In January 2026, something strange happened. The US government loosened its semiconductor export controls on advanced AI chips, allowing case-by-case review of Nvidia H200 shipments to China.[s] Days later, Chinese customs officials instructed border agents not to permit those same H200 chips to enter the country.[s]
The paradox reveals a deeper truth: both Washington and Beijing now treat chip access as an industrial policy weapon, not a trade matter. The result is an accelerating split of global technology into two incompatible systems, each with its own supply chains, standards, and rules.
Semiconductor Export Controls: The Policy Shift
The Bureau of Industry and Security moved from a blanket denial policy to case-by-case review for certain AI chips exported directly to China.[s] To qualify, chips must stay below strict performance thresholds: a Total Processing Performance under 21,000 and memory bandwidth under 6,500 GB/s.[s] Nvidia’s H200 fits; its next-generation Blackwell does not.
One day before this rule took effect, the White House added a 25% tariff on these same chips under Section 232 of the Trade Expansion Act.[s] The combination, looser licensing paired with a new revenue mechanism, reflects an approach that treats semiconductor export controls as both a security tool and a trade lever.
But China is playing its own game. Beijing told domestic tech firms to halt orders for Nvidia H200 devices, using informal guidance to push companies toward domestic alternatives.[s] The controls originally imposed in October 2022 were meant to address narrow national security concerns about AI advances; they have expanded well beyond that scope.[s]
Congress Pushes Back: The MATCH Act
While the executive branch loosened chip controls, Congress moved in the opposite direction with the MATCH Act, the Multilateral Alignment of Technology Controls on Hardware. The bill would impose a country-wide ban on exports of DUV lithography equipment to China, regardless of the specific buyer.[s]
The legislation names five Chinese firms, SMIC, CXMT, YMTC, Hua Hong, and Huawei, as “covered facilities.” All their subsidiaries and affiliates would face presumption-of-denial licensing on equipment sales. Under the bill’s revised April 2026 draft, servicing, spare parts, and technical support for machines already installed would still require export licenses, but those applications would no longer face an automatic policy of denial.[s][s]
The bill gives US allies 150 days to adopt equivalent semiconductor export controls. If the Netherlands and Japan, home to ASML and Tokyo Electron, do not align their rules with Washington’s, the US would expand the Foreign Direct Product Rule to cover any foreign equipment containing American technology.[s]
ASML, a leading Dutch lithography supplier targeted by the proposed controls, drew around 27% of its 2025 revenue from China.[s] The servicing restrictions threaten maintenance contracts for machines already running in Chinese factories.[s]
China’s Counter-Strategy
Beijing’s response has been systematic. On April 7, 2026, the State Council published Order No. 834, creating a unified supply chain security framework monitored by more than 15 agencies, with authority to take legal action against companies deemed harmful to China’s supply chains.[s]
China’s semiconductor self-sufficiency rate has climbed from approximately 33% in 2024 to roughly 50% in 2025. Industry reporting describes a 2030 target of 80% self-sufficiency, with priorities including a fully domestic 7nm equipment line.[s]
In December 2025, China introduced its own “50 percent rule,” requiring fabs to source at least half their equipment domestically by the end of the decade.[s] This indigenization mandate threatens an estimated $18 billion in annual US equipment sales.[s]
SMIC, China’s largest foundry, is producing 7nm chips for Huawei’s Kirin processors and is finalizing a 5nm-equivalent process. Without access to EUV machines, this production relies on older DUV lithography with complex multi-patterning techniques.[s] The workaround costs 40-50% more than TSMC’s EUV-based process, with yields reportedly in the 60-70% range.[s]
The Equipment Gap
Despite progress in chip design and AI applications, China faces a decade-wide gap in manufacturing equipment. The most advanced domestic lithography tool being tested by SMIC is a 28nm-class DUV machine from Yuliangsheng, comparable to an ASML system from 2008.[s]
This gap explains why semiconductor export controls focused on equipment are more effective than controls on finished chips. A lithography machine costs around $200 million and requires years of servicing by the manufacturer. A chip is a commodity that can be rerouted through export control loopholes and intermediary countries.
Congress has also moved to close the “cloud loophole.” The Remote Access Security Act, which passed the House 369-22 in January 2026, targets foreign companies that rent time on advanced GPUs in third-country data centers, effectively importing controlled compute capacity without a single chip crossing a border.[s]
Two Tech Stacks
The pattern echoes Cold War-era technology bifurcation, when CoCom export controls denied the Soviet bloc cutting-edge technology.[s] By 2026, the world is on track to see a bifurcated AI hardware ecosystem split along US-China lines.[s]
Both sides are ring-fencing their tech ecosystems with government support. One stack centers on the US and partners like Taiwan, South Korea, Japan, and Europe. The other centers on China.[s]
The strategic incoherence problem compounds the bifurcation. The executive branch withdrew a February 2026 draft rule that would have further restricted global AI chip sales, while Congress simultaneously pushes the MATCH Act to force multilateral alignment.[s] Each policy reversal reduces US credibility as a reliable supplier and accelerates the self-sufficiency dynamic that semiconductor export controls were designed to prevent.
