US chip controls widened the gap but pushed China to build its own supply chain — resolves 2027-07-21

The US restricts sales of advanced AI chips to China, and the debate is whether that works by slowing China down or backfires by forcing China to build its own chip supply chain. The answer decides whether US chipmakers can win back business in China, and whether the future is one global market or two separate systems for the US and China. One camp looks at the capability gap and the other at China’s faster self-sufficiency, so SecondSource judges each on its own timeline.

SecondSource’s reading: the answer depends on which ledger you read.

On the capability gap, the controls work: China’s AI capability trails by about seven months.

As a catalyst, several sources show the controls backfiring. China is speeding up on three fronts (Huawei Ascend chips designed alongside the models, CXMT memory, and domestic chip design software), and Beijing itself has banned purchases of US chips, taking over the push toward two parallel systems.

On smuggling, both sides’ claims need a discount: the median estimate of cumulative smuggling is about 660,000 H100-equivalent chips.

The US switched to allowing sales for a cut (15% on the H20, 25% on the H200), and that tollbooth carries almost no traffic.

SemiAnalysis
“DeepSeek v4... co-designed in part for Huawei…”

SemiAnalysis
“CXMT... likely to become the largest semiconductor…”

SemiAnalysis
“The Big-3 hold over 85% combined market share…”

CSIS (Center for Strategic and International Studies)
“the controls have clearly accelerated Beijing's…”

Bloomberg
“China is preparing to spend around 2 trillion yuan…”

Reviewed at 12 months, in July 2027: (1) China's purchase ban loosening in substance, together with H200 sales picking up in volume, breaks the self-embargo leg and reopens the judgment that the two-track split is irreversible; (2) CXMT ramping HBM3 yield to 30k wspm by the end of 2026 strengthens the catalysis argument, while stockpiles running out around 2027 with installation stalling demotes the domestic stack to an inference-side stack; (3) a Chinese frontier lab outside the Huawei orbit publicly completing a full flagship training run on domestic silicon breaks the training-side leg of the effectiveness argument; (4) Epoch's capability gap narrowing to within 4 months weakens the effectiveness argument, while widening past 14 months strengthens it; (5) RASA becoming law, combined with smuggling estimates revised down, shrinks the leakage denominator. If none of these trigger and the two-track split keeps deepening, the self-embargo reading stands.

What would overturn it

The tests are listed above. Near-term checkpoints: September 1, 2026 for the purchase ban and trade talks, and October 2026 for CXMT's first post-listing disclosure of HBM capacity and yield.

Correction to the settlement rule

Rewritten on 2026-08-03: the "what would overturn it" line used to say only "see the five checks in the receipt" and did not say what counts as the whole call being overturned. It now states the combining rule: any one of the five weakens or breaks its own leg, and the original reading stands only if none happens while the two-system split keeps deepening. The two mid-course checkpoints are kept. The call itself and its due date are unchanged, and no result had been seen. The original wording stays in our ledger and in the public mirror's version history.