April 16, 2026 | Secure Line Readout
Strengthening Export Controls: A Critical National Security Priority for Congress
April 16, 2026 Secure Line Readout
Strengthening Export Controls: A Critical National Security Priority for Congress
On FDD Action’s latest Secure Line briefing call, experts Craig Singleton and Ryan Fedasiuk assessed why U.S. export controls on advanced AI chips and chipmaking equipment are one of the most consequential national security tools available to Congress. The briefers argued that in the AI race, compute is power — and the United States and its allies still control the choke points that determine who can build and scale frontier AI. Mounting evidence of large-scale chip smuggling, including a recent DOJ indictment involving more than $2.5 billion in diverted Nvidia servers, shows that Chinese actors are going to extraordinary lengths to evade current controls — and reveals where enforcement, allied alignment, and new statutory authorities must be strengthened. The House Foreign Affairs Committee is poised to mark up a slate of export control bills on April 22, including the MATCH Act — a measure targeting deep ultraviolet (DUV) lithography machines already in China that, if enforced, could render Huawei’s DUV lithography fleet obsolete within a few years. With President Trump’s tentatively planned May trip to Beijing approaching, Singleton and Fedasiuk warned that once export controls become transactional at the negotiating table, a long-term strategic advantage risks becoming a short-term diplomatic giveaway.
- Compute is the strategic asset: Advanced AI runs on compute, and the most capable compute still depends on American and allied choke points — the chips themselves, the tools that make them, high-bandwidth memory, advanced packaging, and the cloud and model ecosystems around them. Export controls are one of the few non-kinetic tools available to deny adversaries access to this stack.
- Indigenization predates controls: Arguments that U.S. export controls “caused” China’s push for tech self-reliance get the chronology backwards. Xi’s drive for technological independence predates nearly all U.S. controls. What controls do is force China onto weaker substitutes, raise costs, stretch timelines, and make frontier-scale clusters harder to assemble and refresh.
- Evasion proves controls matter: DOJ indictments and reporting on DeepSeek’s use of smuggled Nvidia Blackwell chips have revealed large-scale smuggling networks moving advanced U.S. chips into China. Leakage does not prove the policy failed — if Beijing could buy openly, it would not need black markets. The remaining gaps are in enforcement capacity, allied alignment, and new evasion methods like model theft and distillation.
- Lithography is the most controllable choke point: China needs three components to scale AI compute — high-bandwidth memory, advanced packaging, and logic chips. Memory and packaging are difficult to control. Logic-chip fabrication via DUV lithography machines, $100 million pieces of equipment that require constant servicing from Dutch manufacturer ASML, is where Congress has real leverage and where the MATCH Act would have the most impact.
- Codify controls in statute: Controls imposed only through executive action are vulnerable to waiver, reinterpretation, and political bargaining across administrations. Statute sets the floor of America’s negotiating position, makes policy durable across administrations, prevents Beijing from treating every summit as a chance to unwind rules, and gives industry the predictability it needs to invest.
- The “revenue for R&D” argument does not hold up: The claim that China revenue is needed to fund next-generation U.S. chip R&D misreads the 2026 compute market. Compute is a sellers’ market — U.S. labs, allies, and customers around the world cannot buy enough. Selling China’s AI labs the exact compute they need to train better models and serve them to the same global markets U.S. labs are trying to reach is, in the briefers’ words, a lose-lose.
The Strategic Picture: Four AI Races, One Compute Moat
Fedasiuk framed the U.S.-China AI competition as four distinct races, each with a different scoreboard. On frontier capability — who can build the best model — the United States is clearly ahead, with Anthropic’s recently unveiled Mythos model (estimated at five to 10 trillion parameters) blowing through leading benchmarks like SWE-bench and demonstrating novel capabilities, such as Project Glasswing’s cyber-vulnerability hunting work, that Chinese models cannot yet match. On adoption — diffusing AI across enterprises and militaries — the picture is murkier, and China’s 2024 “AI Plus” economy initiative under Premier Li Qiang has produced a wave of programs driving domestic uptake, compute credits, and agentic frameworks. On global diffusion, China is positioned to catch up fast: Chinese models’ share of global token usage jumped from roughly 1 percent in 2025 to about 30 percent in 2026. On resilience, the United States retains enduring advantages in adapting to disruption and black-swan events — a lesson the briefers drew from the divergent U.S. and Chinese responses to COVID-19.
