Export controls impact Chinese AI compute

Last updated

Tracker summary: US export controls constrain Chinese access to frontier AI chips, but China adapts through domestic alternatives and workarounds.

At a glance

  • Assessment: Emerging
  • Confidence in assessment: 65%
  • Predicted timing: Ongoing
  • Primary source: Mid 2026: China Wakes Up, archived PDF p. 6; August 2027: The Geopolitics of Superintelligence, archived PDF p. 17

What AI 2027 Predicted

Tracker summary: The scenario describes a dynamic where US export controls genuinely constrain Chinese access to frontier compute (especially NVIDIA GPUs), but China adapts through domestic chip alternatives and creative workarounds.

Scenario source locators: Mid 2026: China Wakes Up: archived PDF, p. 6; August 2027: The Geopolitics of Superintelligence: archived PDF, p. 17.

How We Track This

Evaluate four claims separately: US capability leadership, tightening of controls, effective cutoff from advanced Nvidia chips, and Chinese domestic substitution. A domestic deployment win cannot prove that exports were fully cut off, and a policy restriction cannot prove a capability lead.

Operational definitions (tracker interpretation): “Constrain” requires a documented reduction or delay in usable frontier-chip access attributable to controls, not merely a rule on the books. For each comparison, record the buyer or deployment, chip type, observation period, requested/licensed/delivered quantities or delivery dates, and the evidence connecting a shortfall or delay to the restriction; distinguish licensing from other supply constraints. A documented buyer- or deployment-level causal effect can establish that narrower constraint claim; a national counterfactual is not required. An aggregate national constraint claim additionally requires comparable delivered-compute measurements accounting for licensed imports, third-country access and domestic substitutes. “Adapt” requires documented use of domestic hardware or an alternative access route for an AI workload; record whether it supports inference or training. Certification, orders and shipment plans alone do not establish deployed adaptation or frontier-training parity.

Application to the evidence below: The May 15 H200 report documents non-delivery after approvals, but does not isolate the cause or quantify aggregate lost compute. The BIS notice establishes licensing scope, not a measured impact. The Huawei announcement and GLM-5 report support inference adaptation; they do not establish domestic frontier training. Thus inference adaptation has evidence, but the cited access reports do not isolate a restriction-driven reduction or delay even for a specific buyer. The narrower causal constraint claim and any national aggregate effect remain unresolved.

We monitor:

  • US export control policy updates and enforcement actions
  • Chinese domestic chip development (Huawei Ascend, etc.)
  • Training runs confirmed to use domestic Chinese hardware
  • NVIDIA revenue reports and China-specific GPU production decisions

Current Evidence

Affiliates Rule: status checked November 7, 2026: The November 2025 BIS final rule suspended the expansion of listed-entity controls to majority-owned affiliates from November 10, 2025 through November 9, 2026. A final-rule text filed for public inspection November 5, 2026 provides for extension through January 10, 2027, effective November 9, 2026 and scheduled for publication that day. At the November 7 source check, the extension’s effective date had not yet arrived. The text specifies November 9 as the effective date, not the November 5 filing date. This is not a uniform tightening trajectory, and does not eliminate other advanced-computing licensing restrictions.

Emerging (65%): Active restrictions and domestic inference adaptation are documented, but the cited evidence does not yet establish the causal access effect required to confirm the combined claim. GLM-5’s technical report documents inference optimization across seven domestic chip platforms, including Huawei Ascend. That supports domestic deployment, not entirely domestic training. GLM-5

April 29, 2026: domestic procurement: Reuters’ original reporting, syndicated by AsiaOne, describes Chinese technology firms seeking Huawei chips after DeepSeek V4. The manufacturer’s April 24 announcement independently confirms its DeepSeek V4 adaptation and Ascend inference tooling, not completed production volumes. The previously quoted shipment plan is not treated as verified production.

May 15, 2026: H200 access: CNA/Reuters reports Greer’s statement that export controls were not a major topic in the talks, and reports that deliveries had not followed earlier approvals. This is attributed reporting about that date, not a current shipment census or evidence of a permanent ban.

May 21, 2026: circumvention: AP’s report on the Taiwan investigation describes allegations of forged documents used to route Nvidia-equipped servers to China. An investigation is not a conviction or a measure of aggregate illicit supply.

