On a stage at TechCrunch Disrupt 2026 in San Francisco, Anna Goldie and Azalia Mirhoseini shared a vision that sounds almost like science fiction: rather than humans designing chips to power artificial intelligences, it would be the AIs themselves designing their own circuits — and not just that, but learning from every iteration to create progressively better designs. This was the central thesis of Ricursive Intelligence, a Palo Alto startup born just ten months ago with the ambition to transform the semiconductor industry from the inside out.
Ricursive is not the first company to propose using AI to accelerate chip design. The field had already gained worldwide attention in 2021 when the Google DeepMind team published in the journal Nature the results of AlphaChip, a reinforcement learning system capable of optimizing component placement within a TPU chip — the same technology that now powers successive generations of Google's processors. But AlphaChip was a proof of concept focused on a single step in the workflow: floor planning. Ricursive, established in December 2025 by those same two researchers — Goldie as CEO and Mirhoseini, a computer science professor at Stanford as CTO — wants to go much further.
"The pace of AI progress is dictated by hardware," Goldie told TechCrunch. "Ricursive's mission is to radically accelerate chip design, and ultimately to use AI to design its own silicon substrate." It is not empty rhetoric. The startup raised US$ 335 million in four months, in a US$ 300 million Series A round led by Lightspeed, with participation from Sequoia Capital, DST Global, NVentures (Nvidia's venture arm), Felicis Ventures and others. The valuation? US$ 4 billion.
What sets Ricursive apart is not just the ambitious valuation, but the total scope of what it aims to automate. Designing a modern AI chip is a process of years costing hundreds of millions of dollars. Millions to billions of logic gates must be placed on the silicon wafer at nanometer precision, optimizing power consumption, timing violations, wire length and dozens of other physical constraints. The human process is so expensive and time-consuming that most companies don't even try: they hire a team of hundreds of design engineers, spend two to three years on the cycle and hope for the best. For decades, the industry has been stuck with conventional architectures — hardware innovation moved much slower than software innovation.
Ricursive's proposal is to close what its founders call the "recursive feedback loop" — a cycle where AI models design chips, those chips train better AI models, which in turn design better chips, and so on. The term they use is "designless": just as TSMC enabled fabless companies like Nvidia to emerge, Ricursive wants to enable any company with large-scale workloads to create custom hardware without maintaining a chip design department. The three-phase roadmap is clear: phase one focuses on specific portions of design for existing semiconductor companies; phase two democratizes custom design for companies with massive workloads but no silicon engineering team; phase three is ambitious and almost provocative: "we want to build a frontier AI, period." It wouldn't just be "AI for chip design" — it would be AI designing AI.
But the project has roots in the real world. AlphaChip has been used across four generations of Google's Tensor Processing Units, and the methodology was adopted externally by companies like MediaTek. What Ricursive wants to do is scale the concept from a localized optimization — macro placement — to the entire design flow, from floor planning through design verification. Their system uses a reward signal that classifies how good a design is, then updates the deep neural network's parameters to improve — after thousands of designs, the agent gets better and faster, as Goldie described to EE Times.
One of AlphaChip's most surprising findings was the circular, donut-shaped layout that initially met resistance from engineers but later proved superior to conventional layouts. The idea that AI can discover arrangements no human would ever try is central to Ricursive's proposition. Goldie described that the system is optimizing entire chips, not fragmenting them into dozens of smaller blocks — enabling intricate, wave-like patterns that would be impossible to envision in traditional EDA tools.
The urgency is real. On the Disrupt stage, Goldie estimated that within the next five years there will be demand for at least 10,000 times more compute capacity — and that number would be "maybe conservative." "We don't want to build 10,000 times more data centers," she said. "We think that we need to help by making chips that are more efficient, that use the bandwidth in those data centers more efficiently, that use the wafers and fabs more efficiently."
Still, enormous challenges lie ahead. Ricursive is competing against a mature ecosystem of EDA tools — Cadence, Synopsys and Siemens EDA — that took decades to consolidate. The founders themselves emphasize that they are not an EDA company and will not compete against them. But Ricursive's independence from the traditional toolchain raises questions about integration and adoption: if chip design leaves the established tool flow, how do foundries validate these designs? How does the qualification cycle change? And most importantly: can AI generalize well enough for the "learning across chips" concept to work in practice, or will physical and manufacturing limitations create pitfalls the model cannot foresee?
The market is already moving. Unconventional AI, another AI-for-hardware startup with a US$ 4.5 billion valuation, operates in the same frontier. Nvidia itself, a Ricursive investor, has been developing custom architectures for its own data centers for years. The question is: if AI begins designing better chips, and those chips train better AIs, and those AIs design even better chips — when does that loop stop being an optimization tool and start becoming a form of self-improvement? It is not difficult to imagine scenarios where hardware and software co-evolve to the point where it becomes impossible to separate what was designed by humans from what was generated by machines.
Sources: TechCrunch, EE Times, Data Center Dynamics
✓ Independent sources cross-checked and verified before publishing
