed autonomy works because a human is genuinely in the loop. Define what the robot is allowed to do unsupervised, log every intervention, and treat rising autonomy as something earned per task, not toggled globally.
• The data you accumulate is sensitive in a new way. Egocentric video of people working is a recording of workplaces, faces, documents, and screens. If your data strategy is built on it, consent, retention, and redaction are architecture decisions, not policy documents you write before an enterprise security review.

The reassuring part is that the debuggability improved along with the capability. A world model you can watch predict, per Fan's point that a hallucinated video means a failed action, is a far better safety substrate than a controller that gives no warning before it fails.

What to do this week

• Decide honestly which layer you are entering: foundation model, vertical application, or infrastructure. Only one of the three is realistic for a small team without a research bench, and it is not the first.
• If you are considering a vertical wedge, go watch the workflow in person and time it. The step where a person is doing something repetitive, physically constrained, and expensive is the whole thesis, and you cannot find it from a deck.
• Run the break-even math per robot before you write any code: hardware, deployment, and the human supervision the mixed autonomy phase requires. If the unit does not clear, the wedge is wrong.
• If you are already an AI company, ask whether any of the missing infrastructure Vuong describes (data collection, annotation, evaluation, remote operations) is adjacent to what you already do. That is a software business selling into a hardware market.
• Whatever layer you pick, write down your evaluation plan and your autonomy boundary in the first month. Both get much more expensive to retrofit here than in software.

Working out which layer of a shift you can actually own, what the model is allowed to decide alone, and how to build the trust layer around it, is the same job whichever domain you are in. That operating picture is what we teach in the AI Operating System for Startups.

Sources

• Robotics' End Game: Nvidia's Jim Fan (Sequoia Capital, AI Ascent), the talk this article distills, on the great parallel, world action models, the shift from teleoperation to egocentric video, the neural scaling law for dexterity, and the physical Turing test, physical API, and physical auto research milestones. Fan's role and background: Nvidia's GEAR lab.
• The GPT Moment for Robotics Is Here (Y Combinator's Lightcone) with Physical Intelligence co-founder Quan Vuong, on the changed equation for starting a robotics business, the vertical playbook from workflow to break even to scale, superlinear evaluation cost, hardware heterogeneity, and the missing infrastructure layer. Company background: Physical Intelligence and TechCrunch.
• The Data Layer for the Robot Economy (YC Root Access) with Encord co-founders Eric Landau and Ulrik Stig Hansen, on the compute-versus-data inversion in physical AI, the lower error tolerance of physical systems, and the size of the physical economy. Funding and company details confirmed via Encord's Series C announcement and SiliconANGLE.