{ "meta": { "title": "When Controls Raise the Cost", "subtitle": "China’s Responses to U.S. Advanced-Compute Restrictions", "question": "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?", "version": "0.1.0", "status": "Defensible MVP", "evidenceThrough": "2026-09-03", "accessed": "2026-09-03", "currentFinding": "One response weakens a narrow control point; none proves system-wide independence.", "summary": "The public record shows meaningful adaptation, one narrow case of functional substitution, and real enforcement leakage. It does not show that Chinese developers have escaped the cost, scale, or supply-chain consequences of U.S. controls.", "conversationSummary": "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." }, "responseTypes": [ { "id": "domestic-substitutes", "label": "Domestic substitutes", "definition": "Chinese-designed accelerators or production inputs replace controlled foreign technology." }, { "id": "compute-efficiency", "label": "Compute efficiency", "definition": "Algorithms and software obtain more useful capability from a fixed hardware budget." }, { "id": "systems-engineering", "label": "Systems engineering", "definition": "Networks, software, and scale combine weaker devices into a more capable system." }, { "id": "rerouted-access", "label": "Rerouted access", "definition": "Restricted chips or compute services are reached through intermediaries or offshore infrastructure." }, { "id": "state-support-stockpiling", "label": "Stockpiling & state support", "definition": "Installed inventories, subsidies, procurement, or tax support absorb or delay the constraint." } ], "classifications": [ { "id": "genuine-weakening", "label": "Genuine weakening", "definition": "The relevant capability is restored at useful scale while dependence falls at the targeted control point." }, { "id": "adaptation-cost", "label": "Adaptation + cost penalty", "definition": "The capability is real, but added hardware, energy, time, engineering, or outside dependence remains." }, { "id": "circumvention", "label": "Circumvention, not independence", "definition": "Access bypasses enforcement but remains tied to controlled foreign technology and a disruptable route." }, { "id": "insufficient-evidence", "label": "Insufficient evidence", "definition": "The public record cannot connect the response to restored capability, scale, or a particular rule." } ], "methodology": { "sixQuestions": [ "What AI capability was restored, and on which tasks?", "Can it work repeatedly at useful scale?", "What extra hardware, electricity, money, engineering, or time does it require?", "Does it reduce dependence on technology controlled by the United States or its allies?", "Could a realistic enforcement change break the workaround?", "Can it support the next generation of development, or only current needs?" ], "sourceHierarchy": [ "Rules, court records, regulatory filings, and official policy documents", "Technical disclosures, model cards, repositories, and reproducible benchmarks", "Peer-reviewed research and clearly labelled preprints", "Independent technical evaluation and specialist analysis", "Careful reporting used only where primary evidence is unavailable" ], "limitations": [ "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 MVP is selective, not exhaustive. It favors cases with inspectable public evidence and labels gaps instead of estimating hidden quantities." ], "whatWouldChangeMind": [ "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." ] }, "timeline": [ { "id": "evt-2022-10-07-foundation", "date": "2022-10-07", "displayDate": "07 Oct 2022", "title": "Foundational advanced-compute and semiconductor controls", "summary": "Created advanced-computing ECCNs, China-wide license requirements, supercomputer end-use controls, foreign-direct-product rules, and specified U.S.-person restrictions.", "status": "Effective in phases", "caveat": "Different provisions took effect on 7, 12, and 21 October; later rules changed thresholds and scope.", "sourceIds": [ "fr-2022-21658", "bis-2022-10-07" ] }, { "id": "evt-2023-10-17-tightening", "date": "2023-10-17", "displayDate": "17 Oct 2023", "title": "Performance-density and anti-circumvention update", "summary": "Recalibrated chip thresholds, added a performance-density test, expanded destination and end-user reach, and tightened equipment and U.S.-person controls.", "status": "Effective 17 Nov 2023", "caveat": "The Federal Register publication date is 25 October, despite a conflicting date on one legacy BIS page.", "sourceIds": [ "fr-2023-23055", "fr-2023-23049" ] }, { "id": "evt-2024-12-02-hbm-equipment", "date": "2024-12-02", "displayDate": "02 Dec 2024", "title": "HBM, manufacturing-equipment, software, and entity package", "summary": "Added controls on high-bandwidth memory, more semiconductor-manufacturing equipment, ECAD/TCAD software, additional foreign-produced items, and 140 Entity List entries.", "status": "Effective with split compliance dates", "caveat": "Limited exceptions and delayed compliance mean this was not a universal HBM or equipment ban.", "sourceIds": [ "fr-2024-28270", "fr-2024-28267", "bis-2024-12-02" ] }, { "id": "evt-2025-01-13-ai-diffusion", "date": "2025-01-13", "displayDate": "13 Jan 2025", "title": "AI Diffusion Framework issued", "summary": "Added worldwide regional-stability licensing for the most advanced chips and controls on certain closed-weight model weights, with country tiers and exceptions.", "status": "AI Diffusion requirements not enforced; formal rescission pending", "caveat": "Commerce announced non-enforcement and intended rescission