What This Means
The memory chip supply chain illustrates the stakes. Data centers will account for 70% of all memory chip demand in 2026. Memory prices surged approximately 90% in Q1 2026.[s] China’s two dominant memory producers, CXMT for DRAM and YMTC for NAND, rely heavily on ASML equipment. An abrupt cutoff would amplify price pressure across the AI buildout at exactly the moment both the US and Europe are racing to expand compute capacity.
China’s Ministry of Commerce warned that US legislation would severely disrupt the international economic and trade order.[s] The disruption will not be confined to China. When both sides of a supply chain weaponize restriction, the chain does not hold. It fragments, and everyone pays the cost of building separate systems.
The talent dimension matters as much as the hardware. China graduates more STEM PhDs annually than the US, and improving domestic research conditions are beginning to reverse the historical emigration pattern that benefited American institutions.[s]
Semiconductor export controls buy time. Whether that time is used to build durable advantage or simply to delay an inevitable convergence depends on choices being made now, in Washington, Beijing, and the allied capitals caught between them.
In January 2026, the Bureau of Industry and Security published a final rule (Federal Register Document 2026-00789) shifting export license review for certain advanced computing semiconductors from presumption of denial to case-by-case basis for direct US-to-China shipments.[s] The same week, Chinese customs issued internal guidance blocking H200 imports despite valid BIS licenses.[s]
Both governments are now using semiconductor export controls as industrial policy instruments, accelerating bifurcation of the global AI hardware stack into two parallel, increasingly incompatible systems.
Semiconductor Export Controls: Technical Thresholds
Eligibility for case-by-case review requires chips to meet dual thresholds: Total Processing Performance below 21,000 and Total DRAM Bandwidth below 6,500 GB/s.[s] Nvidia’s H200 scores approximately 15,832 TPP with 4.8 TB/s bandwidth, fitting comfortably. AMD’s MI325X sits at roughly 20,800 TPP with minimal headroom. Blackwell-class devices exceed both thresholds and remain under presumption of denial.
Four conditions govern approval: US supply priority with a 50% volume cap on China shipments relative to domestic sales; documented purchaser compliance programs with end-use monitoring; independent US-based third-party testing before each shipment; and Know Your Customer protocols with remote access safeguards.[s]
The White House announced a 25% Section 232 tariff on qualifying chips one day before the export rule took effect.[s] Reexports from third countries remain under presumption of denial; the relaxation applies exclusively to direct US shipments. The relevant ECCNs include 3A090 for integrated circuits with TPP thresholds, 4A090 for systems containing controlled ICs, and associated 3D001/3E001 categories for software and technology.
Beijing’s counter-move undermined the commercial impact. Chinese authorities instructed domestic tech firms to halt H200 orders, using informal guidance to push adoption of domestic alternatives from Huawei, Cambricon, and Hygon.[s] The original October 2022 controls targeted narrow national security concerns; subsequent modifications have expanded into revenue mechanisms and industrial policy levers.[s]
MATCH Act: Equipment-Layer Controls
The Multilateral Alignment of Technology Controls on Hardware Act operates through two mechanisms. First, a country-wide prohibition on DUV immersion lithography exports to any Chinese destination, regardless of end user.[s] This covers ASML’s NXT:2000i-class scanners and Nikon’s NSR-S631E systems used in SMIC’s 7nm production lines.
Second, statutory designation of SMIC, CXMT, YMTC, Hua Hong, and Huawei as “Covered Facilities,” subjecting all subsidiaries and affiliates to presumption-of-denial licensing on equipment sales, with servicing, spare parts, and technical support for already-installed equipment subject to licensing requirements (the automatic presumption of denial on servicing was softened in the April 2026 revision).[s][s] Unlike current BIS Entity List restrictions requiring case-by-case subsidiary evaluation, the statutory approach would automatically capture any future spin-off or joint venture.
