## The Exodus Anna Goldie and Azalia Mirhoseini spent five years at Google DeepMind building AlphaChip — AI that could design chip layouts in hours instead of months. It worked. Google used it. Then they left. The departure wasn't about money. It was about speed. Google moves at Google pace. The founders wanted to see how far recursive chip design could go, and they couldn't wait for committee approval. Two months later: $4 billion valuation. $300 million Series A led by Lightspeed. The bet isn't on a chip. It's on a loop. ## The Thesis The semiconductor industry has a design problem disguised as a supply problem. Everyone talks about fab capacity. TSMC backlogs. Geopolitical chip wars. But the bottleneck before the bottleneck is design. Traditional chip architecture takes 2-3 years from concept to tape-out. Human engineers manually optimize layouts, simulate performance, iterate. AlphaChip proved this could happen in hours. Ricursive Intelligence is asking: what happens when the AI that designs chips also designs better versions of itself? ## The Temporal Processing Unit The company's first product isn't a chip you can buy. It's a chip architecture optimized for AI inference that was itself designed by AI. They're calling it the Temporal Processing Unit. The recursive part: each generation of TPU gets designed by the previous generation's AI. The architecture improves itself. The humans become reviewers, not designers. Sound familiar? It's the same pattern as Codex and Sonnet 5 — humans shifting from makers to managers. Except this time, the thing being made is the hardware that runs the AI. ## Why This Matters Nvidia's moat is design expertise accumulated over decades. Ricursive's bet is that AI can compress decades into months. If they're right, the chip industry reorganizes around whoever has the best recursive design loop, not whoever has the best human architects. The talent war becomes an algorithm war. If they're wrong, they've still raised $300 million to find out.