I just came off watching Helen Toner talking at a CSIS panel about the window of opportunity for US-China AI cooperation. Her argument, as I understood it: everybody gets stuck on the need for a grand deal between the two countries, when there’s incremental progress to be made everywhere. I think she nailed it.
In many ways, China as an abstract idea is more valuable to American AI companies than the reality of engagement. The world where there’s a nemesis floating just out of your reach, distilling, copying, and acting in bad faith makes the job of convincing everybody that you are the noble national hero much easier.
Inconveniently, that’s just not true. Chinese labs (both the people and the ideas) are not their own ecosystem, but rather a dynamic system in which we, North Americans, are deeply integrated and involved.
In the past week, no less than six people from the Chinese ecosystem have written me about Dwarkesh and Dylan on the pod and what I thought. Wonderful work has been done to translate RedNote and Weibo writing on AI into English. We read Liang Wenfeng’s transcripts from his leaked investor call with rapt attention.
These two ecosystems are not going to cooperate on an awful lot of things, but AI safety is definitely one of them, and it’s the place to start.
Here are 10 things that I think we could do that are simple, effective, and available to us as of today.
A Dedicated AI Safety Cluster in Singapore
I’m shamelessly plugging my project up front; I believe that export controls have been so effective that Chinese labs are choosing capability over safety too often. There is only so much compute to go around over there, and safety work sometimes gets axed when GPUs are tight. A cluster that is free to use for AI safety evals but does not provide gradient updates to models and so cannot be used for capability is an easy lift for the Trump administration. It can be entirely paid for by third parties, something that Trump would love.
Data Center Environmental Parity
The Chinese are masters of green energy, and North American data centers need to get built. Trump should turn the tide on data centers by offering to license Chinese technology to make them more green.
The Chinese would love a story about Americans licensing Chinese technology, and America would love to turn public opinion towards data centers being environmentally friendly. Yes, I know, Chinese anything in American infrastructure is radioactive in D.C. right now; the panels are glass, keep the inverters domestic and move on. Hell, I bet Xi would give us the solar panels for free. Seventy-one percent of Americans don’t want an AI data center next door, mostly over water and power. Trump may not be the biggest environmentalist, but getting data centers to be popular is going to help stock prices, and he cares a lot about that.
Shared Model Licensing Standards
Getting to a grand bargain for AI safety is on the table this September, with Xi due at the White House on the 24th, but before we get there, we can do some small things around the edges; one of the easiest would be getting our open model licenses right. This is the kind of weedsy bureaucracy that China loves to dig into. Revenue sharing between inference providers that are U.S.-based companies and Chinese model makers is ramping up, and there are no signs of slowing down. Having a model licensing agreement that explicitly allows this to go forward would be a good win for both sides of the ecosystem. Trump is going to want a way to get cut in on some of these deals, and that’s good; let him roll these up into tariffs in exchange for 5% or 10%. That money can be reinvested back into either community building, places where data centers exist, or just put into the pockets of the state. I don’t care. Just get the deal done so it’s predictable revenue, and we can build businesses around it.
Secure the Semiconductor Input Supply Chain
China is the key producer of an awful lot of the gases and other inputs into semiconductor manufacturing. Head off this becoming Rare Earths 2.0 by getting a deal done right now.
Schwarzman Scholars for AI Safety
One of the key stories of Chinese AI is the Haidian Cluster being disproportionately represented by Tsinghua University at the highest level. Work on a Schwarzman Scholars-esque initiative for AI safety: moving young people between the two ecosystems and getting a billionaire philanthropist to pay for it all. Personally, I’d hit up Larry Ellison, but your call, big dawg.
AI Safety Red Telephone
Nothing infuriates Americans more than the way the Chinese are sometimes kind of unwilling to pick up the phone in a crisis. D.C. old heads know this to be true and find it inexplicable and problematic.
Have somebody gift two golden telephones and give one to Xi. Ask him to keep it on his desk in case of Terminator. He won’t actually do it, but it’s the thought that counts.
Track Two AI Foreign Policy Summit
Hitting the reset button around AI foreign policy should be a priority; Pax Silica feels like a footnote in Washington, but it’s getting a little existential in Beijing. We should do a 400-person Track Two AI Foreign Policy Summit with all the middle powers in a pleasant location sometime this winter, so that everyone can get together and hash out some of the details. Having a dedicated practice that is explicitly AI foreign policy is something that the folks at ChinaTalk have been banging the drum about for years, but it doesn’t make it any less of a good idea.
Work Out Data Sovereignty and Cross-Border Data Flows
We don’t really have a framework to figure out how weights can move across national boundaries. This is way over my head, but I’m told that people like Samm Sacks are thinking seriously about this. Having a framework for this sooner rather than later is an awfully good idea. From what I understand, this is upstream from copyright law, so it needs to be worked out before we can address things like video gen models doing Mickey Mouse cartoons.
