A startup wants AI to design computer chips, not just run on them
Ricursive Intelligence says AI that designs its own chips could ease a looming compute crunch — but so far it's a pitch, not a shipped product.
What happened: Anna Goldie and Azalia Mirhoseini, the researchers behind Google's AlphaChip project, introduced their new startup, Ricursive Intelligence, at TechCrunch Disrupt. The company trains AI models to design computer chips, test them, learn from the results, and refine the next version — automating a process that normally takes human engineers months or years. Goldie framed the effort as a response to a coming 'compute crisis': demand for AI computing power could rise roughly 10,000-fold over the next five years, an estimate she called possibly conservative. Instead of meeting that demand mainly by building more data centers, the founders want AI to make chips and facilities do more with the same resources.
Why it matters: If AI can squeeze more performance out of existing chips, wafers, and fabrication plants, it could soften the physical and energy toll of the AI buildout — fewer new data centers, less strain on power grids and chip factories. Goldie joked that building enough data centers to match raw demand projections would block out the sun. The bigger pitch is flexibility: if a custom chip can be designed in about a week instead of one to two years, it becomes financially realistic to build specialized chips for niche uses — spacecraft, AR/VR headsets, biological computing — that weren't worth custom silicon before.
How it works, plainly: Ricursive optimizes an entire chip layout as one unit, instead of the usual approach of splitting a chip into dozens of smaller blocks designed separately. That lets the AI try layouts human engineers wouldn't — the founders point to AlphaChip's earlier Google project, which produced a circular, 'donut-shaped' component arrangement that looked wrong at first glance but outperformed conventional designs. The model learns from each design it produces and feeds the results into the next attempt. The founders say they're also exploring wave-like layouts and plan to share more details later.
What's still unproven: The Disrupt appearance was an introduction, not a product launch. There's no mention of named customers, shipped chips, funding totals, or independent benchmarks for Ricursive's technology yet. The founders describe this as part of a broader pattern where humans supply the ideas and AI handles execution — something they say is already happening in software development too. Whether AI-designed chips actually deliver efficiency gains anywhere near the scale of a 10,000-fold demand jump is a claim that, for now, exists only on a conference stage.