AI Chip Export Controls in Autumn 2026: Licences, Tariffs and Beijing's Veto
· TexaHub · Geopolitics & Tech, Semiconductors, Artificial Intelligence, Regional Tech Ecosystems
Washington now licenses H200-class chips to China and favours trusted Gulf buyers, but Beijing's restrictions and domestic chip push limit what US policy alone determines.
No replacement for the diffusion rule
The January 2025 AI Diffusion Rule would have imposed a worldwide licensing requirement on data-centre products such as Nvidia's H200, GB200 and GB300, and would have placed most countries in a "Tier 2" category, according to Nvidia's own description of the rule [4]. In May 2025, the Bureau of Industry and Security (BIS) announced that it would not enforce the rule's new compliance requirements [2].
The promised replacement has not appeared. According to an account by law firm Morgan Lewis, Under Secretary Jeffrey Kessler told Congress on 14 July 2026 that the administration no longer intends to replace the AI Diffusion Rule and is instead focusing on new rulemaking [3]. Nvidia's quarterly report for the period ended 26 July 2026 still described the scope, timing and requirements of a forthcoming replacement rule as uncertain [4]. These two accounts differ on whether a replacement rule is still expected, and nothing in the Federal Register settles the question.
In practice, BIS has said it enforces the worldwide licence requirement for the most advanced accelerators only for destinations in Country Groups D:1, D:4 or D:5 (excluding those also in A:5 or A:6), and for entities headquartered in, or with an ultimate parent in, D:5 or Macau, wherever they are located [15]. Guidance dated 31 May 2026 confirmed that the licence requirement for D:5- and Macau-headquartered entities operating abroad, first introduced on 17 November 2023, is still enforced [2].
From H20 to H200: licences with a price
In April 2025 the US government told Nvidia that its H20 needed a licence for export to China, and Nvidia took a $4.5 billion charge for H20 inventory and purchase obligations [4]. Licences for certain H20 shipments followed in August 2025, and Nvidia reported about $60 million in H20 revenue under them [5]. The company said US officials expected the government to receive 15% or more of revenue from licensed sales, but that no regulation codifying that requirement was published [5].
A BIS final rule effective 15 January 2026 moved licence reviews for exports to China and Macau from a presumption of denial to case-by-case review for chips with a total processing performance (TPP) below 21,000 and total DRAM bandwidth below 6,500 GB/s, naming the Nvidia H200 and AMD MI325X as examples [1]. Applicants must certify that US customers will not face supply delays, that aggregate TPP shipped to China and Macau will not exceed 50 percent of the aggregate TPP shipped to US customers for the same product, and that every shipment is reviewed by a qualified third-party testing lab headquartered in the United States, with the testing done there [1]. The rule keeps a presumption of denial for other applications involving entities headquartered in D:5 destinations or Macau [1].
The 15% expectation for H20 sales has been followed for the H200 by a tariff. A presidential proclamation signed on 14 January 2026 imposed a 25 percent ad valorem duty on certain advanced computing chips from 15 January 2026, while exempting imports for US data centres and several other domestic uses [6]. Nvidia states that because licensed H200s must pass a US inspection before shipment, they are subject to the 25% tariff on import into the United States, and that it has been unable to pass the cost on to customers [4].
Beijing's response: restraint on buyers
US licences have not translated into large sales. Reuters reported on 14 May 2026 that about ten Chinese companies, including Alibaba, Tencent, ByteDance and JD.com, had been cleared to buy H200s, each up to 75,000 chips, with Lenovo and Foxconn approved as distributors, but that buyers had pulled back after guidance from Beijing [8]. Commerce Secretary Howard Lutnick was quoted as saying that "the Chinese central government has not let them" buy the chips because it wants investment to stay focused on domestic industry [8].
Earlier, Reuters reported in November 2025, citing two unnamed sources, that new data-centre projects receiving any state funds had been told to use only domestically made AI chips, and that projects less than 30% complete were told to remove installed foreign chips or cancel planned purchases [9]. In July 2026 TrendForce, citing Reuters and The Information, reported that Beijing was likely to authorise fewer than 200,000 H200s in total, less than half the volume requested, with use reportedly reserved for model training [10]. Nvidia has also said that China's government publicly questioned whether H20 chips contain built-in vulnerabilities, which discouraged customers [4].
Official Chinese statements have been sparse. Asked in January 2026 about reports that H200 purchases would be approved only in special circumstances, Ministry of Commerce spokesperson He Yongqian said, "I am not aware of the situation you mentioned" [11]. The reported conditions therefore rest largely on unnamed sources rather than published Chinese rules.
Nvidia's own disclosures show the result. Its latest quarterly report says H200 sales were restricted by the PRC government, that it took a $0.4 billion H200 charge in the first half of fiscal 2027, and that the fraction of allowed shipments it has made accounted for less than 1% of Data Center revenue in the most recent quarter [4]. Its 26 August 2026 results release said it was not assuming any Data Center compute revenue from China in its outlook [7].
China's domestic supply: Huawei and SMIC
Huawei set out its Ascend roadmap in September 2025: the Ascend 950PR in the first quarter of 2026, the 950DT in the fourth quarter of 2026 using Huawei's HiZQ 2.0 HBM with 144 GB of memory and 4 TB/s of bandwidth, and Atlas 950 SuperPoD systems of up to 8,192 chips [12]. At Huawei Connect on 17 September 2026, rotating chairman David Wang said the Ascend 960DT would be available in the first quarter of 2027, three quarters ahead of the original roadmap, and that Ascend 970 and 980 would follow in 2028 and 2029 [13]. He said more than 1,000 Atlas 900 A3 SuperPoDs had been deployed and that the Atlas 950 SuperPoD was "seeing large-scale commercial use" [13]. Huawei did not publish chip shipment volumes, so independent evidence on how many Ascend parts are reaching data centres is thin.
