The Last Scarce Asset

Section 01

Demand: the scaling run hasn’t happened yet

The state of the art in physical AI trains on a small fraction of frontier-LLM compute. OpenVLA, still a reference open vision-language-action model, used 64 GPUs for two weeks; Grok 3 pre-trained on more than 100,000. The gap between a frontier LLM run and a state-of-the-art robot foundation model is three to four orders of magnitude.

Two constraints keep these models small. The architectures are still maturing, and — the more binding limit — there isn’t enough data to train anything bigger. OpenVLA cycles through its entire dataset 27 times during training, where frontier LLMs see most of their data once. With small datasets, models stay small, and small models need little compute. By LLM standards, the field is still early.

The conditions for a scaling run are forming. DeepMind ships Gemini Robotics; NVIDIA has named robotics its next major market and presented a “Physical AI Data Factory” blueprint at GTC 2026; OpenAI is rebuilding its robotics program; robotics startups have raised $18.8B so far in 2026, more than in all of 2025. If architectures mature enough to absorb LLM-scale compute, balanced scaling implies data demand rises by roughly the same orders of magnitude: closing even half the current gap would require real-world data at 100–1,000× today’s entire public supply. Labs are already buying — XDOF launched with $70M and around 20 customers including frontier labs, General Intuition raised at a $2.3B valuation, and Tesla keeps more than 1,000 Optimus units in-house primarily to collect data.

Section 02

Supply: why physical data is scarce

Language models benefited from an unusual inheritance: thirty years of human activity already transcribed onto scrapeable servers. The indexed web holds roughly 500 trillion tokens, and it cost model builders nothing to accumulate. Nothing comparable exists for the physical world. The geometry of a street, or the way a door handle responds to force, was never uploaded anywhere; it has to be captured, sensor by sensor and place by place, at real marginal cost.

The largest humanoid dataset assembled so far — AgiBot World, collected by 100 robots in a purpose-built Shanghai facility — totals 2,976 hours. YouTube receives that much video every six minutes. Web video is also not a substitute: Meta’s V-JEPA 2 pre-trained on a million hours of it and still needed real robot data for control, because video lacks metric scale, 3D geometry, and action labels.

The chart also leaves out the datasets that built 2D vision — ImageNet’s 14 million images, LAION’s 5.85 billion image-text pairs — because they’re measured in images rather than hours of experience. Meanwhile, the corpora behind today’s 3D-vision foundation models (VGGT, MapAnything) amount to under roughly 1,000 hours-equivalent of real capture.

One empirical result matters for what kind of data to collect. The most careful scaling study in robot learning found that policy generalization follows a power law in the diversity of environments and objects, not in the number of demonstrations collected in any one place. The binding constraint is coverage — many places, many objects, many conditions, refreshed over time — more than raw hours. The sections above cover demand; the rest of the essay is about supply.

Section 03

Four ways to capture reality

If demand develops as argued above, the practical question is how to capture physical data at scale. Four supply models exist today, with very different cost curves.

The first three models share the same limitation: cost scales with coverage. Every new neighborhood is another vehicle trip, another robot, another operator-hour. But coverage — environmental diversity — is exactly what the scaling results reward. Only the fourth model avoids this tension.

Section 04

Why DePIN’s cost curve is different

Decentralized Physical Infrastructure Networks approach capture with an incentive 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 upside tied to the network; the company’s marginal cost for a new location drops to validation and storage. Capex doesn’t disappear — it moves to the edge, spread across thousands of participants who each hold a stake in what they’re building.

There is now a track record to evaluate. Messari counts over 650 DePIN projects and 8.8 million active devices, with a growing share of the sector generating real revenue. Two precedents are most relevant. Helium bootstrapped a global wireless network to roughly one million IoT hotspots — infrastructure that would have cost a telecom billions in capex — by paying contributors in tokens to run radios at home. Hivemapper is the closest precedent for mapping: its dashcam network went from 2% of the world’s roads in March 2023 to 20% within eighteen months, and to roughly a third today. The company reports coverage growth about five times faster than Google Street View achieved with corporate fleets, with over 100 repeat passes where Street View averages one every year or two. These figures are company-reported, but even discounted they suggest decentralized capture can outpace fleets on coverage and freshness.

Token incentives reach places no fleet is ever dispatched to.

Three properties make DePIN a good fit for physical-AI data. First, diversity by construction: supply appears wherever contributors live, which matches the environment-diversity pattern the scaling results reward. Second, steerable supply: token rewards act as a dial, so a demand signal from model training — more indoor retail, more rain, more night scans — becomes a reward boost, and the network re-targets within days. Third, freshness: contributors re-scan because earning is continuous, so the corpus becomes a time series of places rather than a one-pass archive.

