What is the case that export controls on China ARE working as intended — what evidence supports the strategy
Are the Chip Restrictions on China Actually Working? What the Evidence Shows
Based on analysis of a 109-node, 329-edge knowledge graph examining the structural case that export controls on China are functioning as intended.
What This Is About
Since 2022, the United States and its allies have been restricting which computer chips and chip-making tools can be sold to China. The official goal is to prevent China from building the most powerful AI systems and military computing infrastructure. Critics say this is not working — China keeps making AI progress, companies like DeepSeek are releasing impressive models, and smuggling routes exist. Supporters say the controls are working exactly as intended.
A large knowledge graph was built to map out all the evidence, arguments, causes, and effects on the “controls are working” side of this question. Think of it like a very detailed diagram with 109 concepts and 329 arrows connecting them. This document explains what the structure of that diagram reveals — what it says, what it assumes, where its weak points are, and what predictions it makes.
The Three Locks on the Door
The most important structural finding is that the graph does not rely on a single argument. Instead, it describes three separate restrictions that work together, like a door with three different locks. The graph calls this the “Three-Layer Chip Stack Denial Architecture.”
The three locks are:
Lock 1: You cannot build the most advanced chips without special machines. The Netherlands makes the only machines in the world (called EUV lithography machines, made by a company called ASML) that can manufacture the smallest, fastest chips. Those machines are not being sold to China. Without them, China’s main chip manufacturer (SMIC) is stuck making chips that are a couple generations behind — roughly equivalent to where the industry was in 2019. This ceiling appears hard to break without the machines.
Lock 2: The most powerful AI systems need a special type of memory chip called HBM. HBM is only made by three companies: Samsung, SK Hynix, and Micron — all in allied countries. China’s attempt to build its own version (by a company called CXMT) has so far failed to reach production quality. The graph records a specific failure: CXMT’s most advanced attempt did not reach production, and even if it eventually does, those chips would still depend on manufacturing equipment China cannot access.
Lock 3: Connecting thousands of chips together requires advanced packaging technology. Modern AI training does not just use one chip — it uses thousands of chips working together. The technology to connect them (called CoWoS, made primarily by TSMC in Taiwan) is also under allied control. China’s attempt to replicate this at scale has not succeeded.
The key insight: each lock is independently maintained. If China somehow broke through one, the other two would still limit overall capability. This is why the graph treats this as a structural feature rather than a single point of failure.
Why the Gap Gets Bigger Over Time (Not Smaller)
The graph’s central dynamic claim is that the capability gap is not just present — it is widening. Here is the mechanism, explained simply.
Imagine two students studying for an exam. One student has access to all the latest textbooks, the fastest computer, and a full library. The other student has older books and a slower computer. Every year, new topics are added to the exam. Because the first student can study faster, the gap between them grows each year, even if both students are improving. Falling behind in a race where the finish line keeps moving is different from falling behind in a static competition.
The graph identifies at least seven separate processes that all feed into this widening gap: the scaling of AI systems (bigger models need more chips), improvements in chip manufacturing that China cannot access, the software ecosystem built around Nvidia chips (called CUDA, which took 15 years to build and which Chinese chips do not support), the memory chokepoint, the packaging chokepoint, the economic flywheel Nvidia benefits from, and US domestic investment in chip production through the CHIPS Act. Seven independent inputs into one widening trend is a structurally robust claim — it does not depend on any single argument being true.
Why DeepSeek Did Not Disprove This
When DeepSeek released its AI model in early 2025 and it performed nearly as well as American models, many observers concluded that export controls were not working. The graph’s structure tells a different story, and it is counterintuitive.
DeepSeek’s impressive model was built using chips that were purchased before the controls took full effect — pre-ban stockpiles. When DeepSeek tried to train a more advanced successor model on Chinese domestic chips (Huawei’s Ascend), the training failed. The graph records this as direct empirical evidence that the controls are working at the level where it matters most.
There is a distinction the graph treats as central: the difference between training an AI model and running (or “inferring” from) an AI model. Training requires enormous amounts of compute over months. Running a trained model is much cheaper. The controls are primarily a training constraint. DeepSeek demonstrated you can be clever about running models efficiently — but the chips required to train the next generation of frontier models are the ones China cannot easily get.
Think of it like this: imagine China has a talented chef who can make an excellent dish using old ingredients from a stockpile. Everyone says: “Look, the ingredient restrictions are not working!” But what they are missing is that the chef is working through the last of the pre-restriction pantry. When those ingredients run out, the question is what can be made with what is currently available domestically. The graph predicts this test comes in late 2026, when those stockpiles are expected to be depleted.