The connecting thread across all four races is compute, and the gap between the United States and China remains substantial:
- Chip shipments (2025): approximately 800,000 Huawei Ascend 910C units versus approximately five million Nvidia Blackwell units.
- Fabrication process: China’s best at roughly 7nm via inefficient multi-patterning versus leading-edge 4nm.
- Huawei multi-patterning yield rates: sometimes as low as 20 percent relative to Taiwan.
- Mythos model compute footprint: inference requires an estimated 25 to 50 Nvidia Blackwell B200 chips per instance at roughly $35,000 each, with the overall project occupying tens of thousands of chips and potentially as many as 100,000.
- Chinese models’ share of global AI token usage: approximately 1 percent in 2025, approximately 30 percent in 2026.
Three components determine whether China closes that gap: high-bandwidth memory (HBM), advanced packaging, and logic chips. Memory and packaging are genuinely difficult physical problems where China is making progress but where the United States has limited direct leverage. Logic-chip fabrication is where Congress can still act decisively.
Where Controls Fall Short — and How Congress Can Fix Them
The briefers identified three gaps in the current export control regime: enforcement, allied alignment, and new forms of circumvention. On enforcement, mounting evidence shows Chinese actors using smuggling networks, transshipment routes, and shell companies to divert advanced U.S. chips. Existing statutory tools are under-resourced: the Bureau of Industry and Security (BIS) lacks sufficient technical capacity, insiders have too little incentive to report violations, civil penalties often amount to the cost of doing business, and statutes of limitations are too short to unwind sophisticated diversion networks. On allied alignment, controlling U.S. firms tightly while allowing partner suppliers to backfill imposes costs on American companies without fully imposing costs on China — denial has to be multilateral enough to bite. On circumvention, the threat is evolving beyond the physical chip to include model extraction, distillation, and other ways of free-riding on the capital and compute sunk into U.S. frontier systems.
The slate of bills moving through the House Foreign Affairs Committee addresses each gap. Bills like the MATCH Act (H.R. 8170), the STRIDE Act (H.R. 6058), the BIS License Improvement Act (H.R. 8284), and the Interagency Dispute Resolution Act (H.R. 7962) tighten the denial pipe and reduce the chances that disputed cases or ally loopholes quietly become de facto approvals. The BIS STRENGTH Act (H.R. 7003), the Stop Stealing Our Chips Act (H.R. 6322), the ECRA Penalty Increase Act (H.R. 5853), and the Statute of Limitations Extension Act (H.R. 8202) directly target enforcement weaknesses. The Deterring American AI Model Theft Act (H.R. 8283) recognizes that distillation and model extraction function as export-control evasion by other means. And on the positive ledger, the AI Full Stack Export Promotion Act (H.R. 6996) would make the American AI stack more available, secure, and adoptable for allies — so partners have a credible alternative to the Chinese stack.
The MATCH Act: Hitting the Most Controllable Choke Point
The MATCH Act (H.R. 8170), up for HFAC markup on April 22, is one of the most strategically significant bills because it targets the single most controllable choke point in China’s AI supply chain: the DUV lithography machines that print logic chips onto silicon wafers. From 2023 to 2024, Huawei and other Chinese chipmakers imported hundreds of older-generation DUV machines from the Netherlands — equipment that fell below the threshold covered by U.S. and Dutch controls on more advanced extreme ultraviolet (EUV) lithography. Chinese engineers have since upgraded those machines, running silicon through them multiple times in a process called multi-patterning to print more sophisticated logic chips than was previously thought possible. The process is inefficient — Huawei’s yield rates can fall as low as 20 percent relative to TSMC — but it is still capable of producing the millions of logic chips China needs to contest American AI dominance from 2027 onward.
These machines cost roughly $100 million each, ship in dozens of freight containers, and require constant servicing by engineers from ASML, the Dutch company that manufactures them. The MATCH Act would make it illegal to service DUV lithography machines already in China and would tighten export rules on the highest-end DUV machines that currently fall below controlled thresholds. Because components inside these machines are made with American technology by American companies, the bill treats the finished product as a U.S.-origin technology subject to U.S. export jurisdiction. If enacted and enforced, the measure could render Huawei’s existing fleet of DUV machines obsolete within a few years as they fall out of service — a hamstringing of Chinese logic-chip fabrication that no other single policy tool can match.
