The Last Scarce Asset

The coming demand shock

Physical AI's state of the art trains on a rounding error of frontier-LLM compute. OpenVLA — still a reference open vision-language-action model — used 64 GPUs for two weeks. Grok 3 pre-trained on 100,000+. The gap between a frontier LLM run and a SOTA robot foundation model is three to four orders of magnitude.^10–14

This is not frugality. Two ceilings force it: the architectures haven't had their Transformer moment — and, more binding, there isn't enough data to feed a bigger model. OpenVLA recycles its entire dataset 27 times during training; frontier LLMs see most of their data once.¹³ Small data forces small models; small models need small compute. The entire industry is, by LLM standards, warming up.

Now run the tape forward. Every ingredient of the scaling moment is being staged: DeepMind ships Gemini Robotics, NVIDIA declared robotics its next major market and launched a "Physical AI Data Factory" blueprint at GTC 2026, OpenAI is rebuilding its robotics program, and robotics startups have raised $18.8B in 2026 so far — more than in all of 2025.²⁶ When architectures mature enough to absorb LLM-scale compute, balanced scaling drags data demand up by the same orders of magnitude. Closing even half the gap implies demand for real-world data at 100–1,000× today's entire public supply. The labs are already buying: XDOF launched with $70M and ~20 customers including frontier labs; General Intuition raised at $2.3B; Tesla keeps 1,000+ Optimus units on its own floor doing nothing but collecting data.^30,31

There is no Common Crawl for reality

Language models had a miracle underneath them: humanity spent thirty years transcribing itself onto scrapeable servers. The indexed web holds ~500 trillion tokens — a frontier corpus is the equivalent of ~86,000 years of continuous reading, and it was free.^15,16 The physical world granted no such favor. Nobody pre-uploaded the geometry of your street or the friction of a door handle. Physical data is not scraped; it is captured — sensor by sensor, place by place, at real marginal cost.

The largest humanoid dataset ever assembled — AgiBot World, 100 robots in a purpose-built Shanghai facility — totals 2,976 hours. YouTube ingests that every six minutes.^19,21 And the shortcut doesn't work: Meta's V-JEPA 2 pre-trained on a million hours of web video and still required real robot data for control, because video lacks metric scale, 3D geometry, and action labels — the exact dimensions that make physical data physical.¹⁷

And note what is missing from this chart: the celebrated datasets that built 2D vision — ImageNet's 14 million images, LAION's 5.85 billion image-text pairs — are measured in images, not experience, and have no place on an hours axis. Meanwhile the corpora behind today's 3D-vision foundation models (VGGT, MapAnything) total ≲1,000 hours-equivalent of real capture — the entire spatial-AI field trains on less footage than a single television network broadcasts in six weeks.³⁷

One more empirical result sharpens the requirement. The best scaling study in robot learning found that policy generalization follows a power law in the diversity of environments and objects — not the number of demonstrations collected in any one place.²³ The binding constraint is not hours. It is coverage of reality: many places, many objects, many conditions, refreshed over time. Act I established the demand. The rest of this essay is about who can supply it — and what the supplier should do with the position.

The capture problem

Grant Act I and the question stops being philosophical and becomes industrial: how, exactly, do you build the physical world's Common Crawl? There are only four known supply models, and they have very different cost curves.

The first three models share a property that should alarm anyone underwriting them: their cost scales with coverage. Every new neighborhood is another vehicle-trip, another robot, another operator-hour. But coverage — environmental diversity — is precisely the variable the scaling laws reward. The economics and the science point in opposite directions. Except in one model.

DePIN: the only supply curve that bends

Decentralized Physical Infrastructure Networks solve capture with a financial instrument instead of a fleet: a token that converts the network's future demand into present supply. Contributors buy their own sensors, capture where they live, and are paid in equity-like upside; the company's marginal cost of a new location collapses to validation and storage. Capex doesn't disappear — it is distributed to the edge, borne by thousands of participants who each hold a stake in the network they are building.

This is no longer a theory. Messari counts 650+ DePIN projects and 8.8 million active devices, a sector that has crossed from token speculation into revenue fundamentals.²⁷ Two precedents matter most here:

Helium bootstrapped a global wireless network to roughly one million IoT hotspots — infrastructure a telecom would have needed billions in capex to deploy — by paying contributors in tokens to plug radios into their own windowsills.²⁷

Hivemapper is the direct proof for mapping: its dashcam network went from 2% of the world's roads in March 2023 to 20% in eighteen months to roughly one-third today — coverage growth the company benchmarks at five times the pace Google Street View managed with corporate fleets, and with 100+ repeat passes where Google averages one every two years.^28,29 A decentralized network out-mapped the best-capitalized mapping company in history on speed, coverage, and freshness, simultaneously.

Three properties make DePIN uniquely fitted to the physical-AI data problem, mapping one-to-one onto Act I's requirements: diversity by construction — supply emerges wherever contributors live, and since generalization scales with environment diversity, a DePIN network's growth pattern is the scaling law's demand pattern;²³ elastic, steerable supply — token rewards are a dial, so a demand signal from model training ("more indoor retail, more rain, more night scans") becomes a rewards boost and the network re-aims itself within days; and freshness as a native property — contributors re-scan because earning is continuous, turning the corpus into a living time-series of the world rather than a one-pass archive.

