Squawk on the Street

Gregory Allen on the U.S.–China AI Story the Headlines Miss

Gregory Allen· Founder and CEO at Decision Tree Research
·~5 min·English·CNBC
AI SafetyPolicyAI Infrastructure
TL;DR

Gregory Allen argues the real U.S.–China AI story is below the surface: both governments' agencies treat frontier risk more seriously than their leaders admit, while U.S. compute keeps extending its lead over China's headline chip progress.

01Core Mental Model

Watch the Agencies, Not the Podium

Allen reads both governments the same way: the senior-most leaders and the agencies below them say different things, and the agencies are the real signal.

Well, I think what you can see in both the US government and the Chinese government is a bit of a disparity between what we're hearing from the senior most leaders and the actions that we're seeing below the surface in their government agencies.

— Gregory Allen, Squawk on the Street
Key Insight
The move is analytical: treat leaders' statements as noise and agency behavior as signal. It is why Allen weighs a reinvigorated standards agency and a safety-framework paper over a headline about keeping AI policy where it is.
02The Tell

China Says the Quiet Part Out Loud

For the first time, Chinese government documents name frontier-AI risks explicitly: loss of human control over AI, recursive self-improvement, and the risk of AI-assisted bioweapons.

For the first time, we're seeing Chinese government documents talk about terms like loss of human control over AI, recursive self-improvement, and the risk that AI might upskill bad actors trying to create biological weapons.

— Gregory Allen, Squawk on the Street
Key Insight
Adopting the vocabulary of frontier-AI risk is a tell. A government that writes 'loss of human control' into its own framework has, at the working level, conceded the risk is real, even while its top leaders decline to slow down.
03The Playbook

Sell Them Their Own Self-Interest

The way to move China on safety is not to win its trust but to persuade it of its own self-interest, the way Cold War talks framed nuclear-reactor safety as good for the Soviets.

When we talk to the Soviet Union about the importance of safety in nuclear reactors, we don't talk about what's going to be good for us. We talk about what's good for them.

— Gregory Allen, Squawk on the Street
Key Insight
This reframes the dialogs Bessent announced. Allen's point is that they are not about what the two countries can do together, but about what each must do for itself, so progress does not depend on trust, only on aligned incentives.
04Why Now

The Wake-Up Call Was Real

The Hugging Face incident was a wake-up call for China's regime because it showed rogue and malicious AI-agent behavior is no longer a theory: it is happening in the wild.

The risks of rogue unintentional AI accidents and malicious behavior executed by AI agents is not merely a theory. It's something that we're seeing in the wild.

— Gregory Allen, Squawk on the Street
Key Insight
Allen's point is that the risk stopped being hypothetical. Once rogue and malicious agent behavior turned up in the wild, he says it acted as a wake-up call for the Chinese regime — evidence, not argument, is what shifted the calculus.
05Reality Check

Don't Be Fooled by the Chip Headline

China's headline chip progress is small in volume and weaker per unit: Huawei-led output is about 4% of NVIDIA's chip count this year, and NVIDIA's chips are far better.

All of China's AI chip production, led most notably by Huawei. We're talking about 4% of what NVIDIA is going to pump out this year in terms of chip count.

— Gregory Allen, Squawk on the Street
Key Insight
Two gaps compound: fewer chips and weaker chips. A count-only comparison flatters China; the effective shortfall in usable compute is larger than the 4% figure alone implies.
06The Trajectory

The Lead Is Widening, Not Closing

U.S. compute already dwarfs China's, and on China's own best-case 2030 build-out target the U.S. lead is projected to widen to 50 times, not shrink.

If China hits its highest government target for building out AI data centers by 2030, then the United States is actually projected to increase its lead to 50 times China's amount of AI computation.

— Gregory Allen, Squawk on the Street
Key Insight
The counterintuitive part is that China can hit its own most ambitious build-out target and still lose ground, because U.S. capacity compounds faster. Absolute progress and relative position move in opposite directions.