Every machine we are excited about this decade runs on a cathode. The robot arm. The delivery drone. The backup bank behind a data center full of GPUs. The truck, the bus, the grid.
Strip away the software and what is left is a few kilograms of carefully arranged atoms. How well those atoms are arranged decides how far the machine goes, how fast it charges, how long it lives, and what it costs. Physical AI is a materials problem. It just doesn’t look like one from the demo video.
IWe found materials by hand
The way the world makes a new cathode has not changed much since the nineties. A chemist has a hunch. A student loads a furnace. Six weeks later there is a coin cell, a plot, a conversation, and another hunch. Repeat for a career.
It works. It is also how you get one good material a decade. We have spent ten years inside that loop. We love it, and it is too slow for the century we are walking into. The machines are ready. The chemistry is still on foot.
IIA lab that keeps score
Here is what we noticed. Every run through a furnace is a sentence in a language nobody was writing down. What went in. What we did to it. What came out. Composition, temperature, time, how much carbon, the shape of the particles under the microscope, the voltage curve at two in the morning when nobody is watching.
So we started writing it down. All of it. We measure seven things about every batch and we keep the failures, because a failure with its conditions attached is worth more than a success nobody can repeat. We call the record the atlas. It is the most valuable thing we own, and it gets more valuable every week we run the furnace.
IIIThe model and the lab are one machine
Give a model that atlas and something changes. It reads a new powder it has never seen and proposes a recipe. The furnace runs it. The cells report back. The model updates. Round and round, faster than any of us can, and it never forgets a batch.
We built a lab that thinks. The furnace is its hands. The atlas is its memory. We are its first students.
This is the part that keeps us up. A model that has run ten thousand regenerations knows things about iron phosphate that no human chemist has had time to learn, and it can be asked. What happens if we push the carbon. What this feedstock wants. Why the last batch came back better than new. The answers arrive as recipes, and recipes can be run.
IVWhy we started with the humblest cathode
Lithium iron phosphate is the workhorse of the age. It is in the buses, the grid, the data centers, the cheap cars that will put most of the world on electric power. It is also the one nobody could bring back. Dissolve it and there is nothing worth selling.
But it has a secret. It doesn’t wear out. It runs out of lithium. The crystal survives. So we put the lithium back, and the first time we did it properly the material came back better than it left the factory. We checked. We checked again. That was the moment this stopped being a research program.
If a model can learn to do that for every batch of spent cathode on the planet, it can learn to design the next one.
VThe future we are building toward
We think a materials engineer in 2035 will work the way a software engineer works today. She will describe what she needs. This much power. This many cycles. This cost, this footprint, this duty cycle at this temperature. A model will propose. A lab will run. A powder will arrive. Weeks, not decades.
When that is true, the machines of physical AI stop waiting for chemistry to catch up. Robots get the cells they were designed around. Data centers get backup that was tuned to their load trace. The grid gets storage that improves with every generation instead of every decade.
That is the company we are building. It starts with a furnace, an orange powder, and a model that is getting better at its job every single day.
VIWhat we promise
We will regenerate before we design, because that is where the data is. We will publish when it is proven and not before. We will keep the failures. We will build the boring parts well. And when we say better than new, we will mean it in a cell, not on a slide.