Agreement on XXX Deepfakes & Videogen CSAM
Very few people in AI think this is a serious problem, but everyday Chinese and Americans are scared to death of this type of thing. This is very far down the list for 99% of AI people, but it’s the tip of the spear for shared standards. We should consider this the first move. There is nobody who will object to these two countries reducing the amount of CSAM or deepfakes in the world.
Rare Disease Bounty Program
As an easy point of cooperation, the two AI ecosystems should race to cure five rare diseases by 2030. Think of stuff that affects people in both countries and is too small a pool to develop pharmaceuticals for on a dedicated basis, but with the power of bio and AI, we can get something done quickly and tested across the two borders. We could pay out $100 million per rare disease and offer it as a bounty to AI bio companies in both the US and China; point national pride at a friendly race to see whose biologists can save the most people in a three-year period. I know a hundred million dollars doesn’t develop a drug; it doesn’t have to, the bounty is the trophy, and both countries are already paying for the bio. Agree to cut some red tape around rare disease cures and to hold clinical trials in both countries.
I Could Go Ten More
There is so much that could be done in this space that’s simple, direct, and effective. I am sure there are dozens of people in the D.C. news who are laughing at the naivete that this article displays; all I have to say to you is that a better world is possible.
Trump’s inviting his main geopolitical rival to the White House six weeks before the midterms- your playbook’s not running anymore. Maybe, just maybe, we could find a better one instead.
sources & notes (AI generated)
Helen Toner: clip from a CSIS AI panel, posted August 27, 2026, with Mackenzie Arnold, Matt Pearl, and Aalok Mehta. Her framing in the post: people get stuck on the need for a US-China “deal” on AI, when honestly showing what the US is observing and doing is itself a major lever; she suggests it as a Trump-Xi topic “in a few weeks.”
Dwarkesh and Dylan: Dwarkesh Podcast, Dylan Patel, August 25, 2026, on compute concentration and China’s AI buildout.
RedNote as China’s AI discussion hub: Recode China AI, October 30, 2025. Ongoing translation of Chinese AI writing: ChinAI by Jeffrey Ding. Note: ChinAI works mostly from WeChat essays and journals rather than RedNote/Weibo specifically.
Liang Wenfeng leaked investor call: English translation by Fred Gao, July 23, 2026; the call reportedly took place May 20, 2026. DeepSeek has not confirmed authenticity, and the transcript is speech-to-text with a known error rate.
Singapore safety cluster: Pacific Compute, the author’s project; longer argument in “Towards a Singaporean Safety Cluster,” June 4, 2026.
China’s energy advantage: Al Jazeera, May 28, 2026: 430+ GW of wind and solar added in 2025; first large renewable project wired directly to a data center.
US data center opposition: Gallup, May 13, 2026: 71% of Americans oppose an AI data center in their area, with water and energy use the leading reasons.
Inverter security: the “panels are glass” line leans on the May 2025 Reuters reporting on undocumented communication devices found in Chinese-made solar inverters, not panels (Utility Dive, May 14, 2025).
September US-China AI talks: Reuters via AsiaOne, July 22, 2026. Reporting describes a first, definitional dialogue led by Treasury Secretary Bessent; “grand bargain” is this piece’s characterization.
Xi visit: invitation extended May 14, 2026 for a September 24 White House visit (AP); midterms timing analysis from Julian Gewirtz, August 20, 2026. The 2026 midterms are November 3; September 24 is roughly six weeks before.
US intermediaries monetizing Chinese open models: Rest of World, June 17, 2026: DeepSeek at 17% of Vercel token usage in May 2026.
Precedent for the government taking a cut: CNN, August 11, 2025: Nvidia and AMD remit 15% of China chip sales to the US government.
Semiconductor inputs: CSIS, July 17, 2025: China produces ~98% of global gallium. Specialty gases: Global Semiconductor Research, June 29, 2026 on tungsten hexafluoride shortages after Chinese export controls.
Rare earths leverage: Al Jazeera, October 10, 2025: controls extended to 12 of 17 rare earths.
Haidian/Tsinghua: ChinAI #232, August 7, 2023 on the Wudaokou origins of China’s large-model industry.
Schwarzman Scholars: fully funded master’s at Tsinghua, ~200 scholars a year.
Hotline non-answer: CBS News, February 10, 2023: China declined the crisis-line call during the balloon incident.
Pax Silica: US-led supply chain initiative launched December 12, 2025, championed by Under Secretary of State Jacob Helberg; 24 signatories as of June 2026.
ChinaTalk on AI diplomacy: “Cold War Lessons for US-China AI Dialogue,” May 22, 2024. The original “four years” claim could not be verified; softened to “for years”, with found coverage dating from 2024.
Samm Sacks on data and AI sovereignty: Sinica Podcast, July 24, 2026; she is writing a book on US-China relations through the lens of cross-border data flows.