The AI Power Podcast

Gregory Allen is joined by Professor Ben Buchanan of the School of Advanced International Studies to discuss his new book, The Bitter Struggle, which describes his time in the Biden White House setting national AI policy. Greg and Ben discuss the Biden export controls, executive orders, and the National Security Memorandum on AI, alongside safety, similarities and differences with the current Trump administration, and “enduring the thing itself” when facing new and daunting challenges.

Chapters:

0:00 - 7:41: Introduction and Ben's realization that AI growth would require robust AI policy
7:41 -  20:33: Kneecapping China’s AI sector: Justifying the October 2022 export controls
and discussing their implementation
20:33 - 25:12: How Ben and his peers convinced the Biden administration to take seriously the strategic importance of AI
25:12 - 29:58: Critiquing the pitfalls of the Biden AI export controls
29:58 - 33:45: ”Economic warfare” and the importance to act decisively and communicate clearly
33:45 - 46:06: AI export controls and their role in AI safety in the American AI conversation
46:06 -  49:47: The National Security Memorandum on AI
49:47 - 53:51: The intersection of AI national security policy and industrial policy
53:51 - 1:03:17: President Biden’s role in embracing the importance of AI for national security
1:03:17 - 1:15:48: “Spending the lead on safety” and the role of export controls in prioritizing AI safety
1:15:48 - 1:18:46: Thinking about the issue of open weight models
1:18:46 - 1:21:21: How the Biden administration thought about the data center buildout, and what lessons to apply today
1:21:21 - 1:26:06: Learning from the railroad buildout, staying wary of over-reliance on historical analogies, and “enduring the thing itself”

Ben’s book, The Bitter Struggle, co-authored with Tantum Collins releases on November 10, learn more about it here: The Bitter Struggle: Superintelligence, Superpowers, and the Fate of the World

Order via Amazon or Penguin Random House 
https://www.amazon.com/gp/product/B0GNZ324H9?tag=randohouseinc7986-20
https://www.penguinrandomhouse.com/books/821154/the-bitter-struggle-by-ben-buchanan-and-tantum-collins/

Follow Greg on Twitter https://x.com/Gregory_C_Allen
Learn about Decision Tree Research at https://decisiontreeresearch.com

Gregory Allen and Adam Goodwin break down Tuesday’s midterm results and the impact AI had on the outcome. They discuss how the albatross of data centers cost Republican candidates with voters, how Democrats capitalized on these concerns, and what to expect for the industry moving forward. Then, Greg and Adam preview Democrats’ likely legislation agenda towards AI and outline potential commonsense AI regulation for Congress to pass. Finally, Greg discusses breaking new DNA synthesis regulations from the White House that help address the risk at the intersection of AI and biosecurity.
Chapters:
00:00 – 5:04: Greg and Adam react to the the Democratic sweep in Tuesday’s midterm
5:04 -14:08: How much did AI matter in this election? What did voters think about AI and data centers?
14:08 - 26:40: How voter referendums on data centers will impact data center construction
26:40 - 35:05: How AI concerns and AI risk impacted Senate, House, and Governor races across the country 
35:05 - 41:35: Previewing the AI issues Democrats hope to tackle in the new Congress
41:35 - 48:51 Could Congress, and the White House, find their way towards commonsense AI regulation?
48:51 - 54:14: The White House releases regulations to reduce AI misuse and biological risk in DNA synthesis

Director Kratsios’ Announcement of the National Policy for Nucleic Acid Synthesis Screening: https://x.com/mkratsios47/status/2118365268342652934

The Information on Nvidia’s Influence over AI policy in Washington: https://www.theinformation.com/articles/nvidia-uses-soft-power-influence-customers-shape-policy

Jensen Huang is targeted by a Chinese spy, China adopts AI export controls, and AI’s impact on the midterms

Episode description

Adam Goodwin and Gregory Allen examine a shocking New York Times story that Jensen Huang was targeted by Chinese intelligence via the CEO of a partner company, and break down the litany of red flags surrounding diversion efforts to get Nvidia chips into China over the last few years. Then, Adam and Greg break down China’s new AI export controls regime, how it bears a striking resemblance to the Biden era Diffusion Rule, and what strategic objectives Beijing is pursuing by employing this playbook. Lastly, Greg recounts the PHP vulnerability and a Black October that saw record numbers of vulnerabilities found, and the episode finishes with an assessment of how voters feel about AI and its buildout heading into the midterms. 