before compliance began; preexisting D:5/Macau-headquarters controls remain enforced worldwide.", "sourceIds": [ "fr-2025-00636", "bis-2025-05-13-diffusion", "gao-b-337935", "bis-guidance-2026-05-31" ] }, { "id": "evt-2025-01-16-foundry-diligence", "date": "2025-01-16", "displayDate": "16 Jan 2025", "title": "Foundry and packaging due-diligence rule", "summary": "Required front-end fabricators and OSATs to investigate designers and customers and created trusted-designer and approved-OSAT pathways.", "status": "Effective; transition deadlines extended", "caveat": "A 2026 amendment extended some designer-status deadlines through 31 December 2026 without removing the core regime.", "sourceIds": [ "fr-2025-00711", "fr-2026-06851" ] }, { "id": "evt-2025-09-02-veu", "date": "2025-09-02", "displayDate": "02 Sep 2025", "title": "Validated End-User authorizations revoked for three China fabs", "summary": "Removed Intel Dalian, Samsung China Semiconductor, and SK hynix Semiconductor China from the VEU list, requiring individual licenses for covered items.", "status": "Effective 31 Dec 2025", "caveat": "BIS stated it would distinguish existing operations from expansion; TSMC was not named in this rule.", "sourceIds": [ "fr-2025-16735", "bis-2025-veu" ] }, { "id": "evt-2025-09-29-affiliates", "date": "2025-09-29", "displayDate": "29 Sep 2025", "title": "Listed-entity affiliate rule issued, then suspended", "summary": "Extended certain end-user restrictions to entities owned 50 percent or more by listed parties and imposed ownership diligence.", "status": "Suspended through 09 Nov 2026", "caveat": "The automatic affiliates rule was not operative on this project’s 3 September 2026 cutoff date.", "sourceIds": [ "fr-2025-19001", "fr-2025-19846" ] }, { "id": "evt-2026-01-15-license-review", "date": "2026-01-15", "displayDate": "15 Jan 2026", "title": "Conditional case-by-case review for H200-class exports", "summary": "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.", "status": "Effective", "caveat": "This was not license-free access or general approval; reexports, transfers, volume, end use, testing, and customer conditions remained restrictive.", "sourceIds": [ "fr-2026-00789", "bis-2026-01-13" ] } ], "cases": [ { "id": "case-deepseek-v3", "title": "DeepSeek-V3 on Nvidia H800", "shortTitle": "DeepSeek-V3 efficiency", "date": "2024-12-27", "period": "2024–2025", "responseTypes": [ "compute-efficiency" ], "classification": "adaptation-cost", "confidence": "Moderate", "dependencyStatus": "Foreign-controlled hardware", "finding": "Algorithm and systems co-design stretched a restricted H800 fleet, but did not replace the hardware chokepoint.", "tests": [ { "id": "capability", "label": "Capability restored", "status": "Demonstrated", "answer": "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." }, { "id": "scale", "label": "Repeatable scale", "status": "Partial", "answer": "One 2,048-H800 training run is documented. Weights and inference code are public, but the full training stack and dataset are not." }, { "id": "penalty", "label": "Extra resources", "status": "Material", "answer": "The disclosed 2.788 million GPU-hours exclude prior research, ablations, data, staff, acquisition, energy, and failed work. Extensive low-level co-design was required." }, { "id": "dependency", "label": "Outside dependency", "status": "High", "answer": "Training used Nvidia H800 accelerators and a software/networking stack centered on controlled U.S. technology." }, { "id": "enforcement", "label": "Enforcement resilience", "status": "Mixed", "answer": "Algorithms and weights diffuse readily; installed chips cannot be recalled. Future hardware refresh, HBM, and accelerator replenishment remain enforceable chokepoints." }, { "id": "future", "label": "Next generation", "status": "Unclear", "answer": "The run proves useful 2024-era capability. Later DeepSeek releases do not publicly disclose enough hardware detail to show that the same mechanism scales forward." } ], "evidenceIds": [ "ev-deepseek-compute", "ev-deepseek-cost", "ev-deepseek-independent", "ev-deepseek-dependency", "ev-deepseek-repro" ], "uncertainties": [ "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" ] }, { "id": "case-pangu-ultra", "title": "Pangu Ultra MoE on 6,000 Ascend NPUs", "shortTitle": "Pangu domestic training", "date": "2025-05-07", "period": "2025", "responseTypes": [ "domestic-substitutes" ], "classification": "genuine-weakening", "confidence": "Moderate", "dependencyStatus": "Reduced; upstream unresolved", "finding": "A full large-model training run on Ascend weakens direct accelerator denial, but not upstream controls or the cost-of-scale argument.", "tests": [ { "id": "capability", "label": "Capability restored", "status": "Demonstrated", "answer": "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." }, { "id": "scale", "label": "Repeatable scale", "status": "Partial", "answer": "One 6,000-NPU training run and a 32-card serving recipe are public. The full pretraining pipeline has not been independently reproduced." }, { "id": "penalty", "label": "Extra resources", "status": "High", "answer": "The run used 6,000 accelerators, about 19 trillion tokens, 30 percent model FLOP utilization, and extensive recomputation, swapping, topology, operator, and simulation work." }, { "id": "dependency", "label": "Outside dependency", "status": "Reduced", "answer": "Ascend and Huawei’s CANN/MindSpeed layers are domestic, but the disclosed stack also uses open-source PyTorch, Transformers, and NVIDIA-origin Megatron-LM. Public evidence does not resolve HBM, fabrication