Why Smuggling Is Actually Evidence It Is Working
One of the most counterintuitive structural findings is what the graph does with smuggling. When investigators find chips being routed illegally through Southeast Asia, or shell companies buying restricted hardware, the intuitive reading is: “the controls are failing, people are getting around them.” The graph inverts this.
The graph’s logic: if chips were freely available, there would be no need to smuggle them. Criminal organizations only build expensive, risky routes to move goods when there is genuine scarcity and genuine demand. The existence of a black market is proof the legal market is constrained. The graph records multiple documented smuggling networks and enforcement operations as behavioral evidence — not of failure, but of the controls creating real scarcity.
This is a non-obvious structural point that does not show up in most news coverage.
The Four Self-Reinforcing Loops
The graph identifies four feedback loops — processes where effects circle back and reinforce their own causes. This is what makes the structure dynamic rather than static.
The most important: the wider the capability gap gets, the more geopolitical pressure there is to maintain the controls, which keeps the controls in place, which keeps widening the gap. Countries that might consider relaxing their cooperation are reminded by the growing gap of why they signed up in the first place.
A second loop runs through economics: Nvidia’s chips are the best in the world partly because China cannot buy the restricted versions, which lets Nvidia charge premium prices to US and allied AI companies, which gives Nvidia more revenue to develop the next generation of chips, which keeps the gap wide. Export controls created an economic incentive structure that now sustains itself.
A third loop is subtle: when China gets more efficient at AI (like DeepSeek), this actually increases total demand for chips, rather than relieving the shortage. This is a known economic phenomenon called the Jevons Paradox — fuel-efficient cars led people to drive more, not less, and total fuel consumption went up. More efficient AI models make AI more useful and affordable, which means more organizations want to deploy AI at scale, which means more chip demand. The controls are not on demand; they are on supply. Supply cannot meet the new, higher demand.
Where the Argument Has Weak Points
The graph is honest about its vulnerabilities. There are five recorded tensions.
The most significant: the entire structure depends on allied cooperation holding. Japan, the Netherlands, South Korea, and Taiwan all maintain parts of the restriction regime. If any one of them negotiates a carve-out — whether for economic reasons, diplomatic pressure, or because a new government changes priorities — the mechanism weakens. The graph records this as the single most important unmodeled risk. It does not have a scenario where an ally defects.
The second: the current US administration has shown willingness to trade technology access for economic concessions. There is documented evidence of a reversal on design software controls (EDA tools) in exchange for a rare-earth minerals deal. The graph records this as a live contradiction with the “one-way ratchet” claim — the idea that restrictions only escalate, never reverse.
The third: China’s progress on domestic memory chips (CXMT) is real, even if not yet competitive. The graph records both the failure of CXMT’s advanced attempt and the ongoing race. The outcome of this race — whether CXMT achieves production-quality HBM in the next two to three years — will materially affect whether the three-lock structure holds.
What the Graph Predicts
The graph’s structure generates several testable predictions:
- Chinese frontier AI model benchmarks should fall further behind US models through 2026–2027, as the pre-ban chip stockpiles are consumed.
- Chinese efficiency improvements (like DeepSeek) should be followed by more chip procurement attempts, not fewer.
- CXMT will achieve partial HBM production but not full parity, and the outcome will be visible in production volume announcements.
- If the MATCH Act (legislation to codify restrictions on older-generation chip-making machines) fails, China’s chip production ceiling should measurably degrade within 18–24 months.
The Bottom Line
The graph’s most important structural findings, in plain language:
The argument is layered, not single-threaded. It does not depend on one claim being true. Three separate technology chokepoints, each independently maintained, each with its own causal logic.
The gap widens over time by design. The controls are not trying to freeze the current gap — they are trying to ensure that the gap grows, by maintaining US and allied access to advancing technology while restricting Chinese access.
The counterevidence (DeepSeek, smuggling, benchmarks) is already incorporated and addressed. The graph’s structure explains why each of these appears to challenge the thesis but actually supports it under closer examination — particularly the training-versus-inference distinction.
The thesis is time-bounded, not permanent. The graph frames the strategic goal as holding through a specific window (roughly 2022–2030) tied to when AI systems might reach transformative capability thresholds. It does not claim controls will work forever — only that they need to hold long enough.
The single most important unmodeled vulnerability is allied defection. Everything else in the graph — the technology mechanisms, the economic loops, the enforcement waves — depends on Japan, the Netherlands, South Korea, and Taiwan maintaining coordination. The graph has no answer for what happens if one of them breaks ranks.