The main caveat: DePIN’s known failure mode is rewarding volume over quality. The mitigations are structural — validation before rewards vest, demand-side revenue that consumes supply, reward curves tuned to what training actually uses. The networks that survived the 2022–24 shakeout are the ones that built this discipline, and the sector’s compression from 1,000× to 10–25× revenue multiples suggests the market now prices fundamentals.

Section 05

What OVER has built

OVER has been applying this model to 3D mapping since well before the current physical-AI cycle, focused on metric-scale 3D reconstructions of real places rather than street video. Through Map2Earn, anyone with a smartphone can scan a location; the capture is validated, reconstructed into photogrammetry-grade 3D, and registered to the map, and the mapper earns OVR tokens. OVRLand — ownership of 300 m² spatial domains with publishing rights and revenue share — gives contributors a lasting stake in the network rather than piecework wages. The result is, to our knowledge, the largest crowdsourced 3D map:

What matters about this corpus is less its size than its labels. Every capture is a 3D reconstruction with metric scale, camera poses, and spatial registration — the supervision signals that world models, VLA navigation stacks, and sim-to-real pipelines lack in web video (the V-JEPA 2 result above). The same asset serves three kinds of demand: training data for world and geospatial models; digital twins that seed simulation pipelines such as NVIDIA’s data factories; and a live localization service — VPS with centimeter-grade, TEE-encrypted pose estimation — already consumed by robots as an API.

The cost structure follows from the loop. Adding a location doesn’t require dispatching hardware; it requires adjusting a reward curve, because contributors already own the sensors. This is also why coverage is geographically scattered — Toronto, Bangkok, Charleroi, Nghĩa Trụ, Atlanta and Shenzhen all appear in a single month of scans — rather than following a deployment plan.

What licensing alone is worth

Suppose OVER only ever licenses this data. There are recent comparables for what markets pay when data is the bottleneck: Meta paid $14.3B for 49% of Scale AI (implying ~$29B); Surge, bootstrapped to $1.2B in revenue, has reportedly negotiated at $25–30B; Mercor went from $2B to $10B in eight months and is reportedly in talks at $20B. Those valuations are for companies that refine text scraped for free; physical-AI data is proprietary from the moment it’s captured. Licensing at anywhere near those comparables would be a solid business by itself. We think there’s additional upside, because of what happened to the other two inputs over the last year.

Section 06

Research capability is becoming rentable

Until about a year ago, this essay would have ended at licensing. Moving up the value chain required two things a data company couldn’t easily get: a frontier research team and a nine-figure compute budget. Compute went first: an H100-hour that cost over $8 at the 2023 peak rents for $1.50–2.50 today, and Stargate, Colossus, and Meta’s gigawatt campuses are pushing supply further. The research constraint is now easing too, and the change is measurable.

METR tracks the length of software and ML-research tasks (measured in expert-human time) that an AI can complete autonomously at 50% reliability — close to the actual work of training a foundation model. That horizon doubled roughly every seven months from 2019 to 2025; since January 2024, it has doubled about every 105 days.

In May 2026, METR added Claude Mythos Preview — the model class behind Claude Fable 5 — and reported a horizon at or beyond 16 hours, past the range the current benchmark can measure reliably. Concrete examples of what this looks like in practice: Anthropic reports Mythos 5 running a week of largely autonomous research and producing an ML model that outperformed a recent Science-published model at a hundredth of the size; DeepMind’s AlphaEvolve improved on a 56-year-old matrix-multiplication result and recovers around 0.7% of Google’s global compute; on METR’s RE-Bench, AI agents score four times human experts at two-hour research budgets. On the open-weights side, Moonshot’s Kimi K3 — 2.8 trillion parameters, third on GDPval behind Claude Fable 5 and GPT-5.6 — rents at $15 per million output tokens, with the open-closed gap around 3–4 months and capability prices falling roughly 10× per year.

The specialized capability needed to train a state-of-the-art model — the scarcest input of the last decade — increasingly has a market price. To be clear, the claim is about market structure, not about human researchers becoming obsolete: the engineering capability required to train a competitive physical-AI model is ceasing to be a differentiator, because every serious team works with frontier research agents and training recipes diffuse through open weights within months. When an input is available to everyone at a price, it stops being a moat. The practical barrier between data vendor and AI lab — people and capex — is much lower than it was.

Section 07 · Conclusion

Three tiers of value

Taken together, the argument prices OVER’s position in three tiers. The demand and supply case justifies the first two on its own; the fall in research and compute costs opens the third — and the third is where foundation-model economics live.

Of a physical-AI model’s three inputs — compute, research capability, and data — the first two can now be rented. Diverse, metric-scale data of real places cannot: it has to be accumulated over years, and that is the asset OVER holds.

OVER Research · info@ovr.ai

Section 08

What could break this