Commerce has 150 days post-enactment to negotiate equivalent country-wide semiconductor export controls with the Netherlands and Japan. If negotiations fail, the bill directs expansion of the Foreign Direct Product Rule to cover any foreign-manufactured tool incorporating US-origin software, technology, or components, a near-zero de minimis threshold given that virtually every advanced chipmaking tool relies on some American intellectual property in EDA software, metrology subsystems, or process control algorithms.[s]
ASML derived 27% of 2025 revenue from China versus approximately 16% from the US.[s] The servicing restrictions threaten the company’s long-term maintenance contracts.[s] The diplomatic calculus is further complicated by EU member-state control authority; the Netherlands cannot unilaterally impose extraterritorial controls, and codifying restrictions through US legislative mandate differs fundamentally from negotiated diplomatic alignment.
China’s Indigenization Architecture
State Council Order No. 834, published April 7, 2026, creates a unified supply chain security framework monitored by MOFCOM, MIIT, and over 13 additional agencies, with authority to pursue legal action against companies deemed harmful to China’s supply chains.[s]
Semiconductor self-sufficiency has progressed from approximately 33% in 2024 to roughly 50% in 2025. Industry reporting describes a 2030 target of 80%, with priorities including a fully domestic 7nm equipment line and stable 14nm production capacity.[s]
China’s December 2025 “50 percent rule” requires fabs to source at least half their equipment domestically by decade’s end.[s] Foreign fabs operating in China, including SK Hynix and Samsung facilities, received temporary one-year waivers, underscoring continued foreign-equipment dependency.
SMIC’s 7nm process, used for Huawei’s Kirin 9000S, relies on Self-Aligned Quadruple Patterning on older DUV lithography since EUV machines are blocked. The 5nm-equivalent process in development carries a 40-50% cost premium over TSMC’s EUV-based production, with 7nm yields reportedly in the 60-70% range, far below industry-standard yields.[s] This “sanction tax” is tolerable for national security applications but commercially unviable for volume consumer products.
Equipment Gap Quantified
The most advanced domestic lithography tool under SMIC testing is Yuliangsheng’s 28nm-class DUV scanner, a subsidiary of SiCarrier. This system is comparable to ASML’s Twinscan NXT:1950i from 2008, several generations behind the NXT:2000i used for 7nm/5nm production.[s] ASML’s CEO has publicly cited a 10-15 year technology gap, and current domestic tooling substantiates that assessment.
Equipment-layer semiconductor export controls are more effective than finished-chip controls because manufacturing tools create capability dependencies. Lithography machines require ongoing servicing, software updates, spare parts, and process tuning. Once installed, equipment participates in production for many years; the MATCH Act’s servicing restrictions would degrade existing capacity over time as machines require regular maintenance to sustain yield.
The Remote Access Security Act (passed House 369-22 in January 2026) closes the “cloud loophole” where foreign entities access controlled GPU compute by renting time in third-country data centers, effectively importing capacity without physical chip movement.[s] One cited case involved a Shanghai-based startup accessing Blackwell systems through an Indonesian data center arrangement.
These export control loopholes explain the legislative push to move from entity-specific restrictions to country-wide equipment bans. A bipartisan February 2026 letter to Commerce and State argued that entity-specific controls could not substitute for countrywide restrictions on critical chokepoint tools.[s]
Bifurcation Dynamics
The pattern parallels Cold War-era CoCom controls that denied the Soviet bloc cutting-edge technology, but on a far larger economic scale.[s] By 2026, the AI hardware ecosystem is splitting along US-China lines, with each side building distinct tech stacks: US-centered (Taiwan, South Korea, Japan, Europe) versus China-centered.[s]
Strategic incoherence compounds the problem. The executive branch withdrew a February 2026 draft rule that would have extended multilateral chip restrictions, while Congress pushes legislation requiring exactly that multilateral alignment.[s] Each policy reversal reduces US credibility as a stable supplier, accelerating the self-sufficiency dynamic that semiconductor export controls were designed to delay.
Memory supply chain stress illustrates near-term risks. Data centers will account for 70% of 2026 memory demand; Q1 2026 prices surged approximately 90%.[s] CXMT (DRAM) and YMTC (NAND) depend on ASML equipment. An abrupt cutoff would amplify price pressure during a historic AI-driven supercycle, undermining both US AI buildout and European data center expansion plans.
MOFCOM’s warning that US legislation would severely disrupt the international economic and trade order reflects this mutual vulnerability.[s] The fragmentation cost falls on everyone. When both sides weaponize supply chains, the chain does not hold; it bifurcates, and participants must fund parallel systems.
The talent dimension is equally consequential. China produces more STEM PhDs annually than the US, and improving domestic research conditions are reversing historical emigration patterns.[s] The competition for skilled AI researchers may prove as decisive as the hardware contest.
Semiconductor export controls buy time. Whether that time enables durable structural advantages or delays an inevitable technology convergence depends on sustained investment, multilateral coordination, and policy coherence, all of which remain uncertain in the current triangular standoff between the White House, Congress, and allied governments.
