Arm CEO Rene Haas on AI: Nvidia Lessons, Intel’s Decline and the US-China Chip War

Summary

Arm CEO Rene Haas joins the All-In hosts to explain why Nvidia dominates AI training, how Arm CPUs sit alongside every accelerator (Grace Blackwell uses 72 Arm cores), and why inference and edge or physical AI will push compute toward dedicated, energy-efficient chips. He diagnoses Intel's decline as missing mobile and under-investing in EUV, leaving TSMC with a compounding leading-edge flywheel that is hard to catch. The conversation also covers US industrial policy on rare-earth refining and onshore fabs, export controls (David Sacks warns that licensing regimes would slow chip innovation while Haas warns that cutting off China would create a parallel ecosystem), and Haas's optimism about US-China cooperation on AI guardrails.

• Nvidia's lead stems from GPUs fitting parallel training workloads since AlexNet; competition is emerging from Google TPUs and hyperscaler custom silicon.
• Arm supplies the CPU alongside AI accelerators, licenses IP for custom chips, and hinted it may go further into building chips itself.
• Haas expects compute to split into training chips, distilled mixed training-inference chips, and dedicated low-power inference chips at the edge.
• Physical AI and robotics could become a larger chip market than data centers by unit volume, replacing repurposed automotive ADAS chips.
• Intel was punished for missing mobile and under-investing in EUV; TSMC's leading-edge customer flywheel compounds and is hard to catch.
• The rare-earth constraint is refining capacity rather than mineral access, requiring decades-long industrial policy and a rebuilt US manufacturing culture.
• David Sacks argues export-control licensing regimes would slow semiconductor innovation; Haas says cutting China off would spawn a parallel ecosystem.
• Haas is optimistic the US and China can cooperate on AI safety guardrails rather than treating AI purely as an arms race.