The researchers behind AlphaChip are building a company on the claim that AI should design entire chips, not assist humans block by block. Ricursive Intelligence founders Anna Goldie and Azalia Mirhoseini told a Disrupt audience on Wednesday that their system designs chips, learns from each iteration and improves subsequent designs, optimising the whole chip rather than following the traditional approach of partitioning it into dozens of smaller blocks.
Their demand framing is deliberately extreme. Goldie put expected growth in compute demand at at least 10,000 times over the next five years and argued the answer cannot be 10,000 times more data centers: chips must use bandwidth, wafers and fabs more efficiently. The pair pointed to AlphaChip’s earlier unconventional layouts, including a circular arrangement engineers initially doubted, as evidence that AI search finds designs humans would not propose.
The economic claim is about selectivity. A design cycle of six months to two years forces companies to be extremely selective about which chips to build; compress design toward a week, they argue, and custom chips for space, augmented and virtual reality, and biological computing become financially sensible, with better energy efficiency as the byproduct.
Treat this as a thesis statement, not a result. No product performance, customer or tape out was disclosed on stage. The verification path is concrete, though: a Ricursive designed chip, fabricated and benchmarked against a conventionally designed equivalent. Until that exists, the company belongs in the same category as every AI for science claim this cycle, promising direction, unproven magnitude.