AlphaChip's Creators Want AI to Design AI Chips: Ricursive Intelligence Bets Compute Demand Grows 10,000-Fold

The people who taught machines to floorplan Google's TPUs now want to hand them the whole chip. Anna Goldie and Azalia Mirhoseini, the researchers behind AlphaChip, took the stage at TechCrunch Disrupt on Wednesday to lay out the thesis behind their startup Ricursive Intelligence: as AI demand explodes, the bottleneck is not fabs or power alone, but the slow, human-bounded process of designing silicon — and AI can compress it.

Goldie framed the scale of the problem in one number. "There is this understanding that in the next five years we are going to have at least 10,000x more demand for compute, and that's somewhat potentially conservative," she said. "So we don't want to build 10,000 times more data centers. We think that we need to help by making chips that are more efficient, that use the bandwidth in those data centers more efficiently, that use the wafers and fabs more efficiently, because even if we make [data centers] in space or something, I don't know, you won't see the sun anymore."

The founders' credentials come from AlphaChip, the reinforcement-learning system they previously developed that placed chip components on a silicon die. One of its outputs became an industry legend: a circular, donut-shaped arrangement that surprised the engineers who reviewed it but ultimately proved successful. The pair argue such counterintuitive layouts are exactly where machine search beats human intuition — and where their new company is aiming.

The technical approach differs from the traditional flow in a fundamental way. Conventional chip design partitions a chip into dozens of smaller blocks that are optimized separately and then stitched together. Ricursive Intelligence says it optimizes the entire chip as one problem, letting its systems produce intricate patterns — including wave-like designs the founders say they plan to reveal more about later — while learning from each design iteration to improve the next.

The economic argument is about cycle time. "If it takes six months or a year or two years to design a new chip, then you really have to be extremely selective about the chips that you design," Goldie said. "But if you can design a chip in like a week, and then say for manufacturing, you can design custom chips for all sorts of different types of applications in like space and AR, VR, and biological computing, it makes sense financially to do that, and we can be more energy efficient." Week-long design cycles would let specialized silicon serve markets that today cannot justify a custom part.

For those uneasy about AI designing its own hardware, Goldie offers the same framing now standard in software: "I think the way everything is shifting is that humans bring their great ideas and creative thoughts, and execution is handled by AI. And this is happening in chip design, software development, and other places."

Investors have already bought the thesis at scale: Ricursive Intelligence raised $335 million at a $4 billion valuation earlier this year, only four months after being founded, as TechCrunch reported in February. What the company has not yet published is independent benchmark data showing its AI-designed chips outperforming conventionally designed ones — the efficiency claims remain the company's own. But if the AlphaChip track record is any guide, the donut-shaped future of chip design may arrive sooner than the next fab comes online.