Chip Controls Are Not a Switch: How to Read China’s Race for AI Hardware

US controls on advanced semiconductor technology destined for China are often condensed into an overly neat picture: Washington turns off a tap and Beijing runs out of chips for artificial intelligence. The real rules are more complicated, and the distinction matters for companies, research and public debate.

On December 2, 2024, the Commerce Department’s Bureau of Industry and Security announced a package covering 24 types of semiconductor-manufacturing equipment, three software tools, high-bandwidth memory and 140 additions to the Entity List. The measure aimed to restrict Chinese capacity to produce advanced semiconductors with AI and military applications.

That news does not mean China has lost all computing capacity, nor that it will instantly recover it by making one domestic chip. An AI chip is not an isolated object. It emerges from design, software tools, machinery, materials, fabrication, memory, packaging, networks and programs that use it. A control can strongly affect one link while leaving others available, subject to licensing or replaceable by less efficient alternatives.

What changed — and what the announcement does not say

BIS described the December 2024 controls as a strengthening of earlier rules. Its release lists manufacturing equipment, design and production software, HBM memory and entities added to a list of parties facing requirements. This matters because large-scale AI needs more than processors: it needs fast memory, interconnection, power and systems able to train or run workloads at scale.

But the regulatory text does not say that all US technology in China is banned. Part 744 of the Export Administration Regulations establishes controls linked to particular products, destinations, end uses and end users. That detail changes the analysis. A licensing system may deny many transactions, approve others with conditions or pursue diversion. To know what happens in a case, identify the product, buyer, intended use and applicable rule.

The first lesson is linguistic. “Restricted” is not the same as “eliminated”; “subject to a license” does not mean “approved”; and “one company on a list” does not describe an entire national economy. The right verb comes from the document, not the headline’s intensity.

A chip is not a single capability

When someone says China can or cannot make AI chips, ask: make which part? A useful way to read the claim is to separate the supply chain into seven layers.

First is design: architecture, circuits and electronic design automation software. A firm can design an accelerator without owning a factory. Second is fabrication: turning that design into wafers with extremely precise processes, equipment and materials. Third is memory and packaging: a processor needs to communicate with memory and components at sufficient bandwidth; advanced packaging can determine whether a system scales in practice.

Fourth is system integration: boards, networks, cooling, servers and data centers. Fifth is software: compilers, libraries and tools that turn a model into work executable on particular hardware. Sixth is access to capacity: even with a chip available, an organization needs volume, power, space and reliable operations. Seventh is adoption: using that capacity in products, research or industrial processes in a way that creates value.

Progress in one layer does not remove a bottleneck in another. A company may demonstrate a prototype without producing it at volume. It may make a chip without memory or software that uses it efficiently. It may buy computing capacity without the data or teams prepared to use it. That is why self-sufficiency is not a binary label; it is a collection of capabilities, costs, performance levels and dependencies.

Why equipment and memory are part of the controls

Controls do not focus only on finished processors because industrial advantage is decided both before and after the wafer. BIS included manufacturing equipment, software tools and high-bandwidth memory. The agency itself says HBM is critical to AI training and inference at scale; system performance does not depend only on the number of operations a processor advertises.

This pattern helps readers assess a performance number. A manufacturer may announce operations per second, but practical use depends on numerical precision, memory, interconnection, power consumption, software availability and model type. Two chips with a similar slide-deck number can perform very differently in training, inference or smaller workloads.

The rules also seek to limit diversion through controls tied to destination, end user and exporter knowledge. That is why the documents discuss certifications and records — terms less dramatic than “chip war,” but central to compliance. The object of a rule is not exhausted by silicon; it includes tracking a transaction and requiring accountability from exporters.

A license is neither an open door nor a total wall

BIS license-application rules show why a uniform-wall metaphor is inaccurate: different products, recipients and uses enter different regimes and require information suited to the case.

The existence of licenses should not be used to conclude that access has no consequence. Seeking one adds uncertainty, compliance cost and risk of denial; certain uses or users can be directly barred. Companies make investment decisions over years. If a critical component depends on a regulatory decision, finding local alternatives or redesigning systems can be rational even before a particular transaction is rejected.

The useful question is not “is there a blockade, yes or no?” It is: “which component, at which technical threshold, for which user, under which review, with which alternative available?” That wording may be less convenient, but it allows readers to compare successive announcements without losing the thread.

China’s response also has layers

Chinese public policy shows an explicit intention to integrate AI and production. An official notice published on January 9, 2026 and signed by eight government bodies sets out an AI-plus-manufacturing action. It is evidence of policy priority and coordination. On its own, it is not evidence that a manufacturer has matched a particular foreign technology or can produce it at any scale.

To assess an industrial response, separate announcement, demonstration and scale. An announcement can commit subsidies, public procurement, test centers or training. A demonstration can show that a chip functions on a task. Scale requires reproducible yields, suppliers, factory output, costs, customers and software support. Mixing these stages makes every release look like a final victory.

There is also partial substitution. If access to a leading component is difficult, an organization can use earlier-generation chips, more units, smaller models, more efficient algorithms or distributed services. That can keep applications running, but it does not guarantee the same cost, speed or capability. “It can continue doing AI” and “it can do it at the same technical frontier” are different claims.

A mental table for the next headline

Before sharing a chip story, fill in four boxes. Product: processors, memory, machinery, software or servers? Scope: one sale, a category, a military use, a company or a whole destination? Mechanism: prohibition, license requirement, end-user control or list addition? Result: does the source show design, prototype, production, volume, performance or adoption?

If one box is empty, the conclusion should be more modest. A report naming only a prototype does not prove mass production. A rule governing one category does not explain an entire computing fleet. An industrial-policy release does not measure performance. When all four boxes are documented, a reader can identify a real shift without adopting either government’s narrative.

The race is not decided by one announcement

US controls seek to alter which advanced technology can reach specified uses and users in China; China’s response seeks to reduce dependencies and expand domestic capacity. Both dynamics are real. Neither supports an instant conclusion about who is “winning” AI.

The story worth following is more concrete: which rules are published, which licenses are granted or denied, which products are made with verifiable performance, which bottlenecks remain and which applications are deployed. That is the difference between geopolitics told as a scoreboard and an explanation readers can reuse next year.