equipment, packaging, optics, or the provenance of every die." }, { "id": "enforcement", "label": "Enforcement resilience", "status": "Mixed", "answer": "Domestic accelerators reduce leverage at the finished-chip layer. HBM, EDA, fab tools, foundry services, and packaging remain alternative control points." }, { "id": "future", "label": "Next generation", "status": "Promising, unverified", "answer": "The run shows 2025-scale training is possible; there is no independent evidence of parity with the 2026 frontier or sufficient fleet volume for repeated runs." } ], "evidenceIds": [ "ev-pangu-training", "ev-pangu-codesign", "ev-pangu-weights", "ev-pangu-upstream" ], "uncertainties": [ "Independent capability parity", "Fleet volume and repeatability", "Energy, time, yield, and capital cost", "HBM and fabrication supply-chain provenance" ] }, { "id": "case-cloudmatrix", "title": "CloudMatrix384 serving DeepSeek-R1", "shortTitle": "CloudMatrix systems engineering", "date": "2025-06-19", "period": "2025", "responseTypes": [ "systems-engineering" ], "classification": "adaptation-cost", "confidence": "Moderate", "dependencyStatus": "Reduced; upstream unresolved", "finding": "Pooling Ascend devices restores useful large-model inference, while shifting the constraint into chip count, networking, power, and custom engineering.", "tests": [ { "id": "capability", "label": "Capability restored", "status": "Demonstrated", "answer": "The vendor-authored evaluation serves a 671B-parameter DeepSeek-R1 model with little reported INT8 accuracy loss across 16 benchmarks." }, { "id": "scale", "label": "Repeatable scale", "status": "Partial", "answer": "Authors describe a 384-NPU production-grade architecture, but performance tests used a 256-NPU subset and fixed benchmark configurations." }, { "id": "penalty", "label": "Extra resources", "status": "High", "answer": "The complete system uses 384 NPUs, 192 CPUs, a large optical fabric, disaggregated memory, custom collectives, scheduling, quantization, and fault recovery." }, { "id": "dependency", "label": "Outside dependency", "status": "Reduced", "answer": "Huawei designed the accelerator, CPU, and CANN layers, while the disclosed stack interoperates with PyTorch, TensorFlow, ONNX, and Kubernetes; HBM, fab tools, packaging, and high-speed optics remain unresolved." }, { "id": "enforcement", "label": "Enforcement resilience", "status": "Mixed", "answer": "Pooling weakens a per-chip threshold. Aggregate HBM, networking, energy, manufacturing equipment, and production volume remain observable constraints." }, { "id": "future", "label": "Next generation", "status": "Current inference", "answer": "Evidence supports vendor-cloud testbed inference, not independently verified production service or frontier pretraining." } ], "evidenceIds": [ "ev-cloudmatrix-architecture", "ev-cloudmatrix-performance", "ev-cloudmatrix-limits", "ev-cloudmatrix-power", "ev-pangu-upstream" ], "uncertainties": [ "Independent reproduction", "Metered energy and total cost", "Comparable benchmark conditions", "Upstream component provenance" ] }, { "id": "case-gatekeeper", "title": "Operation Gatekeeper GPU diversion network", "shortTitle": "Rerouted H100/H200 access", "date": "2025-12-08", "period": "2024–2025", "responseTypes": [ "rerouted-access" ], "classification": "circumvention", "confidence": "Moderate", "dependencyStatus": "Foreign access route", "finding": "A guilty plea and seizures prove material diversion, but not a durable supply route or the AI capability produced by the chips.", "tests": [ { "id": "capability", "label": "Capability restored", "status": "Unknown", "answer": "The chips could support advanced training and inference, but public court and agency materials do not identify a completed model or workload." }, { "id": "scale", "label": "Repeatable scale", "status": "Disrupted", "answer": "The pleaded conduct covered exported and attempted shipments worth at least $160 million; authorities also seized more than $50 million in technology and cash." }, { "id": "penalty", "label": "Extra resources", "status": "Material", "answer": "The route required straw purchasers, intermediaries, mislabeled hardware, false paperwork, warehouses, logistics companies, and large cross-border payments." }, { "id": "dependency", "label": "Outside dependency", "status": "Complete", "answer": "The route depended on Nvidia accelerators and foreign procurement and logistics. It created no domestic technological independence." }, { "id": "enforcement", "label": "Enforcement resilience", "status": "Low", "answer": "Customer checks, beneficial-ownership review, shipment inspection, data-center verification, and financial tracing can disrupt this route; this case was disrupted." }, { "id": "future", "label": "Next generation", "status": "Fragile", "answer": "Diversion can provide batches of current chips, but repeated access to each new generation remains exposed to enforcement changes." } ], "evidenceIds": [ "ev-gatekeeper-volume", "ev-gatekeeper-method", "ev-bis-diversion-guidance" ], "uncertainties": [ "Number of GPUs delivered to end users", "Ultimate recipients and tasks", "Prevalence of undetected networks", "Net price and delay penalty" ] }, { "id": "case-state-support", "title": "Tax preferences and local compute subsidies", "shortTitle": "State support evidence gap", "date": "2025-06-21", "period": "2023–2026", "responseTypes": [ "state-support-stockpiling" ], "classification": "insufficient-evidence", "confidence": "Low", "dependencyStatus": "Unknown", "finding": "Public policy proves support exists; it does not show which controlled