Geopolitics

Washington's control regime and Beijing's answer to it are both unfinished.

The shape of the contest

In September 2022 National Security Advisor Jake Sullivan called securing "as large a lead as possible" in technologies such as AI a national security imperative, and the October rules that followed debuted his doctrine of a small yard with a high fence^[1]. Those rules went beyond the old aim of staying two technology nodes ahead of China and tried to push China's capability below its current level^[2]. Licensing once asked who was buying and why. The 2022 rules added where the item was going, with a license required for any advanced AI chip bound for China and for tools bound for Chinese fabs past set technology thresholds^[3]. Most applications were presumed denied, though equipment for the China fabs of firms headquartered in the United States or allied and partner countries was reviewed case by case^[3].

Chokepoints and who holds them

• Equipment. CSET's ETO Supply Chain Explorer, TechInsights 2025 revenue by parent-company headquarters^[4].
• Materials. Country shares of photoresist and silicon wafers as Japan's Ministry of Economy, Trade and Industry published them in December 2025, from Fuji Keizai's materials survey, by parent-company headquarters^[5]. Japan held 77 percent of photoresist sales in 2021 and 78 percent in 2023, and 52 percent of silicon wafer sales against 53 percent, so neither position is moving.
• Taiwan. The 78 percent is the Taiwan Semiconductor Industry Association's 2025 count, as the Stimson Center gives it^[6]. The 69 percent is the Semiconductor Industry Association and Boston Consulting Group's 2022 measurement of logic capacity below 10 nm, the most recent the report gives^[7]. The 92 percent quoted more often is the same measure for 2019, from the two groups' April 2021 report^[8].
• Memory. SK hynix's US listing prospectus splits 2025 high-bandwidth memory revenue between SK hynix at 63.2 percent, Samsung at 19.3 and Micron at 17.4, so Korea holds 82.5; by the first quarter of 2026 Korea was down to 76.9^[9]. The underlying figures are IDC's, relayed by the prospectus.
• Design. CSET's ETO Supply Chain Explorer, from WSTS and SIA figures on 2022 sales of logic chips and microprocessors, still the latest year it gives for this segment^[10]. On a wider measure that counts every type of chip, memory and analog included, US-headquartered firms took 53.4 percent of global chip sales in 2025^[11].
• China's mature nodes. The 33 percent is the 2022 industry measurement of logic capacity at 28 nm and above, and the same report projects 37 percent by 2032^[7]. In a February 2025 filing the Semiconductor Industry Association counted every type of chip made at 28 nm and above, memory and analog included, and put China's share of that capacity at 33 percent in 2023, up from 19 percent in 2015^[12].

Advanced packaging is missing because nobody publishes a capacity split and TSMC will not give one. No country is self-sufficient, so each side's controls can hurt the other.

The US export-control campaign, 2018 to today

Commerce built the regime one rule at a time, each closing a gap the last one left.

• 17 August 2020, the Foreign Direct Product Rule. The Entity List, Commerce's roster of buyers that need a license, cut Huawei and 68 affiliates off from US technology in May 2019^[13], so Huawei bought foreign chips instead. Commerce then extended the rule to any foreign item made with US technology and bound for Huawei^[14]. That reach makes American controls extraterritorial.
• 7 October 2022, the first country-wide rules. Two new export classifications, 3A090 and 3B090, capped exportable AI chips and set fab thresholds at FinFET or 16/14 nm logic, 18 nm half-pitch DRAM and 128-layer NAND^[3]. They also required a license for US persons to support chip development or production at those advanced Chinese fabs^[3].
• 17 October 2023, the density fix. Nvidia had designed its China-specific chips around the 2022 interconnect-speed threshold, so Commerce's Bureau of Industry and Security (BIS) dropped that parameter and now identifies controlled chips by total processing performance, performance density, and whether they are designed or marketed for data centers^[15]. A data-center chip now needs a license at 4,800 points of that performance score, or at 1,600 with a density of 5.92^[15].
• 2 December 2024, the widest rule. 24 equipment types, three software tools, the first controls on high-bandwidth memory and 140 Entity List additions^[16].
• 15 January 2025, AI Diffusion. A rule requiring a license for the top band of accelerators, 3A090.a and 4A090.a, to any destination worldwide, then reopening it by exception: License Exception AIA to the nineteen destinations named in supplement no. 5 to part 740, License Exception LPP in capped volumes elsewhere, nothing to China^[17]. BIS said on 13 May that it would not enforce the framework, two days before the compliance date^[18]. It withdrew none of the text. The worldwide license requirement, License Exceptions AIA, ACM and LPP, supplement no. 5 and the 790 million TPP per-country allocation all stand in the Code today, and BIS says it enforces the requirement only against Country Groups D:1, D:4 and D:5 and against firms headquartered in D:5 or Macau^[19][20]. No tiered replacement has appeared; bilateral deals took the place of the caps.

