When Controls Raise the Cost

A capable model released under export restrictions is evidence of technical progress. To understand what it says about the restrictions, we need to ask which hardware made it possible, what resources it required, and whether the result can be repeated.

The question

How have Chinese AI developers responded to U.S. advanced-compute restrictions since October 2022? I examine what each response changes about capability, resource requirements, and dependence on outside technology.

The distinction matters because a response can restore useful capability while leaving a particular restriction consequential. More efficient training may extend the use of an existing chip fleet. Domestic accelerators may replace finished chips while retaining dependencies further up the supply chain. Rerouting an import may restore access without changing the underlying technology.

Research design

The accompanying application compares five cases selected to examine different responses. Each faces the same six questions: restored capability, repeatability at useful scale, resource requirements, outside dependence, exposure to enforcement, and relevance to further development.

The unit of analysis is the response in each case. The cases are not a representative sample of Chinese AI activity, and their outcomes cannot be added together into a measure of policy success. Claims are linked to source passages, with counterevidence and uncertainties recorded alongside the assessment.

Five responses

• Compute efficiency. DeepSeek-V3 demonstrates how model and systems design can stretch an H800 fleet. Its reported training still relied on Nvidia accelerators.
• Domestic substitution. Huawei reports training Pangu Ultra on 6,000 Ascend NPUs. This supports substitution at the finished-accelerator level, while upstream supply and comparable training costs remain unresolved.
• Systems engineering. CloudMatrix384 combines accelerators to support inference. The September review added vendor evidence on deployment and production serving; independent performance and cost comparisons remain limited.
• Rerouted access. Operation Gatekeeper documents diversion of controlled Nvidia chips and disruption of a supply route. The public record does not establish the completed AI workloads or the durability of that access.
• State support. Tax preferences and compute subsidies can shift who bears a cost. The reviewed record does not connect that support to a named recipient’s restored capability or net cost.

What the evidence supports

My current assessment is that the Pangu case supports narrow control-point weakening, conditional on Huawei’s reported result. Across the five cases, the evidence does not establish broad supply-chain independence. That conclusion is narrower than saying the controls have succeeded: continuing dependence does not tell us the size of the restriction’s effect.

A benchmark reflects choices about the model, data, hardware, and implementation. Vendor disclosures and enforcement records answer different questions. I treat each as evidence for the proposition it can support, without extending it to costs, scale, or outcomes the source does not establish.

What would change the assessment

Independently reproduced training results, sustained deployment data, comparable total costs, and documented access to upstream inputs would help distinguish durable substitution from a result that remains difficult to repeat. Acquisition dates and inventories would also clarify whether a response depends on chips obtained before a restriction took effect.

The version 0.2 research record contains 21 evidence records and 35 sources. Its evidence cutoff is 3 September 2026. The methodology explains the classification criteria and the limits of the comparison.