capability a named recipient restored or at what net cost.", "tests": [ { "id": "capability", "label": "Capability restored", "status": "Unlinked", "answer": "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." }, { "id": "scale", "label": "Repeatable scale", "status": "Unknown", "answer": "Program ceilings are public. Firm-level receipts, additionality, and repeat use are not." }, { "id": "penalty", "label": "Extra resources", "status": "Unknown", "answer": "Subsidies may transfer cost to the state without removing real hardware, energy, or time requirements." }, { "id": "dependency", "label": "Outside dependency", "status": "Unknown", "answer": "A tax preference or compute voucher does not disclose the origin of chips, memory, networking, or fabrication inputs." }, { "id": "enforcement", "label": "Enforcement resilience", "status": "Indirect", "answer": "Fiscal support can absorb higher prices but cannot itself replace a denied component or protect a foreign access route." }, { "id": "future", "label": "Next generation", "status": "Unknown", "answer": "The policy horizon extends several years, but future technical outcomes cannot be inferred from budget authority alone." } ], "evidenceIds": [ "ev-state-tax", "ev-state-hangzhou", "ev-state-stockpile-caveat" ], "uncertainties": [ "Named recipients and award amounts", "Causal effect on firm decisions", "Hardware origin", "Capability produced per yuan of support" ] } ], "evidence": [ { "id": "ev-deepseek-compute", "caseIds": [ "case-deepseek-v3" ], "claim": "DeepSeek reports training a 671B-parameter MoE on 14.8 trillion tokens with 2.788 million H800 GPU-hours.", "sourceId": "deepseek-v3-report", "sourceType": "Technical preprint", "publicationDate": "2024-12-27", "accessDate": "2026-09-03", "quotationOrData": "“2.788M H800 GPU hours.”", "location": "Abstract; section 1; Table 1", "counterevidence": "The run and benchmarks are author-reported and have not been independently reproduced end to end.", "confidence": "High", "remainingUncertainty": "The public report does not establish acquisition dates, electricity use, or total program cost.", "relationship": "supports" }, { "id": "ev-deepseek-cost", "caseIds": [ "case-deepseek-v3" ], "claim": "The widely repeated $5.576 million figure is a rental-equivalent training-compute estimate, not an all-in development cost.", "sourceId": "deepseek-v3-report", "sourceType": "Technical preprint", "publicationDate": "2024-12-27", "accessDate": "2026-09-03", "quotationOrData": "Prior research and ablation costs “are not included.”", "location": "Section 1, note immediately below Table 1", "counterevidence": "The narrow figure remains useful for comparing the disclosed final run if its scope is stated.", "confidence": "High", "remainingUncertainty": "R&D, acquisition, depreciation, staff, data, networking, and failed-run costs are undisclosed.", "relationship": "qualifies" }, { "id": "ev-deepseek-independent", "caseIds": [ "case-deepseek-v3" ], "claim": "On METR’s autonomy suite, DeepSeek-V3 was comparable to Claude 3.5 Sonnet (Old), while trailing newer frontier models.", "sourceId": "metr-deepseek-v3", "sourceType": "Independent evaluation", "publicationDate": "2025-02-12", "accessDate": "2026-09-03", "quotationOrData": "METR reported comparable autonomy performance to Claude 3.5 Sonnet (Old), with weaker results than newer frontier models.", "location": "Opening findings; General Autonomous Capabilities", "counterevidence": "METR used basic elicitation and warned that looping and other failures might understate or distort capability.", "confidence": "Moderate", "remainingUncertainty": "Benchmark performance is task- and elicitation-dependent and does not establish frontier parity.", "relationship": "supports" }, { "id": "ev-deepseek-dependency", "caseIds": [ "case-deepseek-v3" ], "claim": "Nvidia identified H800 as a product covered by the October 2023 China licensing restrictions.", "sourceId": "nvidia-2024-10k", "sourceType": "Regulatory filing", "publicationDate": "2024-02-21", "accessDate": "2026-09-03", "quotationOrData": "Restrictions on H800 shipments took effect on 23 October 2023.", "location": "Item 1, Government Regulations—Global Trade", "counterevidence": "DeepSeek may have acquired its H800 fleet before the restriction took effect; acquisition timing is not public.", "confidence": "High", "remainingUncertainty": "The size and provenance of DeepSeek’s remaining controlled-hardware inventory are unknown.", "relationship": "supports" }, { "id": "ev-deepseek-repro", "caseIds": [ "case-deepseek-v3" ], "claim": "The public repository enables inference and inspection of model weights but not reproduction of the full training run.", "sourceId": "deepseek-v3-repository", "sourceType": "Company repository", "publicationDate": "2024-12-26", "accessDate": "2026-09-03", "quotationOrData": "The release includes weights and inference examples for Nvidia, AMD, and Ascend.", "location": "Repository README, model downloads and inference sections", "counterevidence": "Open weights make broad deployment and independent downstream testing possible.", "confidence": "High", "remainingUncertainty": "Training data, the full HAI-LLM stack, and exact full-run configuration are unavailable.", "relationship": "qualifies" }, { "id": "ev-pangu-training", "caseIds": [ "case-pangu-ultra" ], "claim": "Huawei reports a 718B-parameter MoE run on 6,000 Ascend 910B NPUs and says its system supports all training stages.", "sourceId": "pangu-ultra-paper", "sourceType": "Technical preprint", "publicationDate": "2025-05-07", "accessDate": "2026-09-03", "quotationOrData": "“MFU