May 26, 2026: domestic certification: The China Information Technology Security Evaluation Center’s original notice and Annex 1 lists nine AI training/inference chip products at security/reliability Level I, including Ascend 310 and 910. Certification is not a performance benchmark and does not establish deployed quantities or Nvidia parity.

May 31, 2026: offshore purchasing entities: BIS guidance confirms continued licensing requirements for covered advanced-computing items for entities headquartered in, or ultimately parented from, Country Group D:5 or Macau even when located elsewhere, subject to applicable exceptions. It clarifies a pre-existing requirement, rather than creating a new blanket ban or requiring bona fide data-center operators to halt ongoing use.

Counterevidence & Limitations

SubcheckDated evidence and verdictLimit
US AI leadNo direct comparable frontier-capability measurement established by the trade-policy evidence hereNot established by the controls or certification evidence
Controls tightenSeptember 29, 2025 affiliate expansion, then November 10, 2025 suspension and the announced November 2026 extensionMixed direction, not continuous tightening
China is cut off from advanced Nvidia chipsMay 15, 2026 H200 reporting and May 21, 2026 alleged server smuggling describe non-delivery and alleged circumventionLicensing, deliveries and illicit access differ; a full cutoff is not established
Domestic alternativesHuawei’s April 24, 2026 DeepSeek V4 adaptation announcement describes Ascend inference support; the May 26 certification list documents domestic productsDeployment/adaptation, not proof of all-domestic frontier training or performance parity
  • Inference compatibility, certification and procurement reports do not establish domestic frontier-training performance or the amount of usable compute lost because of controls.
  • Restrictions, licensing decisions and enforcement change over time. Evidence must be dated rather than summarized as uniformly tightening policy.

What Would Change Our Assessment

  • Confirm the scoped claim: Document a buyer- or deployment-level reduction or delay in usable frontier-chip access attributable to controls, alongside documented domestic or alternative-route adaptation. Record the buyer, chip, period, quantities or delivery dates and causal evidence. National totals are required only for a national aggregate claim.
  • Weaken the constraint assessment: Verified licensed deliveries that close the same documented access shortfall, or removal of its identified policy barrier, weaken that case. Approval alone is not delivery; full Chinese compute parity is not required to count as counterevidence.
  • Keep subclaims separate: US capability leadership, complete cutoff, domestic frontier training and aggregate lost compute each require their own evidence.

