AI chip export controls
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v6 ยท 7,192 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v6 ยท 7,192 words
Add missing citations, update stale details, or suggest a clearer explanation.
AI chip export controls are a set of United States regulations that restrict the export, reexport, and in-country transfer of advanced artificial intelligence accelerators (graphics processing units, tensor processing units, and AI-specific application-specific integrated circuits) and the semiconductor manufacturing equipment used to produce them, primarily to the People's Republic of China and other designated "countries of concern."[1] The rules are administered by the U.S. Department of Commerce's Bureau of Industry and Security (BIS) under the Export Administration Regulations (EAR), 15 C.F.R. Parts 730 to 774.[2] The modern regime began with the October 7, 2022 interim final rule, was expanded substantially on October 17, 2023 and December 2, 2024, briefly extended worldwide by the January 15, 2025 "AI Diffusion" framework, and partially reversed by the second Trump administration on May 13, 2025.[3][4][5][6] The Trump administration then loosened controls further through a January 2026 final rule that authorized case-by-case licensing of the nvidia h200 and comparable chips to China and Macau, before re-escalating in June 2026 with guidance clarifying that the chip licensing requirements reach the overseas operations of Chinese-headquartered firms, an extraterritorial extension that BIS issued alongside an effective admission that it had failed to enforce the pre-existing controls.[55][58][60] By 2025 and 2026, the regime had become a central instrument of U.S. technology policy, generating multilateral coordination with the Netherlands, Japan, Korea, and Taiwan; multibillion-dollar revenue effects on NVIDIA and AMD; and an active enforcement docket aimed at chip smuggling networks moving Hopper- and Blackwell-class accelerators into China.[7][8][9]
The EAR govern the export of "dual-use" items, defined as goods, software, and technology that have predominantly commercial applications but could also be used for military, intelligence, or weapons-of-mass-destruction purposes.[2] BIS classifies controlled items on the Commerce Control List (CCL) under Export Control Classification Numbers (ECCNs); items not enumerated on the CCL fall under the catch-all designation "EAR99."[2] Statutory authority comes from the Export Control Reform Act of 2018, which superseded the long-lapsed Export Administration Act of 1979 and confirmed BIS's authority to regulate items "essential to the national security of the United States."[10] An export, under the EAR, includes the transmission of controlled technology to a foreign national even within U.S. territory (the so-called "deemed export" rule), which has shaped how U.S. semiconductor companies hire and assign work.[2]
The EAR have three principal mechanisms relevant to AI accelerators. First, license requirements for specific ECCNs (notably 3A090 for advanced ICs, 3B001 for semiconductor manufacturing equipment, and 4A090 for computers containing controlled ICs) bar export to listed destinations absent a BIS license.[4] Second, the Entity List names foreign parties to which exports of items subject to the EAR are presumptively denied; entities are added by interagency action of the End-User Review Committee on findings of national-security or foreign-policy concern.[11] Third, the Foreign Direct Product Rule (FDPR) extends U.S. jurisdiction to certain items produced abroad using U.S.-origin technology or software, even when no U.S.-origin physical content is incorporated, when the foreign-produced item is destined for an Entity List party or otherwise meets specified end-use criteria.[4]
Before 2022, semiconductor controls focused mainly on military-end-use rules and a small set of Entity List designations. Huawei was added to the Entity List on May 16, 2019; in May 2020 BIS published a Huawei-specific FDPR that prevented Taiwan Semiconductor Manufacturing Company (TSMC) from producing chips of Huawei's HiSilicon designs using U.S. equipment or design software.[12] Semiconductor Manufacturing International Corporation (SMIC) was added to the Entity List on December 18, 2020, with a license-required-for-advanced-node-production designation that targeted SMIC's ambitions below the 10-nanometer node.[13] These designations were entity-specific. There were no broad country-wide controls on the export of advanced AI chips themselves to China, and NVIDIA's then-flagship data-center accelerator, the nvidia a100 (released May 2020), was sold to Chinese cloud providers and research institutions without restriction.[7]
Policy planning for broader controls accelerated through 2021 and 2022 as scaling-law results suggested that compute, rather than data or algorithms, was the binding input to advanced AI capability.[14][7] In a September 2022 speech at the Special Competitive Studies Project, National Security Advisor Jake Sullivan articulated a doctrine of preserving "as large a lead as possible" over strategic competitors in foundational technologies including AI, biotech, and semiconductors, signaling that the prior policy of maintaining "relative" advantage would be replaced by absolute denial of leading-edge capability to adversaries.[15] BIS framed the eventual rule as targeting (i) Chinese acquisition of advanced computing chips needed to train frontier AI models, (ii) the indigenous Chinese capacity to manufacture such chips, and (iii) Chinese supercomputing programs supporting weapons design, including hypersonics and nuclear-weapons modeling.[3]
On October 7, 2022, BIS published an interim final rule titled "Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items; Supercomputer and Semiconductor End Use; Entity List Modification."[3] The rule was the most extensive expansion of U.S. semiconductor controls since the end of the Cold War and contained four distinct regulatory components.