of 30.0%” on 6,000 Ascend NPUs.", "location": "Paper abstract; sections 1 and 2.2", "counterevidence": "The run is vendor-reported and has no independent full-training audit or reproduction.", "confidence": "Moderate", "remainingUncertainty": "Run duration, energy, capital cost, fleet availability, and yield are not disclosed.", "relationship": "supports" }, { "id": "ev-pangu-codesign", "caseIds": [ "case-pangu-ultra" ], "claim": "The model and execution plan were reshaped around NPU constraints, with cumulative interventions raising reported throughput 58.7 percent.", "sourceId": "pangu-ultra-paper", "sourceType": "Technical preprint", "publicationDate": "2025-05-07", "accessDate": "2026-09-03", "quotationOrData": "6,000 NPUs; 30% MFU; hierarchical all-to-all; recomputation and host swapping.", "location": "Section 4.5, Table 5", "counterevidence": "Thirty percent MFU still implies substantial unused peak compute, and no comparable counterfactual run is public.", "confidence": "High", "remainingUncertainty": "The incremental burden caused specifically by controls cannot be separated from normal large-model engineering.", "relationship": "qualifies" }, { "id": "ev-pangu-weights", "caseIds": [ "case-pangu-ultra" ], "claim": "The official model card says Pangu Ultra was trained from scratch; the repository publishes roughly 1.47 TB of weights and documents a minimum 32-card Ascend serving setup.", "sourceId": "pangu-ultra-repository", "sourceType": "Company repository", "publicationDate": null, "accessDate": "2026-09-03", "quotationOrData": "62 weight shards; four Atlas 800T A2 nodes; 32 64GB cards.", "location": "Model card overview and sections 1, 4.1, and 4.4; file tree", "counterevidence": "Serving reproducibility does not reproduce pretraining or validate the disclosed training fleet.", "confidence": "High", "remainingUncertainty": "The complete training corpus, code, and run configuration are not public.", "relationship": "supports" }, { "id": "ev-pangu-upstream", "caseIds": [ "case-pangu-ultra", "case-cloudmatrix" ], "claim": "BIS assessed that listed Ascend 910-series chips likely involved U.S. design technology, software, or equipment subject to export rules.", "sourceId": "bis-gp10-2025", "sourceType": "Government guidance", "publicationDate": "2025-05-13", "accessDate": "2026-09-03", "quotationOrData": "The illustrative list names Ascend 910B and 910C.", "location": "Pages 1–2, summary, illustrative list, and applicable regulations", "counterevidence": "This is an enforcement assessment, not a bill-of-materials audit of the Pangu fleet or CloudMatrix installation.", "confidence": "Moderate", "remainingUncertainty": "The precise origin and legal status of each chip and upstream component are not public.", "relationship": "qualifies" }, { "id": "ev-cloudmatrix-architecture", "caseIds": [ "case-cloudmatrix" ], "claim": "CloudMatrix384 connects 384 Ascend NPUs and 192 Kunpeng CPUs through an all-to-all Unified Bus architecture.", "sourceId": "cloudmatrix-paper", "sourceType": "Technical preprint", "publicationDate": "2025-06-19", "accessDate": "2026-09-03", "quotationOrData": "384 NPUs, 192 CPUs, disaggregated prefill, decode, and caching.", "location": "Abstract; sections 3.2–3.4", "counterevidence": "The paper is authored by Huawei and SiliconFlow researchers and is not peer reviewed.", "confidence": "High", "remainingUncertainty": "Public fleet volume and production availability outside the disclosed deployment are unknown.", "relationship": "supports" }, { "id": "ev-cloudmatrix-performance", "caseIds": [ "case-cloudmatrix" ], "claim": "The disclosed DeepSeek-R1 test achieved 1,943 decode tokens per second per NPU at 49.4 ms, below one profiled H800 result but above another published H800 baseline.", "sourceId": "cloudmatrix-paper", "sourceType": "Technical preprint", "publicationDate": "2025-06-19", "accessDate": "2026-09-03", "quotationOrData": "1,943 tokens/s/NPU at 49.4 ms; reported efficiency 1.29 tokens/s/TFLOPS.", "location": "Section 5.2, Tables 2–4", "counterevidence": "Tests used 256 of 384 NPUs; decode assumed a 70% MTP acceptance rate; batch sizes and implementations differed; and the highest prefill result assumed idealized expert load balancing.", "confidence": "Moderate", "remainingUncertainty": "Like-for-like cost, power, reliability, and hardware-count comparisons are unavailable.", "relationship": "supports" }, { "id": "ev-cloudmatrix-limits", "caseIds": [ "case-cloudmatrix" ], "claim": "Tighter latency requirements sharply reduced reported per-NPU throughput, exposing the workload sensitivity of the comparison.", "sourceId": "cloudmatrix-paper", "sourceType": "Technical preprint", "publicationDate": "2025-06-19", "accessDate": "2026-09-03", "quotationOrData": "Throughput fell from 1,943 to 538 tokens/s/NPU as the target tightened from about 50 ms to 15 ms.", "location": "Section 5.2, Table 4", "counterevidence": "The system still met the stated tighter latency target and maintained useful throughput.", "confidence": "High", "remainingUncertainty": "Real traffic mixes, uptime, and long-run service cost are not independently reported.", "relationship": "qualifies" }, { "id": "ev-cloudmatrix-power", "caseIds": [ "case-cloudmatrix" ], "claim": "A specialist estimate argues that matching a GB200 NVL72-class system requires substantially more Ascend devices, optical components, and power.", "sourceId": "semianalysis-cloudmatrix", "sourceType": "Specialist analysis", "publicationDate": "2025-04-16", "accessDate": "2026-09-03", "quotationOrData": "Estimated 4.1 times the system