On 9 April 2025 the government told Nvidia it needed a license to ship the H20 to China, to the D:5 group of arms-embargoed countries, and to any company headquartered or ultimately owned there, so no Chinese buyer could route the order through a foreign subsidiary^[21]. Nvidia took a $4.5 billion charge for excess H20 inventory and purchase obligations^[22]. Licenses came back in August 2025, and the same filing says US officials expected the government to receive 15 percent of the revenue from licensed H20 sales, though no regulation codifies that requirement^[22]. Nvidia's annual report for the year to 25 January 2026 put its H20 revenue under those licenses at about $60 million^[23]. Beijing restricted those sales, and Nvidia said in its filing for the quarter ended 26 July 2026 that it had been unable to sell its H20 inventory^[24].

Early 2026 set the current settlement.

• 13 January 2026, effective 15 January. BIS announced on 13 January that it would move the H200 and AMD MI325X to case-by-case review, on conditions that included independent third-party testing in the United States^[25]. The rule, at 91 FR 1684, took effect on publication two days later and caps exports to China or Macau at 50 percent of the total processing performance of the same chips shipped to US customers^[26].
• 14 January 2026, effective 15 January. A national-security proclamation under Section 232 put a 25 percent tariff on the same class of chips and exempted almost every American use, from data centers to startups^[27]. The H200 licenses require each China-bound chip to pass an inspection in the United States before shipment, so every licensed H200 is imported first and pays the tariff^[24]. Nvidia said in its filing for the quarter ended 26 July 2026 that it had been unable to pass any of that tariff on to its customers^[24].
• 12 February 2026. Applied Materials agreed to pay about $252 million, the statutory maximum, over about $126 million of ion implanters routed through Korea to an Entity-Listed Chinese customer^[28].

Eight rule changes that shaped the regime

1. 2019-05-16Huawei and 68 affiliates added to the Entity List
2. 2022-10-07First country-wide controls on AI chips and fab tools
3. 2023-10-17Performance density added; A800 and H800 captured
4. 2024-12-0224 tool types, HBM controls, 140 entity listings
5. 2025-01-15AI Diffusion rule published, three country tiers
6. 2025-05-13BIS stops enforcing the AI Diffusion rule
7. 2025-08-11US officials seek 15% of licensed H20 sales
8. 2026-01-13H200 and MI325X case-by-case review announced, effective 15 January

Source: BIS press releases, Federal Register rules, a CSET explainer and Nvidia filings, cited in the text

Cloud and remote access

Renting a chip is not exporting it. An export, at 15 CFR 734.13, is a shipment or transmission out of the United States or a release of technology or source code to a foreign person in the United States, so a Singapore data center selling time on an H200 already installed there moves nothing across a border^[29].

The American rules reach that trade through the firms holding the chips. A January 2026 license for China or Macau is conditioned on the applicant getting the consignee's know-your-customer procedures and its list of Infrastructure-as-a-Service customers in Belarus, China, Cuba, Iran, Macau, North Korea, Russia and Venezuela, "necessary to prevent unauthorized remote access from end users described in paragraph (dd)(1)(iv)", and on a promise that no algorithm trained on the chips is served back to those users^[30].

The diffusion rule addressed the other end, the output. It classified frontier model weights as ECCN 4E091, warned American cloud providers that training such a model for the US subsidiary of a foreign-headquartered customer raises a red flag that the weights will leave without a license, and barred its validated end users from training one outside the nineteen listed destinations without BIS authorization, while saying that API access to a model, and rented capacity for inference, "are not prohibited"^[17].

So the controls reach the chip, frontier model weights, and the promises a licensed consignee signs. Compute sold by the hour on chips lawfully installed abroad sits outside them.

Allies: the Netherlands, Japan, Korea and Europe

Unilateral tool controls would only transfer the sales, so the campaign depends on The Hague and Tokyo.