Update History

DateUpdate
2026-11-092026-11-09 · Correction: status Confirmed (90%) → Emerging (65%). The test accepts buyer- or deployment-level effects attributable to controls without requiring a national counterfactual. Domestic inference adaptation is documented, but the cited access reports do not isolate the restriction-driven shortfall; policy existence alone does not confirm the combined claim.
2026-11-07Evidence reviewed: Anthropic’s June 30 notice documents lifting the June model controls, while the chip-access evidence does not isolate aggregate controls-attributable lost compute. Adaptation and the existence of restrictions remain distinct from their measured effect. Correction: the December 2025 “mixed” label was not a defined status and does not establish a transition to one. Correction: the Affiliates Rule was suspended from November 10, 2025 through November 9, 2026; the November 5 public-inspection extension takes effect November 9 through January 10, 2027. Restrictions, measured effects and inference adaptation remain separate evidence checks.
2026-09-06Current evidence and assessment explanation reviewed. Clarified measurement scope, source interpretation and limitations.
2026-06-08BIS issued guidance clarifying that advanced-computing license requirements apply to entities headquartered in, or ultimately parented from, Country Group D:5 or Macau even when the entity is located abroad. This reinforces the active constraint side of the export-control dynamic while leaving the China adaptation evidence unchanged.
2026-06-01Added reporting that China certified nine domestic AI processors for government procurement. This reinforces the confirmed pattern of export controls pushing Chinese procurement toward domestic alternatives, while not changing the assessment of the remaining performance gap.
2026-05-25AP reported that Taiwanese authorities are investigating three people suspected of using forged documents to smuggle servers containing advanced Nvidia chips to China. This adds another incremental enforcement/circumvention data point supporting the confirmed export-control dynamic.
2026-05-23Added CNA/Reuters reporting that semiconductor export controls were not a major topic in May US-China talks and that no H200 deliveries had occurred despite earlier US clearances for some Chinese buyers. This modestly reinforces the active-constraint picture while preserving the caveat that policy is fluid.
2026-05-04Reuters reported Chinese cloud and tech firms scrambling to secure Huawei AI chips after DeepSeek V4, with Huawei planning roughly 750,000 Ascend 950PR units in 2026 and mass production beginning in April. This strengthens evidence for both constraint and adaptation under export controls. Confidence adjusted 0.85 → 0.90.
2026-04-13Mixed signals on export control effectiveness. CFR analysis (Dec 2025) argues controls ARE working — Huawei “falling further behind,” recommends maintaining restrictions (CFR). However, Trump admin approved H200 exports to China (Dec 2025), partially loosening controls. DeepSeek V4 on Huawei chips shows constrained self-sufficiency. Reports of Nvidia Blackwell chip smuggling for V4 training suggest controls partially circumvented. Alibaba opening datacenter with 10K own chips + China Telecom — domestic alternatives growing. Overall: controls still constraining but effectiveness gradually eroding. No status or confidence change.
2026-04-06DeepSeek V4 confirmed to run exclusively on Huawei chips (Reuters, The Information). Alibaba, ByteDance, and Tencent ordered hundreds of thousands of Huawei chips in preparation. This represents the most significant evidence of China achieving domestic AI chip self-sufficiency — a direct response to US export controls as the scenario predicted. Meanwhile, the US trade war is creating pressure on America’s AI dominance, with Trump administration policies complicating the export control regime (WebAndITNews). The dual dynamic (export controls constrain + China adapts) continues to play out exactly as predicted. No status or confidence change. 2026-11-07 · Source correction: The WebAndITNews interpretation is withdrawn as support. Reuters, April 3 supports only the narrower pre-release report of Huawei compatibility, code adaptation and advance chip orders, not confirmed exclusivity, self-sufficiency or a demonstrated loss of US AI dominance. Historical wording and links remain for provenance.
2026-03-23Supermicro co-founder and two others charged with smuggling $2.5B in Nvidia AI chip servers to China via Southeast Asia shell companies (CNN, NBC News). FDD analysis finds Chinese firms routinely circumvent controls by purchasing through US-based entities (FDD). Meanwhile ByteDance secured Nvidia chip access for global ops outside China (Breitbart). Confidence adjusted 0.80 → 0.85 — the dual dynamic of enforcement and circumvention is playing out more dramatically than expected. 2026-11-07 · Source correction: The Breitbart citation is superseded by Reuters, March 13, 2026, reporting WSJ’s account of ByteDance’s planned Malaysian deployment through Aolani Cloud. Aolani said it complies with export controls; overseas access is not itself evidence of unlawful circumvention. Historical wording and links remain for provenance.
2026-03Clear evidence of both impact (Chinese labs face compute constraints) and adaptation (Huawei Ascend, smuggling networks, cloud workarounds). Dynamic playing out as predicted.
2025-12Trump reversed H200 chip export controls to China, allowing Nvidia to resume H200 sales in exchange for 25% of revenue going to the US government. The H200 is approximately 6× more powerful than the H20. This is a significant policy reversal from the “small yard, high fence” regime the AI 2027 essay assumed would persist. Status downgraded to mixed.
2025-10BIS 50% Rule takes effect October 1. GAIN AI Act passes in Senate NDAA amendment (October 9), requiring chipmakers to prioritize U.S. customers before exporting. Export control regime tightening procedurally, though Chinese labs continue releasing frontier-competitive models. 2026-11-07 · Correction: The original BIS rule became effective September 29, 2025, not October 1; September 30 was publication day.
2025-09BIS 50 Percent Rule announced (effective September 29), extending export control restrictions to all companies 50%+ owned by blacklisted entities — closing a key enforcement gap. DeepSeek reveals R1 training cost of $294,000, demonstrating China can build competitive reasoning models with dramatically lower compute — counterevidence to the “compute scarcity constrains China” mechanism.
2025-07Updated US export controls further restrict AI chip sales to China. NVIDIA develops China-specific variants within compliance limits.
2025-04US imposes indefinite export license requirement on Nvidia H20 chips to China (April 9). Nvidia expects $5.5B in charges. The H20 was the highest-performance chip permitted under prior rules; this closes that route. Strengthens the prediction’s core mechanism.