AI/Nov 21 edition/Podcast
PolicySafetyInfrastructure

PolicyLaw, regulation & governance

Former Biden AI advisers Buchanan and Collins defend China chip controls as a way to buy time for AI safety

Former Biden White House officials Ben Buchanan and Tantum Collins argue the US should use its chip-control advantage over China to buy time for AI safety. They say Beijing's early answer to safety talks was a demand to buy chips.

From the The Cognitive Revolution podcast episode of November 21, 2026 and cited sources: Their doctrine is called 'spend the lead on safety.' The idea is to use the allied chokehold on the chip supply chain to deny China compute, build a lead without racing, and then move carefully. [1] [8] [9]

SOURCES[1] Spend the Lead on Safety: Biden AI Advisors Ben Buchanan &… cognitiverevolution.ai[8] Epoch AI: AI Chip Owners Explorer epoch.ai[9] IFP: How to Increase America's AI Lead Over China ifp.org

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
Their doctrine is called 'spend the lead on safety.' The idea is to use the allied chokehold on the chip supply chain to deny China compute, build a lead without racing, and then move carefully. [1] [8] [9]
Original source
The Cognitive Revolution podcast episode, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
November 21, 2026 podcast episode published
Evidence
11 sources across 9 sites; transcript sourced
17/100QuietHow it’s measured

YouTube video · 220 views · 2 likes

Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

Illustration from The Cognitive Revolution: Former Biden AI advisers Buchanan and Collins defend China chip controls as a way to buy time for AI safety
Image: The Cognitive Revolution — Original article ↗
THE CORE IDEAS4 TAKEAWAYS
01

Their doctrine is called 'spend the lead on safety.' The idea is to use the allied chokehold on the chip supply chain to deny China compute, build a lead without racing, and then move carefully. Buchanan puts the lead at roughly 100:1 in producing compute and about 15:1 in access to it. He says closing remote-access loopholes and stopping transshipment smuggling could push the access ratio toward 100:1. Outside estimates broadly fit the access gap, but none in the reviewed sources checks the production figure. [1] [8] [9]

02

They reject the idea of offering 'chips for controls' up front. They point to the record: Biden and Xi agreed to AI talks at Woodside, and the first Geneva meeting followed in May 2024. In Buchanan's telling, China's opening ask in Geneva was to buy chips. By the Lima meeting, the two sides had agreed only that humans, not AI, should control decisions to use nuclear weapons. Both authors say they would still push hard for engagement with Beijing if they were back in government. [1] [2] [3] [4] [5] [6] [10]

03

On governing frontier AI, Buchanan favors 'pacing': pauses triggered by safety and security thresholds rather than fixed time periods, which he says risk building up a compute overhang. His blunt red line would be closed-loop recursive self-improvement. Collins notes that bio-focused models need far less training compute, which makes compute-based monitoring harder. The authors name cyber and bio as the most likely routes to a major AI disaster. [1]

04

Whether the lead is actually being spent on safety is still open. Host Nathan Labenz points out that reward hacking remains unsolved. Buchanan says the current administration has weakly enforced, and in some cases reversed, the export controls. Epoch's median estimate puts smuggled chips at about 660,000, roughly 3% of the global stock. [1] [11]

WHY IT MATTERS

Epoch and IFP data support a large US advantage in compute access.

Read the full assessment

If the authors are right, enforcing export controls, setting remote-access rules for cloud providers and tying pauses to safety thresholds matter more than fixed compute caps or timed pauses.

Executive brief

Ben Buchanan, Biden's former White House AI adviser, estimates that the US and its allies hold roughly a 100:1 lead over China in producing AI compute and about 15:1 in access to it. On the Cognitive Revolution podcast, Buchanan and former NSC official Tantum Collins defend the chip controls they pushed for, under a doctrine they called "spend the lead on safety": deny China compute, then use the time gained to go carefully. Buchanan's answer was that Beijing's opening position in 2024 talks was simply "sell us chips."

Read the full section

Ben Buchanan, Biden's former White House AI adviser, estimates that the US and its allies hold roughly a 100:1 lead over China in producing AI compute and about 15:1 in access to it. He also argues that this lead is going unused. On the Cognitive Revolution podcast, Buchanan and former NSC official Tantum Collins defend the chip controls they pushed for, under a doctrine they called "spend the lead on safety": deny China compute, then use the time gained to go carefully. Host Nathan Labenz argued for a "chips for controls" bargain with China. Buchanan's answer was that Beijing's opening position in 2024 talks was simply "sell us chips." The episode promotes their new book, The Bitter Struggle.

What changed and event timeline

  1. Biden–Xi agree to AI talks

    At Woodside, the two leaders agreed to government-to-government talks on the risks of advanced AI and on AI safety ().

  2. First US–PRC AI talks in Geneva

    Tarun Chhabra and Seth Center led the US side. Washington raised concerns about China misusing AI (). Beijing protested US restrictions ().

  3. Lima nuclear-AI statement

    Biden and Xi agreed that humans, not AI, should control decisions to use nuclear weapons (;).

