Compute demand is heading somewhere absurd.
Ricursive Intelligence is developing AI that can design chips, learn from each design iteration, and improve the next one. The startup was founded by Anna Goldie and Azalia Mirhoseini, who are betting that chip design cannot keep moving at the pace of a calendar measured in years.
Goldie says there is an understanding that demand for compute will grow at least 10,000x in the next five years, a figure she described as potentially conservative. That forecast creates a problem beyond buying more hardware: the chips themselves need to use bandwidth, wafers, and fabs more effectively.
AI changes the chip design trade-off
Traditional chip design forces teams to choose their projects with care because each new design can take six months, a year, or two years. “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.
Ricursive’s approach is to let AI handle more of the design and execution process. Goldie said humans bring “their great ideas and creative thoughts,” while AI handles execution—a shift she also connects to software development and other fields.
The important change is not just speed. AI can search for layouts that challenge traditional chip design methods, including arrangements engineers might not consider through established approaches. Those unconventional layouts can outperform traditional ones, opening the door to chips that might not have been possible before.
AlphaChip, an AI system developed by Ricursive, produced a circular, “donut-shaped” arrangement of chip components. It is an unusually visible example of the system’s ability to explore designs outside familiar patterns—because apparently even chip components are tired of being arranged in tidy little grids.
Custom chips for a wider range of uses
Ricursive argues that faster design could make custom chips practical for more applications. Goldie said that if a chip can be designed in like a week and then manufactured, companies could create custom hardware for space, AR, VR, and biological computing when the economics make sense.
That timeline would change the number of ideas worth pursuing. A design process measured in weeks allows teams to consider chips for specific applications instead of limiting development to projects large enough to justify years of work.
Goldie also linked the approach to energy efficiency. Custom chips built for particular applications can support more efficient computing, while AI-designed hardware can help data centers get more value from their bandwidth, wafers, and fabs.
Ricursive’s AI can optimize entire chips instead of splitting them into smaller blocks and optimizing each piece separately. That broader view matters because a chip’s performance depends on how its components work together, not only on how well each individual block performs.
The company is developing AI-designed chips against a demand curve that calls for far more computing power and a design process that currently moves too slowly. If its systems can learn across iterations, discover better layouts, and reduce the time needed to create custom hardware, chip design becomes less about selecting a few safe projects and more about testing many targeted ones.
The promise is clear, even if the engineering remains formidable: a chip designed in like a week could serve space, AR, VR, or biological computing without requiring every application to wait years for hardware. The compute crisis will not be solved by optimism alone, but Ricursive is trying to make the silicon itself learn faster.
