AI That Designs Its Own Chips: Ricursive's Anna Goldie and Azalia Mirhoseini

Why It Matters

By using AI to automate and accelerate chip design, Recursive could dramatically reduce time‑to‑market and cost, unlocking custom silicon for more companies and accelerating the overall AI hardware race.

Key Takeaways

• AI agents now generate superhuman chip layouts for real silicon.
• Recursive aims to accelerate design, democratize chips, then build its own.
• Their tools claim 100,000× speedup over traditional design loops.
• Custom AI‑optimized chips could cut time‑to‑market and costs dramatically.
• Organic, curved placement patterns outperform conventional regular layouts.

Summary

Recursive Intelligence, founded by former Google Brain researchers Anna Goldie and Azalia Mirhoseini, is building AI systems that design semiconductor chips. Their flagship technology, AlphaChip, demonstrated that deep reinforcement‑learning agents can produce chip layouts that surpass human experts and has already been taped‑out in multiple generations of Google’s TPU, data‑center CPUs, and autonomous‑vehicle chips.

The company divides its roadmap into three phases: first, accelerating physical design and verification, which currently consume up to a year and cost hundreds of millions per delayed Nvidia‑class chip; second, democratizing chip creation by offering a platform that takes a workload description and delivers a GDSII‑ready design; and third, vertically integrating to fabricate its own AI‑optimized silicon. Recursive claims its tools run 100,000 × faster than conventional EDA loops, with a static‑timing analysis engine that is a thousand times quicker while maintaining commercial‑tool fidelity.

During the launch, the team highlighted that AI‑generated placements resemble organic, curved shapes rather than the regular grids favored by human designers, reducing wire length and boosting performance. A demo showed an outer‑loop reinforcement‑learning optimizer improving chip speed after just a few iterations, thanks to the ultra‑fast analysis engine. The staff blends LLM experts from projects like Gemini and Grok with veteran chip architects, creating a rare cross‑disciplinary talent pool.

If Recursive’s vision succeeds, chip design cycles could shrink from months to weeks, lowering development costs and enabling smaller firms to order custom silicon tailored to specific AI models. This could trigger a “Cambrian explosion” of specialized processors, intensifying competition in the AI hardware market and reshaping the economics of large‑scale model training.

Original Description

At AI Ascent 2026, Anna Goldie and Azalia Mirhoseini, co-founders of Ricursive Intelligence, introduce the company and the thesis behind it: AI should design the chips that train AI. The two have spent the last decade building the foundations for this together at Google Brain, DeepMind, Anthropic, and Stanford, including AlphaChip, the deep reinforcement learning system that has shipped on the last four generations of Google's TPUs. They walk through Ricursive Intelligence's three-phase plan: first, accelerating chip design with AI tools that run a hundred thousand times faster than today's commercial software; second, becoming the "design-less" platform that lets any company with a meaningful workload commission custom silicon, just as TSMC enabled the fabless era; and third, vertical integration into their own chips and models. Plus why AI-generated chip layouts come out looking organic and curved instead of the rigid grids human engineers produce, and what that says about how AI is going to redesign the rest of physical engineering.