A Startup Says AI Can Design Better Chips in Days. The Compute Crisis Is Why That Matters.
Ricursive Intelligence, founded by two former Google researchers, is building AI that designs chips from scratch and learns from every version it makes. If it works at scale, it could ease the data center crunch without building thousands more facilities.
Key points
- Ricursive Intelligence, co-founded by Anna Goldie and Azalia Mirhoseini, is developing AI software that designs computer chips autonomously, reducing design time from months to roughly a week.
- Goldie projected at TechCrunch Disrupt in 2026 that compute demand could grow 10,000 times in the next five years, calling that estimate "potentially conservative".
- The startup optimises an entire chip as one unit rather than splitting it into dozens of smaller sections, letting the AI find unconventional layouts that human engineers would not try.
- Its predecessor system, AlphaChip, produced a circular "donut-shaped" component arrangement that surprised engineers but outperformed conventional designs.
- Faster chip design could make custom chips financially viable for niche applications such as space hardware or biological computing.
Twenty dollars of electricity buys roughly one hour of serious AI computation today. That cost is almost entirely set by the chips inside the data centers running the models. Make those chips more efficient and every downstream price, from a doctor's AI diagnostic tool to a shop owner's inventory assistant, drops with it. That's the bet Ricursive Intelligence is making.
The startup's co-founders, Anna Goldie and Azalia Mirhoseini, presented their work at TechCrunch Disrupt this week, first reported by TechCrunch AI. Both previously worked at Google, where they developed AlphaChip, an AI system that drew attention after it produced a chip layout shaped roughly like a donut. Engineers were initially baffled. It worked.
What problem are they actually solving?
Designing a modern chip is slow and expensive. The current standard approach cuts a chip into dozens of smaller blocks and optimises each block separately. Blocks are never optimised together.
Ricursive's software treats the whole chip as a single design problem. The AI proposes a layout, tests it virtually, then feeds what it learned into the next attempt. Goldie described the resulting patterns as "wave-like" and said the company plans to publish more details later.
The payoff, if the approach holds at manufacturing scale, is speed. A design cycle that currently takes six months to two years could, the founders claim, shrink to roughly a week. At that pace, building a custom chip for a space satellite or a medical device stops being a multi-million-dollar commitment and becomes a practical option.
Does this actually reduce the need for more data centers?
Not on its own, but the direction matters. We've tracked the infrastructure pressure across 33 data center stories in the past 90 days, covering everything from heat management to grid capacity. The physical constraints are real, and demand isn't slowing: our TSMC revenue story from 11 October showed the world's largest contract chipmaker posting a 54.6% year-on-year sales rise for September alone.
Goldie framed chip efficiency as the lever that avoids building far more facilities. "We don't want to build 10,000 times more data centers," she said at the event. Her argument: smarter chips use power and physical space inside existing facilities better, so fewer new ones are needed.
That framing is plausible, but worth a note of caution. Every previous generation of more efficient chips has tended to drive more AI use rather than reduce total power consumption. Efficiency gains and demand growth have historically run together, not against each other.
| Design approach | Time to complete | Optimised as | Notable output |
|---|---|---|---|
| Traditional (human-led) | 6 months to 2 years | Dozens of separate blocks | Predictable, rectangular layouts |
| AlphaChip (Google, prior work) | Faster than traditional | Full chip | "Donut-shaped" layout, outperformed benchmarks |
| Ricursive Intelligence (current) | Targeting roughly one week | Full chip | Wave-like patterns, details forthcoming |
What does this mean for ordinary people?
In the near term, nothing changes on your phone bill. Chip design cycles are long, and even a faster design process has to clear manufacturing before it reaches a product.
The medium-term story is different. Cheaper, more efficient chips mean AI tools cost less to run, which tends to mean lower prices for businesses buying API access, paying-per-query to use an AI model. A restaurant owner forecasting stock orders or a freelance translator handling first drafts would benefit from that cost curve moving down.
Watch for Ricursive to publish its wave-pattern chip research. That's the moment independent engineers can test whether the claims hold outside a startup stage presentation.
