Global AI Governance

AI export controls are government restrictions on the transfer of advanced AI-relevant items, principally high-performance computing chips, semiconductor manufacturing equipment, and related technology, to specified destinations or end users. In practice the central actor is the United States Bureau of Industry and Security (BIS), whose rules target the compute supply available to China and other designated destinations.

Training and deploying frontier models depends on concentrated, physically controllable inputs (see TECH-HW-001 and ECON-COMP-001). Governments therefore treat chips and fabrication tools as strategic. Policy has oscillated between tightening, rescission and conditional licensing since 2022, and is entangled with trade negotiations, so that the same instrument serves security, industrial and bargaining purposes. This is the geopolitical dimension of the compute-governance idea in GOV-INTL-001.

7 October 2022. BIS issued an interim final rule, "Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items; Supercomputer and Semiconductor End Use; Entity List Modification", published in the Federal Register on 13 October 2022 (87 FR 62186). It added Export Control Classification Numbers (ECCNs) 3A090 (high-performance integrated circuits) and 4A090 (computers and assemblies containing them) to the Commerce Control List, expanded foreign-produced-item licence requirements for 28 Entity List entities in China, and informed the public that certain U.S.-person activities that support the development or production of certain ICs in the PRC require a licence. Its provisions took effect in phases: 7, 12 and 21 October 2022 (Federal Register, 2022; Ref 9).

Later rounds. Two further rules were published on 25 October 2023, both effective 17 November 2023: "Implementation of Additional Export Controls: Certain Advanced Computing Items; Supercomputer and Semiconductor End Use; Updates and Corrections" (88 FR 73458) and "Export Controls on Semiconductor Manufacturing Items" (88 FR 73424). A rule on foreign-produced direct product additions and refinements for advanced computing and semiconductor manufacturing items was published on 5 December 2024 (89 FR 96790; effective 2 December 2024). Only the titles, dates and effective dates were checked from Federal Register metadata; the contents of these three rules were not read.

AI Diffusion Rule (January 2025) and its status. The interim final rule "Framework for Artificial Intelligence Diffusion" was published on 15 January 2025 (90 FR 4544), effective 13 January 2025, with a general compliance date of 15 May 2025 (Federal Register, 2025a; Ref 10). On 13 May 2025, according to a law-firm client alert, BIS announced the rescission of the rule and issued three guidance documents, including guidance that Huawei Ascend 910B, 910C and 910D chips are presumed to implicate General Prohibition 10 (Crowell & Moring, 2025) [VERIFICATION NEEDED: the BIS guidance documents themselves were not read]. As of 18 October 2026 no Federal Register document rescinding the rule was found in a search of Federal Register metadata for Bureau of Industry and Security documents, and the eCFR (version of 1 October 2026) still contains the rule's provisions, including License Exception Artificial Intelligence Authorization (15 CFR 740.27), ECCN 4E091 (AI model weights) and the related licence requirement in 15 CFR 742.6(a)(13) (eCFR, 2026). The Crowell alert describes the rescission as formal; the Federal Register and eCFR records checked do not show it [VERIFICATION NEEDED: whether any formal rescission or replacement rule exists, and how BIS treats the codified text].

H20 (April 2025). In a Form 8-K filed 15 April 2025, Nvidia reported that on 9 April 2025 the U.S. government informed it that a licence is required to export its H20 integrated circuits (and circuits achieving the H20's memory bandwidth, interconnect bandwidth, or combination thereof) to China including Hong Kong and Macau, and to D:5 countries or companies headquartered in them, and on 14 April that the requirement would be in effect for the indefinite future (NVIDIA, 2025; Ref 16). Nvidia's Form 10-Q for the quarter ended 26 April 2026 states that its first quarter of fiscal year 2026 (quarter ended 27 April 2025) included a USD 4.5 billion charge for H20 excess inventory and purchase obligations (NVIDIA, 2026).

