Part 2: Eric Schmidt on AI's Biggest Threats: China's Strategy, Cyberwar and Biotech Risks
In part two of Eric Schmidt week, Dave Blundin presses Schmidt on whether America wins the AI race against China. Schmidt argues the US leads on the path to AGI and super intelligence because hardware export controls have slowed China's frontier training, but China is pulling ahead on energy, humanoid robotics and cheap hardware-driven industries the way it already won EVs and solar. Schmidt lays out his three-threats framework (misinformation, cyber, bio) for the eventual AI 'wakeup call' crisis, walks through why export-control thresholds like Biden's 10^26 flop rule are losing relevance as training gets more efficient, and warns distillation lets open-source labs (and China) catch up to closed frontier models cheaply. He closes with founder advice: barrier to entry is near zero so the moat is building a self-reinforcing learning loop and a platform with network lock-in, and reflects on the scale of the current moment versus the PC and internet eras.
Schmidt says the 'San Francisco consensus' path from agentic AI to recursive self-improvement to AGI/superintelligence favors the US because hardware export restrictions on China are working; China is instead pursuing broad classical AI adoption rather than an AGI race, easing Schmidt's earlier fear of a preemptive superintelligence race.
China winning robotics, EVs and solarRobotics▶ 3:29
Schmidt says China is repeating its EV and solar playbook in humanoid robotics: cheap, high-volume hardware (stepper motors, components) will let China dominate inexpensive humanoid robots even though US software stays superior; he cites Unitree's $6,000 R1 robot shipping in December.
Schmidt says China has solved its electric power problem (172 GW of solar added in one year) while the US needs 92 GW of new capacity by 2030 for data centers (equivalent to dozens of nuclear plants) and is effectively building none; current US policy promotes oil and gas but has hobbled solar and wind, which he calls an error that could leave frontier AI training dependent on Gulf-state electricity deals.
Three threats framework: misinformation, cyber, bioAI▶ 10:15
Schmidt lays out three near-term AI risk categories he believes are well understood: open-source-enabled misinformation (fake video/news, mainly a threat to democracies), cyberattacks (since code-writing AI is equally good at writing exploits), and biological risk (modifying existing pathogens to evade detection); he expects one of these to trigger a 'mini crisis' that forces governments to seriously address AI downsides.
Nuclear treaty analogy for AI governanceGeopolitics▶ 9:14
Schmidt describes long conversations with Henry Kissinger about how 1950s-era nuclear nonproliferation treaties, negotiated over ~15 years after the bomb was used, kept the world safe, and asks whether an analogous governance framework is possible for AI even though there has been no equivalent 'AI Hiroshima' to force the issue yet.
Biden vs Trump/Sacks AI policy approachGeopolitics▶ 12:04
The Biden-era approach tried to contain AGI to five major labs and require registration of training runs above 10^26 flops; that policy was scrapped and replaced by the Trump administration's David Sacks-authored approach, which prioritizes speed and winning the race over proliferation controls, while still funding security research framed around China.
Schmidt says the 10^26-flop reporting threshold (which he helped set somewhat arbitrarily) is becoming moot because training is getting dramatically more compute-efficient — moving from FP16 to FP8 and now toward 4-bit floating point — and both the US and China likely already know roughly where each other's large training runs are via espionage and visible data-center footprints.
Schmidt is more worried about open-source groups distilling frontier capability cheaply than about large state/corporate training runs, which are countable; he cites DeepSeek R1/R2 reaching 80-90% of top closed-model quality via distillation, and Dave Blundin notes distillation/transfer learning runs about 1% of original training cost.
Closed US models vs open Chinese modelsGeopolitics▶ 16:21
Schmidt relays a friend's prediction that the biggest US models will eventually never be released publicly (companies will distill down smaller versions to protect them from distillation), while China's biggest models stay open-source; because open models are free, most of the world outside wealthy Western countries may standardize on Chinese models for cost reasons alone, which worries him geopolitically.
Founder advice: near-zero barrier to entryEconomy▶ 19:29
Schmidt tells founders the barrier to starting a company is now effectively zero — you can register a company, pay AI tools like Google or Claude to write your code, and use contract manufacturers for hardware — but that means you're competing against literally everyone, so speed and compression of time become the decisive edge.
