In the part of Palo Alto where former DeepMind researchers congregate, Anna Goldie and Azalia Mirhoseini have spent the last few months convincing some of the most disciplined investors in the industry that a chip can design a better chip, which can then design a better chip, and so on, until the loop bends the cost curve of silicon itself. The pitch worked. Ricursive Intelligence, founded in 2025, raised a $300 million Series A at a $4 billion valuation roughly two months after coming out of stealth [TechCrunch, Jan 2026]. By February, total disclosed capital had risen to about $335 million [TechCrunch, Feb 2026].
The wedge is unusually clean for a deeptech company at this stage. Ricursive is not trying to build chips. It is building the AI tools that lay them out, and selling that capability to the people who already make silicon. "Ricursive is building AI tools that design chips, not the chips themselves," TechCrunch noted in February, adding that the company is "not a wannabe Nvidia competitor. In fact, Nvidia is an investor" [TechCrunch, Feb 2026]. Intel is on the cap table too, alongside Lightspeed Venture Partners, Felicis, DST Global, and Sequoia Capital. When your two most strategically paranoid potential customers both write checks into the same round, the product question and the go-to-market question start to answer themselves.
The technical lineage is the other reason the round closed quickly. Goldie (CEO) and Mirhoseini (CTO) led the team behind AlphaChip at Google DeepMind, a reinforcement learning method for chip floor-planning that, per the company, has been used in four generations of Google's TPU [TechCrunch, Jan 2026]. Several of the AlphaChip contributors followed them out: Ebrahim Songhori joined as a founding member of technical staff, with Hao Chen and Jiwoo Pak also on board [Data Center Dynamics]. Additional early hires include Eva Dyer, Chandrahas T. (ex-Intel), Hao Zhou (ex-Google DeepMind), and Young-Joon Lee (ex-Apple) [LinkedIn profiles, 2026]. For a company barely a year old, that is an unusually concentrated dose of the people who have actually shipped production AI-designed silicon.
Electronic design automation is one of the quieter oligopolies in technology. Synopsys and Cadence Design Systems sit between every fabless designer and every fab, and their tools price accordingly. The category has been waiting for an AI-native challenger that does not just bolt models onto an existing flow but rebuilds the inner loop around them. The recursive framing the company uses, an AI that designs the silicon that trains the next AI [PRNewswire], is the marketing version of a real engineering claim: if each generation of the tool shortens design cycles even modestly, the compounding is the product.
The back of envelope is worth doing in plain text. A leading-edge SoC tape-out is widely reported to cost in the high hundreds of millions of dollars and take roughly 18 to 24 months of engineering time. Call it 20 months and $500M of fully loaded design cost as a round-number baseline (estimated). If a recursive design loop trims 15 percent off the schedule and 10 percent off the engineering hours, that is roughly three months and $50M per tape-out (estimated). Across the dozen or so companies doing leading-edge custom silicon today, the addressable savings sit comfortably in the single-digit billions per year (estimated), which is the kind of number that supports both a $4B valuation and a credible path to a much larger one. The catch, as always with EDA, is that customers will pay a fraction of the value created, not the whole thing.
Funding at a glance
Series A (Jan 2026) | 300 | $M
Disclosed total by Feb 2026 | 335 | $M
Post-money valuation | 4000 | $M
[TechCrunch, Jan 2026; TechCrunch, Feb 2026]
The bears will say that Synopsys and Cadence are not standing still, that both have been integrating machine learning into their flows for years, and that the switching costs in EDA are legendary. A two-month-old company with a $4B price tag is, on that view, paying for a lot of execution it has not yet done. The New York Times framed the broader recursive-AI thesis as a Silicon Valley enthusiasm with as many skeptics as believers [NYT, Jan 2026]. The bull answer is in the cap table and the lineage: Nvidia and Intel rarely co-invest in design tooling startups, and the AlphaChip team is one of the very few groups on earth with a public, multi-generation track record of AI-designed production silicon shipping at volume [TechCrunch, Jan 2026]. The question is not whether the underlying technique works. It is whether Ricursive can wrap it in a product that a chip design VP will trust on a critical path.
The next twelve months should clarify three things. First, the first named external customer, which will tell the market whether Nvidia and Intel are strategic investors or strategic users. Second, hiring velocity, particularly whether more of the AlphaChip alumni still inside Google DeepMind cross the street. Third, any disclosed benchmark against an existing EDA flow on a real block, not a toy one. A Series B in late 2026 or early 2027 at a step-up valuation would be the conventional signal that the wedge is widening. A quiet year would be the more interesting one to read carefully.
The company to beat, in the end, is Synopsys. Ricursive does not need to replace it to be a generational outcome. It needs to make the inner loop of chip design something an AI does by default, and then sell that loop to everyone Synopsys already sells to. If the recursive flywheel spins even half as fast as the founders believe, the incumbent's moat starts looking less like a moat and more like a long, expensive software contract waiting to be renegotiated.