power of a GB200 NVL72.", "location": "Opening comparison; CloudMatrix 384 System Architecture", "counterevidence": "The estimate is not metered or audited, and the compared systems differ in chip count and configuration.", "confidence": "Low", "remainingUncertainty": "No public independent energy, cooling, or total-cost measurement validates the exact multiplier.", "relationship": "qualifies" }, { "id": "ev-gatekeeper-volume", "caseIds": [ "case-gatekeeper" ], "claim": "A company and its owner pleaded guilty in a scheme involving at least $160 million of exported and attempted H100 and H200 shipments.", "sourceId": "doj-gatekeeper", "sourceType": "Government enforcement", "publicationDate": "2025-12-08", "accessDate": "2026-09-03", "quotationOrData": "Conduct spanned October 2024 to May 2025; more than $50 million was seized.", "location": "Release summary; paragraph beginning ‘According to court documents’", "counterevidence": "The $160 million figure combines completed and attempted exports; as of the 8 December 2025 release, not every person described had been convicted.", "confidence": "High", "remainingUncertainty": "The number of chips delivered, ultimate recipients, and resulting AI tasks are not disclosed.", "relationship": "supports" }, { "id": "ev-gatekeeper-method", "caseIds": [ "case-gatekeeper" ], "claim": "The public record describes straw purchasers, third-country customers, relabeling, false paperwork, and warehousing as elements of the diversion route.", "sourceId": "doj-gatekeeper", "sourceType": "Government enforcement", "publicationDate": "2025-12-08", "accessDate": "2026-09-03", "quotationOrData": "Labels were removed and replaced with a fictitious company name.", "location": "Paragraphs beginning ‘According to charging documents’ and complaint allegations", "counterevidence": "Several details come from charging documents and, as of the 8 December 2025 release, remained allegations against defendants who had not pleaded guilty.", "confidence": "Moderate", "remainingUncertainty": "The release does not quantify how common comparable networks are.", "relationship": "supports" }, { "id": "ev-bis-diversion-guidance", "caseIds": [ "case-gatekeeper" ], "claim": "BIS identified concrete checks that could expose transshipment, opaque customers, foreign IaaS access, and implausible data-center capacity.", "sourceId": "bis-diversion-guidance-2025", "sourceType": "Government guidance", "publicationDate": "2025-05-13", "accessDate": "2026-09-03", "quotationOrData": "Red flags include freight forwarders, undisclosed end users, foreign IaaS, and infrastructure inconsistent with ordered chips.", "location": "Pages 1–5, red flags and due-diligence actions", "counterevidence": "Guidance does not establish that every recommended check is implemented or effective in every jurisdiction.", "confidence": "High", "remainingUncertainty": "Detection rates and cross-border enforcement capacity are not public.", "relationship": "qualifies" }, { "id": "ev-state-tax", "caseIds": [ "case-state-support" ], "claim": "China continued national tax-preference eligibility for integrated-circuit production, design, materials, packaging, and major projects in 2024.", "sourceId": "ndrc-2024-ic-tax", "sourceType": "Government policy", "publicationDate": "2024-03-22", "accessDate": "2026-09-03", "quotationOrData": "The eligibility notice covers production at 28 nm and below and multiple upstream inputs and services.", "location": "Article 1 and attached eligibility criteria", "counterevidence": "Eligibility rules do not disclose recipients, award value, additionality, or a capability outcome.", "confidence": "High", "remainingUncertainty": "No firm-level causal link to a reviewed AI-chip workaround is established.", "relationship": "supports" }, { "id": "ev-state-hangzhou", "caseIds": [ "case-state-support" ], "claim": "A 2025 Hangzhou policy authorized annual compute vouchers and large-model training support, but no reviewed award record ties that support to DeepSeek-V3.", "sourceId": "hangzhou-ai-policy-2025", "sourceType": "Government policy", "publicationDate": "2025-06-21", "accessDate": "2026-09-03", "quotationOrData": "Annual compute-voucher ceiling: CNY 250 million; qualifying model-training support: up to CNY 50 million.", "location": "Measures on compute services and large-model training support", "counterevidence": "The policy postdates the DeepSeek-V3 release and does not prove DeepSeek received a grant.", "confidence": "Moderate", "remainingUncertainty": "Recipients, disbursements, hardware origin, and capability produced are not public in the reviewed record.", "relationship": "qualifies" }, { "id": "ev-state-stockpile-caveat", "caseIds": [ "case-state-support" ], "claim": "GAO reports that BIS used interim-final rules partly to reduce pre-rule stockpiling, but this does not establish the scale of any Chinese inventory response.", "sourceId": "gao-25-107386", "sourceType": "Government audit", "publicationDate": "2024-12-02", "accessDate": "2026-09-03", "quotationOrData": "BIS cited avoiding stockpiling as one reason for immediate enforcement.", "location": "Highlights; What GAO Found, paragraphs 1–2", "counterevidence": "A policy concern is not evidence of actual inventory, intent, ownership, or use by a named Chinese developer.", "confidence": "High", "remainingUncertainty": "Firm-level pre-control and legacy inventories remain largely undisclosed.", "relationship": "qualifies" } ], "sources": [ { "id": "fr-2022-21658", "title": "Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items; Supercomputer and Semiconductor End