Ding Xuexiang brands Chinese AI developers traitors for wanting chips that work.
It is true that China’s semiconductor engineers are doing genuinely creative work under extraordinary constraint. The strategy Huawei calls “diaohua” — fine-carving older deep-ultraviolet lithography machines, the tools that etch circuit patterns onto silicon wafers, to squeeze performance from technology a generation behind — represents real ingenuity under duress. Huawei and its partners have published some 200 patents in the last two years on production workarounds, error-reduction techniques, and software compensations, sometimes using AI to tune processes that would otherwise require equipment Beijing cannot buy. Meituan recently released an AI model it says rivals Google’s flagship, trained entirely on domestic silicon. These are not trivial accomplishments.
The trouble is that ingenuity under constraint is not the same thing as closing the gap, and the coercion required to keep companies on domestic chips tells you exactly how large that gap remains.
Huawei’s best AI chip delivers roughly a quarter of the computing power of Nvidia’s top offering. China’s total AI computing capacity stands at about 14 percent of the United States’, according to Bernstein estimates. Huawei plans to ship 1.5 million chips this year — double its 2025 volume — and aims to produce 750,000 of its latest Ascend 950 chips. These are big numbers. They are also, in context, insufficient: Moonshot AI, the Beijing startup whose Kimi K3 model this month closely trails top American systems, had to suspend premium sign-ups and scramble for capital after two days because it lacked the computing capacity to serve its own users.
When a company builds a product that competes with the best in the world and cannot afford to let people use it, the bottleneck is not marketing.
The coercion makes the engineering picture legible. Ding’s committee, formed in early 2023, drew on China’s best companies and labs to master each element of the chip supply chain — the same all-out mobilization Beijing used for its atomic bombs, hydrogen bombs, and satellites in the 1960s. Dependence on foreign AI chips has dropped from 90 percent in 2021 to less than 60 percent in 2025, according to Morgan Stanley data, with a target of 25 percent by 2030. The government first allowed sales of Nvidia’s downgraded H20 chip, then privately told companies to halt purchases, then required commitments to use more domestic silicon. At each step, the ratchet tightened — and at each step, the gray market tightened with it. Prices for smuggled Nvidia chips have more than doubled in six months, according to Chinese distributors.
The gray market is the market’s verdict. When companies pay double for illegally imported chips rather than use the ones their government is handing them, they are telling you — in the only language left to them — what the engineering reality actually is. Ding calls this treason. The engineers smuggling hardware call it survival, and the benchmark results are on their side.
The state has answers. A roughly $48 billion semiconductor fund raised in 2024. Salary caps scrapped at equipment maker Naura Technology to attract talent. Huawei’s plan to bundle hundreds of thousands of chips into massive clusters that function as a single computer, leveraging networking expertise to compensate for weaker individual silicon — a strategy that works up to a point, and that point is set by physics, not ambition. The company has outlined a method for stacking circuits to pack more computing power into a single chip footprint, aiming for chips by 2031 that could outperform anything currently in production. Eric Xu, Huawei’s deputy chairman, frames the entire effort as forced self-reliance: “If the U.S. hadn’t forced our country, our company and our industry into a corner, we would never have done something like this.”
He is not wrong about the forcing. U.S. export controls cut off China’s access to extreme-ultraviolet lithography machines — the next-generation equipment needed to manufacture the most advanced chips — and the controls have been effective. They have also been clarifying. The underlying reporting on this campaign closes with the observation that if Washington controls the top technology, “the rest of the world will depend on Washington’s favor to make progress in fields such as medicine, robotics, autonomous driving and military strategy.” That sentence, buried at the bottom of a story about Chinese coercion, is the actual story. Export controls are not just a national-security instrument; they are a dependency-maintenance instrument. The distinction matters because the framing — “security” — obscures the structural reality: whoever controls the semiconductor supply chain controls who gets to compute, and on what terms.
This is not a new pattern. It is the latest instance of what Harold Innis called the staples economy — the political economy organized around a single critical resource whose supply determines who has leverage over whom. Wheat and timber and fur gave way to oil; oil gave way to telecommunications bandwidth; bandwidth now gives way to semiconductor fabrication capacity. The staple changes. The structure of dependency does not. The alternative to both Washington’s leverage and Beijing’s coercion is a supply chain that no single state controls — which is precisely why neither is building one.
Chris Miller, the Tufts historian who wrote the definitive account of semiconductor geopolitics, puts the manufacturing challenge plainly: chip fabrication is “orders and orders of magnitude more complex than almost every other type of manufacturing.” Bernstein’s Qingyuan Lin views the strategy as “more about finding a way to prevent the gap from widening quickly” than about catching up; SemiAnalysis’s Ray Wang estimates that matching the needed lithography technology could take at least a decade and as many as fifty years. Brookings fellow Kyle Chan captures the scale: “It’s trying to catch up with a bullet train.”
The honest assessment is not that China cannot build a domestic semiconductor industry. It is that the industry being built under coercion runs at a quarter the power, serves a fraction of the demand, and depends on the state threatening its own companies to maintain adoption. Moonshot AI, which the White House has accused of stealing from Anthropic, built a model that competes with the best in the world and cannot afford the silicon to run it. That is the shape of the constraint.
Ding calls resistance treason. The engineers smuggling Nvidia chips through the gray market are running the numbers. The numbers do not care about patriotism.