Recursive Intelligence founders Anna Goldie and Azalia Mirhoseini contend that more efficient AI‑driven chip design could alleviate the looming compute crisis. Speaking at TechCrunch Disrupt on Wednesday, they outlined how their startup is building AI that designs chips, learns from each iteration, and refines subsequent designs to automate and speed up the chip‑creation process.
“We anticipate at least a 10,000‑fold increase in compute demand over the next five years, and that estimate may even be conservative,” Goldie remarked on stage. “Building ten thousand additional data centers isn’t feasible. Instead, we need chips that are far more efficient—making better use of data‑center bandwidth, wafer utilization, and fab capacity—because even if we moved data centers into orbit, the environmental toll would be unsustainable.”
Although the notion of AI designing its own chips may unsettle some, Goldie argues that entrusting chip design to AI frees humans to concentrate on application development and exploring novel capabilities that expand what computing can achieve.
“I envision a future where humans supply the creative vision and AI handles execution,” Goldie said. “This shift is already evident in chip design, software engineering, and other domains, and it’s precisely what Recursive Intelligence is pursuing for chip design.”
The founders note that AI is overturning conventional chip‑design methodologies by uncovering unconventional layouts that can surpass traditional approaches. For instance, AlphaChip—the AI system they previously developed—produced a circular, “donut‑shaped” arrangement of components that initially surprised engineers but ultimately proved highly effective.
They explained that they are optimizing chips holistically rather than partitioning them into dozens of smaller blocks. This strategy enables AI to generate intricate patterns, such as wave‑like designs, which they intend to elaborate on in future disclosures.
The founders also contend that AI‑crafted chips will unlock the possibility of creating chip types that were previously unattainable.
“If designing a new chip takes six months, a year, or even two years, you must be extremely selective about which chips to pursue,” Goldie said. “But if we can produce a chip design in a week, we could afford to create custom chips for a wide range of applications—space systems, AR/VR, biological computing, and more—while remaining financially viable and energy‑efficient.”
