Anna Goldie

Ask Anna Goldie what her company does and the answer is a single word she will happily spell out. "We're recursive. We're AI for chip design and chip design for AI, so recursive self-improvement." It is the kind of sentence that sounds like a slogan until you remember she has the receipts. Before Ricursive, before the $4 billion valuation and the Nvidia check, Goldie spent years inside Google teaching reinforcement learning to do something most engineers considered stubbornly human: arrange the millions of components on a chip so they run fast, cool, and cheap.

The result was AlphaChip. Working with Azalia Mirhoseini, Goldie reframed chip floorplanning as a game, the same trick that turned AlphaGo loose on a Go board. A model placed components one by one, scored the layout, and learned. What had taken human engineers weeks or months collapsed into hours. The system did not stay in a lab. It laid out multiple generations of Google's Tensor Processing Units - the silicon underneath a large share of the world's AI - and the work landed as a co-first-authored paper in Nature.

That is the strange specific worth holding onto: the chips that train the models that write your email were partly arranged by another model. Goldie did not just publish the idea. She shipped it into production hardware that millions of people use without ever knowing the layout was machine-drawn.

Ricursive's ambition is to take that capability out of one company's basement and hand it to everyone. The semiconductor world already went "fabless," where companies design chips and let foundries manufacture them. Goldie wants the next step: "designless." Her pitch is that an end-to-end AI system can handle the whole pipeline - component placement, routing, verification, closing the design - so that any electronics maker, not just the giants with armies of EDA engineers, can spin up custom silicon. "We want to enable any chip to be built in an automated and very accelerated way," she says.

Goldie did not start in hardware. At MIT she earned dual bachelor's degrees - computer science and linguistics - an unusual pairing that says something about how she thinks. Language is structure under constraint; so is a chip. She added a master's in EECS, did thesis work on dialogue systems back in 2011, and then went to Stanford for a PhD in computer science with the NLP group. The throughline was always sequential decision-making: how an agent, human or model, makes one choice after another toward a goal.

At Google Brain she co-founded and led the ML for Systems team, the group that asked whether machine learning could improve the computers it runs on. That question produced AlphaChip. When Brain merged into Google DeepMind in 2023, she carried the work forward as a Senior Staff Research Scientist. Along the way she spent time at Anthropic, contributing to large language model efforts, and her early career even includes a stint at TripAdvisor. MIT Technology Review named her to its Innovators Under 35 in 2021.

There is a partnership at the center of this story that reads almost like a buddy film. Goldie and Mirhoseini met at Stanford, where Mirhoseini taught computer science. From there their careers ran in near-perfect parallel: they started at Google Brain together, left on the same day, joined Anthropic at the same time, and founded Ricursive together. Colleagues took to calling them "A&A." Jeff Dean, one of Google's most senior engineers, joked that their chip project was "chip circuit training" - a nod to the workout routine the two shared. When you build a company with someone, that kind of shorthand is the real moat.

The two were good enough that the recruiting calls got loud. Goldie has recounted, with a laugh, getting "those weird emails from Zuckerberg making crazy offers." She and Mirhoseini said no. They wanted to build the loop themselves.

Ricursive launched in 2025 with a $35 million seed round led by Sequoia Capital at a roughly $750 million valuation - an eye-watering number for a company barely out of the gate. Then it accelerated. Within four months the lab announced a $300 million Series A led by Lightspeed at a $4 billion valuation, with Nvidia among the investors. Total raised: $335 million. Headcount at that point: somewhere around nineteen people.

The math is the point. A team you could seat at two dinner tables, valued like a mid-size public company, because the bet is not on a product feature but on a flywheel. If AI designs better chips, the chips train better AI, which designs even better chips. Investors are not pricing today's revenue. They are pricing the loop.

What makes Goldie unusual is not that she is a brilliant researcher who became a founder. It is the through-line that connects a linguistics degree, a dialogue-system thesis, AlphaGo's playbook, and a wafer of silicon. She keeps finding the same shape - a sequence of constrained decisions - in places other people see unrelated problems. AlphaChip was that pattern recognition made physical. Ricursive is the wager that the pattern keeps paying off, one layer down, where the compute actually lives.