Deep Dives
Fedasiuk walked through the three supply-chain components China needs to unlock in order to achieve the chip independence required to compete in frontier AI. The first is high-bandwidth memory (HBM). China has a memory champion, CXMT, which is working to produce a high-end form of HBM3 by the end of 2026. This is a very hard physical project and one to monitor, but there is no easily controllable lever for Congress here.
The second is advanced packaging — the process of stacking memory and logic chips onto a silicon wafer to produce a finished accelerator. Very few companies in the world are capable of this; SMIC, which contracts with Huawei, is making progress. Like memory, packaging is more something for Congress to be aware of than something it can directly control.
The third, and the one where Congress retains real leverage, is logic-chip fabrication. Modern AI logic chips are printed onto silicon wafers using lithography machines made almost exclusively by one company in the world: ASML in the Netherlands. This is the choke point the MATCH Act targets — and the one where U.S. policy can most directly constrain China’s ability to build large-scale compute.
The briefers compared the 2026 AI market to the auto industry: most of the world wants a Toyota Corolla, not a Ferrari. For most white-collar workloads — perhaps 70 to 80 percent of a desk worker’s job — a 14-to-80-billion-parameter model running on local hardware is good enough. Chinese labs have focused on architectural innovations in these small-to-mid-sized open-weight models, releasing them free of charge and undercutting U.S. frontier labs at the low end of the value chain.
Chinese labs can pursue this strategy because they are not trying to make money in the short term. Each has a different theory of commercialization — ByteDance integrates AI with TikTok as a data play; Alibaba’s Qwen group rents out compute; others are bankrolled by state subsidies. The strategic implication is that China’s open-source push is a force multiplier that narrows the technical gap, spreads Chinese models and standards globally, and reduces China’s compute burden by letting a wider ecosystem optimize and deploy. The U.S. response cannot be limited to blocking chips; it has to include protecting compute, preventing model theft, accelerating allied adoption of the American stack, and ensuring Chinese open models do not become the default infrastructure layer in the developing world.
Beyond physical chip smuggling, a new front in export-control evasion is emerging: model theft and distillation. Reporting on DeepSeek and other Chinese labs has suggested that distillation techniques are being used to accelerate catch-up, effectively free-riding on the capital and compute sunk into U.S. frontier systems. Singleton argued Congress should treat this as an end run around export controls rather than a separate issue. Legislation aimed at model theft recognizes that the strategic purpose of chip controls — denying China the ability to build frontier capability on the back of U.S. investment — can be undone if model weights and capabilities can be extracted or copied. Effective U.S. policy needs to address both the hardware and the intellectual inputs.
On March 19, the Department of Justice unsealed an indictment detailing an operation to smuggle more than $2.5 billion worth of AI servers containing Nvidia chips to Chinese customers in violation of U.S. export controls. Separately, Reuters reporting on February 23 indicated that DeepSeek trained its latest model on Nvidia Blackwell chips — Nvidia’s most advanced AI chip — that were smuggled into China despite controls barring their export. These cases underscore two points the briefers emphasized: first, that controls are working, because China would not need smuggling networks if it could buy openly; and second, that enforcement authorities, penalties, and timelines have not kept pace with the sophistication of the diversion networks being built to evade them.
Opponents of export controls argue that revenue from selling chips to China is needed to fund the next generation of U.S. research and development, and that controls will hinder the diffusion of the American AI stack. The briefers rejected both premises. Properly designed controls do not choke off global diffusion; they steer it — redirecting access away from high-risk users while pushing adoption across allies and friends. And the revenue argument misreads the 2026 compute market, which is a sellers’ market. U.S. labs, enterprises, and allied customers cannot buy enough memory or chips as it is; the constraint on chip revenue is supply, not demand. Selling China’s AI labs the precise compute they need to train better models — models they then offer for free to the same global publics that U.S. labs are trying to reach — is, as Fedasiuk put it, a lose-lose. The companion argument that the U.S. can get China “addicted” to American chips fails against the clear evidence that Xi’s indigenization strategy predates controls and is a permanent fixture of Chinese industrial policy.