On July 17, 2026, Xi Jinping stood in front of the biggest names in Chinese AI and told them, without naming a single country, that nobody should place its own security above everyone else's. He didn't need to name one. Nine days earlier, the U.S. House Foreign Affairs Committee had advanced a bill giving Congress veto power over chip export licenses. Six months earlier, Washington had tightened those licenses again. Two years into an export-control regime built on a single chokepoint (the machines that print the world's most advanced chips) the question circulating this summer isn't really whether China can build its own version of that machine. It's whether the machine was ever the point.
Section 01
The Version Everyone Knows
The containment logic behind the October 2022 export controls was simple and, on its own terms, sound. Cut off extreme ultraviolet (EUV) lithography, the technology used to print transistors smaller than about seven nanometers, and you cut off China's path to frontier-class chips. Only one company in the world builds EUV machines: ASML, in the Netherlands. Block the sale, and the theory was that Chinese fabs would stay stuck a decade behind, no matter how much money or talent Beijing threw at the problem.
Four years on, the headlines read like the theory is failing. DeepSeek's arrival in January 2025 already dented confidence that the gap was as wide as assumed. By April 2026, four Chinese labs released open-weights frontier-class models in a twelve-day window, priced low enough to make the Western equivalents look economically indefensible (a story this site covered at the time in The Flood). By July, food-delivery giant Meituan had trained a 1.6-trillion-parameter model end to end on 50,000 domestic Chinese chips, with zero Nvidia hardware in the pipeline (see The Homegrown Model). Chinese chip production rose 36 percent in the first four months of 2026. And this summer brought a fresh round of coverage suggesting Huawei's chip design division had found a way past the lithography wall entirely. In May, Nvidia CEO Jensen Huang said his own company had effectively conceded China's AI chip market to Huawei.
Put those facts next to each other and the popular reading writes itself: the fence slowed China down for a couple of years, and now Chinese engineers have found a way around it, or maybe even through it. Give it a few more years and the gap closes. Maybe it reverses.
That reading isn't unreasonable. Almost every individual fact in it is true. What it gets wrong is which fact is doing the work.
Section 02
What the Evidence Actually Shows
Start with the machine everyone means when they say China has "cracked lithography." There are two different machines, and 2026 has been a good year for exactly one of them.
A state-backed manufacturer in Shanghai began mass-producing immersion deep ultraviolet (DUV) lithography tools this year. The first units are due to ship to three customers: SMIC, Hua Hong Semiconductor, and memory maker ChangXin Memory Technologies (CXMT). Volumes are modest by global standards, roughly five machines in 2026, climbing to about twenty in 2027, and SMIC has reportedly been testing a similar domestic immersion tool internally since September 2025. For comparison, ASML alone shipped 131 immersion DUV systems worldwide in 2025. This is real progress. It is not, by itself, the chokepoint technology the export controls were built around.
| Technology | Wavelength | Enables | Global maker(s) | China's status |
|---|---|---|---|---|
| Immersion DUV | 193nm | Down to ~7nm with multi-patterning | ASML, Nikon, Canon | Mass production began 2026: ~5 units, first deliveries to SMIC, Hua Hong, CXMT |
| EUV | 13.5nm | Below ~7nm, down to ~2nm | ASML (sole manufacturer) | Prototype only, Shenzhen, 2025 — no commercial units shipping |
Both technologies etch circuit patterns onto silicon by shining light through a mask. DUV light has a wavelength of 193 nanometers; EUV light is 13.5 nanometers, small enough to draw dramatically finer lines. Chipmakers can push DUV down to roughly 7-nanometer transistors by exposing the same layer multiple times, a slower and more error-prone workaround called multi-patterning. Below that, EUV becomes close to mandatory. ASML is the only company on Earth that builds EUV machines, and U.S., Dutch, and Japanese export licensing has kept every one of them out of mainland China since 2019.
EUV itself tells a much smaller story. Researchers in Shenzhen secretly built a prototype EUV machine in 2025, an important research milestone and, by most technical assessments, years away from anything a fab could run in production. As of this writing, no Chinese-built EUV tool is shipping to any customer, domestic or otherwise. The specific claim in circulation this summer, that China cracked EUV, does not hold up against the public reporting.
What actually moved in 2026 was something adjacent to lithography rather than a breakthrough in it. Huawei's semiconductor division unveiled a chip architecture it calls LogicFolding, built on a broader design principle its engineers describe as the Tau Scaling Law. Instead of shrinking transistors, which requires EUV, the approach folds circuits that would normally sit flat on a single layer into a vertical, three-dimensional stack, using lithography tools China already has. The company claims a 53.5 percent increase in transistor density, roughly 238 million additional transistors per square millimeter of chip area. Huawei's public roadmap says its highest-end chips will reach the transistor density associated with a 1.4-nanometer node, a tier that currently requires EUV's most advanced generation, by 2031.
Independent analysts caution against reading the roadmap as a done deal. Cost, power, heat, and system integration remain major challenges for this approach, especially in cloud AI servers running at full load. In the short term, China may narrow the gap with global leaders, but a technology gap with the most advanced nodes will still remain, even as it closes.
None of this is nothing. A country that cannot legally buy the world's best lithography machine just demonstrated it can push transistor density up by more than half through architecture alone, using tools it already owns. But that is a different achievement than the one being reported this summer. Beating a chokepoint by routing around it is not the same as removing it.
Section 03
Why the Gap Between Story and Reality Exists
Three things explain why the "China cracked EUV" reading spread faster than the facts underneath it.
The first is ordinary information friction. China's semiconductor sector is opaque by design, and independent verification of any specific manufacturing claim is genuinely hard. Chinese firms have every incentive to project unstoppable momentum, to investors and to Washington alike, and Western coverage, primed by three years of "China caught up faster than expected" stories, tends to fold ambiguous news into that same pattern. One of the more careful technical assessments published this summer was literally subtitled parsing signal from noise. That subtitle is doing real work.