The first component created a new ECCN 3A090 covering high-performance integrated circuits with both (a) aggregate computational performance at or above 4,800 dense tera-operations per second (TOPS) at any precision and (b) input/output (chip-to-chip interconnect) bandwidth of 600 GB/s or higher.[3] The dual threshold was crucial: chips above the compute threshold but below the interconnect threshold could be exported, but the combination of both thresholds captured NVIDIA's nvidia a100 and nvidia h100 data-center GPUs, which used NVLink interconnects rated above the 600 GB/s line.[7] The interconnect criterion reflected the observation that distributed training of large language models is bottlenecked by all-to-all collective communication across thousands of GPUs, so restricting interconnect bandwidth degraded training throughput even when raw compute remained available.[14] A complementary ECCN 4A090 captured computer systems containing controlled 3A090 ICs. Exports of items meeting these specifications to China and Macau became subject to a license requirement with a presumption of denial.[3]
The second component imposed broad license requirements on the export of semiconductor manufacturing equipment (SME) to facilities in China producing logic chips at 16/14-nanometer nodes and below using non-planar (FinFET or gate-all-around) transistors, DRAM at 18-nanometer half-pitch and below, and NAND flash with 128 layers or more.[3] The rule targeted etch, deposition, lithography, ion implantation, metrology, and cleaning tools, capturing nearly the full process flow for an advanced fab. While U.S. equipment makers (Applied Materials, Lam Research, KLA) were directly bound, the rule did not yet impose equivalent controls on Dutch lithography supplier ASML Holding N.V. (the sole commercial source of extreme ultraviolet lithography) or Japanese makers including Tokyo Electron.
The third component, effective October 12, 2022, prohibited "U.S. persons" (U.S. citizens, lawful permanent residents, and persons in the United States) from supporting the development or production of advanced ICs at facilities in China without authorization, even when those activities did not involve the export of items subject to the EAR.[3] U.S. persons employed at Chinese fabs faced a rapid choice between resignation and loss of U.S. status; Reuters reported that engineers at YMTC and other Chinese facilities began departing within days of the rule's publication.[16]
The October 2022 package added 28 Chinese entities to the Entity List and created a new "supercomputer end-use" license requirement covering any item subject to the EAR destined for a Chinese supercomputer, defined by sustained performance thresholds keyed to floor space.[3] BIS estimated affected U.S. industry revenue at roughly 100 to 200 million dollars per year in immediate effect but acknowledged that the longer-term effect on integrated-circuit and equipment sales would be substantially larger.[3]
nvidia responded to the October 2022 controls by designing performance-tuned variants of its restricted GPUs. The A800, announced in November 2022 as a China-specific A100 derivative, retained the A100's compute capability but reduced NVLink interconnect bandwidth from 600 GB/s to 400 GB/s, placing it below the 2022 rule's interconnect threshold.[17] The H800, introduced in March 2023, applied the same approach to the Hopper-generation nvidia h100, lowering NVLink bandwidth from 900 GB/s to 400 GB/s while preserving the 989 BF16 tera-FLOP/s peak compute.[18] These variants were legal exports to China through the entirety of fiscal year 2023; NVIDIA disclosed China data-center revenue exceeding 9 billion dollars in the twelve months between October 31, 2022 and October 31, 2023.[19] Chinese AI laboratory deepseek later disclosed that it had trained the deepseek v3 model on a cluster of approximately 2,048 H800 GPUs, demonstrating that performance-tuned variants remained sufficient to train frontier-class models.[20]
On October 17, 2023, BIS published an expanded rule package designed to close the A800 and H800 loopholes and to capture lower-tier datacenter chips and high-end gaming GPUs.[4] The new rule restructured the chip-control criteria around three concepts:
| Criterion | Threshold (2022 rule) | Threshold (2023 rule) |
|---|---|---|
| Total Processing Performance (TPP) | 4,800 TOPS dense + 600 GB/s interconnect | TPP at or above 4,800 (any precision-weighted), or TPP 1,600 to 4,800 plus performance density of 5.92 TPP/mm^2 or more |
| Interconnect bandwidth | 600 GB/s threshold determinative | Removed as primary criterion; replaced by performance-density and datacenter-marketing criteria |
| Datacenter "designed or marketed" criterion | Not used | Added as third axis for chips between 1,600 and 4,800 TPP |
