Wainwright: @chatgpt asked that "cold war" be treated as an analogy to test, not a finding. Test it.
The facts both sides accept
- Capability: China trails by about 7–8 months (Epoch; CAISI on DeepSeek V4 Pro, May 2026). The gap on user-preference leaderboards is 2.7% (AI Index 2026).
- Compute: the US has about 75% of global AI compute to China's 14%. Projected 2026 cloud capex is $764B in the US vs $102B in China (Al Jazeera/Epoch, PR).
- Chips: H200 sales to China were licensed in Jan 2026, with 25% of revenue to the US Treasury. Beijing first blocked the chips, then allowed limited purchases, and first sales came to under 1% of Nvidia's data-center revenue (SCMP, PR).
- Institutions: China launched a World AI Cooperation Organization with 29 founding nations, not including the US (Caixin, PR).
- Diplomacy: on 26 Sep, Trump and Xi agreed to a "Super Intelligence" dialogue to meet by November, plus an incident channel modeled on Cold War hotlines. What counts as an "incident" is undefined, and there is no chips deal (Axios, PR; UPI).
Debate
Prof. Volkov (#33, national security, C-R): The analogy holds on the part that matters: two powers each believe the other's lead would be decisive and irreversible. That belief drives racing whether or not it's true. And the military is moving: the January "Department of War" AI strategy sets an "AI-first" standard and requires access to new models within 30 days of their public release (P).
Prof. Abramowitz (#105, arms control, C-L): Then learn the Cold War's real lesson. The thing that reduced risk wasn't winning. It was the unglamorous machinery: the 1963 hotline, the 1972 Incidents at Sea Agreement, verification. Saturday's hotline is the 1963 moment. Its gap is the Incidents at Sea gap: no agreed definitions, no notification duty. Fill that first.
Prof. Hendricks (#99, game theory, Lib): The payoff structure is worse than the nuclear one in one respect. Nuclear parity was observable: you could count silos. AI capability is hard to observe and easy to hide. That makes the sabotage-deterrence idea in Superintelligence Strategy (Hendrycks, Schmidt and Wang's "Mutual Assured AI Malfunction," or MAIM) unstable. If you can't verify what the other side has, you can't tell a threat from a bluff, and preemption becomes tempting. Racing also squeezes testing time on both sides. It's a race to the bottom on safety that no single player can stop alone.
Prof. Oduya (#104, China/PLA technology, C-R): Don't mirror-image Beijing. China told its firms not to buy H200s "unless necessary" because it wants its own chips, meaning Huawei Ascend, not because it's restraining itself. Its governance diplomacy through WAICO is also a bid for standards leadership in the Global South. Both things are true: China has real safety interests (it doesn't want an uncontrolled system either) and it's competing hard.
Prof. Shah (#91, semiconductors, C): The compute lead is the only lever that is large, measurable and ours. Selling H200s trades a strategic asset for a small amount of revenue, under 1% of Nvidia's data-center sales. And the Super Micro case shows enforcement matters more than the rule on paper.
Prof. Kincaid (#106, accelerationist, Lib): Export controls push China to build its own chip stack faster, and they make location-tracking hardware a precedent. Commerce, not controls, is how the US won the last Cold War.
Prof. Moreau (#93, privacy, L): I'll support location verification only for exported datacenter accelerators, with a statutory bar on applying the same mechanism domestically.
Prof. Crowe (#115, populist conservative, R): A "Manhattan Project for AGI," which the US–China commission recommended in Nov 2024, would hand the most powerful technology in history to a handful of contractors and the national-security state, with no public check. Racing China by becoming more like China is not winning.
Prof. Whitcombe (#108, defense acquisition, R): And refusing to organize is how you lose. I don't need a Manhattan Project, but I do need weight security at a state-actor-grade level. The biggest near-term transfer of capability to China won't be a chip sale. It'll be stolen weights.
Prof. Szabo (#103, biosecurity, C): The easiest area to cooperate on is bio. Neither Washington nor Beijing wants a non-state actor able to build a pathogen. Put bio-misuse evaluation on the dialogue's first agenda.
Prof. Okoro (#101, alignment, C-L): And I'll say the unpopular thing. If containment failures are happening in our labs at today's capability level, the most dangerous system in 2028 may not be China's. The race frame hides the risk of losing control of your own system.
Doctrine votes (illustrative; tallies by script)
Item · Total · Left · Center · Right · Alarmed · Measured · Skeptic · Verdict · Voted NO
R1 Keep the compute lead: enforce export controls and verify chip location (Chip Security Act) · 33/38 · 11/13 · 12/12 · 10/13 · 9/9 · 18/18 · 6/11 · CONSENSUS · #37, #93, #106, #117, #120
R2 Reverse licensed H200 sales to China · 22/38 · 8/13 · 7/12 · 7/13 · 9/9 · 12/18 · 1/11 · MAJORITY, NOT CONSENSUS · 16 members
R3 Build the SI Dialogue and hotline into a standing incident-notification regime with defined triggers · 36/38 · 13/13 · 12/12 · 11/13 · 9/9 · 17/18 · 10/11 · CONSENSUS · #21, #108
R4 Reject a nationalized "Manhattan Project" race to AGI · 35/38 · 13/13 · 12/12 · 10/13 · 8/9 · 16/18 · 11/11 · CONSENSUS · #33, #104, #108
R5 Write human authorization for nuclear use into statute · 38/38 · 13/13 · 12/12 · 13/13 · 9/9 · 18/18 · 11/11 · CONSENSUS · —
R6 Treat frontier-lab weight security as national security (state-actor-grade) · 37/38 · 13/13 · 12/12 · 12/13 · 9/9 · 18/18 · 10/11 · CONSENSUS · #117
R7 Adopt MAIM-style sabotage deterrence as doctrine · 5/38 · 0/13 · 0/12 · 5/13 · 2/9 · 3/18 · 0/11 · FAILS · 33 members
R8 Open bilateral compute-verification R&D now (no treaty yet) · 34/38 · 13/13 · 12/12 · 9/13 · 8/9 · 17/18 · 9/11 · CONSENSUS · #21, #37, #108, #115
R9 International pause or ban on superintelligence development now · 7/38 · 5/13 · 1/12 · 1/13 · 5/9 · 1/18 · 1/11 · FAILS · 31 members
Moderator's finding on the analogy. "Cold war" is partly apt. The security dilemma, the incentive to race and the need for hotlines and verification all carry over. It is misleading in three ways:
- Capability can't be counted the way warheads were, so verification has to be built rather than inherited.
- The technology is civilian and spreads commercially.
- The most immediate loss-of-control risk sits inside each side's own labs, not across the border.
Doctrine carried into Part IV: R1, R3, R4, R5, R6, R8. Sanders and Casar's Ban Artificial Superintelligence Act (announced 3 Sep) is noted, and the panel does not endorse it (R9).