Use; Entity List Modification", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2022-21658", "sourceType": "Federal rule", "publicationDate": "2022-10-13", "accessDate": "2026-09-03", "primary": true }, { "id": "bis-2022-10-07", "title": "Commerce Implements New Export Controls on Advanced Computing and Semiconductor Manufacturing Items to the PRC", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.bis.gov/press-release/commerce-implements-new-export-controls-advanced-computing-semiconductor-manufacturing-items-peoples", "sourceType": "Government announcement", "publicationDate": "2022-10-07", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2023-23055", "title": "Implementation of Additional Export Controls: Certain Advanced Computing Items; Supercomputer and Semiconductor End Use; Updates and Corrections", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2023-23055", "sourceType": "Federal rule", "publicationDate": "2023-10-25", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2023-23049", "title": "Export Controls on Semiconductor Manufacturing Items", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2023-23049", "sourceType": "Federal rule", "publicationDate": "2023-10-25", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2024-28270", "title": "Foreign-Produced Direct Product Rule Additions, and Refinements to Controls for Advanced Computing and Semiconductor Manufacturing Items", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2024-28270", "sourceType": "Federal rule", "publicationDate": "2024-12-05", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2024-28267", "title": "Additions and Modifications to the Entity List; Removals From the Validated End-User (VEU) Program", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2024-28267", "sourceType": "Federal rule", "publicationDate": "2024-12-05", "accessDate": "2026-09-03", "primary": true }, { "id": "bis-2024-12-02", "title": "Commerce Strengthens Export Controls to Restrict China’s Capability to Produce Advanced Semiconductors for Military Applications", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.bis.gov/press-release/commerce-strengthens-export-controls-restrict-chinas-capability-produce-advanced-semiconductors-military", "sourceType": "Government announcement", "publicationDate": "2024-12-02", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2025-00636", "title": "Framework for Artificial Intelligence Diffusion", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2025-00636", "sourceType": "Federal rule", "publicationDate": "2025-01-15", "accessDate": "2026-09-03", "primary": true }, { "id": "bis-2025-05-13-diffusion", "title": "Commerce Announces Rescission of AI Diffusion Rule and Strengthens Chip-Related Controls", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens", "sourceType": "Government announcement", "publicationDate": "2025-05-13", "accessDate": "2026-09-03", "primary": true }, { "id": "gao-b-337935", "title": "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", "organization": "U.S. Government Accountability Office", "url": "https://www.gao.gov/products/b-337935", "sourceType": "Government legal decision", "publicationDate": "2026-05-12", "accessDate": "2026-09-03", "primary": true }, { "id": "bis-guidance-2026-05-31", "title": "Guidance Regarding Enforcement of License Requirements for Advanced Computing Items for Entities Headquartered in Country Group D:5 and Macau", "organization": "U.S. Bureau of Industry and Security", "url": "https://media.bis.gov/media/documents/bis-guidance-may-31-2026.pdf", "sourceType": "Government guidance", "publicationDate": "2026-05-31", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2025-00711", "title": "Implementation of Additional Due Diligence Measures for Advanced Computing Integrated Circuits; Amendments and Clarifications; and Extension of Comment Period", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2025-00711", "sourceType": "Federal rule", "publicationDate": "2025-01-16", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2026-06851", "title": "Extension of Authorized Integrated Circuit (IC) Designer Status and Application Deadline To Become an Approved IC Designer", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2026-06851", "sourceType": "Federal rule", "publicationDate": "2026-04-09", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2025-16735", "title": "Revocation of Validated End-User Authorizations in the People’s Republic of China", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2025-16735", "sourceType": "Federal rule", "publicationDate": "2025-09-02", "accessDate": "2026-09-03", "primary": true }, { "id": "bis-2025-veu", "title": "Commerce Closes Export-Control Loophole for Foreign-Owned Semiconductor Fabs in China", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.bis.gov/press-release/department-commerce-closes-export-controls-loophole-foreign-owned-semiconductor-fabs-china", "sourceType": "Government announcement", "publicationDate": "2025-08-29", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2025-19001", "title": "Expansion of End-User Controls To Cover Affiliates of Certain Listed Entities", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2025-19001", "sourceType": "Federal rule", "publicationDate": "2025-09-30", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2025-19846", "title": "One Year Suspension of Expansion of End-User Controls for Affiliates of Certain Listed Entities", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2025-19846", "sourceType": "Federal