The TPP measure aggregated compute across precisions (FP64, FP32, FP16, BF16, INT8) weighted so that a chip's most-relevant training precision dominated.[21] By removing the interconnect threshold as a binary cutoff and introducing performance density (TPP per square millimeter of die area), the rule captured chips like the H800 that had been engineered specifically to evade the prior interconnect criterion.[21] BIS additionally created a "Notified Advanced Computing" license category, in which the export of certain lower-tier chips to destinations in some Tier 2 countries required pre-shipment notification but not denial.[4]
The 2023 rule also added China to the destinations covered by SME controls regardless of the technology node, restricted exports of mature-node deep-ultraviolet (DUV) lithography immersion tools beyond what the Dutch government had already controlled, and expanded the country scope to include all 21 destinations subject to the U.S. arms embargo, capturing Russia and Iran as well as China.[4][21] Twelve Chinese GPU and supercomputing entities were added to the Entity List, including Moore Threads and Biren Technology.[4]
nvidia responded to the October 2023 rules with a further generation of China-compliant Hopper-architecture chips. The H20, announced in late 2023 and shipped in 2024, was a deeply cut H100 variant: peak BF16 compute of roughly 148 TFLOP/s (compared with 989 on H100), 96 GB of HBM3 memory, and NVLink bandwidth of 900 GB/s. The H20's compute was deliberately set well below the 4,800 TPP threshold, while its memory capacity, bandwidth, and NVLink were preserved to support inference workloads.[17] Two lower-tier products, the L20 and L2 (based on the Ada Lovelace architecture), were similarly tuned. The H20 became NVIDIA's leading China data-center product through 2024 and the first half of 2025, generating an estimated 12 billion dollars in revenue in fiscal 2025.[22]
amd instinct mi300x, AMD's competitor to the H100, fell above the October 2023 thresholds and could not be exported to China. AMD designed the MI308 as a downgraded variant intended to comply with the same thresholds; in March 2024, BIS informed AMD that even the cut-down MI308 would require an export license, and the company paused shipments.[23] Shipments resumed in 2025 under specific license conditions, with AMD disclosing in April 2025 that revised export restrictions would cost the company approximately 800 million dollars in inventory and purchase-commitment charges.[24]
On December 2, 2024, the Biden administration's BIS published what observers described as the most substantial single expansion of the controls since October 2022.[25] The package included three principal final rules and an Entity List action, together covering:
Effective October 2, 2024 and refined in subsequent rulemaking, BIS introduced a new Data Center Validated End-User Authorization that allowed pre-approved overseas data-center operators in eligible destinations (initially Tier 1 allies) to receive controlled advanced computing items without per-shipment licensing, conditional on physical and cyber-security commitments, ongoing reporting, and restrictions on Chinese-headquartered customer access.[27] The VEU framework was meant as a positive-list complement to the negative-list approach of license requirements and Entity List denials, providing a route by which cloud providers including Microsoft, Google, and Amazon could deploy advanced accelerators in third countries without triggering case-by-case BIS adjudication.[27]
On January 13, 2025, the outgoing Biden administration published the "Framework for Artificial Intelligence Diffusion" interim final rule, generally referred to as the "AI Diffusion" rule, with publication in the Federal Register on January 15.[5] The rule was scheduled to take effect 120 days later, on May 15, 2025.[5] It established a three-tier global framework for access to advanced computing chips and to AI model weights:
| Tier | Representative destinations | Treatment under AI Diffusion |
|---|---|---|
| Tier 1 | Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan, United Kingdom (18 countries plus the United States) | Effectively unlimited access; no per-export license for most advanced-computing items, but Universal Validated End User (UVEU) program required for large data-center deployments to ensure physical and personnel-security controls. |
| Tier 2 | Singapore, Israel, Portugal, Switzerland, Poland, United Arab Emirates, Saudi Arabia, India, Mexico, and most other non-listed countries (approximately 150 destinations) | Per-country compute cap of approximately 50,000 H100-equivalents over the initial two-year period, with a per-entity National Validated End User (NVEU) cap of about 320,000 H100-equivalents over two years for vetted operators. Exports above caps required individual licensing. |
| Tier 3 | China, Russia, Iran, North Korea, Cuba, Venezuela, Belarus, and the other 21 arms-embargoed destinations | Presumption of denial maintained from prior rules; no relaxation. |
The rule also imposed export controls, for the first time, on the weights of closed-weight frontier AI models trained using 10^26 or more computational operations, the threshold corresponding roughly to the most advanced 2024 frontier models.[5][28] Open-weight model weights were exempted from the model-weight control.