Learning loops and platform lock-in as the moatEconomy▶ 19:34
Schmidt's core startup philosophy: build everything around continuous learning rather than fixed specification, because a system that learns compounds into a quasi-monopoly; the resulting path to the next trillion-dollar company is self-replicating reinforcement learning, and durable value comes from building a platform others depend on, which creates network lock-in the way Microsoft and Google once did.
This decade as a historic inflection pointAI▶ 23:56
Schmidt compares the arrival of non-human (AI) intelligence to the historical significance of electricity or fire, arguing the next 10 years will shape the next 100 more than any prior period, and that countries and companies that embrace it aggressively will become the big winners while slow adopters lose.
Predictions made
openEric Schmidt: China will dominate the market for inexpensive humanoid robots, similar to how it came to dominate EVs and solar panels.
“the arrival of humanoid robots um is likely to be dominated by China”
Your call:
openEric Schmidt: US frontier AI training may end up depending on electricity secured through deals with Gulf states like Saudi Arabia and the UAE due to domestic power shortfalls.
“American intelligence, is actually being developed in kingdoms and that may be the only fall back we have”
Your call:
openEric Schmidt: One of misinformation, cyberattacks, or a biological incident will cause a crisis significant enough to force governments to seriously address AI's downside risks.
“most people believe that one of those three will create some kind of mini crisis”
Your call:
openEric Schmidt: The biggest US AI models will eventually never be publicly released, with companies instead distilling smaller public versions to protect the frontier model.
“the biggest models will never be released and that the companies will distill their own models down”
Your call:
openEric Schmidt: Most governments and countries outside the wealthy West will standardize on open Chinese AI models over closed US models, mainly because they are free.
“eventually two or three years, you'll start to have various forms of reinforcement learning which are self-replicating”
Your call:
openEric Schmidt: The next trillion-dollar company will emerge from self-replicating reinforcement learning systems, succeeding the current Anthropic/OpenAI-scale tier.
“That's the most likely path for the next trillion dollar company”
Your call:
Numbers that matter
China added 172 GW of solar capacity in one yearSchmidt cites this as evidence China has effectively solved its electricity problem.
US data centers will need 92 GW of new electricity capacity by 2030From Schmidt's congressional testimony on US electricity demand for AI; a large nuclear plant is 1-1.5 GW, implying dozens of plants' worth of capacity are needed.
Unitree's R1 humanoid robot costs $6,000, shipping in DecemberCited by Schmidt as an example of China's inexpensive humanoid robot hardware.
Biden-era policy set the AI training-run reporting threshold at 10^26 flopsSchmidt says he helped set this number somewhat arbitrarily, with no better basis available at the time.
DeepSeek R1 (and upcoming R2) reach roughly 80-90% of the quality of top closed frontier modelsCited as evidence that distillation lets open-source models rapidly catch up to closed ones.
Distillation/transfer learning costs about 1% of original model training costDave Blundin's rule of thumb for how cheaply a model's capability can be replicated.
Moving from FP32 to FP4 yields roughly an 8x performance boost, compounding with ~100x from transfer learningDave Blundin's estimate of combined efficiency gains from lower-precision training plus distillation.
Worth digging into
🕳️ The 10^26-flop training-run reporting threshold
Schmidt admits the number was picked somewhat arbitrarily and is being outrun by efficiency gains (FP16 to FP8 to FP4), raising the question of whether any flop-based compute governance threshold can stay meaningful.
🕳️ David Sacks' AI policy document vs the Biden AI executive order
Schmidt frames this as the key policy pivot from containment/proliferation control to speed-and-win-the-race, with real consequences for global AI governance.
🕳️ Distillation defense and query-masking
Schmidt says there's no good current solution to stop a rival from masking queries as normal users to distill a model, which is a live, unresolved security problem for frontier labs.
🕳️ China's 92 GW vs US energy buildout gap
The claimed gap between US data-center electricity needs (92 GW by 2030) and near-zero new nuclear plant starts is a concrete, checkable infrastructure bottleneck with direct implications for the AI race.