How the judgments are made.

01 / The question

What did the response actually change?

Since October 2022, which Chinese responses have actually weakened U.S. controls on advanced AI chips, and which have only made Chinese AI development more expensive, slower, or more dependent on outside technology?

Export controls can impose costs while Chinese developers continue to release capable models. A developer may use more chips, a larger network, more electricity, substantial engineering, a fragile supply route, or state support. Each mechanism has different implications for the control being tested.

Progress after a rule cannot, by itself, establish what would have happened without it. The record must show what capability returned, whether it can be repeated, and which dependencies remain. Reported resource requirements do not measure the extra cost caused by controls without a credible comparison.

02 / The common test

Six questions for every case.

1. 01

   What AI capability was restored, and on which tasks?

2. 02

   Can it work repeatedly at useful scale?

3. 03

   What extra hardware, electricity, money, engineering, or time does it require?

4. 04

   Does it reduce dependence on technology controlled by the United States or its allies?

5. 05

   Could a realistic enforcement change break the workaround?

6. 06

   Can it support the next generation of development, or only current needs?

03 / The interpretation

A judgment has a defined scope.

Each classification applies to the control point examined. Finished-chip substitution does not establish independence in memory, manufacturing equipment, fabrication, packaging, or optics.

04 / Sources

Start close to the recorded claim.

1. 01

   Rules, court records, regulatory filings, and official policy documents

2. 02

   Technical disclosures, model cards, repositories, and reproducible benchmarks

3. 03

   Peer-reviewed research and clearly labelled preprints

4. 04

   Independent technical evaluation and specialist analysis

5. 05

   Careful reporting used only where primary evidence is unavailable

A vendor paper is primary evidence of what the vendor reported. Independent reproduction, comparable workload assumptions, and complete cost measurements are separate questions.

Each record includes a claim, source, dates, exact location, quotation or data, counterevidence, confidence, and remaining uncertainty. Citation checks connect each record to its cases. Source interpretation still requires human judgment.

05 / Limits

What this record cannot establish.

• Exact chip inventories, acquisition dates, manufacturing yields, energy use, subsidies, and total engineering costs are usually private.
• The Pangu and CloudMatrix performance evidence is vendor-authored and has not been independently reproduced at full scale.
• Benchmark gains mix the effects of algorithms, data, hardware, and implementation; they do not isolate the causal effect of export controls.
• Enforcement cases reveal detected schemes, not the prevalence of undetected diversion or the capability ultimately produced.
• Rules changed repeatedly. Each response must be matched to the rule and license policy in force at that time.
• This study is selective, not exhaustive. It favors cases with inspectable public evidence and labels gaps instead of estimating hidden quantities.

06 / Revising the finding

Evidence that would change the assessment.

• Repeated, independently audited frontier-scale training on domestically fabricated accelerators with disclosed HBM, yield, power, cost, and volume would move the domestic-substitute finding toward broad control failure.
• Metered CloudMatrix deployments showing cost and power parity—not only per-TFLOPS efficiency—would weaken the current cost-penalty judgment.
• Evidence that diversion or offshore access reliably supplies next-generation compute after realistic customer, ownership, and data-center checks would show a durable enforcement failure.
• Audited firm-level data showing no material delay, redesign, inventory drawdown, or cost increase after a matched rule change would challenge the claim that controls still impose friction.
• Conversely, documented project delays, unmet accelerator demand, falling training scale, or sustained upstream shortages would strengthen the cost-imposition finding.