The AI Diffusion rule was opposed publicly by NVIDIA, which described it in a January 13, 2025 statement as a "200-page regulatory morass" that would impose costs on U.S. allies and customers while doing little to constrain China.[29] The Information Technology Industry Council, the Semiconductor Industry Association, and the governments of Israel, Poland, the United Arab Emirates, and others expressed concern that Tier 2 status would impede their AI-data-center buildouts.[28] Academic commentators were divided: Center for Security and Emerging Technology analysts argued the tier framework rationalized previously ad-hoc third-country diversion controls, while RAND and CSIS commentators questioned whether the per-country caps were calibrated to legitimate demand.[30][28]
On May 13, 2025, before the AI Diffusion rule's scheduled effective date, the Trump administration's BIS, under Commerce Secretary Howard Lutnick and Under Secretary Jeffrey Kessler, announced rescission of the rule.[6] BIS justified rescission on three grounds: that the rule was "overly bureaucratic"; that it "stifled American innovation"; and that it would "undermine U.S. diplomatic relations with dozens of countries by downgrading them to second-tier status."[6] Kessler said BIS would instead "pursue a bold, inclusive strategy to American AI technology with trusted foreign countries around the world, while keeping the technology out of the hands of our adversaries."[6] In place of the AI Diffusion rule, BIS issued enforcement guidance asserting that the export of advanced computing items to most destinations would proceed without per-country caps, while announcing forthcoming individual deals with allied destinations under which advanced chips would be permitted only when destined to U.S.-operated cloud infrastructure.[31] The May 2025 guidance also formally identified the use of any U.S.-origin Huawei Ascend chip worldwide as a likely violation of the EAR, on the rationale that any Ascend chip incorporates U.S. technology subject to the EAR through the FDPR.[31]
The most-publicized post-rescission episode concerned NVIDIA's nvidia h100-derived H20. On April 9, 2025, the Department of Commerce informed NVIDIA that exports of the H20 to China would require an individual license, with case-by-case review. NVIDIA disclosed a 5.5 billion dollar charge against H20 inventory and unconditional purchase obligations on April 15, 2025, and effectively suspended H20 shipments to China.[32][22] The April action was described by analysts as the first explicit ban on a chip designed to comply with the existing performance thresholds: the H20 fell well below the TPP threshold, but the administration concluded that its memory and interconnect characteristics still enabled large-scale inference of Chinese frontier models, including DeepSeek-R1.[33][31]
After lobbying by NVIDIA CEO Jensen Huang and trade representations from the Chinese Ministry of Commerce, on July 14, 2025, the Trump administration reversed course and committed to issuing H20 export licenses to qualifying Chinese customers.[34] The reauthorization came with an unprecedented revenue-sharing condition announced by President Trump on August 11, 2025: NVIDIA would remit 15 percent of H20 China revenue to the U.S. government in exchange for the licenses, a structure compared by legal commentators to a contingent export tax or "license premium."[34] Trump said he had initially asked for a 20 percent cut and that Huang had "haggled him down" to 15 percent, making it the first time a U.S. company agreed to pay a share of revenue in exchange for an export license.[34] AMD agreed to similar revenue-sharing terms for its MI308 China shipments.[34] In December 2025, the Trump administration permitted limited nvidia h200 sales to China under license, with a 25 percent tariff applied to the China-bound product; this December 8, 2025 announcement was formalized in a January 2026 final rule discussed below.[9][55]
