Continuing DUV lithography exports risks handing America's billion-unit chip capacity edge to China
A US think tank estimates that if advanced DUV lithography machines remain available, China's chip manufacturing could match America's by the 2030s; tightening controls now would lock in a 100-fold capacity gap.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Think tanks should anchor forward-looking research in supply-chain reality
Khan says that when he was researching semiconductor policy at Georgetown's CSET, AI policy was not yet anyone's concern, leaving ample time for forward-looking research aimed one to two years ahead. Now, with daily AI news, no one is willing to spend time on this work. He believes CTS must fill this gap in two ways: develop policy ideas that will actually matter one to three years out, and ground research in real supply-chain details—from manufacturing equipment to tokens to robotics—because physical-world constraints are predictable and reward deep investigation in ways abstract policy debates don't.
— Saif KhanAmerica's eight-month lead hides the mechanisms nobody can explain
It's widely said that American models lead Chinese ones by about eight months, but Khan says almost no one explains the mechanism behind this figure—how much comes from training larger models, how much from Chinese companies' improvements via distillation or reinforcement learning, or how US acceleration or deceleration would transmit to China. He cites an example: when GLM 5.3 flash was revealed to be trained on tens of trillions of tokens, commentators immediately declared ‘export controls are useless’, but this judgment lacks systematic explanation.
— Saif KhanDUV remains China's final unbreakable chokepoint
Khan explains that beyond EUV lithography, DUV immersion lithography is the second-best technology currently used to manufacture advanced chips—Huawei's 7-nanometer chips depend on it. Only Japan and the Netherlands, both US allies, can produce these machines, making them America's most effective leverage against China, which cannot manufacture them. DUV is the hardest link in the entire semiconductor equipment supply chain for China to achieve self-sufficiency in. The report estimates that China won't have commercially viable domestic DUV machines until the mid-2030s, making it the final bottleneck in the supply chain to overcome.
— Saif KhanExport control loopholes quietly doubled China's DUV inventory
When export controls took effect in 2023, lower-tier DUV immersion machines were still permitted for sale to China, when inventory stood at approximately 170–180 units. Since then, this inventory has roughly doubled. The report models a scenario: if DUV exports were completely cut off today, by the mid-2030s the US could produce about one billion H100-equivalent AI chips annually, while China's domestic capacity would be locked at tens of millions; but if these machines continue available, the capacity ranges could overlap, effectively surrendering America's entire lead.
— Saif KhanRobotics, not software, will be AI's next inflection point
Khan argues that the true turning point in AI over the next few years may not be recursive self-improvement in software, but automation of the physical world—AI combined with robotics, scaled deployment of robots learning to manufacture, eventually forming self-sufficient supply chains. This direction is nearly blank territory in Washington; he says the policy world has a handful of people seriously studying the robotics-AI intersection, with almost no equivalent work in China. This could be the next critical node in US-China AI competition.
— Saif KhanAI policy's unresolved divide: safety-first vs. diffusion-focused worldviews
Khan categorizes AI policy disagreements into two mindsets: a ‘safety-first’ view holds that the critical actors are the few AI companies developing frontier models that could pose catastrophic risks or shift global power, and such change often occurs outside existing institutions; a ‘diffusion-focused’ view instead holds that what matters is how AI adoption co-evolves with existing institutions, with wealth redistribution and AI's role in the economic system at the core. He emphasizes these are not opposed but complementary—even preventing bioweapon risks leaves questions unresolved if we don't answer what role humans play in an AI economy.
— Saif KhanCoding AI agents compressed think-tank staffing without sacrificing depth
Khan says that over the past nine months, programming tools like Claude Code and Codex have allowed small teams like CTS to query large datasets and conduct deep analysis rapidly; the technical depth the DUV report achieves owes greatly to this. He believes that work once requiring fifteen to twenty people at a think tank can now be done by a smaller team with equivalent output density and analytical depth, which is why CTS has invested heavily in such tools.
— Saif KhanRSI is plausible; how it reshapes the growth curve remains uncertain
When asked if he's more skeptical of recursive self-improvement than Dario Amodei, Khan denies being a skeptic: he believes AI companies are fully capable of assessing whether they're achieving AI research automation, and it's quite plausible that much AI research will be significantly automated within years. His real uncertainty lies elsewhere—whether automating AI research merely sustains the current capability growth curve or pushes it steeper. This year has seen milestones like autonomous cyber attacks, agentic capabilities, and solving Navier-Stokes equations—a millennium-prize problem. But Khan says the real-world impact of research automation remains hard to judge.
— Saif KhanIn their own words · checked verbatim
Right now, it's still the case that two countries, two U.S. allies, Japan and the Netherlands, are the sole producers of these tools. And it is really the most effective choke point that we still have available. China can't make it.
Saif Khan13:21
like someone's written for 10 years where like I learned something on every page, like I genuinely learned something on every page of this report.
Jordan14:21
So if you look back to around the time when those controls were first introduced, there were probably like, um, upper, like 100, like maybe like 170, 180 tools installed in China around that time. But because we still allow some of these tools to go, that stockpile has, uh, roughly doubled since then.
Saif Khan17:30
If we stop the flow of DOV, DOV immersion tools today, we project that China is probably going to be capped out at something like 10 million, uh, per year in their indigenous production.
Saif Khan17:30
the real way that the world is going to transform is through physical automation and robotics and marrying AI capabilities with robotics and starting to deploy robots at scale, having them learn how to do manufacturing and start to ultimately, even at some point, perhaps create self-sustaining supply chains.
Saif Khan31:18
you have the most important actors really being frontier AI companies that are developing really powerful AI systems that are going to pose catastrophic risks or just kind of change the balance of power in the world in a significant and very discreet way, oftentimes outside of existing institutions.
Saif Khan33:27
I would not at all say I'm a skeptic. The companies are in a really good position to know whether they are automating their AI research. And I think they actually plausibly in the next couple of years could succeed in substantially automating a big portion of it.
Saif Khan55:36
Figures
| US model capability lead | Approximately 8 months | 7:04 |
| China's DUV machine inventory when 2023 controls took effect | Approximately 170–180 units | 17:30 |
| Current China DUV inventory vs. 2023 | Approximately doubled | 17:30 |
| China domestic AI chip annual capacity under export cutoff scenario | Approximately 10 million units/year | 18:34 |
| China domestic chip capacity under full-autonomy scenario | Tens of millions units/year scale | 18:34 |
Glossary
- DUV immersion lithography
- Second only to EUV among advanced lithography techniques; Huawei uses it to fabricate 7-nanometer chips.
- EUV lithography
- The leading lithography technology for manufacturing the most advanced chips today.
- RSI (recursive self-improvement)
- The concept of AI automating its own research improvements and potentially accelerating capability growth.
- MATCH Act
- A US Congressional bill under discussion to restrict semiconductor equipment exports to China.
- IFP (Institute for Progress)
- The policy research institution funding and supporting CTS, the think tank launched by this episode's guest.
How to listen
Those tracking US-China AI and semiconductor export controls, chip supply chain policy, or interested in how think tanks use empirical methods for policy research.
The opening discussion of CTS's founding and talent competition can be skipped.