Ricursive Intelligence wants AI to take on more of the work of designing computer chips. Founders Anna Goldie and Azalia Mirhoseini say the system can learn from each design iteration and improve what it produces, with the aim of speeding up and automating chip design.
Their pitch is that better chip efficiency could help address growing demand for computing without relying only on building more data centers. The approach also challenges familiar design methods, as AI may find layouts that engineers would not have chosen.
Efficiency as an answer to rising compute demand
At TechCrunch Disrupt on Wednesday, Goldie said demand for compute could grow at least 10,000x in the next five years, a projection she described as potentially conservative. She argued that meeting that demand should involve using computing infrastructure more efficiently, rather than simply multiplying the number of data centers.
That efficiency, in the founders’ view, must extend across several parts of the system. Goldie pointed to chips that use bandwidth in data centers more effectively, as well as more efficient use of wafers and fabs. The idea is to get more computing value from the resources involved in making and operating chips.
This framing makes chip design part of a wider infrastructure challenge. If demand rises as Goldie expects, improving how chips are designed could help address pressure on data centers and manufacturing capacity. The company’s proposed contribution is to automate design work so that more efficient options can be explored.
Letting AI explore unfamiliar layouts
The founders say AI can challenge traditional chip design by finding arrangements that differ from conventional layouts and may perform better. AlphaChip, an AI system they previously developed, produced a circular, “donut-shaped” arrangement of chip components. Engineers were initially surprised, but the design ultimately proved successful.
Ricursive’s approach, they explained, is to optimize an entire chip instead of splitting it into dozens of smaller blocks. Working at that broader level can allow the AI to produce more intricate patterns, including wave-like designs. The founders said they plan to share more about those patterns in the future.
The example suggests why AI-designed layouts may look unfamiliar: the system can search across possibilities that depart from established design habits. An unconventional appearance alone does not establish that a design is useful, but the founders point to AlphaChip’s circular layout as an example that surprised engineers and succeeded.
From selective projects to custom chips
Design time affects which chips are worth pursuing. Goldie said that when a new chip takes six months, a year or two years to design, teams have to be highly selective about what they build. If the design process could take about a week, she argued, custom chips for a wider range of applications could make financial sense.
Goldie named space, AR, VR and biological computing as possible areas for such chips. The argument is that a faster process could make it practical to tailor chip designs to different needs, while also pursuing greater energy efficiency.
These are possibilities the founders described, rather than outcomes the article establishes. Their case depends on AI shortening design work enough to make custom chips practical and on those designs delivering useful efficiency. If that happens, applications that would not justify a long design effort could become candidates for specialized hardware.
Changing the division of work
Goldie sees AI handling more execution while people contribute ideas and creative thinking. She said this shift is happening in chip design, software development and other areas, and described Ricursive as applying that model to chip design.
For Ricursive, automating design is meant to give people room to work on applications and explore what computing might make possible. The founders’ broader claim is that quicker iteration and less conventional layouts could help make chips more efficient and more varied, offering one way to respond to the compute crisis they describe.