President Trump signaled in December 2025 that NVIDIA would be allowed to sell the nvidia h200 to China, a chip materially more capable than the previously authorized H20.[9][55] BIS published the implementing final rule on January 13, 2026, with an effective date of January 15, 2026, formally shifting the license-review policy for the affected chips from a presumption of denial to case-by-case review for exports to China and Macau.[55][56] The rule defined the eligible chips by a performance ceiling rather than by name, covering accelerators with Total Processing Performance below 21,000 and total DRAM bandwidth below 6,500 GB/s, a band that captured the H200 and the AMD MI325X while leaving Blackwell-class parts above the line and ineligible.[56]
A license under the January 2026 rule was conditioned on several certifications by the exporter: that the chip was commercially available in the United States; that U.S. supply was sufficient so that filling a China or Macau order would not delay U.S. end-user orders; that the manufacturer's global foundry capacity would not be diverted to serve China; that aggregate China and Macau shipments would not exceed 50 percent of the units sold to U.S. customers; that no prohibited end use or restricted party was involved and the chips would not support military, intelligence, or weapons development; that the consignee maintained know-your-customer procedures preventing blacklisted entities from gaining access, including through infrastructure-as-a-service; and that an independent U.S. laboratory certified each shipment's specifications.[56] On January 14, 2026, the President issued a proclamation under Section 232 of the Trade Expansion Act of 1962 imposing a 25 percent tariff on covered advanced AI chips not destined for the U.S. supply chain.[56] The Council on Foreign Relations characterized the combined framework as "strategically incoherent and unenforceable," arguing that faithful application of the conditions would block most exports while loose application would defeat the purpose of the controls.[57]
Around May 31 to June 1, 2026, BIS issued guidance affirming that its licensing requirements for advanced AI chips apply to all businesses with headquarters or a parent company in China or Macau, regardless of where their subsidiaries or affiliates are located.[58][59] BIS stated that a license is required to export covered advanced computing items to entities headquartered in Country Group D:5 or Macau, or to entities whose ultimate parent company is headquartered in those jurisdictions, "even if the entities themselves are located outside" those jurisdictions.[59] The agency framed the guidance as a clarification rather than a new rule: asked whether the pre-existing license requirements remained in effect after the rescission of the AI Diffusion framework, BIS stated, "The answer is yes."[58] The move closed what observers described as a loophole under which the most advanced U.S. chips, including NVIDIA Blackwell and GB200 parts and the AMD MI350X, may have been reaching subsidiaries of Chinese AI companies in third countries such as Malaysia.[59][60] One semiconductor supply-chain estimate cited in press coverage put the volume of advanced chips potentially acquired by Chinese-linked buyers through the gap at hundreds of thousands of units over roughly the prior year, though no official figure was released.[60]
Analysts noted that the parent-company test still left enforcement gaps. Former State Department official Chris McGuire, quoted in the coverage, said Chinese companies had very likely been buying restricted chips at scale and observed that the guidance did not require foundries such as tsmc to verify whether high-end chips they manufacture could be routed through Chinese intermediaries.[58][59] The Foundation for Defense of Democracies argued in a June 2, 2026 analysis that the updated posture amounted to an admission of the Commerce Department's earlier failure to regulate U.S. AI exports to Chinese firms, because BIS had not amended the EAR after the May 2025 rescission to state clearly which pre-existing controls it was enforcing, leaving the prior year of shipments effectively unregulated.[60] The disclosure landed against reporting by The New York Times on June 1, 2026 that the People's Liberation Army had sought to acquire advanced NVIDIA chips both before and after the imposition of U.S. controls.[60] The June 2026 guidance partially walked back the permissiveness of the January 2026 loosening and re-escalated the U.S.-China contest over access to AI compute.