• The Netherlands. The Dutch began licensing advanced deep-ultraviolet tools on 1 September 2023, widened that in September 2024^[31] and reached measurement and inspection tools on 1 April 2025^[32].
• Japan. Controls on advanced manufacturing equipment took effect on 23 July 2023 and apply to all destinations and name no country^[33].
• South Korea. BIS revoked the Validated End-User status that let Samsung's and SK hynix's China subsidiaries and Intel Semiconductor (Dalian) receive US equipment without individual licenses, effective 31 December 2025^[34]. BIS said it would license those fabs to keep running, but not to expand capacity or upgrade technology^[35]. SK hynix, which bought the Dalian fab from Intel, received an annual BIS license in December 2025 to ship approved US equipment to its China fabs during 2026^[36].

A BIS rule effective 29 September 2025 extended Entity List restrictions to any company at least 50 percent owned by one or more listed firms^[37]. On 30 September the Dutch economy minister invoked the Goods Availability Act against Nexperia, the Nijmegen chipmaker that supplies European car lines, citing governance failures that risked moving technology out of Europe^[38]. Neither lasted: BIS suspended its rule for a year on 10 November^[39] and the minister suspended his order on 19 November^[40]. Europe holds a chokepoint at the leading edge and none at the mature nodes its car industry runs on.

The MATCH Act, introduced on 8 April 2026, would ban sales of immersion deep-ultraviolet scanners to China, name CXMT, Hua Hong, Huawei, SMIC and YMTC in statute, and give allies 150 days to match before the US extends the Foreign Direct Product Rule alone^[41].

China's response

Five bodies run Beijing's side. The Ministry of Industry and Information Technology (MIIT) sets substitution targets, and in 2024 reportedly told Chinese telecom operators to strip foreign chips out of their networks by 2027^[42]. The National Development and Reform Commission (NDRC) plans the compute build-out and co-issued the national computing-power network plan^[43]. The Ministry of Commerce (MOFCOM) writes and administers the export controls, including the mineral rules below^[44]. The State Administration for Market Regulation (SAMR) is the antitrust authority, and opened its Nvidia investigation in December 2024^[45]. The Cyberspace Administration of China (CAC) runs the security reviews, and in September 2025 told firms including ByteDance and Alibaba to cancel Nvidia orders^[46].

The retaliation itself targets the inputs. Micron was first, failing a CAC review on 21 May 2023 and losing its critical infrastructure buyers^[47]. Mineral controls followed.

• 3 December 2024. Notice 46, from MOFCOM, banned exports of gallium, germanium, antimony and superhard materials to the United States, one day after the American chip rules^[48].
• 4 April 2025. Announcement 18 put seven medium and heavy rare earths under licensing, from samarium to yttrium, as metal, oxide, compound or magnet material^[44].
• 9 October 2025. A further announcement reached foreign-made goods containing Chinese rare earths worth 0.1 percent or more of their value, putting a Chinese export license between a foreign firm and its own product^[49]. MOFCOM suspended it on 7 November 2025 until 10 November 2026^[50].
• November 2025. After Trump and Xi met at the Asia-Pacific summit, the ministry suspended the gallium ban for a year^[51].

Nothing was repealed; the gallium ban and the October rules are paused, and the April licensing regime stands.

The blunter instrument is demand. Chinese regulators discouraged customers from buying, importing or using Nvidia's data-center products, and endorsed performance-per-watt and bandwidth standards for accelerators in Chinese data centers^[24]. When Washington licensed H200 sales in February 2026, Beijing restricted them, and Nvidia's shipments came to less than 1 percent of data-center revenue^[24]. That quarter's data-center revenue was $89.0 billion, and the next quarter's guidance assumed no China compute revenue^[52].

Chinese tool vendors held 10 to 15 percent of their home market before 2022, took 25 percent in 2024 and 35 percent in 2025, and hold about 40 percent of domestic etch and deposition^[42]. At ASML, China's share of net system sales fell from 41 percent in 2024 to 33 percent in 2025^[53].

Substitution stalls at the leading edge. Without extreme-ultraviolet scanners, SMIC's most advanced process, N+3, reaches the density of TSMC's 6 nm-class process by printing its tightest layers in several passes on deep-ultraviolet tools and by redesigning its standard cells^[54]. In September 2025 SemiAnalysis expected Huawei to ship about 805,000 Ascend units in 2025, with dies made by both TSMC and SMIC^[55]. It judged memory the binding constraint. By its count, a stockpile of roughly 13 million memory stacks, most bought before the December 2024 rules took effect, was enough for about 1.6 million Ascend 910C packages, and it expected CXMT, China's own memory maker, to make only about 2 million stacks in 2026, enough for 250,000 to 300,000 more^[55]. Huawei said in September 2025 that its Ascend 950 chips would carry high-bandwidth memory of its own design, HiBL 1.0 on the 950PR due in the first quarter of 2026 and HiZQ 2.0 on the 950DT due in the fourth quarter, and it gave no figure for how much of that memory it would make^[56].