The second is that the fence itself moved more than once in 2026. A rule published on January 15 tightened total processing power thresholds and shifted licensing for key AI chips, including Nvidia's H200 and AMD's MI325X, from a presumption of denial to case-by-case review. Then, mid-year, the same administration quietly approved higher-tier chip exports and paused further restrictions ahead of a presidential visit to Beijing, prioritizing trade talks over the chip fight, even as the Commerce Department reaffirmed in June that its licensing rules reach Chinese-headquartered firms wherever in the world they operate. Congress moved the opposite direction over the same months: the Chip Security Act, which would require firms to verify that licensed chips stay where they are supposed to go, and the AI OVERWATCH Act, which would hand lawmakers a veto over future export licenses, both advanced out of committee in January. Applied Materials was fined $252 million in February for illegally exporting chipmaking equipment to China. When the people building the fence cannot agree amongst themselves how high it should be, everyone reading the same set of facts from outside is going to reach different conclusions depending on which week they're looking at.
The third reason is the one that matters most, and it is the one the "China cracked EUV" narrative misses entirely. China did not answer the lithography question at the World AI Conference. It changed the question being asked.
Xi's keynote in Shanghai laid out four principles for what he called a just and equitable system of global AI governance: openness and shared innovation, risk-awareness without letting any one country's definition of security override everyone else's, respect for cultural diversity in how AI gets deployed, and multilateralism through the United Nations. He did not claim chip parity. He claimed the argument.
"We should together oppose the practice of overstretching the concept of national security in the field of artificial intelligence, and of placing one's own security above that of other countries."Xi Jinping · Opening Ceremony, 2026 World AI Conference & High-Level Meeting on Global AI Governance · Shanghai, July 17, 2026
The day before that speech, representatives of 29 countries signed the agreement founding the World Artificial Intelligence Cooperation Organization (WAICO) in Shanghai, with UN Secretary-General António Guterres and the deputy prime ministers of Kazakhstan, Laos, and Pakistan in the room. WAICO is the institutional child of the Global AI Governance Initiative Xi first proposed in 2023. Alongside it, China pledged 5,000 AI training and seminar opportunities for developing countries over the next five years, new AI application cooperation centers with ASEAN, the League of Arab States, the African Union, the Community of Latin American and Caribbean States, the Shanghai Cooperation Organisation, and BRICS, and its MAZU AI-powered weather-warning system for 30 more countries.
| Jurisdiction | Legal status, Aug 2026 | Core philosophy | July 2026 milestone |
|---|---|---|---|
| United States | No comprehensive federal law; 1,500+ competing state bills active | Innovation-first, security-driven export controls | Great American AI Act remains a discussion draft |
| China | Global AI Governance Initiative (2023) plus new treaty body | "Development and safety in equal measure"; opposes "overstretching" national security | WAICO founded in Shanghai, 29 nations plus the UN Secretary-General |
| European Union | Binding, risk-tiered AI Act, in force | Precautionary, rights-based, enforceable | Continued phased enforcement, standing apart from both |
Set against WAICO, the American governance picture looks thin. There is still no comprehensive federal AI law. The bipartisan discussion draft known as the Great American AI Act has not passed. Trump's June 2 executive order sets voluntary standards for frontier model security rather than binding ones. More than 1,500 AI-related bills are moving through U.S. statehouses this year, on top of roughly 150 that became state law in 2025, none of them coordinated with each other. The European Union runs a third model entirely, a binding, risk-tiered law that resembles neither Washington's patchwork nor Beijing's new multilateral body. Three governance philosophies, all active in the same summer, mostly not speaking to one another in the same room, except at WAICO, which the United States did not join.
The chip story and the governance story are not two unrelated things that happen to overlap in 2026. They are the same strategy read at two different altitudes. On the manufacturing floor, China is routing around a chokepoint it cannot yet remove. In the rulebook, it is trying to write a new one before Washington finishes arguing about whether it wants one at all.
Section 04
What the More Accurate Version Allows Us to Understand
The instinct that export controls only delayed China, rather than stopping it, holds up reasonably well against the record. The controls didn't prevent China's AI buildout; they redirected it. Blocked from shrinking transistors through EUV, Chinese engineers pushed on packaging and architecture instead, and that redirection is producing real, measurable output: chip production up more than a third in early 2026, and a chip executive at the world's most valuable semiconductor company publicly conceding a slice of market share to a domestic Chinese rival his own firm is legally barred from selling into anyway.
The "just a matter of time" framing needs a date attached before it means very much, though. Huawei's own roadmap puts density parity with an advanced node at 2031, five years out, and independent analysts still expect a gap with the most advanced global nodes to persist even then. Almost every technology gap in history has eventually closed given enough time. That observation is also compatible with a gap lasting long enough to shape everyone reading this today.
The more consequential race this summer might not be measured in nanometers at all. A treaty organization with 29 founding members and a UN Secretary-General in the room is a real asset in the fight over whose AI rules the rest of the world adopts by default, in something like the way whoever controls the developer ecosystem ends up shaping the assumptions baked into billions of everyday AI interactions. The United States does not have an equivalent body to send its own deputy prime ministers to. It has a discussion draft, a voluntary executive order, and fifteen hundred uncoordinated bills moving through fifty statehouses.
Xi's line about a symphony, not a solo performance, was written for a room full of AI researchers and diplomats on the banks of the Huangpu River. Read against the rest of 2026, it sounds less like an appeal to cooperation for its own sake and more like a country that ran the numbers on a fabrication race it could not win outright, and decided the governance race was the one still open to lead.
Buy me a coffee