The honest caveat: DePIN's known failure mode is incentive misalignment — rewarding volume over quality. The mitigation is architectural: validation before rewards vest, demand-side revenue burning supply, reward curves tuned to what training actually consumes. The networks that survived the 2022–24 shakeout are the ones that built this discipline, and the sector's multiple compression (1,000× → 10–25× revenue) shows the market now prices fundamentals.²⁷

OVER: the asset — and the floor

OVER has been running this playbook since before "physical AI" had a name — aimed at the hardest, highest-value layer: not road video, but metric-scale 3D reconstructions of real places. Through Map2Earn, anyone with a smartphone scans a location; the capture is validated, reconstructed into photogrammetry-grade 3D, and registered to the map; the mapper earns OVR tokens. OVRLand — ownership of 300 m² spatial domains with publishing rights and revenue share — gives contributors a durable stake in the network's success, not just piecework wages. The result is the world's largest crowdsourced 3D map:³²

What makes this corpus AI-grade rather than merely large is that it carries the labels web video lacks. Every capture is a 3D reconstruction with metric scale, camera poses, and spatial registration — the exact supervision signals world models, VLA navigation stacks, and sim-to-real pipelines starve for (recall V-JEPA 2's million web-video hours failing to substitute for grounded data). The same asset serves three demand curves at once: training data for world and geospatial models, digital twins that seed simulation pipelines like NVIDIA's data factories, and a live localization service — VPS with centimeter-grade, TEE-encrypted pose estimation — that robots consume as an API, today.³²

Note the cost structure the flywheel implies. When a fleet company wants location #275,001, it dispatches hardware. When OVER wants it, it adjusts a reward curve — the capex was already bought, by someone who owns a piece of the outcome. That is why OVER's coverage map looks like the world (Toronto, Bangkok, Charleroi, Nghĩa Trụ, Atlanta, Shenzhen — this month's scans alone) rather than like a fleet-deployment plan.³²

Pricing the floor

Suppose OVER only ever licenses this asset. We already know what markets pay when a platform's bottleneck is data — even when the raw material is free: Meta paid $14.3B for 49% of Scale AI (~$29B); Surge, bootstrapped to $1.2B of revenue, has negotiated at $25–30B; Mercor repriced from $2B to $10B in eight months and is reportedly in talks at $20B.^33–35 Those are the multiples for refining scraped text. Physical AI's raw material is proprietary from the first byte. Licensing at these comparables is a real business — and it is the floor, not the thesis. What raises the ceiling is what happened to the other two inputs over the last twelve months.

The twist: intelligence just became infrastructure

Until about a year ago, this essay would have ended at licensing. A data vendor's ceiling was a data vendor's multiple, because moving up the value chain required the two things no data company could get: a frontier research team and a nine-figure compute budget. The compute wall fell first — an H100-hour that cost $8+ at the 2023 peak rents for $1.50–2.50 today, while Stargate, Colossus and Meta's gigawatt campuses race to make FLOPs abundant.^24,25 The research wall is falling now, and it is measurable.

METR tracks how long a software or ML-research task (in expert-human time) an AI completes autonomously at 50% reliability — precisely the substance of training a foundation model. That horizon doubled every seven months from 2019 to 2025. Since January 2024 it has doubled every ~105 days.^2,3

Read the right edge. In May 2026, METR added Claude Mythos Preview — the model class behind Claude Fable 5 — and reported it at or beyond 16 hours, past the point where the benchmark can measure at all. The model outgrew the ruler.³ The receipts are concrete: Anthropic reports Mythos 5 running a week of largely autonomous research, designing and training an ML model that outperformed a recent Science-published model at 1/100th the size;⁴ DeepMind's AlphaEvolve broke a 56-year-old matrix-multiplication record and recovers ~0.7% of Google's global compute;⁵ on METR's RE-Bench, AI agents score 4× human experts at two-hour research budgets.⁶ And the day before this essay was finished, Moonshot released Kimi K3 — 2.8 trillion parameters, open weights, third on GDPval behind only Claude Fable 5 and GPT-5.6 — at $15 per million output tokens, with the open-closed gap at ~3–4 months and capability prices falling ~10× per year.^7–9

To be precise: this is not a claim that human researchers are obsolete. It is a narrower claim about market structure — the engineering intelligence required to train a SOTA physical-AI model is ceasing to be a differentiator, because every serious team works with frontier research agents and the recipes diffuse through open weights within months. When everyone has the same intelligence on tap, it prices like electricity: essential, and margin-free. Which means the wall between "data vendor" and "AI lab" — a wall made of people and capex — is gone.

The value ladder

Put the three acts together and OVER's position prices in tiers. Acts I and II justify the first two rungs on their own. Act III unlocks the third — and the third is where foundation-model economics live.

What could break this