Robot brain builders push past their GPT-2 moment

Unitree's stock chart tells the story physical AI investors keep avoiding. China's leading robot maker went public and hit a $66 billion valuation, then lost nearly half of it this week, TechCrunch reported. Analysts point to one issue: the robots keep getting more physically capable, but they still don't know how to do work that creates real value.

That tension showed up everywhere at last week's Actuate conference, a gathering for developers building AI "brains" for robots. The event has tripled in size since it launched in 2023, pulling in 1,500 attendees, according to organizer Foxglove, a company that helps physical AI model builders manage and visualize their data. A booth sign for Avala, another infrastructure player, put the risk in plain language: it promised to solve "the robotics data crisis," the shortage of high-quality training data that keeps generalized, any-task robots out of reach.

Harry Mellsop, founder of the simulation-tools startup Antioch, frames the moment bluntly: physical AI is in its "GPT-2 era," referencing the OpenAI model that came before ChatGPT. Getting past that point takes more data and more compute, especially GPUs optimized for ray tracing, the kind used to build high-fidelity simulations.

Autonomous vehicles sit furthest ahead of the pack, partly because cars driven by humans generate usable data automatically, and partly because avoiding contact is a simpler problem than manipulating the physical world. Much of the tooling driving robot model-building traces back to AV companies; Foxglove itself was founded by former employees of Cruise, GM's shuttered self-driving unit. Car companies are now betting that same tooling lets them compete with dedicated humanoid makers. Tesla is already running that play with Optimus, and both Wayve and Uber have opened robotics labs aimed at humanoid form factors.

"I think you need to start in vehicles. Manipulation robotics is like self-driving five years ago," said Alex Kendall, CEO of Wayve. "The data infrastructure, the simulation, ML ops infrastructure will probably be shared, but the specific world model for the simulator will be a different post-training. There's going to be a lot more commonality than not, but then there's going to need to be some differences for different embodiments." Kendall argues it's too early to commit to one hardware platform, since sensors and other components keep advancing and a general model needs to stay hardware-agnostic.

Théophile Gervet, CEO of the vertically integrated humanoid startup Genesis AI, which raised a $105 million seed round this year, disagrees. "We're too early in this wave for a brain strategy to work," he told TechCrunch. "There's lots of opportunities to co-design hardware and AI." Gervet points to companies already getting robots into the field by staying narrow: Gritt building solar farms, Agility deploying robots in industrial settings, Bedrock running excavators autonomously, while general-purpose humanoids stay stuck in the lab. "No customer cares about the general purpose robot that works at 80% success rate," he said. "But if you're building for a narrow vertical on top of GPT-2, you're going to get crushed by the company building on GPT-4."

• Unitree's IPO valuation: $66 billion, before losing nearly half its value
• Actuate conference attendance: 1,500, up from a third of that size in 2023
• Genesis AI's 2026 seed round: $105 million
• Wayve's target for consumer autonomy: under $1,000 of hardware for eyes-off driving

Bedrock CTO Kevin Peterson said the company started with excavation to understand "manipulation in the wild," with plans to build an intelligence layer spanning multiple construction machines. Foxglove announced a new product this week too, built on Nvidia's open-weight Cosmos world model, letting engineers search dense visual and lidar data with natural-language queries to speed up debugging.

Sam Altman recently said physical AI's "ChatGPT moment" sits a few years out. Kendall isn't convinced investors are the right audience to watch for it: the largest robot deployment on Earth remains consumer vacuum bots, and a real breakthrough would need to excite regular people, not investors who are already sold. Gervet describes what that would look like: "manipulation that just works out of the box, where you can talk to a robot in natural language and have it do any basic task, pushing, pulling, closing a laptop, cleaning up a table, and it works... let's say 80% plus out of the box. That's roughly your ChatGPT experience."

Intokened covered a related bet on the humanoid side of this race when XPeng raised $900 million at a $6.3 billion valuation for its IRON robot, the same investor enthusiasm now colliding with Unitree's stock chart. Whether the breakthrough looks like Kendall's cheap car autonomy, Gervet's reliable manipulation, or something neither of them has named yet, the conference floor made one thing clear: nobody in the room thinks the brain problem is solved.

Nothing here should be taken as financial advice — just information to consider.