H200 policy shift (January 2026). BIS published a final rule, "Revision to License Review Policy for Advanced Computing Commodities" (91 FR 1684), on 15 January 2026, effective the same day. It changes the licence review policy for exports to China and Macau from a presumption of denial to case-by-case review for advanced computing commodities with a total processing performance (TPP) below 21,000 and a total DRAM bandwidth below 6,500 GB/s, such as the Nvidia H200 or AMD MI325X, subject to conditions. The rule's stated conditions include that the product is commercially available in the United States, that supply and foundry capacity for U.S. customers is not diverted, that aggregate shipments to China and Macau do not exceed 50% of the total shipped to U.S. customers, that the consignee applies know-your-customer procedures, and that every shipment is reviewed by an independent third-party testing laboratory in the United States. The presumption of denial is maintained for entities headquartered in, or with a parent headquartered in, Macau or a Country Group D:5 destination (Federal Register, 2026a; Ref 11). Separately, Proclamation 11002 of 14 January 2026 (published 20 January 2026, 91 FR 2443), issued under section 232 of the Trade Expansion Act of 1962, imposes a 25 percent ad valorem duty on imports of certain advanced computing chips and derivative products, when the importation does not contribute to the buildout of the U.S. technology supply chain, effective for goods entered on or after 12:01 a.m. eastern standard time on 15 January 2026 (Federal Register, 2026b; Ref 12). Nvidia's quarterly filing (period ended 26 April 2026) says that beginning in February 2026 the U.S. government granted licences allowing small shipments of H200 products to specific China-based customers, that the licence requires inspection in the United States, that H200s shipped under the programme are subject to a 25% tariff upon importation into the United States, that it had generated no revenue under the programme to date, and that it did not know whether China would allow imports (NVIDIA, 2026). Law-firm summaries (Baker McKenzie, 2026; Mayer Brown, 2026) were not read in full and were not relied on for the facts above.

Affiliates rule and trade-truce suspensions. BIS published the interim final rule "Expansion of End-User Controls to Cover Affiliates of Certain Listed Entities" on 30 September 2025 (90 FR 47201); it extends Entity List restrictions to entities at least 50 percent owned by listed parties. A final rule published 12 November 2025 (90 FR 50857) suspends it for one year: the changes are stayed from 10 November 2025 until 9 November 2026 and are added back to the EAR effective 10 November 2026 (Federal Register, 2025b; Ref 13). No extension or other later BIS document on this rule was found in Federal Register metadata through 18 October 2026, so, absent a future extension, the suspension ends within about three weeks of this review date. On the Chinese side, law-firm summaries report that on 7 November 2025 China's Ministry of Commerce and customs authority issued Announcement No. 70 of 2025 suspending, until 10 November 2026, export-control measures announced on 9 October 2025 (rare earths, lithium batteries and related items) following the 30 October 2025 Trump-Xi meeting (China Briefing, 2025) [VERIFICATION NEEDED: the MOFCOM announcements were not read; an extension reported by one source to January 2027 could not be confirmed and conflicts with the 9 November 2026 end date in the Federal Register].

Other 2026 BIS rules. A final rule effective 10 July 2026 (91 FR 43034, published 14 July 2026) moves the United Arab Emirates from Country Groups D:3 and D:4 to A:5 and, per its summary, gives the UAE Government and approved commercial entities licence-free access to advanced computing items (Federal Register, 2026c; Ref 14). A rule effective 7 April 2026 (91 FR 17851) extends authorized integrated circuit designer status and the deadline to become an approved designer; only its title and date were checked.

What is controlled. Performance-based thresholds on chips (aggregate processing performance and interconnect/memory characteristics), manufacturing tools, and in some cases model-related items. As of the eCFR version of 1 October 2026, ECCN 3A090.a covers integrated circuits with one or more digital processing units having a "total processing performance" of 4800 or more, or a TPP of 1600 or more together with a "performance density" of 5.92 or more; the licence requirement for 3A090.a applies to any destination worldwide under 15 CFR 742.6(a)(6)(iii)(A) (eCFR, 2026; Ref 15). Thresholds for other ECCNs (3A090.b and .c, 4A090) and for manufacturing equipment were not checked. The October 2022 rule used different, non-TPP criteria; TPP-based criteria are described here only as currently codified.

Mechanisms. (1) Licence requirements and review policies (presumption of denial versus case-by-case); (2) Entity List designations; (3) foreign direct product rules extending U.S. jurisdiction to non-U.S.-made items; (4) end-use and U.S.-person controls; (5) end-user diligence and diversion guidance. Controls can be combined with tariffs and licence conditions such as inspection and know-your-customer requirements.

Objectives. Stated objectives include preventing military end uses and slowing frontier AI development by adversaries; critics argue controls spur indigenous substitution. The empirical effect on Chinese AI capability is Contested and is not assessed from a primary study here.

Allied-country measures (for example, Dutch and Japanese equipment controls) are not detailed because no source was retrieved [VERIFICATION NEEDED if added].

Examples: the H20 and H200 licensing episodes above; guidance naming Huawei Ascend chips; and chip-smuggling enforcement, motivating bills such as the Chip Security Act, which the House Foreign Affairs Committee describes as directing the Department of Commerce to set up security mechanisms for advanced chips exported abroad (House Foreign Affairs Committee, 2026; Congress.gov, 2026; Refs 17, 18). Quantities of chips shipped are fast-changing and not reported here; see FACT-001.