China's share of revenue at the four largest chip-tool makers, peak fiscal year and latest%

China's share of revenue at the four largest chip-tool makers, peak fiscal year and latest

ASML, 2024	41 %
ASML, 2025	33 %
Applied Materials, FY2024	37 %
Applied Materials, FY2025	30 %
Lam Research, FY2024	42 %
Lam Research, FY2026	34 %
Tokyo Electron, FY2024	44.4 %
Tokyo Electron, FY2026	34.1 %

Source: ASML Q4 2025 investor presentation, Applied Materials and Lam Research 10-K filings, Tokyo Electron Q4 FY2026 results

Industrial policy: CHIPS, Rapidus, Big Fund and the rest

The CHIPS and Science Act authorized about $52.7 billion^[42]. By July 2025 Commerce had awarded $30.9 billion of direct funding to 19 companies for 40 projects, a median 14.2 percent of each project's capital cost^[57]. Commerce counted a federal tax credit, the Advanced Manufacturing Investment Credit, when it sized those grants^[57]. Since the One Big Beautiful Bill Act of 2025, the credit covers 35 percent of qualified investment in an advanced chip plant placed in service after 2025, up from 25 percent^[57].

On 22 August 2025 the government converted $5.7 billion of unpaid Intel grants and $3.2 billion of Secure Enclave defense money into an $8.9 billion purchase of Intel stock, equivalent to a 9.9 percent stake^[58].

Tariffs became the second instrument. The 2026 US-Taiwan trade agreement cuts the reciprocal tariff on Taiwanese goods to 15 percent in exchange for at least $250 billion of Taiwanese investment in US chip production^[59]. On 16 July 2026 TSMC announced another $100 billion for four more Arizona fabs at 2 nm or more advanced, bringing its total Arizona investment to $265 billion^[60].

The European Commission says the EU still depends on other countries for advanced chip manufacturing, and its Chips Act 2.0 proposal, adopted on 3 June 2026, adds measures to support advanced chip production in the EU^[61]. Japan backs a national champion instead. Rapidus began prototyping 2 nm gate-all-around transistors at its IIM-1 plant in July 2025 and measured their first electrical characteristics^[62]. In February 2026 it raised another 267.6 billion yen, about $1.7 billion, from a government agency and 32 companies, aiming for 2 nm mass production by 2027^[63]. By June 2026 it had raised 424.95 billion yen in equity, the latest 150 billion yen from a state agency^[64].

China spends more, and differently. The latest round of its National Integrated Circuit Industry Investment Fund, the Big Fund, reached $47 billion^[65]. CSIS puts cumulative state funding since 2014 at an estimated $150 billion and calls the CHIPS Act substantially smaller in scale^[42]. The money has produced a domestic tool industry and no leading-edge node.

CHIPS Act direct funding awarded, by firm, as of July 2025$B

CHIPS Act direct funding awarded, by firm, as of July 2025

Intel	7.9 $B
Micron	6.4 $B
TSMC	6.6 $B
Samsung	4.7 $B
Texas Instruments	1.6 $B
GlobalFoundries	1.6 $B
SK hynix	0.5 $B
Amkor	0.4 $B
GlobalWafers	0.4 $B
Hemlock Semiconductor	0.3 $B

Source: GAO-26-107882, Semiconductors: Information on Projects Funded to Strengthen U.S. Supply Chain

Taiwan

Taiwan is the concentration risk that no policy has reduced. Its foundries hold 78 percent of the world's contract chipmaking on the industry association's 2025 count, and TSMC makes its 2 nm chips only in Taiwan^[6]. Below 10 nm the island held 69 percent of the world's logic capacity in 2022, on a 2024 count by the Semiconductor Industry Association and Boston Consulting Group^[7]. The packaging is on the island too. Asked in October 2025 how much CoWoS capacity he would add, TSMC's chairman said only that it was "working very hard to narrow the gap between the demand and supply", with wafer and packaging capacity both "very tight"^[66].