Chapters:
00:00 – 18:49: Jensen Huang is targeted by a Chinese spy, Nvidia insists chip diversion isn’t happening despite examples to the contrary
18:49 -35:37: China’s new AI export controls assert authority over the flow of Chinese AI tech
35:37 - 54:24: Black October saw a seminal moment in AI cybersecurity risk, the PHP software vulnerability, and lessons from Anthropic’s bug bounty program
54:24 - 1:02:15: The impact of AI on voters’ opinions going into the midterms

Mackenzie Hawkins’ Bloomberg piece from 2025 mentioned by Greg: https://www.bloomberg.com/news/features/2025-12-22/nvidia-partner-megaspeed-draws-china-chip-smuggling-concerns-in-us
ChinaTalk’s breakdown of Chinese AI export controls: https://www.chinatalk.media/p/mofcoms-new-ai-rules

Saif Khan, founding director of the Center for Technology and Statecraft (CTS), and Nick Brown, CTS co-founder and senior fellow, join Gregory C. Allen to unpack their landmark 100+ page report, DUV Immersion Lithography: The Foundation of the US AI Advantage: Closing Gaps in Export Controls to China. Their central finding is that EUV gets the headlines, but the older immersion DUV machines are what China actually uses to make AI chips. China’s stockpile has grown from roughly 50–100 immersion tools in October 2022 to around 340 today. It imported about 90 a year in 2024 and 2025, close to two-thirds of ASML’s global immersion market.
They explain why ASML’s NXT:1980, which can be legally exported to China, is the machine to watch. They also cover how etch and deposition controls are masking the full power of China’s lithography stockpile, and why Chinese indigenization announcements, including the new 28nm prototype, don’t change the strategic picture before the 2030s. Then they lay out their scenario models. Even in an extreme case, China’s current stockpile can’t match US AI chip production through 2035, but continued imports could close the gap or even tip it. Finally, Saif and Nick make the case for a countrywide ban on immersion DUV, the levers the US could pull without new legislation, and what the MATCH Act in the Senate NDAA could mean for the fight.
Read the report at techstatecraft.org/duv.
Chapters
00:00 – 05:35: Meet the Center for Technology and Statecraft and the people behind the report
 05:35 – 08:22: What immersion DUV is, and why it matters more than EUV for China’s AI chips
 08:22 – 12:56: From Cold War controls to ASML’s near-monopoly: how lithography export controls evolved
 12:56 – 17:29: The counterfactual: where would Huawei and SMIC be without EUV controls?
 17:29 – 22:45: Three tiers of DUV, the NXT:1980 loophole, and why countrywide beats case-by-case licensing
 22:45 – 25:18: What if the US had banned all immersion DUV in 2022? From 50+ machines to ~340
 25:18 – 30:02: Why China hasn’t yet unlocked the full power of its lithography stockpile
 30:02 – 37:26: Can China build its own? Prototype, commercialization, scale-up, and the 2030s
 37:26 – 45:40: The scenario models: when stockpile alone isn’t enough, and when imports tip the balance
 45:40 – 50:13: Flops per gigawatt, 90 machines a year, and 100,000 Huawei chips per tool
 50:13 – 55:11: Monday morning at BIS: allied controls, light sources, FDPR, de minimis, and the MATCH Act
 55:11 – 55:27: Where to find the report

Jack Conlin and Gregory Allen examine how a new AI incident led to a shift in AI safety policy from the White House’s Super Intelligence Task Force, and CAISSI gets more direct access to the White House. 

Greg and Jack discuss the landmark shift in AI policy that takes safety more seriously, but still fails to address the growing risk of Chinese models. Then, Anthropic reports it has stopped distillation from 90-99% of attempts, cutting off a core avenue that feeds Chinese AI development. Will this milestone prove to be durable, and how will it affect US AI leadership and strategy?

Chapters:
00:00 – 10:10: Anthropic agents cause White House shift to tougher de facto AI regulation and oversight
10:10 – 14:06: CAISSI gets a name change and now reports to the White House
14:06 – 22:38: Focusing on US model safety without addressing Chinese models is inconsistent and jeopardizes American AI leadership
22:38 – 32:06: Anthropic has stopped 90-99% of adversarial distillation, will it lead to a durable US advantage?