How Ricursive Intelligence’s Founders are Using AI to Shape The Future of Chip Design

AlphaChip creators Anna Goldie & Azalia Mirhoseini launch Ricursive Intelligence to automate end-to-end chip design and enable a ‘designless’ era of custom silicon

• AlphaChip used in 4 successive TPU generations; performance delta vs. baseline grew each generation — AI scales with training data.
• Chip floor planning involves graphs with millions of nodes per block; humans divide a chip into dozens of blocks, each owned by a team — AlphaChip replaced human layouts block by block.
• AlphaChip generated curved, donut-shaped macro placements humans found too risky — turned out to reduce wire length, power, and timing violations.
• The critical bottleneck isn’t fabrication but design: chip design cycle is asymmetric vs. model iteration speed, blocking co-design of hardware and AI.
• Target customers phase 1: Nvidia, AMD, ARM, MediaTek (accelerate existing design cycles); phase 2: any company running workloads at scale (inference alone ~$100B+ spend this year).
• LLMs insufficient for chip design alone — large portion is combinatorial graph optimization, not language/code; Ricursive will use hybrid AI (LLMs + custom graph optimizers).
• Synthetic data is the binding constraint; customer data stays siloed — Ricursive claims synthetic data can exceed customer-shared data by orders of magnitude.
• AlphaChip controversy came not from physical designers whose jobs were threatened, but from researchers who had built prior methods — ‘bitter lesson’ reaction.
• ‘Fabless’ was once unthinkable; ‘designless’ is the next shift — companies spending hundreds of millions on in-house chip teams won’t need them.
• Recursive self-improvement loop: AI designs better chips → faster chips run better AI → better AI designs chips faster.

Guests: Anna Goldie (Ricursive Intelligence co-founder, ex-Google Brain), Azalia Mirhoseini (Ricursive Intelligence co-founder, ex-Google Brain) · 2026-01-14 · Watch on YouTube