The Data Bottleneck
Physical AI is bottlenecked not by models or hardware, but by data — and the gap is orders of magnitude wide.
Physical AI Has No Internet
Large language models were built on decades of internet content — trillions of tokens, freely available, spanning virtually every domain of human knowledge. With that foundation in place, the remaining challenge was largely computational: scale compute, improve reasoning, iterate.
Physical AI has no such foundation. The physical world has no search engine index; there is no "internet of robotic interactions" to train on. The total volume of available robot training data is less than one-millionth the scale of the text corpora behind modern LLMs. This is not a long-tail problem — the field lacks foundational data at every level.
How Far From the GPT-1 Moment?
Language models arrived in three steps, and each one says something about where robotics stands.
GPT-1 proved a direction — Transformer plus pre-training works. It was validation, not scale.
GPT-2 trained on roughly 10 billion tokens: about 500,000 hours of reading at a human pace. Emergent behavior appeared for the first time.
GPT-3 trained on roughly 300 billion tokens, about 15 million hours. Without that volume there is no ChatGPT and none of what followed. At that stage the lever was not curation — it was volume.
Robotics has not had its GPT-1 moment yet, and what remains unsettled is the data itself:
- Which pre-training data distribution is the right one
- What the sim / real / ego ratio should be
- How action resolution should be modeled
- How long the temporal horizon needs to be
Models are converging and algorithms are maturing. The structure of the data has no consensus behind it.
Carry the language comparison across and the target is stark. If 15 million hours of text produced GPT-3, a genuinely general-purpose Physical AI foundation model plausibly needs on the order of 100 million hours of human manipulation data — our estimate, and one that stays daunting even with an order of magnitude of slack. Volume comes first here, exactly as it did for language.
Three bars have to be cleared at once, and today's best datasets fall orders of magnitude short of each:
| Dimension | What generalization requires | Where the field stands today |
|---|---|---|
| Data Volume | 100M+ hours of manipulation data (order-of-magnitude estimate) | \~2,000 hrs — the industry leader, Open X-Embodiment |
| Scene Diversity | 10,000+ distinct physical environments | single digits — traditional lab collection |
| Atomic Skills | 1,000+ robotic manipulation skills | \~700 — a current benchmark, RT-1 |
The distance between these columns is the distance between narrow, brittle demos and a general-purpose robotic brain — and it is why demand for structured manipulation data should compound over the next two to three years rather than level off.
Why the Gap Persists
Compute, hardware, and algorithms are ready. What Physical AI starves for is data — and three structural forces keep it scarce.
Algorithms starved for data. Model architectures and robot hardware are mature enough to learn general manipulation. What is missing is the massive, everyday interaction data needed to trigger a GPT-1-style step change — the moment scale unlocks emergent generalization.
Trapped in physical collection. Collecting robot data in the real world is hardware-bound: specialized rigs, safety systems, and one trained operator per robot, in the same room, in real time. This caps both volume and coverage and makes foundation-model scale structurally impossible.
Failing at generalizable diversity. Real-world generalization demands compounding diversity across objects, scenes, embodiments, and viewpoints that hand-collected datasets can never reach. Simulation and ego-centric capture are the only paths that scale to it.
Blind to their own failures. Even with more demonstrations, a policy trained only on expert trajectories sees only the states an expert visited — so the situations where it is weakest are precisely the ones least represented in its training data (the covariate-shift problem, where errors compound over a task). Closing this requires on-policy corrective data: humans supervising the policy's own execution and correcting it exactly where it fails. That correction data has traditionally required one human continuously supervising one robot — making it a burning, unmet need across academia and industry.
Every one of these barriers points the same way: the answer is not a better model, but a fundamentally different way to produce data — one that generates diversity on demand, collects it without hardware, and turns each policy's failures into the next batch of data.
Next: The Compounding Data Engine →