Language models were built on a resource that turned out to be finite: the written internet. The physical world has no equivalent reservoir, and the reason why is the most under-discussed constraint in AI right now.
The last decade of progress in AI rested on a quiet piece of luck. Humanity had spent thirty years writing things down on the internet, for its own reasons, and that accumulated text turned out to be exactly the substrate a transformer needed. Nobody collected it deliberately. It was already there, and it was free.
That reservoir is close to exhausted. Work published at ICML in 2024 estimated the stock of public human text at roughly three hundred trillion tokens and projected that models would be trained on datasets equal to the entire available stock somewhere between 2026 and 2032. As a result, frontier labs are now paying for data. One publisher deal reportedly runs above $250 million over five years; others sit in the single-digit millions annually. Priced data is the tell that free data ran out.
We should be precise about the claim, because the same organization that produced the exhaustion estimate published a follow-up arguing that power, not data, is the binding constraint on scaling to 2030. Both can be true. The claim we are making is narrower: the supply of public human text is finite and nearly all of it has been used, not that AI runs out of data in general.
Now consider the physical world.
There is no crawl for reality
Ask what the equivalent corpus looks like for a system that has to act in the world rather than write about it, and the answer is uncomfortable. There is no crawl.
Well, parts of the physical world havebeen crawled. For example, Google Street View has photographed most of the world's roads on a repeating cycle for twenty years, which is a genuine crawl of the physical world and a genuinely useful one. But what it produced is a record of what places look like, not of what happens in them. A panorama of an intersection gives you the geometry, the sightlines, the signage and the surface. It does not tell you that four hundred people have nearly been hit there, or what the drivers did in the second before each one.
Nobody spent twenty years incidentally recording how vehicles behave at wet roundabouts at dusk, what a cyclist does when a car door opens, or how a delivery van reacts when a child steps between parked cars.
Some of it exists in fragments: dashcam clips uploaded to social platforms, crash databases assembled after the fact, insurance claims files written by adjusters. All of these are downstream of the event and stripped of everything that made it teachable. A claims file tells you a collision occurred. It does not tell you what the driver saw four seconds earlier, what they did, whether it worked, or what the road looked like at the moment they made the decision.
There is a further problem with the fragments, and it is the one that matters if you intend to compute anything. A clip uploaded to a social platform has been re-encoded, cropped, and stripped of a lot of very useful information: no calibration, no ego motion, no synchronized clock, no depth, no reliable timestamp. You can watch it. You generally cannot register it against anything else, which means you generally cannot measure with it. The distinction between footage you can watch and a record you can compute on is not a detail. It is the difference between an archive and an instrument.
The written internet was a by-product of people communicating. There is no equivalent by-product of people driving, because until recently nobody was recording what the driver saw and what they did about it.
The appearance of the world can be photographed on a schedule you control. The behavior of the world cannot. That distinction is the whole of this essay.
Why you cannot simply generate more
The obvious answer is synthesis. If the data does not exist: create it, render it! This works better than skeptics expect for common situations and worse than optimists hope for the situations that matter.
A simulator can only produce what its designers specified. It renders the scenarios somebody thought to write down in the combinations somebody thought to combine. That is genuinely useful for coverage: you can run a known scenario ten thousand times with varied parameters and find the edges of a policy. Parts of the autonomous driving community call this “fuzzing”.
But the events that determine whether a system is safe are, almost by definition, the ones nobody anticipated.
The clearest example on record is a fatality. A federal investigation into a pedestrian killed by a development vehicle found that the system detected her 5.6 seconds before impact and reclassified her repeatedly, as vehicle, then bicycle, then other, without ever settling on pedestrian. The finding, verbatim:
The system never classified her as a pedestrian, or correctly predicted her path, because she was crossing at a location without a crosswalk, and the system design did not include consideration for jaywalking pedestrians.
The remedy proves the point by inversion. Afterwards, the software was changed so that a pedestrian detected outside a crosswalk could be assigned the goal of crossing mid-block. Jaywalking became a possible pedestrian goal. It had not been one.
And the detail that should stop anyone who builds scenario libraries: that same investigation recorded that the company's human safety drivers were trained on closed courses using scenarios explicitly designed to teach them to watch for jaywalking pedestrians and encountered jaywalkers routinely in testing.
The humans knew. The scenario library did not.
Two more, both from recall filings. (1) One operator's own defect description reads as a scenario specification that did not exist until it had to: a collision-detection subsystem that could pull a vehicle over after a collision, in circumstances involving a pedestrian positioned low on the ground in the path of the vehicle. Nobody had written that down. (2) Another recalled software after a vehicle mispredicted the motion of a backwards-facing pickup being improperly towed at an angle across two lanes, and a second vehicle from the same fleet struck the same truck minutes later, because the whole fleet shares one specification.
There is a deeper version of the problem, and it now has a literature. A 2024 Nature paper showed that models trained recursively on their own output degrade in a way the authors called irreversible. A follow-up paper at ICLR the following year found that as little as one percent synthetic contamination was enough to break the benefit of a larger dataset, which means this is not a dosage problem you can dilute your way out of.
