Public transcript

Part 2: Eric Schmidt on AI's Biggest Threats: China's Strategy, Cyberwar and Biotech Risks

By: Peter H. Diamandis
Length: 25 min
Language: English
Text from: the video's captions
Transcribed: 14 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

In this second part of Eric Schmidt week on the Moonshots podcast, Dave Blondon interviews Eric Schmidt about whether America will win the AI race against China, the energy constraints facing US data centers, and the threats of misinformation, cyber and biological attacks. Schmidt argues the US leads in AI software and hardware but China will dominate inexpensive robotics and has solved its electricity problem, while the US needs roughly 92 gigawatts of new power by 2030. He closes with founder advice and a reflection on the historic importance of non-human intelligence.

Key points

Schmidt believes the US will win the AI race because hardware restrictions on China will prevent it from competing at the frontier toward AGI and superintelligence.
China is likely to dominate humanoid robotics the way it won solar and electric vehicles, exemplified by Unitree's $6,000 R1 robot.
The US needs about 92 gigawatts of new electricity by 2030 to power expected data center demand, while almost no nuclear power plants are being started in America.
Schmidt identifies three major AI threats: misinformation, cyber attacks, and biological risks, with modified undetectable pathogens being his biggest worry.
He is more concerned about open-source groups operating in the shadows than big companies, since open-source models like DeepSeek's R1 and R2 appear to reach 80-90% of top closed models.
Distillation and transfer learning cost roughly 1% of original training, and the biggest US models may eventually never be released, while Chinese models stay open-source and free.
For founders, the barrier to entry is now effectively zero, so the winning strategy is to build systems around learning that can scale from one to infinity.
Schmidt compares the arrival of non-human intelligence to electricity or fire, saying the next 10 years will determine the next hundred.

Questions it answers

00:00Intro

What is this episode about?

Peter Diamandis introduces part two of Eric Schmidt week, in which Dave Blondon interviews Schmidt on US versus China in AI and how to avoid crisis during hyperexponential growth.

Audio is choppy because Schmidt was on hotel Wi-Fi.
The recording was made about a month before release.

01:41Winning the AI race

Is America going to win the AI race?

Schmidt says it looks like the US will win, because hardware restrictions on China will prevent it from competing at the frontier toward AGI and superintelligence, and China is instead adopting AI classically across products and services.

The 'San Francisco consensus' path runs from agentic AI to AGI and superintelligence.
Schmidt's earlier fear of a pre-emptive superintelligence race seems unfounded for a few years.

03:38Robotics and energy

Where is China ahead, and does America have an energy vulnerability?

China will likely dominate inexpensive humanoid robots as it did EVs and solar, while the US leads in software and chip architectures. America needs 92 gigawatt of new electricity by 2030 for data centers, and without more energy it cannot fully exploit its AI lead.

Unitree's R1 robot costs $6,000.
A big nuclear plant produces 1-1.5 gigawatt and almost none are being started in the US.
Training may end up happening in Saudi Arabia and the UAE.

08:42AI disasters and treaties

What happens after an AI wake-up call?

Schmidt draws on Henry Kissinger's use of the nuclear bomb to negotiate proliferation treaties over 15 years, noting governments act reactively and we lack an equivalent triggering event in AI.

Nuclear proliferation treaties are why we are alive today, in Schmidt's view.
Governments tend to act only after a crisis.

10:13Three threats

What are the biggest AI threats right now?

Misinformation, cyber attacks, and biological risks. Schmidt is most worried about modifying an existing pathogen so it cannot be detected but remains dangerous.

If you can write code, you can write cyber attacks.
Bio is the easiest CBRN threat for a small group and the hardest to contain.

12:11Proliferation policy

How did the Biden and Trump approaches differ on AI safety?

Biden's approach contained AGI to five companies and required reporting of training runs above 10^26 flops; the David Sachs document under Trump prioritizes winning the race but still funds study of nation-state security threats.

The 10^26 threshold was a made-up number.
Training locations are likely known to both governments through espionage.

14:55Open source and distillation

Why does open source create a proliferation problem?

Open-source groups can operate in the shadows and patch together good-enough models; DeepSeek R1 and R2 look like 80-90% of top closed models, and distillation costs about 1% of original training, so the biggest US models may never be released while free Chinese models become the global standard.

Training is getting cheaper via FP8 and FP4 precision.
Most countries may standardize on Chinese models because they are free, not better.

18:31Founder advice

What should founders do in this era?

The barrier to entry is effectively zero, so founders should build learning-based systems that can scale from one to infinity, hire networks of smart people, and build platforms others depend on for network lock-in.

Founders should 'know nothing, learn everything'.
Platforms with strong interconnection create network lock-in, as with Microsoft and Google.

Notes

US vs China in the AI Race

Schmidt believes the US will win, following the "San Francisco consensus" path from agentic AI to recursive self-improvement to AGI and superintelligence.
Hardware restrictions imposed by the Trump and Biden administrations look likely to prevent China from competing at the frontier.
After visiting Shanghai, he concludes China is playing a different game: adopting AI classically into every product and service rather than racing for AGI.
His earlier fear of a pre-emptive superintelligence race now seems unfounded for a few years.

Robotics and Energy

China is doing to robotics what it did to EVs and solar, installing 172 gigawatt of solar last year; Unitree launched the R1 humanoid robot for $6,000.
The US will keep high-end niches and superior software, but the world will be "awash" in inexpensive Chinese robots.
US data centers need 92 gigawatt of new electricity by 2030; a big nuclear plant produces 1-1.5 gigawatt and effectively zero are being started in America.
Without more energy, the US cannot fully exploit its AI lead; fallback may be training in Saudi Arabia and the UAE.

Threats and Proliferation

Three obvious threats: misinformation, cyber, and bio. Schmidt is most worried about modified pathogens that evade detection.
The Biden 10^26 training-run reporting threshold was a made-up number; training is getting cheaper via FP8 and FP4 precision.
He worries more about open-source groups in the shadows than big companies; DeepSeek R1 and R2 look like 80-90% of top closed models.
Distillation costs about 1% of original training; the biggest US models may never be released, while free Chinese open-source models may become the global standard.

Founder Advice and This Moment

The barrier to starting a company is effectively zero; you compete against everyone, and time is compressed.
Build systems around learning, scale from one to infinity, and build platforms others depend on for network lock-in.
Non-human intelligence is as consequential as electricity or fire; the next 10 years will shape the next hundred.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.