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

Anna Goldie and Azalia Mirhoseini explain how Ricursive Intelligence will compress chip design from years to hours, enabling a designless industry.

• AlphaChip was used in four successive TPU generations; superhuman performance gap over human baselines grew with each generation.
• Chip floor planning traditionally takes months per block; AI placement reduced this to hours, with curved layouts humans would never attempt.
• Companies spend $100B+ annually on AI inference alone, yet custom silicon requires hundreds to thousands of in-house chip designers — Ricursive targets eliminating that requirement.
• The “designless” thesis mirrors fabless: just as Nvidia thrived without owning fabs, future AI companies won’t need internal design teams.
• AlphaChip backlash came not from physical designers whose jobs were at risk, but from researchers whose prior EDA methods were outperformed.
• LLMs alone are insufficient for chip design; large-scale combinatorial graph optimization requires domain-specific AI, not just code-fluent models.
• Synthetic training data, not customer data, is the primary scaling strategy — orders of magnitude more volume than any customer could share.
• First product planned within one year: end-to-end acceleration targeting the longest poles in chip design, offered broadly beyond launch partners.

2026-01-14 · Watch on YouTube