The honest counterweight: another group showed that accumulating synthetic data alongside the original real data avoids model collapse. That is a real result and it does not rescue the position it appears to. Accumulation still requires the original real data to exist and to keep growing. That is the supply claim, restated.
The most authoritative summary we can point to is the 2026 AI Index, which reports global AI compute capacity growing 3.3× per year and states plainly that synthetic data is still not replacing real data in pre-training.
Compute is growing at 3.3× a year. The real-world data that has to be fed into it is not growing at anything like that rate, and no amount of compute produces more of it.
The supply curve is set by physics
Here is the constraint that makes this different from every other data problem in AI.
Rare road events arrive at the rate the world produces them, and the numbers are more brutal than intuition suggests.
The Strategic Highway Research Program's SHRP 2 Naturalistic Driving Study, then the largest study of its kind, followed more than 3,100 volunteer drivers in instrumented vehicles. Researchers ran trigger algorithms across 5,512,900 trip files, validating 1,549 crashes and 2,705 near-crashes. A peer-reviewed analysis drew on 905 crashes involving injury or property damage, describing them as the basis for the first large-scale direct analysis of causal factors using crashes alone.
The 100-Car Naturalistic Driving Study, run by the Virginia Tech Transportation Institute for NHTSA, is starker still. A hundred instrumented vehicles, roughly two million miles, 42,300 hours of continuous driving data. It produced fifteen police-reported crashes.
Set that against the population rate. US federal figures for 2023 give about 6.14 million police-reported crashes against 3,247 billion vehicle miles. This translates to roughly one per 529,000 miles, and a fatality once per 79 million. At an average annual mileage, the typical driver is about forty years from their next reported crash.
You cannot pay to make crashes happen sooner. More sensors give more simultaneous observation, but the tail fills at the rate reality supplies it. For scale: Waymo reports 220.6 million rider-only miles through March 2026, the largest fully autonomous record anyone publishes. American drivers cover that distance about every thirty-six minutes.
The same constraint binds everyone, us included. Rare events show up in proportion to exposure, and nobody has enough of it yet.
This means capital behaves differently here than it does almost anywhere else in technology. A well-funded competitor can buy sensors, hire engineers and stand up a network in a year. What they cannot buy is the decade of elapsed exposure during which the rare things happened to somebody who was recording.
It also means the resource does not deplete the way the text corpus did. The written internet was a fixed stock, and it ran out. Road experience is a flow. It arrives only as fast as the world produces new combinations, new infrastructure, new vehicle types and new failure modes, which is slow. It also never stops arriving. An input you cannot accelerate and cannot exhaust behaves differently from anything else in AI, and on a long enough horizon that difference favors whoever started first.
The archive does not lose value, because what happened remains what happened. What decays is a model's assumption that the world still works that way. The record is history and it stays true. A system's understanding of the present is perishable, which is the argument of a different essay.
What actually counts as supply
Volume is the wrong unit. A record containing ten billion miles of uneventful motorway driving is very large and carries very little information. What makes a mile valuable is whether something happened, whether a human responded to it, and whether we know how it ended.
The event. What actually occurred, in context, including road conditions.
The response. What the person noticed, attempted, or failed to notice. This is the part that separates a benign scene from a critical one, and it is absent from almost every dataset in this field.
The outcome. What happened next. Loss, near miss, intervention, cost. Without it, an event is a picture rather than a lesson.
A record with all three is experience. A record with only the first is footage, and the automotive industry has a great deal of footage.
The uncomfortable implication
We would say that, of course. We have been recording driving data for over ten years and the argument flatters us.
If this is right, then the bottleneck on physical AI, at least in the domain of driving, is not model architecture nor compute. Architectures keep improving and compute is plentiful. The bottleneck is the supply of real data, with outcomes attached, covering the events that matter most, and those events arrive on the world's schedule rather than ours.
That is a claim about driving specifically. It holds here because you cannot manufacture the data by driving badly on purpose: a crash, or another driver's reaction to a near miss, is not something anyone can ethically stage at scale. It does not follow automatically for every domain of physical AI, and we are not going to claim that it does.
Which suggests the organizations that will matter in this field are not necessarily the ones with the largest models. They are the ones that started measuring earliest and never stopped.
There is a second-order effect worth calling out. As architectures converge, and our recent experience at CVPR suggests they are indeed converging, the model stops being the differentiator. Several teams building autonomy stacks today would describe their approach in nearly the same words, and the published gaps between them are narrowing. When the method becomes common property, the remaining variable is what you trained it on. That is an uncomfortable position for a field that has spent over a decade competing on cleverness, and it is why the supply question is about to get much less academic.
What would change our mind
A demonstration that synthetic or simulated experience substitutes for recorded experience on the tail specifically; not on average performance, where it plainly helps, but on the rare, outcome-attached events. The AI Index sentence above is the strongest evidence we have that it does not yet. If that sentence changes in a future edition, our position should change with it.
What we are least sure about
The claim that outcome-attached physical experience cannot be manufactured. We have found no source that establishes it. It is an argument from the absence of a counterexample, and we should say so rather than dress it as a finding. What is documented is that nobody has done it. That is not the same as nobody being able to.