The constraint is manufacturing. Huawei itself has said the Chinese mainland will lag in process nodes "for a relatively long time" [12]. SMIC, China's largest foundry, reported second-quarter 2026 revenue of $3,005.6 million, utilisation of 93.7% and monthly capacity of 1,096,500 8-inch-equivalent wafers, with China accounting for 90.2% of revenue [14]. SMIC does not break out how much of that capacity goes to AI accelerators, so the figures show a fully loaded foundry rather than a measured AI chip output.
What it means for buyers elsewhere
For the Gulf, the policy has shifted from blanket caps to approvals for named entities. In November 2025 Commerce authorised exports to G42 in the UAE and Humain in Saudi Arabia; The National reported the approvals as covering up to 35,000 Blackwell GB300 chips combined, subject to security and reporting requirements [16]. A BIS rule effective 10 July 2026 moved the UAE from Country Groups D:3 and D:4 to A:5 but kept licence requirements for advanced computing items except for UAE government agencies, G42 and Core42, and named US companies including Amazon, Google, Microsoft, OpenAI and Oracle and their subsidiaries [15]. The G42 and Core42 authorisation expires automatically on 6 April 2027 unless BIS gives notice otherwise, and BIS said it intends to review licence applications involving MGX favourably [15]. Saudi Arabia was not covered by that rule.
In Asia, transit hubs have added their own controls. Malaysia has required a Strategic Trade Permit since 14 July 2025 for exports, transshipments and transits of high-performance US-origin AI chips, with at least 30 days' advance notification [17]. The May 2026 BIS guidance means that Chinese-headquartered operators outside China still need US licences for top-end accelerators wherever they operate [2].
In our analysis, buyers in Europe and other allied markets outside the enforced country groups face fewer US licensing hurdles now than the diffusion rule would have imposed. The main exposures are the ownership of the end user, the remote-access conditions written into licences, and legislation such as the GAIN AI Act and Remote Access Security Act, which Nvidia lists as possible sources of new restrictions [4]. Washington is also promoting bundled exports: the American AI Exports Program, which opened for proposals on 1 April 2026, says approved consortia may receive priority for export control licence reviews and prioritised access to federal credit programmes [18]. Whether buyers outside the United States and China face more or less friction will depend on rulemaking that, as of 1 October 2026, has not been published.
Sources
[1] Revision to License Review Policy for Advanced Computing Commodities (91 FR 1684) — Bureau of Industry and Security, Federal Register, 15 Jan 2026
[2] Guidance Regarding Enforcement of License Requirements for Advanced Computing Items for Entities Headquartered in Country Group D:5 and Macau — US Bureau of Industry and Security, 31 May 2026
[3] BIS Upgrades UAE Export Control Status, with AI Chip Access Limited to Approved Entities — Morgan Lewis, 21 Jul 2026
[4] NVIDIA Corporation Form 10-Q for the quarter ended July 26, 2026 — NVIDIA / US SEC, quarter ended 26 Jul 2026
[5] NVIDIA Corporation Form 10-K for the fiscal year ended January 25, 2026 — NVIDIA / US SEC, fiscal year ended 25 Jan 2026
[6] Adjusting Imports of Semiconductors, Semiconductor Manufacturing Equipment, and Their Derivative Products Into the United States — The White House, 14 Jan 2026
[7] NVIDIA Announces Financial Results for Second Quarter Fiscal 2027 — NVIDIA, 26 Aug 2026
[8] U.S. clears H200 chip sales to 10 China firms as Nvidia CEO looks for breakthrough: Reuters exclusive — Reuters via BNN Bloomberg, 14 May 2026
[9] Exclusive-China bans foreign AI chips from state-funded data centres, sources say — Reuters via Investing.com, 5 Nov 2025
[10] [News] China Reportedly to Allow NVIDIA H200 Imports, but Approvals May Fall Below 200K Chips — TrendForce, 9 Jul 2026
[11] MOFCOM responds to question on China only approves purchases of H200 chips under special circumstances — Global Times, 22 Jan 2026
[12] Groundbreaking SuperPoD Interconnect: Leading a New Paradigm for AI Infrastructure — Huawei, 18 Sep 2025
[13] Advancing the Agentic World, Building a Solid Silicon Foundation — Huawei, 17 Sep 2026
[14] SMIC Reports Unaudited Results for the Three Months Ended June 30, 2026 — SMIC / HKEX announcement, 13 Aug 2026
[15] Enhanced Favorable Treatment for the United Arab Emirates Under the Export Administration Regulations — Bureau of Industry and Security, Federal Register, 14 Jul 2026
[16] US approves export of Nvidia AI chips to UAE and Saudi Arabia — The National, 20 Nov 2025
[17] MITI Imposes Permit Requirement On US-Origin AI Chip Exports, Transshipments To Close Regulatory Gaps — Bernama, 14 Jul 2025
[18] Department of Commerce Announces New American AI Exports Program Phase — International Trade Administration, US Department of Commerce, 16 Mar 2026