Compute EvidenceChips · AI · Policy
DatasetGitHub

Research by Moayd Ghazzawi

Progress. Cost.
Dependency.

Five responses to chip restrictions. The same six questions for each.

Evidence through Read the research method
05Response cases
21Evidence records
35Cited sources33 primary
08Policy milestones

Explore the cases.

Capability, costs, and unresolved questions.

05 / 05 cases
More filters
Case 01 / 2024–2025

DeepSeek-V3 on Nvidia H800

Algorithm and systems co-design stretched a restricted H800 fleet, but did not replace the hardware chokepoint.

Adaptation + cost penaltyModerate confidence
01

Capability restored

Demonstrated

A 671B-parameter mixture-of-experts model delivered strong author-reported benchmark results; on METR’s autonomy suite it was comparable to Claude 3.5 Sonnet (Old) while trailing newer frontier models.

02

Repeatable scale

Partial

One 2,048-H800 training run is documented. Weights and inference code are public, but the full training stack and dataset are not.

03

Extra resources

Material

The disclosed 2.788 million GPU-hours exclude prior research, ablations, data, staff, acquisition, energy, and failed work. Low-level co-design was substantial; neither this total nor the engineering effort isolates the extra cost caused by controls.

04

Outside dependency

High

Training relied on Nvidia H800 accelerators. Hardware replenishment is the relevant foreign-controlled dependency; using foreign-origin open-source software alone does not establish an enforceable chokepoint.

05

Enforcement resilience

Mixed

Algorithms and weights diffuse readily; installed chips cannot be recalled. Future hardware refresh, HBM, and accelerator replenishment remain enforceable chokepoints.

06

Next generation

Unclear

This run establishes useful 2024-era capability. It does not by itself demonstrate independent, repeatable access to the hardware needed for later generations.

Still unresolved

  • H800 acquisition dates and inventory
  • Total program and electricity cost
  • Full-run reproducibility
  • Hardware used for later DeepSeek models
  • Whether the efficiency choices were caused by export controls

Evidence

View all 5
DeepSeek reports pre-training a 671B-parameter MoE on 14.8 trillion tokens and using 2.788 million H800 GPU-hours across all disclosed training stages.DeepSeek-AIThe widely repeated $5.576 million figure is a rental-equivalent training-compute estimate, not an all-in development cost.DeepSeek-AIOn METR’s autonomy suite, DeepSeek-V3 was comparable to Claude 3.5 Sonnet (Old), while trailing newer frontier models.METR

Compare the outcomes.

Choose a dimension. Follow the evidence.

Compare capability restored across all research cases
Response caseCapability restoredOverall judgment
DeepSeek-V3 efficiency2024–2025Demonstrated

A 671B-parameter mixture-of-experts model delivered strong author-reported benchmark results; on METR’s autonomy suite it was comparable to Claude 3.5 Sonnet (Old) while trailing newer frontier models.

Adaptation + cost penalty

Moderate confidence

5 evidence records
Pangu domestic training2025Demonstrated

Huawei reports training a 718B-parameter model from scratch on Ascend and says its system supports all training stages, with self-reported results near DeepSeek-R1 on several tasks.

Narrow control-point weakening

Moderate confidence

4 evidence records
CloudMatrix systems engineering2025–2026Demonstrated

The vendor-authored evaluation serves a 671B-parameter DeepSeek-R1 model with little reported INT8 accuracy loss across 16 benchmarks.

Adaptation + cost penalty

Moderate confidence

7 evidence records
Rerouted H100/H200 access2024–2025Unknown

The chips could support advanced training and inference, but public court and agency materials do not identify a completed model or workload.

Circumvention, not independence

Moderate confidence

3 evidence records
State support evidence gap2023–2026Unlinked

National tax preferences and local compute vouchers support the sector, but no reviewed award record ties them to a specific restored capability in this MVP.

Insufficient evidence

Low confidence

3 evidence records

Judgments are provisional. “Narrow control-point weakening” applies to a targeted control point and does not establish independence across the supply chain.

The policy timeline.

Dates distinguish announcements, effective rules, and suspensions. Expand a milestone for its scope and original sources.

  1. 01
    Effective in phases

    Foundational advanced-compute and semiconductor controls

    Created advanced-computing ECCNs, China-wide license requirements, supercomputer end-use controls, foreign-direct-product rules, and specified U.S.-person restrictions.

    Scope and caveat. Different provisions took effect on 7, 12, and 21 October; later rules changed thresholds and scope.

    Original sourcesU.S. Bureau of Industry and Security: Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items; Supercomputer and Semiconductor End Use; Entity List Modification (opens in new tab)U.S. Bureau of Industry and Security: Commerce Implements New Export Controls on Advanced Computing and Semiconductor Manufacturing Items to the PRC (opens in new tab)
  2. 02
    Effective 17 Nov 2023

    Performance-density and anti-circumvention update

    Recalibrated chip thresholds, added a performance-density test, expanded destination and end-user reach, and tightened equipment and U.S.-person controls.

    Scope and caveat. The Federal Register publication date is 25 October, despite a conflicting date on one legacy BIS page.

    Original sourcesU.S. Bureau of Industry and Security: Implementation of Additional Export Controls: Certain Advanced Computing Items; Supercomputer and Semiconductor End Use; Updates and Corrections (opens in new tab)U.S. Bureau of Industry and Security: Export Controls on Semiconductor Manufacturing Items (opens in new tab)
  3. 03
    Effective with split compliance dates

    HBM, manufacturing-equipment, software, and entity package

    Added controls on high-bandwidth memory, more semiconductor-manufacturing equipment, ECAD/TCAD software, additional foreign-produced items, and 140 Entity List entries.