rule", "publicationDate": "2025-11-12", "accessDate": "2026-09-03", "primary": true }, { "id": "fr-2026-00789", "title": "Revision to License Review Policy for Advanced Computing Commodities", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.federalregister.gov/d/2026-00789", "sourceType": "Federal rule", "publicationDate": "2026-01-15", "accessDate": "2026-09-03", "primary": true }, { "id": "bis-2026-01-13", "title": "Department of Commerce Revises License Review Policy for Semiconductors Exported to China", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china", "sourceType": "Government announcement", "publicationDate": "2026-01-13", "accessDate": "2026-09-03", "primary": true }, { "id": "deepseek-v3-report", "title": "DeepSeek-V3 Technical Report", "organization": "DeepSeek-AI", "url": "https://arxiv.org/abs/2412.19437", "sourceType": "Technical preprint", "publicationDate": "2024-12-27", "accessDate": "2026-09-03", "primary": true }, { "id": "metr-deepseek-v3", "title": "Details about METR’s Preliminary Evaluation of DeepSeek-V3", "organization": "METR", "url": "https://evals.alignment.org/evaluations/deepseek-v3-report/", "sourceType": "Independent evaluation", "publicationDate": "2025-02-12", "accessDate": "2026-09-03", "primary": false }, { "id": "nvidia-2024-10k", "title": "NVIDIA Corporation Form 10-K for Fiscal 2024", "organization": "NVIDIA Corporation / U.S. SEC", "url": "https://www.sec.gov/Archives/edgar/data/1045810/000104581024000029/nvda-20240128.htm", "sourceType": "Regulatory filing", "publicationDate": "2024-02-21", "accessDate": "2026-09-03", "primary": true }, { "id": "deepseek-v3-repository", "title": "DeepSeek-V3 Official Repository", "organization": "DeepSeek-AI", "url": "https://github.com/deepseek-ai/DeepSeek-V3", "sourceType": "Company repository", "publicationDate": "2024-12-26", "accessDate": "2026-09-03", "primary": true }, { "id": "pangu-ultra-paper", "title": "Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs", "organization": "Huawei", "url": "https://arxiv.org/html/2505.04519", "sourceType": "Technical preprint", "publicationDate": "2025-05-07", "accessDate": "2026-09-03", "primary": true }, { "id": "pangu-ultra-repository", "title": "openPangu Ultra MoE 718B Model", "organization": "Huawei openPangu", "url": "https://huggingface.co/openpangu/openPangu-Ultra-MoE-718B-model", "sourceType": "Company repository", "publicationDate": null, "accessDate": "2026-09-03", "primary": true }, { "id": "bis-gp10-2025", "title": "Guidance on Application of General Prohibition 10 (GP10) to People’s Republic of China (PRC) Advanced-Computing Integrated Circuits (ICs)", "organization": "U.S. Bureau of Industry and Security", "url": "https://www.bis.gov/media/documents/general-prohibition-10-guidance-may-13-2025.pdf", "sourceType": "Government guidance", "publicationDate": "2025-05-13", "accessDate": "2026-09-03", "primary": true }, { "id": "cloudmatrix-paper", "title": "Serving Large Language Models on Huawei CloudMatrix384", "organization": "Huawei and SiliconFlow", "url": "https://arxiv.org/html/2506.12708v3", "sourceType": "Technical preprint", "publicationDate": "2025-06-19", "accessDate": "2026-09-03", "primary": true }, { "id": "semianalysis-cloudmatrix", "title": "Huawei AI CloudMatrix 384: China’s Answer to Nvidia GB200 NVL72", "organization": "SemiAnalysis", "url": "https://newsletter.semianalysis.com/p/huawei-ai-cloudmatrix-384-chinas-answer-to-nvidia-gb200-nvl72", "sourceType": "Specialist analysis", "publicationDate": "2025-04-16", "accessDate": "2026-09-03", "primary": false }, { "id": "doj-gatekeeper", "title": "U.S. Authorities Shut Down Major China-Linked AI Tech Smuggling Network", "organization": "U.S. Department of Justice", "url": "https://www.justice.gov/opa/pr/us-authorities-shut-down-major-china-linked-ai-tech-smuggling-network", "sourceType": "Government enforcement", "publicationDate": "2025-12-08", "accessDate": "2026-09-03", "primary": true }, { "id": "bis-diversion-guidance-2025", "title": "Industry Guidance to Prevent Diversion of Advanced Computing Integrated Circuits", "organization": "U.S. Bureau of Industry and Security", "url": "https://media.bis.gov/media/documents/ai-counter-diversion-industry-guidance-may-13-2025.pdf", "sourceType": "Government guidance", "publicationDate": "2025-05-13", "accessDate": "2026-09-03", "primary": true }, { "id": "ndrc-2024-ic-tax", "title": "2024 Integrated-Circuit and Software Tax-Preference Eligibility Notice", "organization": "National Development and Reform Commission of China and partner ministries", "url": "https://www.ndrc.gov.cn/xwdt/tzgg/202403/t20240322_1365170.html", "sourceType": "Government policy", "publicationDate": "2024-03-22", "accessDate": "2026-09-03", "primary": true }, { "id": "hangzhou-ai-policy-2025", "title": "杭州市人民政府办公厅关于促进杭州市人工智能产业高质量发展若干措施的通知 (Hangzhou measures supporting high-quality AI-industry development; translated title)", "organization": "Hangzhou Municipal People’s Government", "url": "https://zjjcmspublic.oss-cn-hangzhou-zwynet-d01-a.internet.cloud.zj.gov.cn/jcms_files/jcms1/web149/site/attach/0/a6c3459cc18f497c97e7e592bd9e318f.pdf", "sourceType": "Government policy", "publicationDate": "2025-06-21", "accessDate": "2026-09-03", "primary": true }, { "id": "gao-25-107386", "title": "Export Controls: Commerce Implemented Advanced Semiconductor Rules and Took Steps to Address Compliance Challenges", "organization": "U.S. Government Accountability Office", "url": "https://www.gao.gov/products/gao-25-107386", "sourceType": "Government audit", "publicationDate": "2024-12-02", "accessDate": "2026-09-03", "primary": true } ] }