I only noticed the connection between their data engine and this community sale after seeing @axisrobotics’ official update. The official announcement states that the $AXIS community sale has already oversubscribed. The window remains open until September 28, 13:00 SGT. Allocations will be distributed proportionally based on accepted commitments, and any unused USDC will be automatically refunded. Nine months ago, they started from scratch building this Physical AI data infrastructure—and now they generate over 100,000 trajectories daily. Public statistics show more than 5.17 million verified trajectories covering 6,885 tasks, contributed by nearly 200,000 participants. The numbers themselves aren’t the point. What’s more compelling is how they’re using this data. Traditional robotics data collection is limited by hardware availability and centralized workflows, making scalability difficult. Axis uses browser-based simulation: global contributors generate robotic arm demonstrations, the system automatically validates and scores them, and approved trajectories are recorded on-chain. Their process doesn’t start by collecting massive amounts of data upfront. Instead, they identify what the current model lacks, then targetably collect only the needed data types, train, and reassess. They summarize this闭环 (closed loop) as: Model → Evaluate → Select → Collect → Train. One research update quantified this logic: the π0.5 policy trained on AXIS data improved success rates on LIBERO-Plus from 83.9% to 88.8%. The related paper has been accepted by the Physical World Models workshop at IROS 2026, evaluated across over 1,800 tasks and 1.5 million trajectories. It’s not the sheer volume of data that matters—it’s how it improves policy performance. The public sale terms are clear: $100M valuation, fixed price of $0.10 per token, selling 10 million tokens (1% of total supply), capped at $1M, paid in USDC on Base, with a minimum purchase of $100 and maximum of $100,000 per person. Vesting: 10% at TGE, followed by a 6-month cliff, then the remaining 90% linearly unlocked over another 6 months. The window is still open—oversubscription simply means later participants receive a proportional share, not that they’re locked out. The underlying engine is already running. I’m still using the contributor portal alongside @KaitoAI’s interface: https://t.co/eT4RE35jjM. The ultimate scale of this data flywheel will depend on whether future models and task coverage continue to align effectively.