UpTrajectory Review

Russell Brandom's TechCrunch piece is less a story than a diagnosis: the world-model race — the push to build AI systems that can simulate physical reality well enough to train robots, predict outcomes, and generate video — has become one of the most secretive corners of an industry already notorious for opacity. The players are flush with capital and press attention, yet neither founders nor the data suppliers feeding them will say what they're actually constructing. For a reader who hasn't followed this niche, world models are the infrastructure bet behind much of the current robotics and generative-video boom: instead of scraping text, these companies ingest massive amounts of real-world footage and sensor data to teach machines how the physical world behaves. That makes the data itself — who supplies it, on what terms, at what price — as strategically important as the models.

For a small-business operator, this secrecy matters more than it might appear. If you run a company whose work generates visual or spatial data — a logistics firm with warehouse camera footage, a construction company with drone surveys, a retailer with in-store video, a medical or dental practice with imaging archives — you are sitting on an asset that well-funded AI companies may soon want, and the terms of those deals are being set right now, in private, without any public benchmark for what the data is worth. Brandom's observation that even the data suppliers won't talk suggests nondisclosure agreements and competitive pressure are suppressing price discovery. That is a market condition, and it has a direct implication: the first sellers in a silent market almost always underprice.

What is genuinely notable here is the inversion of the usual startup narrative. Companies at this stage typically broadcast aggressively to recruit talent, attract capital, and intimidate rivals. Instead, the world-model cohort has gone quiet in a way that suggests the moat is not the architecture but the data pipeline and the exclusive supplier relationships. We are skeptical of any framing that treats this as mere corporate discretion; when an entire category coordinates on silence, it usually signals either antitrust-adjacent caution, a land grab for proprietary datasets before regulation catches up, or both. Brandom is right to flag it as a story in itself, because journalism cannot evaluate claims — about capabilities, about safety, about market power — that companies refuse to make on the record.

The downstream effects cut unevenly. Large data holders — studios, sports leagues, industrial sensor networks — can negotiate exclusivity premiums and lock competitors out. Small operators, by contrast, face take-it-or-leave-it offers with NDAs attached, and may not realize that selling footage non-exclusively today could undercut a future revenue stream or expose them to liability if the data is used in ways they didn't anticipate. There is also a labor and creative angle: if world models are trained on footage of real workplaces, the workers in that footage are producing training data without compensation or consent frameworks. The cost of all this secrecy is ultimately borne by the least powerful parties in the transaction.

What to watch: whether any supplier breaks ranks and discloses deal terms, which would force repricing across the market; whether regulators, particularly in Europe where data-protection rules bite harder, start asking what these companies actually hold; and whether any world-model firm publishes a credible capability benchmark rather than a demo reel. In the meantime, if your business generates visual or sensor data, treat it as a balance-sheet asset. Get legal advice before signing any data-licensing agreement, demand transparency about downstream use, and compare notes with peers — because the companies buying your data are coordinating, and you should be too.

“good luck getting anyone — from the founders to their own data suppliers — to tell you what they're actually building” — TechCrunch

Takeaway: If your business generates visual or sensor data, treat it as a strategic asset and get legal advice before signing any NDA-bound licensing deal with an AI company.