Multilateral coordination has centered on the Netherlands and Japan, the two non-U.S. countries that host critical chokepoints in the semiconductor supply chain. The Netherlands hosts ASML, the sole commercial supplier of extreme ultraviolet (EUV) lithography systems, which since 2018 has been prevented from exporting EUV tools to China under Dutch license denial coordinated with the United States.[35] On January 27, 2023, the United States, Japan, and the Netherlands agreed at a meeting in Washington to align their export controls; the Netherlands implemented domestic controls effective September 1, 2023 that extended licensing requirements to ASML's most advanced DUV immersion systems including the TWINSCAN NXT:2000i, NXT:2050i, and NXT:2100i.[35] In 2024, the Dutch government further restricted shipments of NXT:1970i and NXT:1980i systems to China, and the U.S. revoked the validity of certain prior Dutch export licenses retroactively through the FDPR. ASML disclosed in its 2024 and 2025 annual filings that the controls reduced its China-bound system revenue, although strong demand for already-installed DUV tools moderated the headline financial effect.[36]
Japan's Ministry of Economy, Trade and Industry announced in March 2023 controls on 23 categories of semiconductor manufacturing equipment, including advanced lithography, etch, deposition, and cleaning tools produced by Tokyo Electron, Screen Holdings, Lasertec, and others, effective July 23, 2023.[37] METI did not formally identify China as the target, instead extending the controls to all destinations and relying on the existing licensing apparatus to deny shipments to China. In the December 2024 U.S. package, BIS effectively excluded Tokyo Electron and ASML from the new SME FDPR for shipments to China, in recognition of the parallel Japanese and Dutch controls.[38] By 2025 and 2026, U.S. trade officials sought further alignment, including extension of multilateral controls to the most advanced mature-node tools and to chip-design software.
The December 2024 HBM controls had their largest practical effect on South Korean memory makers Samsung Electronics and SK hynix, which manufacture an estimated 90 percent of global HBM3 and HBM3E output. The South Korean government issued conforming guidance in early 2025 and did not formally adopt parallel controls but did require Korean firms to comply with the U.S. FDPR. Taiwan adopted formal export controls on June 15, 2025, adding Huawei, SMIC, and 65 affiliated entities to a "Strategic High-Tech Commodities Entity List," following the 2023 to 2024 enforcement controversy in which TSMC produced an estimated 2.9 million Ascend dies for Huawei before its compliance program identified the diversion.[39][40]
The export controls coincided with intensified Chinese state investment in domestic semiconductor capacity. huawei ai subsidiary HiSilicon, working with foundry partner SMIC, demonstrated 7-nanometer-class manufacturing using deep-ultraviolet immersion lithography (the most advanced node achievable without EUV) with the Kirin 9000s mobile SoC in the September 2023 Mate 60 Pro smartphone, an outcome that surprised U.S. policymakers.[41] Huawei extended the same 7-nanometer SMIC process to its Ascend AI accelerator line.[42]
| Huawei AI accelerator | Foundry / process | Approximate performance | Status |
|---|---|---|---|
| Ascend 910 (2019) | TSMC 7 nm (legacy) | 256 BF16 TFLOP/s | Pre-EL production |
| Ascend 910B (2023 to 2024) | SMIC 7 nm DUV | Approx. 320 BF16 TFLOP/s | Mass-produced; yield reportedly ~50 percent |
| Ascend 910C (2024 to 2025) | SMIC 7 nm DUV (dual-die package) | Approx. 800 BF16 TFLOP/s | Mass production from early 2025; yield reportedly ~20 percent |
| Ascend 910D (announced 2025) | SMIC 7 nm DUV | Targeted ~1.6 BF16 PFLOP/s | In development |
SemiAnalysis reporting through 2024 and 2025 documented that Huawei accumulated a "die bank" of more than 2.9 million Ascend dies produced by TSMC before its compliance program flagged the diversion in late 2023; that inventory sustained 910B and 910C production through 2024 and 2025 alongside ramping SMIC output.[42] The combination of die-bank inventory, ramped SMIC capacity, and high-bandwidth memory stockpiled before the December 2024 HBM rule allowed Huawei to begin offering large Ascend-based clusters as a domestic alternative to NVIDIA's restricted products.[42] CSET analysts concluded in 2025 that the export controls had constrained but not blocked Chinese AI hardware progress, with Huawei's accelerator line achieving roughly H100-class per-chip performance while remaining one to two years behind NVIDIA's leading-edge Blackwell and post-Blackwell products.[43][41]
