Eric Schmidt on the US–China AI Race and Avoiding a Global Crisis

Summary

• Eric Schmidt’s base case is that America likely wins the top-end AGI race, while China pursues AI in “every product, every service, everything.” He says the hardware restrictions imposed by the Trump and Biden administrations appear likely to prevent China from competing in that AGI space. He now considers his fear of a pre-emptive attack between superintelligent rivals “not well-grounded in facts” and thinks the US will be okay for a few years.
• China may dominate physical AI even if America retains better software and the most sophisticated systems. Schmidt sees China repeating its EV strategy in humanoid robotics, with Unitree’s $6,000 R1 arriving in December. His blunt forecast: the world will be “awash in inexpensive Chinese robots.”
• Electricity—not chips—could become the binding constraint on America’s AI advantage. Schmidt said China put in 172 GW of solar last year—“I think is the number”—while his calculation says 92 GW must be built in the US by 2030 for data centers. A large nuclear plant contributes only 1–1.5 GW, and effectively zero nuclear plants are getting started in America. If domestic supply fails, US AI training—“American intelligence”—may have to happen in Saudi Arabia or the UAE, possibly as the only fallback.
• Schmidt says many expect misinformation, cyber or biology to produce the mini-crisis that finally forces coordinated AI policy. He is most worried about biology: biological techniques might modify an existing pathogen so it remains dangerous but evades detection. Blundin says this can be done by three people in a basement and is exceptionally difficult to contain.
• Falling compute costs make model-proliferation controls structurally fragile. The 10^26-FLOP reporting threshold was a number Schmidt’s group “made up because we had no better number.” Blundin says FP4 offers roughly 8× the performance of FP32 and cites a current rule of thumb that distillation or transfer learning costs about 1% of original training while reaching the same destination. DeepSeek R1—and R2 coming—suggests open models can reach “80% or 90%” of leading closed models.
• Free Chinese models could capture global standards even without winning on absolute quality. Query masking may make distillation difficult to stop. One of Schmidt’s friends thinks US companies may keep their biggest models closed and distill their own models, while China’s biggest models could be open. Schmidt worries that governments without Western-level resources may standardize on Chinese systems “not because they’re better, but because they’re free.”
• The startup opportunity is enormous, but near-zero entry costs mean founders compete against everyone continuously. Schmidt wants products built around learning loops that can accelerate into quasi-monopolies, followed in two or three years by self-replicating forms of reinforcement learning. The investment test is not merely zero to one: “Show me how you’re going to build a system that goes from zero to infinity.”

Deep dive

Not yet available upstream; scheduled sync will retry.