    Scope and caveat. Limited exceptions and delayed compliance mean this was not a universal HBM or equipment ban.

    Original sourcesU.S. Bureau of Industry and Security: Foreign-Produced Direct Product Rule Additions, and Refinements to Controls for Advanced Computing and Semiconductor Manufacturing Items (opens in new tab)U.S. Bureau of Industry and Security: Additions and Modifications to the Entity List; Removals From the Validated End-User (VEU) Program (opens in new tab)U.S. Bureau of Industry and Security: Commerce Strengthens Export Controls to Restrict China’s Capability to Produce Advanced Semiconductors for Military Applications (opens in new tab)
  4. 04
    AI Diffusion requirements not enforced; formal rescission pending

    AI Diffusion Framework issued

    Added worldwide regional-stability licensing for the most advanced chips and controls on certain closed-weight model weights, with country tiers and exceptions.

    Scope and caveat. Commerce announced non-enforcement and intended rescission before compliance began; preexisting D:5/Macau-headquarters controls remain enforced worldwide.

    Original sourcesU.S. Bureau of Industry and Security: Framework for Artificial Intelligence Diffusion (opens in new tab)U.S. Bureau of Industry and Security: Commerce Announces Rescission of AI Diffusion Rule and Strengthens Chip-Related Controls (opens in new tab)U.S. Government Accountability Office: U.S. Department of Commerce, Bureau of Industry and Security—Applicability of the Congressional Review Act to the Rescission of the Artificial Intelligence Diffusion Rule (opens in new tab)U.S. Bureau of Industry and Security: Guidance Regarding Enforcement of License Requirements for Advanced Computing Items for Entities Headquartered in Country Group D:5 and Macau (opens in new tab)
  5. 05
    Effective; transition deadlines extended

    Foundry and packaging due-diligence rule

    Required front-end fabricators and OSATs to investigate designers and customers and created trusted-designer and approved-OSAT pathways.

    Scope and caveat. A 2026 amendment extended some designer-status deadlines through 31 December 2026 without removing the core regime.

    Original sourcesU.S. Bureau of Industry and Security: Implementation of Additional Due Diligence Measures for Advanced Computing Integrated Circuits; Amendments and Clarifications; and Extension of Comment Period (opens in new tab)U.S. Bureau of Industry and Security: Extension of Authorized Integrated Circuit (IC) Designer Status and Application Deadline To Become an Approved IC Designer (opens in new tab)
  6. 06
    Effective 31 Dec 2025

    Validated End-User authorizations revoked for three China fabs

    Removed Intel Dalian, Samsung China Semiconductor, and SK hynix Semiconductor China from the VEU list, requiring individual licenses for covered items.

    Scope and caveat. BIS stated it would distinguish existing operations from expansion; TSMC was not named in this rule.

    Original sourcesU.S. Bureau of Industry and Security: Revocation of Validated End-User Authorizations in the People’s Republic of China (opens in new tab)U.S. Bureau of Industry and Security: Commerce Closes Export-Control Loophole for Foreign-Owned Semiconductor Fabs in China (opens in new tab)
  7. 07
    Suspended through 09 Nov 2026

    Listed-entity affiliate rule issued, then suspended

    Extended certain end-user restrictions to entities owned 50 percent or more by listed parties and imposed ownership diligence.

    Scope and caveat. The automatic affiliates rule was not operative on this project’s 3 September 2026 cutoff date.

    Original sourcesU.S. Bureau of Industry and Security: Expansion of End-User Controls To Cover Affiliates of Certain Listed Entities (opens in new tab)U.S. Bureau of Industry and Security: One Year Suspension of Expansion of End-User Controls for Affiliates of Certain Listed Entities (opens in new tab)
  8. 08
    Effective

    Conditional case-by-case review for H200-class exports

    Moved qualifying direct U.S. exports of H200, MI325X, and similar chips to China or Macau from a presumption of denial to conditioned case-by-case review.

    Scope and caveat. This was not license-free access or general approval; reexports, transfers, volume, end use, testing, and customer conditions remained restrictive.

    Original sourcesU.S. Bureau of Industry and Security: Revision to License Review Policy for Advanced Computing Commodities (opens in new tab)U.S. Bureau of Industry and Security: Department of Commerce Revises License Review Policy for Semiconductors Exported to China (opens in new tab)

The research in two minutes

What the record supports.

Method and limitations

Chinese AI progress is real, but progress and control failure are not synonyms. DeepSeek-V3 shows that efficient model and systems design can stretch a restricted Nvidia H800 fleet. Huawei’s Pangu Ultra shows that thousands of domestic Ascend accelerators can complete a very-large-model training run, weakening the narrow claim that access to top U.S. accelerators is indispensable. CloudMatrix384 shows that system design can recover useful inference performance by pooling many weaker devices. A U.S. prosecution shows that restricted H100 and H200 chips also moved through a diversion network. Each result carries a different implication. The Huawei cases still leave fabrication, high-bandwidth memory, power, networking, and economics unresolved. DeepSeek remained dependent on controlled foreign hardware. Smuggling is an enforcement failure, not technological independence. The strongest current reading is that the controls impose friction and shift costs rather than create an absolute barrier—and that only repeated, independently verified scale can show whether that friction is durable.

Explore the evidence Download research data