Three bets

Bet hard or get out. Eidon's multimodal dataset sat in an awkward middle, not bespoke enough to stand on uniqueness and not large enough to stand on size, and egocentric robotics data is drifting into the same limbo. Three bets have better odds. Scale: a lab consumes ~500,000 hours a month, huge for any single collector and puny next to TikTok or YouTube, whose owners are a feature launch away from doubling the industry's output. Specificity: maximize information per hour (touch, force, contact) instead of hours, though the closer data matches an embodiment the more the robotics company wants to own collection itself. Synthetic: real data as the seed of a GPU-generated corpus, which buyers can bring in-house even more easily than physical collection. The window to start, scale, and exit a new data company is closing if it hasn't already. For those thinking about getting in, I'd get out.