• Arguments for: slows access to leading-edge compute by designated actors; compute is physically concentrated and traceable.
• Arguments against: smuggling and cloud access can circumvent controls; policy reversals create uncertainty for suppliers; controls may accelerate domestic alternatives; they impose revenue costs on suppliers (Nvidia's disclosed H20 charge is one dated example).
• Unresolved: effectiveness; stability of the January 2026 policy given the absence of reported revenue under the H200 programme as of Nvidia's April 2026 quarter and the Chinese import stance (Nvidia states it does not know whether China will permit imports); whether the diffusion framework is formally rescinded or replaced.

As of 2026-10-18 the policy is in flux. The 2022 baseline persists with amendments; the Diffusion Rule was announced as rescinded in May 2025 but its text remained in the eCFR on 1 October 2026 and no rescinding Federal Register document was found; H200-class chips (TPP below 21,000 and DRAM bandwidth below 6,500 GB/s) may be licensed to China and Macau case by case with conditions since 15 January 2026, with a 25% section 232 duty; and the one-year suspension of the affiliates rule is scheduled to end on 9 November 2026 unless extended. Claims about the strategic effect of the controls are Contested.

No forecast is registered. Items to watch: the scheduled end of the affiliates-rule suspension on 9 November 2026 (a Federal Register API search of Bureau of Industry and Security documents on 2026-10-18 found no extension or replacement notice) and of the reported Chinese suspensions on 10 November 2026, a formal rescission or replacement of the diffusion rule, congressional chip legislation, and Chinese import decisions. Dated updates belong in TIME-001.

• Foundational Concepts:
  • ECON-COMP-001 The Economics of Compute and Scaling
  • TECH-HW-001 AI Hardware
• Related:
  • GOV-US-001 United States AI Policy: the wider US policy context.
  • GOV-CN-001 China's AI Regulation
  • GOV-INTL-001 International AI Governance and Compute Governance
  • ECON-INFRA-001 Energy, Data Centers and Semiconductor Supply Chains: the supply chain being controlled.

International AI governance comprises the treaties, intergovernmental declarations, summit processes and scientific-assessment bodies through which states coordinate policy on AI. Compute governance is a sub-approach that uses control of the computational hardware needed to train and run AI systems as a point of visibility and intervention, because compute is detectable, excludable and quantifiable (Sastry et al., 2024).

No binding global AI treaty with broad membership is in force as of 2026-10-18. International coordination consists mainly of non-binding declarations, summit series, a UN scientific panel and dialogue process, and one regional-plus treaty (the Council of Europe Framework Convention) that has not yet entered into force. National regimes, described in GOV-EU-001, GOV-US-001 and GOV-CN-001, differ substantially, and coordination has been shaped by those differences and by geopolitical competition (GOV-EXPORT-001).

Summit series.

• Bletchley Park, 1-2 November 2023. The Bletchley Declaration was signed by 28 countries and the European Union (29 participants); New Zealand joined on 23 October 2024 (UK Government, 2023). It states that "particular safety risks arise at the 'frontier' of AI", says substantial risks may arise from intentional misuse or unintended issues of control, names cybersecurity and biotechnology as areas of special concern, and refers to "potential for serious, even catastrophic, harm, either deliberate or unintentional".
• Seoul, May 2024. The AI Seoul Summit was co-hosted by South Korea and the UK on 21-22 May 2024 [VERIFICATION NEEDED: dates and the Seoul Declaration's signatory list were taken from search-result excerpts of government copies; the primary page could not be opened]. The UK Government's "Frontier AI Safety Commitments" page lists 16 original organisations, with four more added (Magic, Minimax, 01.ai, NVIDIA) by its update of 7 February 2025 (UK Government, 2025). These are voluntary company commitments, not state treaty obligations.
• Paris, 10-11 February 2025. The "Statement on Inclusive and Sustainable Artificial Intelligence for People and the Planet", published on the Élysée website, lists 65 signatories: 63 countries, the African Union Commission and the European Union. The United States and the United Kingdom are not listed (Élysée, 2025). Media reports of 58 or 61 signatories differ from the official list.
• New Delhi AI Impact Summit (February 2026). The "New Delhi Declaration on AI Impact" is described as non-binding. Brookings (12 March 2026) reports that the United States joined 91 countries and international organisations in signing it, with all G7 countries, China and the EU among the endorsers. The count changed over time: search-result excerpts of Indian government (PIB) releases give 88 as of 21 February 2026, 91 by 24 February and 92 in later releases, and other outlets give 89 [VERIFICATION NEEDED: PIB pages returned access-denied; current official count not confirmed]. The agenda shifted toward adoption and development. Switzerland announced a follow-on summit in Geneva in 2027 (Brookings, 2026).