The silicon shield argument holds that this concentration deters invasion: destroying TSMC would cost an aggressor more than the island is worth. The counterargument is that staying irreplaceable gives Taipei reason to slow the diversification Washington wants. Taiwan's rules keep TSMC's newest process at home first, so what TSMC builds in Arizona has run a generation behind what it builds in Taiwan, and Stimson expected 2 nm production there in 2028 at the earliest^[6]. TSMC said in October 2025 that it was "preparing to upgrade our technologies faster to N2 and more advanced process technologies in Arizona"^[66]. In July 2026 it announced four more fabs there at 2 nm or more advanced^[60].

Arizona narrows the gap in wafers, and TSMC has two advanced packaging fabs planned there^[66]. But a fab without its Taiwanese mask shops, chemical suppliers and trained crews is not a substitute. The shield is real and weakening, and diversification is not keeping up.

Where the chips go: the Gulf, Southeast Asia, and smuggling

The Gulf has been let inside the controls. In November 2025 Commerce authorized G42 of the United Arab Emirates and HUMAIN of Saudi Arabia to buy up to 35,000 Nvidia GB300s each under security and reporting conditions^[67]. On 10 July 2026 it moved the Emirates from Country Groups D:3 and D:4 into A:5, the tier that holds Washington's close allies, and named the Emirati government alongside G42 and US-headquartered AI companies as consignees who can take controlled AI chips and systems without a license. Licenses survive for the top-tier items, 3A090.a and 4A090.a, so the favorable treatment is tied to the named buyers^[19]. Saudi Arabia still works license by license.

Southeast Asia is where chips get past the controls. Since 14 July 2025 Malaysia has required a permit to export, tranship or transit any high-performance US-origin AI chip, plus 30 days' notice under its Strategic Trade Act^[68]. On 25 March 2026 the Justice Department announced charges against three people accused of using the names of Thai companies to order servers with export-controlled chips for China, including a $170 million order that was never completed^[69].

The House select committee on China cites estimates that a median 140,000 chips were smuggled to China in 2024^[70]. Congress's answer, the Chip Security Act, would require location verification on controlled chips within 180 days of enactment^[71]. The House Foreign Affairs Committee passed it 42-0 on 26 March 2026^[72]; it has not become law.

Who gains and who pays

• Taiwan and Korea. Nothing in the policy mix moves leading-edge logic, memory stacking or advanced packaging off an island and a peninsula before 2030; reshoring adds capacity elsewhere without subtracting it there.
• The Netherlands and Japan. Their tool makers build the controlled machines, so they gain a say in who gets what, while China counts for less of their sales. Tokyo Electron's China share fell from 44.4 percent in the year to March 2024 to 34.1 percent in the year to March 2026, and the company said the share dropped to 26.8 percent in the quarter to March 2026 because spending on leading-edge chips grew faster than spending on older ones^[73].
• The United States. Its firms took 65 percent of the world's logic chip sales in 2022^[10]. Counting every type of chip, they took 53.4 percent of global sales in 2025, their highest share since 1984^[11]. Commerce expects the CHIPS-funded projects to lift the US share of global leading-edge logic chip manufacturing from 0 percent in 2022 to 20 percent by 2030^[57].
• China. Mature nodes, sub-leading-edge tools, DRAM and NAND are going domestic on a visible timetable. Extreme-ultraviolet machines, their resist and high-bandwidth memory at scale are not, and in September 2025 SemiAnalysis expected China, without smuggled memory, to make fewer Ascends in 2026 than in 2025^[55].

The Gulf gains without making anything, since named Emirati buyers can now receive advanced computing items without a license^[19]. And Washington's instruments now reach past licensing. Every licensed H200 bound for China pays a 25 percent tariff on its way through US inspection^[24]. The government agreed in August 2025 to take a 9.9 percent stake in Intel^[58]. Sales of the H200 to China pass through a case-by-case queue^[26].

What to watch

• November 2026. Whether China renews the one-year suspension of its gallium ban or lets it expire^[51].
• The MATCH Act's 150-day deadline. If it passes, allied divergence on deep-ultraviolet tools gets a statutory deadline^[41].
• CXMT's memory ramp. In September 2025 SemiAnalysis judged that domestic memory would cap Chinese accelerator volume before logic wafers did^[55].
• SMIC's advanced-node ramp. In September 2025 SemiAnalysis conservatively estimated SMIC's capacity at 7 nm and below at 45,000 wafers a month by the end of 2025, 60,000 in 2026 and 80,000 in 2027^[55]. SMIC makes those wafers without extreme-ultraviolet machines, so its tightest layers need several passes^[54]. Yield is the open question.
• Arizona's lag. How far Arizona's newest process stays behind Taiwan's as TSMC's four new fabs there, at 2 nm or more advanced, start production^[60].