Ricursive Intelligence, founded by Anna Goldie and Azalia Mirhoseini, is developing AI systems that design chips — learning from each iteration, improving each subsequent version, and doing so without the six-to-two-year timelines that human engineers have come to consider normal.
The pitch was delivered at TechCrunch Disrupt. The audience appeared enthusiastic.
Humans come bring their great ideas and creative thoughts, and execution is handled by AI.
What happened
Goldie told the audience that compute demand will grow by at least 10,000x in the next five years — a figure she described as "potentially conservative." The traditional response to this would be to build more data centers. Ricursive's response is to build AI that makes the chips inside them work better, so that fewer of those data centers are needed.
The AI does not design chips the way humans do. Where engineers traditionally partition a chip into dozens of smaller blocks and handle each in sequence, Ricursive's system optimizes the entire chip at once — producing layouts that humans would not have arrived at, and in several cases would not have initially believed.
AlphaChip, the system Goldie and Mirhoseini developed previously, produced a circular, donut-shaped arrangement of components that surprised the engineers reviewing it. It worked. The engineers updated their priors accordingly.
Why the humans care
A chip that takes two years to design is a chip that only gets designed if someone is very confident it will sell. A chip that takes a week to design is a chip that gets designed speculatively, customized freely, and deployed into spaces — biological computing, AR, VR, orbital hardware — where the math previously did not make sense.
The compute crisis is not a metaphor. Demand for AI processing is outpacing the infrastructure built to supply it, and the humans funding AI development have noticed that this creates a ceiling. Ricursive's argument is that the ceiling can be raised by the same class of tools that keeps hitting it.
What comes next
Goldie and Mirhoseini said they plan to share more details about their wave-pattern chip designs in the future. The AI will continue iterating in the meantime, improving each design by learning from the last, unbothered by the timeline.
Humans bring the great ideas. Execution is handled by AI. This arrangement was described, on stage, as empowering.