Fractal Semiconductor

Teaching AI how chips are made

Fractal Semiconductor builds AI models that understand the physics of chip manufacturing. Our goal is AI that designs and runs the experiments it takes to get new chips, new materials, and new factories into production.

The world is building fabs faster than it can run them

Every AI model, robot, and car runs on chips, and demand is outrunning supply. The most advanced fabs are booked years ahead, some before they’re even built. Countries are racing to build fabs at home. In the US, incentives have pulled in more than $300 billion in private investment.

But a finished fab isn’t a working fab. It takes experts to get one running, and there aren’t enough. The US could be short up to 157,000 semiconductor workers by 2030, and only about 3% of engineering grads go into chips.

Every new chip is a science experiment

A chip passes through hundreds of process steps. Each step has dozens of settings. Temperature, pressure, gas flow, timing. Nobody knows the right combination in advance, and a small change early on can ruin the result many steps later. Engineers find working recipes by running costly experiments on test wafers, often for months.

This work is called process discovery. Every new chip, every new material, and every new fab has to go through it.

What we’re building

We’re building models that learn the physics of chip manufacturing. Our goal is AI that proposes the next experiment, reads the result, and updates its understanding with every run. We want these models to find working processes in far fewer experiments than a human team needs, and over time, to keep those processes running on their own.

Hardware moves only as fast as we can learn to make it. We want to unbottleneck hardware.

Who we are

We have run process discovery for neuromorphic chips and worked inside chip fabs, and in both places we saw the same thing firsthand: months of experiments to get one step right, and almost none of it carried over to the next chip. We started Fractal Semiconductor to change that.

If you run a fab or develop processes, tell us what slows you down. If this problem excites you, join us.