United Nations. The General Assembly adopted resolution A/RES/79/325 by consensus on 26 August 2025, establishing an Independent International Scientific Panel on AI (40 members) and a Global Dialogue on AI Governance. The panel's co-chairs, elected 3 March 2026, are Yoshua Bengio and Maria Ressa. A preliminary report was released on 1 July 2026, and the first Global Dialogue was held in Geneva on 6-7 July 2026; the second is scheduled for May 2027 in New York (United Nations, 2026). The panel states it is not a regulatory body and does not set rules.

Council of Europe. The Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225) is the first binding international AI treaty. The Dutch government treaty database lists the European Union as ratifying on 15 May 2026, and no other ratification, alongside signatures from 20 other parties (Albania, Andorra, Armenia, Bosnia and Herzegovina, Canada, Georgia, Iceland, Israel, Japan, Liechtenstein, Moldova, Montenegro, North Macedonia, Norway, San Marino, Switzerland, Ukraine, the United Kingdom, the United States, Uruguay); the earliest listed signature is dated 5 September 2024 (Dutch Ministry of Foreign Affairs treaty database; snapshot date not stated). Search-result excerpts describing Article 30 say entry into force requires five consent-to-be-bound expressions including at least three by Council of Europe member states, followed by a three-month period. On that basis the Convention was not in force on 2026-10-18 [VERIFICATION NEEDED: coe.int Treaty Office pages returned 403 and the Convention text was not read; current ratification count and Article 30 wording not confirmed first-hand].

Why compute. Sastry et al. (2024) argue that compute is a tractable governance lever for three reasons: it is detectable (physical, energy-intensive facilities), excludable (it can be withheld, as with export controls), and quantifiable (it can be counted in floating-point operations). They identify three uses: increasing regulatory visibility, allocating resources toward beneficial uses, and enforcing restrictions on irresponsible or malicious development. They also warn of risks to privacy, the economy and power concentration if designed badly.

Thresholds. Regulatory thresholds based on training compute include the EU AI Act's presumption of systemic risk above $10^{25}$ floating-point operations (Article 51(2); see GOV-EU-001) and California's SB 53 definition above $10^{26}$ (secondary sources; see GOV-US-001). A cumulative-FLOP threshold is a proxy: it can be crossed by efficient large runs or avoided by smaller, more capable models, and is revised periodically. How FLOPs are counted is itself a measurement choice (see ECON-COMP-001).

Mechanisms discussed in the literature. Compute accounting and reporting, know-your-customer duties for cloud providers, export controls, and hardware-enabled monitoring. Sastry et al. describe which are piloted and which need research; this article does not claim any international verification regime is operating.

Safety institute networks are covered in ORG-SAFETY-001.

Concrete applications of compute governance are national: the EU threshold and US export controls. Internationally, the UN panel has produced a preliminary report (1 July 2026) and a thematic brief on AI agents and misalignment (advance unedited version dated 21 September 2026 per the UN FAQ). No multilateral compute-monitoring body exists in the sources reviewed.

• Advantages: shared risk vocabulary, evidence pooling, low commitment cost.
• Limitations: non-binding outputs; divergent positions (the US reportedly rejects binding international AI rules and new global governance bodies, per Brookings, 2026); inconsistent signatory lists; summit agenda drift between safety and adoption.
• Contested: whether international verification of compute use is technically and politically feasible; whether compute thresholds track risk. Evidence is limited to proposals and early pilots (Sastry et al., 2024).

As of 2026-10-18: declarations exist but no binding multilateral AI-safety regime with major-power membership is in force. A UN panel and dialogue are operating. The Council of Europe treaty awaits entry into force. Compute governance is implemented only through national and EU measures.

Scheduled events: the second Global Dialogue (2027) and a Geneva summit (2027). These are announcements, not forecasts. No quantified forecast is registered here; see FCST-REG-001. Proposals for international compute verification remain hypothetical.

• Foundational Concepts:
  • GOV-LAW-001 Global AI Governance & Regulation
  • SAFE-XRISK-001 Existential Risk from AI: risk framing behind frontier-focused declarations.
• Related Methods & Technologies:
  • TECH-HW-001 AI Hardware: the hardware being governed.
  • ECON-COMP-001 The Economics of Compute and Scaling
• National regimes:
  • GOV-EU-001 The EU AI Act
  • GOV-US-001 United States AI Policy
  • GOV-CN-001 China's AI Regulation
  • GOV-EXPORT-001 Export Controls and AI Geopolitics