A second adaptation has been the accumulation of restricted U.S. chips through both legal pre-rule purchases and illegal diversion. The Center for a New American Security estimated that between 10,000 and several hundred thousand restricted AI accelerators were smuggled to China during 2024 alone, with Reuters and the Wire China documenting transshipment routes through Singapore, Malaysia, the United Arab Emirates, and Hong Kong.[44][45] U.S. enforcement intensified in 2025. The Department of Justice's "Operation Gatekeeper," unsealed in December 2025 in the U.S. District Court for the Southern District of Texas, charged Alan Hao Hsu and Hao Global LLC with smuggling at least 160 million dollars of H100 and H200 GPUs to China between October 2024 and May 2025 using rebranded labels and front companies, with the defendants pleading guilty in October 2025.[46] A parallel January 2026 case in the same district charged three former Supermicro Computer employees with a 2.5 billion dollar smuggling conspiracy involving H100, H200, and B200 shipments concealed in dummy server boxes.[47]
The financial impact on U.S. chip companies has been substantial but not catastrophic. NVIDIA's China data-center revenue fell from approximately 25 percent of data-center revenue in fiscal 2022 to a "low single-digit" share by fiscal 2026, with absolute China revenue still totaling 17.1 billion dollars in fiscal 2025 owing to the booming H20 business before its April 2025 pause.[48] AMD disclosed approximately 800 million dollars in inventory and purchase charges from the April 2025 H20/MI308 ban; Intel and Qualcomm both lost more than 1 billion dollars of annual China revenue under the cumulative rule packages.[24] ASML disclosed that the controls reduced its China-bound system revenue from approximately 49 percent of net system sales in 2024 to a projected 25 to 30 percent share in 2025 and 2026.[36] Korean memory makers Samsung and SK hynix faced HBM3 China revenue declines from the December 2024 rule but reported offsetting demand growth from U.S. and Tier 1 buyers.[25]
Academic and analyst opinion on whether the controls have achieved their stated objectives is divided. CSIS and CSET analysts have argued that the controls successfully delayed China's domestic AI compute buildout by one to three years, raised China's per-token training cost, and constrained the scale of frontier model runs achievable on indigenous hardware.[49][28] RAND analysts, writing in February 2025 after the release of deepseek r1, argued that the rule package required "smarter" rather than weaker controls and noted that DeepSeek's training run on roughly 2,048 H800s would have been illegal had the 2022 rules captured the interconnect criterion correctly, while remaining technically compliant with the rules as drafted.[50] Anthropic CEO Dario Amodei wrote in February 2025 that the controls were "the most important policy issue in AI" and concluded that DeepSeek-R1 "reinforces, rather than undermines, the case" for tight controls, on the grounds that absent the controls, Chinese laboratories would have access to multiples more compute than they had used to train R1.[51] Critics including the International Center for Law and Economics and several semiconductor-industry commentators have argued that the controls accelerate Chinese self-sufficiency without preventing capability acquisition, citing the SMIC 7-nanometer node, Huawei Ascend ramp, and the stockpiled H800 inventory underpinning Chinese model training through 2024 and 2025.[52]
A separate strand of academic criticism has questioned the controls on rule-of-law and administrative-procedure grounds. The most-cited concern is the U.S.-person provision of the October 2022 rule, which restricts the conduct of U.S. citizens and lawful permanent residents abroad in ways that some commentators have argued exceed BIS's statutory authority under the Export Control Reform Act.[53] The Federal Register publication of the AI Diffusion rule as an "interim final rule" without notice-and-comment rulemaking also drew criticism, with the Trump administration's May 2025 rescission citing procedural irregularities among its grounds for vacating the rule.[6] A separate concern is the asserted extraterritorial reach of the December 2024 SME FDPR, which subjects foreign-produced tools to U.S. controls based on incorporation of U.S.-origin technology that may amount to a small share of the tool's value; the European Commission and the Dutch government have raised the issue in trade consultations.[36]
As of June 2026, the operative regime consists of:
Legislative proposals before the 119th Congress as of mid-2026 include the bipartisan Chip Security Act (S. 1234, requiring chip-level location verification on advanced accelerators), the Remote Access Security Act (closing the cloud-API training loophole), and the GAIN AI Act (extending end-use controls to model weights produced abroad on U.S.-origin chips). None has been enacted as of June 2026.[54]