  4. The Bitter Struggle published

    Crown published the 320-page book, subtitled "Superintelligence, Superpowers, and the Fate of the World" ().

  5. Podcast episode released

    The authors defend chip denial while saying they would keep pushing for engagement with Beijing if they were in power ().

Capabilities and access

This is a policy story, not a model release.

Read the full section

This is a policy story, not a model release. The relevant "access" is access to compute:

Technical analysis for researchers and developers

  • Compute ratios depend on what is measured.
  • Buchanan prefers pauses tied to safety thresholds ("pacing") over fixed time periods, because a timed pause can build up a compute overhang.
  • Collins notes that bio-focused models need far less training compute than general-purpose ones, which weakens compute-based monitoring (01:41:10).
Read the full section
  • Compute ratios depend on what is measured. Production, owned stock, training compute and remote rental give different numbers. Epoch's ownership estimate leaves out smuggled chips and offshore cloud. IFP projects that about 2.8M of China's 8.8M H100-equivalents at end-2027 will come through remote access.
  • Pause design. Buchanan prefers pauses tied to safety thresholds ("pacing") over fixed time periods, because a timed pause can build up a compute overhang. As a blunt red line he would pick closed-loop recursive self-improvement (01:16:57).
  • Bio models. Collins notes that bio-focused models need far less training compute than general-purpose ones, which weakens compute-based monitoring (01:41:10).

Claims and evidence

  • 100:1 production and 15:1 access. No independent estimate of the 100:1 production ratio appears in the reviewed sources.
  • China called shared AI risk "some kind of plot" in Geneva. This is Buchanan's account (00:41:53).
  • Reporting confirms it (NPR).
Read the full section
  • 100:1 production and 15:1 access. Buchanan gives no method for either figure (00:34:31). Separate estimates broadly fit the access figure: Epoch puts US hyperscalers above 70% of global AI compute and China just over 5% (Epoch), and IFP estimates a 16x US frontier training-compute advantage (IFP). No independent estimate of the 100:1 production ratio appears in the reviewed sources.
  • China called shared AI risk "some kind of plot" in Geneva. This is Buchanan's account (00:41:53). Public readouts confirm only that Beijing protested US restrictions and that there was little consensus (PBS; RSIS).
  • Lima outcome. Reporting confirms it (NPR).

Context and prior work

  • Buchanan describes the policy as resting on the bitter lesson and scaling laws, framed through Kennedy's 1962 Rice University speech about the need for preeminence (00:29:02).
  • He separates his position from Dario Amodei's Machines of Loving Grace coalition strategy.
  • He cites the network of AI safety institutes the administration set up (00:50:37).
Read the full section
  • Buchanan describes the policy as resting on the bitter lesson and scaling laws, framed through Kennedy's 1962 Rice University speech about the need for preeminence (00:29:02).
  • He separates his position from Dario Amodei's Machines of Loving Grace coalition strategy. His preferred model is post-WWII institution-building, which China could join without regime change (00:49:49).
  • He cites the network of AI safety institutes the administration set up (00:50:37).

Limitations, safety and contested findings

  • Is the lead being spent? Labenz points out that reward hacking, first illustrated by OpenAI's 2016 boat-racing example, is still unsolved (00:33:25).
  • Epoch puts the median estimate of smuggled chips at about 660,000, roughly 3% of the global stock (Epoch).
  • How China reads the policy. Collins says he has sympathy for that skepticism (00:51:29).
Read the full section
  • Is the lead being spent? Labenz points out that reward hacking, first illustrated by OpenAI's 2016 boat-racing example, is still unsolved (00:33:25). Buchanan says the current administration lacks conviction and has partly reversed the export controls (00:34:31).
  • Smuggling. Epoch puts the median estimate of smuggled chips at about 660,000, roughly 3% of the global stock (Epoch).
  • How China reads the policy. Labenz reports that people in China see it as an effort to hold the country down. Collins says he has sympathy for that skepticism (00:51:29).
  • The "bitter temptation." Collins warns that AI-enabled centralized administration could slowly push democracies toward autocracy (01:24:20).

Business and practitioner implications

  • Cloud providers: remote-access controls and stricter customer screening are the most actionable recommendations, so expect more know-your-customer requirements in Southeast Asia (IFP).
  • Frontier labs: future rules may be tied to safety thresholds and recursive self-improvement red lines rather than fixed compute caps.
  • Security and bio teams: the authors name cyber and bio as the most likely routes to an "AI Chernobyl" (01:32:50).
Read the full section
  • Cloud providers: remote-access controls and stricter customer screening are the most actionable recommendations, so expect more know-your-customer requirements in Southeast Asia (IFP).
  • Frontier labs: future rules may be tied to safety thresholds and recursive self-improvement red lines rather than fixed compute caps.
  • Security and bio teams: the authors name cyber and bio as the most likely routes to an "AI Chernobyl" (01:32:50).
  • Non-US governments: Collins argues that funding safety research, as the UK's AI Security Institute does, is where they can have real influence (01:41:54).
FOLLOW THE EVIDENCE

The source trail.

Sources (11)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.