Robots Got a Base Model and a Ticker Symbol in the Same Week

Two things happened close together, and together they tell you where robotics is heading. A startup said it built a general model for physical AI, the way GPT is a general model for text. And a humanoid maker filed to go public with real customers and real revenue. One is a research bet. One is a business fact. Let me catch you up on both, and on why this matters even if your product never touches a robot.

One model, many bodies

The old way to build a robot was to gather huge piles of task-specific data for one machine in one room. Pim de Witte, CEO of General Intuition, wants to kill that. His pitch: train one base model that understands how things move through space and time, then let others build on it. "The generalization of the model itself is the product," he said.

The proof point is small but loud. The company trained on millions of hours of video game data, then fine-tuned on just eight minutes of real robot data to run a four-legged robot. It handled a room with people walking by and objects it had never seen, using only a front camera. That is a big drop from the hundreds of thousands of hours the old approach needed.

The money follows the thesis. General Intuition raised $320 million at a $2.3 billion valuation. If this holds, robotics starts to look like the app layer you already know: a shared base, then your fine-tuning on top.

The part that already ships

While one camp sells the future, Agility Robotics is selling the present. It plans to go public through a SPAC that values it around $2.5 billion and raises more than $620 million, the largest raise in humanoid robotics so far. CEO Peggy Johnson points to over $300 million in booked, multi-year revenue, roughly 1,000 robots rented monthly to names like Amazon, Toyota, and GXO Logistics.

The robot, Digit, is boring on purpose. It moves heavy totes in warehouses. It has bird legs so it can reach low shelves without knees hitting the racking. Johnson said the real edge is safety, which is expensive and slow. "You can't build your robot and then make it safe. That's a redesign," she said. She thinks a robot in your home is "10-plus years" out.

The demo problem

The headlines run ahead of the work, and it helps to name the gap. Surgeons at UC San Diego had two humanoid robots remove a gallbladder in a preclinical trial. Real result, real promise for remote or understaffed care. Also real caveats: the robot had to be recalibrated a few times mid-procedure, and latency is still a problem when a human controls it from far away.

Then there is the stage version. Hyundai and Boston Dynamics put their Atlas robot on the pitch at the World Cup, where it mimicked a player's celebration and handed the ball to the ref. Great marketing. Boston Dynamics said the training method behind the dance is close to how they teach industrial tasks, which is a fair point. It is still a controlled show, not a job.

When a vendor pitches you, sort the surgery from the halftime act. Ask what broke and how often, not what looked smooth on camera.

Why the video game bet matters to you

The more interesting idea sits under the General Intuition story: text models are weak at space and time. As de Witte's team argues in a TechCrunch Equity episode, video game data carries something the open internet does not. It records the action, which button a human pressed and when. That action data is what teaches a model to reason about movement.

That is a lesson about your own data, not just theirs. The value is in the sequence of choices your users make, not the static screenshots of the result. If you own a product, you likely have logs of what people did in what order. That is the kind of data these models are hungry for.

The deep cut

Both camps agree on one split, and it decides where you build. Johnson said it plainly: "The LLMs had the entire internet to train on. When you think about the physical AI of humanoids, that doesn't quite exist yet." There is a semantic layer, high-level instructions a model can already handle, and a physical layer, the balance and grip that takes a decade of real deployment to earn.

For your team that means the reasoning layer is becoming a commodity you rent. Agility is "LLM-agnostic," swapping in Claude and Gemini. The defensible part is the physical execution and the operating data behind it. Read your own roadmap the same way. If your edge is a smart layer that anyone can rent, it is not an edge. Your moat is the messy, real-world data and workflow only you have. Fund that.

Three questions for your team

• Where in our product are we sitting on action data, the sequence of what users did and when, that we treat as logs instead of an asset?
• If the reasoning layer becomes a rented commodity, what part of our work stays hard to copy, and are we spending on it?
• When a vendor demos, do we have a standard way to separate a real result with caveats from a staged show?