Ricursive Intelligence is developing AI that can design chips, learn from each attempt, and apply those lessons to later designs. Co-founders Anna Goldie and Azalia Mirhoseini described the approach onstage at TechCrunch Disrupt on October 14, presenting faster and more efficient chip development as one response to rising demand for computing power.

Goldie said the industry could face at least a 10,000-fold increase in demand for compute over the next five years, a projection she characterized as potentially conservative. Rather than trying to meet that increase only by building vastly more data centers, she argued that the industry should also make chips more efficient in their use of bandwidth, semiconductor wafers, and fabrication capacity. TechCrunch reported the forecast as Goldie’s view and did not provide an independent basis for the figure.

Ricursive’s premise is a feedback loop: AI helps produce a chip design, learns from what happened, and uses that experience to improve the next one. The company wants to automate and accelerate a process that remains difficult and time-consuming, while allowing human designers to concentrate on new applications and higher-level ideas. Goldie framed that division of labor as part of a broader shift in which people supply creative direction and AI increasingly handles execution.

Abstract chip components shifting from a grid into a circular layout.
The founders say AI can discover workable chip arrangements that human engineers may not naturally propose.

The founders pointed to AlphaChip, an AI system they previously developed, as evidence that machine-generated layouts can depart from conventional engineering instincts. They said AlphaChip produced a circular, doughnut-like arrangement of components that surprised engineers but ultimately worked. The example illustrates the possibility that an optimization system may discover successful structures a human team would not naturally propose.

Ricursive says it is optimizing complete chips instead of relying on the traditional practice of separating a design into dozens of smaller blocks. Goldie and Mirhoseini said the whole-chip approach can produce complicated, unconventional patterns, including wave-like layouts. They indicated that more information about those designs would be shared later, so the TechCrunch report does not provide technical results or comparative performance data for them.

Faster design could also change which chips are financially worth making. Goldie argued that when development requires six months, a year, or two years, companies must be highly selective about the hardware they pursue. If a design could instead be completed in roughly a week before manufacturing, she said, specialized chips might become viable for a wider set of uses, including space systems, augmented and virtual reality, and biological computing.

Specialized chips branching toward space, mixed-reality, and biological applications.
Shorter development cycles could make purpose-built chips practical for a broader range of applications.

That promise remains different from a demonstrated production cycle. The report does not state that Ricursive has delivered a manufacturing-ready chip in a week, quantify energy savings, or name customers using its designs. It also does not detail how the company validates a layout, accounts for manufacturing constraints, or measures improvement from one iteration to the next. Those missing benchmarks will matter because an unconventional pattern is useful only if it can be verified, fabricated, and operated reliably.

The company’s argument is nevertheless consequential for the AI industry. More capable models require hardware, while new hardware can take far longer to develop than software. Ricursive is trying to narrow that mismatch by applying AI to the physical infrastructure on which AI itself depends. If its systems can reliably shorten design cycles and improve resource use, the effect would reach beyond a single accelerator by making more specialized computing architectures practical.

For now, TechCrunch’s report establishes the founders’ strategy and the design ideas they presented, not proof that the approach has solved the compute shortage. Ricursive must still show that whole-chip optimization produces repeatable gains under real fabrication and operating constraints. Its bet is that the route to more computing power is not simply more data centers, but a faster learning process for designing the silicon inside them.