Ricursive Intelligence founders pitch AI-designed chips to ease the compute crisis
Anna Goldie and Azalia Mirhoseini, the founders of Ricursive Intelligence, took the stage at TechCrunch Disrupt on Wednesday to describe their startup's work. The company is developing AI that can design chips, learn from each iteration, and improve subsequent designs. The stated goal is to automate and accelerate the chip design process. The founders believe more efficient AI chip design could help solve the compute crisis.
Goldie framed the problem around demand. She said there is an understanding that over the next five years there will be at least 10,000x more demand for compute, and that this is "somewhat potentially conservative". Her conclusion was that the industry does not want to build 10,000 times more data centers. Instead, she argued, chips need to be more efficient: they should use the bandwidth in data centers more efficiently, and use wafers and fabs more efficiently. She added that even if data centers were built in space, "you won't see the sun anymore".
Goldie acknowledged that some people may be frightened by the idea of AI designing its own chips. Her answer is that letting AI handle chip design lets humans focus on developing applications and exploring new possibilities that expand what they can do with computing. In her words, humans bring the ideas and creative thoughts, and execution is handled by AI, a shift she said is also happening in software development and other fields.
The founders said AI is challenging traditional chip design methods by finding unconventional layouts that can outperform conventional ones. Their example was AlphaChip, the AI system they previously developed, which produced a circular, donut-shaped arrangement of chip components. It initially surprised engineers but ultimately proved successful.
They also explained that they optimise entire chips, rather than following the traditional approach of partitioning a chip into dozens of smaller blocks. This whole-chip approach lets the AI create intricate patterns, including wave-like designs, about which the pair said they plan to share more in the future.
Finally, the founders argued that AI-designed chips will make it possible to create all sorts of chips that might not have been possible before. Goldie's reasoning was economic: if a new chip takes six months, a year or two years to design, companies must be extremely selective about which chips they build. If a chip could be designed in about a week, it would become feasible to design custom chips for varied applications such as space, AR and VR, and biological computing, if it makes financial sense, and to be more energy efficient.
Key facts
- Ricursive Intelligence is developing AI that designs chips, learns from each iteration and improves later designs, with the goal of automating and accelerating chip design.
- Goldie cited an understanding that compute demand will grow at least 10,000x over five years, and said building 10,000 times more data centers is not the answer.
- The founders optimise entire chips instead of partitioning them into dozens of smaller blocks, and say this allows intricate, wave-like patterns they plan to show later.
- As an example of unconventional layouts, they pointed to a donut-shaped component arrangement produced by AlphaChip, their earlier AI system.
- Goldie's week-long chip design is a hypothetical: if design took about a week instead of six months to two years, custom chips for many applications would make financial sense.
Why it matters
The pitch ties chip design to the compute shortage from an unusual angle. Instead of building ever more data centers, the founders argue for making each chip, and the bandwidth, wafers and fabs around it, work more efficiently. Their second claim is about speed: if design time falls from six months or years to something like a week, small markets such as space, AR and VR, and biological computing could justify custom silicon.
Who it affects
Chip designers are the direct audience, since the proposal shifts execution of the design work to AI while humans supply ideas. Data center operators and anyone worried about compute supply are the indirect audience, because the argument is that efficiency gains in chips could reduce how much new capacity is needed.
How to use it
There is nothing to use yet. The talk described a startup's direction, not a released product. No product launch date, customer or manufacturing partner is named.
How solid is it
This is a founders' conference talk. The 10,000x demand figure is presented as an "understanding" voiced by Goldie, with no underlying study or source cited. The week-long design time is a hypothetical, not a reported result. No measured performance, power or area figures for any AI-designed chip are given, and no quantified outperformance of the donut-shaped AlphaChip layout is given. Only Goldie is directly quoted.
Risks and caveats
Details of the wave-like designs are not given; the founders said they plan to share more in the future. No funding amount, valuation, investors or headcount for Ricursive Intelligence is given. Goldie herself noted that some people may be scared by AI designing its own chips, and her response was that humans would keep setting the ideas while AI handles execution.
“So we don’t want to build 10,000 times more data centers.”
— Anna Goldie, Ricursive Intelligence, at TechCrunch Disrupt