David Feith testified before the US Senate Committee on Banking, Housing, and Urban Affairs on artificial intelligence and the American dream.
Written Testimony
Chairman Scott, Ranking Member Warren, and members of the Committee: Thank you for the opportunity to testify today.
I will focus my remarks on AI competition with China and the importance of export controls to preserving America’s technological, economic, and national-security advantages.
1. A Historic Opportunity
We are meeting amid a historic opportunity playing out on three levels.
First, even in the broadest terms of human history and American civilization, the implications of the AI industrial revolution are enormous. AI will be the driving force behind economic growth, innovation, and military affairs. It is hard to overstate the need to get AI policy right, especially in terms of strategic competition with China.
Second, we are at a hinge point in the Trump administration’s AI policy, which rightly recognizes that America must win the AI race with China. Anthropic’s powerful Mythos model has underscored both the risks and opportunities of frontier AI capabilities in areas like cyber. At the same time, recent news of chip smuggling and export-control loopholes suggest that vast quantities of the most advanced U.S. AI technology have reached China, potentially undermining years of bipartisan policy. These episodes show why we need clearer strategy, stronger rules, tighter enforcement, and robust oversight.
Third, this Committee now has a rare legislative opportunity: Members here and across Congress have introduced a historic package of bills to strengthen America’s export-control system, bolster national security, and assert Congress’s role in this key area of policy. The work on this legislation has been notably bipartisan, and I look forward to discussing these bills as you consider whether to mark them up and pass them into law.
2. The Strategic Stakes: The Age of Personalized Power
What is at stake geopolitically in the contest over AI? One essential feature is the personalization of technology – powerful, precise, transformative, potentially very positive, but also potentially very threatening. AI offers personalized medicine. But it also offers personalized warfare. And it offers personalized state power.
In medicine, AI can analyze massive datasets and detect patterns beyond unaided human cognition – as with Google DeepMind’s protein-folding breakthrough, which modeled the structure of virtually every known protein. AI can also enable individualized therapies tailored to each specific patient. The promise is extraordinary.
But there is also personalized warfare. During the 12-day war in Iran last year, Israel’s first-night strikes against dozens of Iranian military and nuclear leaders showed the emergence of a new era of intelligence-fused precision warfare. Never before was such individualized targeting possible. Ukraine’s Operation Spiderweb offered another glimpse: autonomous drones launched deep inside Russia to strike strategic aircraft on the ground. This is new in world affairs. And nations that master both its offensive and defensive dimensions will hold enormous advantages.[1]
Then there is personalized state power. AI dramatically expands the capacity of governments to track, influence, or coerce specific individuals: voice data, biometric scans, geolocation, purchases, social networks, medical records, behavioral patterns. The tyrants of the 20th century could only dream of the total personal reach that modern AI systems have made possible.
The question is how this power will be used. Will AI-enhanced governments operate within constitutional systems, subject to checks and balances, elections, public debate, and civil liberties? Or will authoritarian systems, especially Beijing’s, integrate AI into governance in ways that eliminate restraints – and then export those systems abroad?
The AI contest is about more than economic or military advantage. It is about whether this transformative technology will serve systems that restrain power and protect human freedom – or systems that concentrate power and stifle freedom.
3. The Good News: America’s AI Compute Advantage
The U.S.-China competition in AI will be decided by factors including talent, capital, energy, data, and computing power (or “compute”). China has real strengths in many of these areas. It has world-class engineers, enormous energy resources, sophisticated AI companies, and a government willing and able to mobilize resources behind strategic technologies.
But the decisive factor is likely to be compute – and there the United States and our allies enjoy a substantial advantage.
Today, China’s best AI chips are much less capable than America’s – and, thanks in significant part to U.S. and allied export controls, they are falling further behind. This matters because frontier AI increasingly rewards scale. The ability to train larger models, run more experiments, and deploy more powerful systems depends on access to vast quantities of advanced compute.
Publicly available data and roadmaps show that Nvidia’s leading AI chips are five to eight times more powerful than the best offerings from China’s Huawei, and this gap could widen to 17 times by 2027.[2] Huawei’s latest chip, the Ascend 950PR, which has been on the market for three months, still lags Nvidia’s H100 from 2022.[3]
China also struggles to produce advanced chips in large quantities. According to estimates from the Institute for Progress, Huawei is likely to produce only about 1%-2% as much advanced AI computing capacity as the United States in 2026. IFP estimates U.S. production in 2026 at roughly 6.89 million B300-equivalent chips, compared with only 62,000 to 160,000 B300-equivalents from Huawei.[4]
Nor are these merely American assessments. Chinese AI companies themselves routinely acknowledge compute shortages. Leaders at Alibaba, Tencent, DeepSeek, and other firms have publicly described limited access to advanced chips as the major constraint on their ambitions to dominate AI.[5]
The conclusion is simple: Compute is the lifeblood of the AI economy, and China has less of it, at lower quality, than the United States and its allies. That makes compute China’s principal chokepoint in the AI race. But this advantage is not self-executing. It will last only if U.S. and allied policy preserves it. That is what export controls are for.
4. The Less-Good News: Policy and Enforcement Drift
Unfortunately, U.S. export-control policy of late has not been equal to that task.
The Trump administration’s topline policy remains that America’s most advanced chips, including Blackwells, should not go to China – a point President Trump made clear in November. The White House’s July 2025 AI Action Plan was similarly direct: The United States should deny advanced AI compute to foreign adversaries, strengthen export controls on semiconductor manufacturing equipment, including key subsystems, and press allies to align with U.S. controls rather than backfill them.
These policies haven’t yet been realized, however.
a. Advanced chip policy has become less clear.
The administration initially moved in the right direction on chip sales to China. In April 2025, it blocked exports of Nvidia’s H20 – a China-specific chip whose enhanced memory bandwidth makes it useful for large-scale AI deployment. But in July, the administration reversed course and allowed H20 exports, then in December also gave a green light for H200 exports. H200s are not Blackwells, but they are still highly capable chips, far beyond anything China can produce at scale.
The Commerce Department later applied relatively restrictive licensing rules to H200 exports, and China hadn’t received any as of April.[6] But that may reflect Chinese government resistance as much as U.S. discipline: Some in Beijing may want to force Chinese labs to rely on Huawei, while others may be holding out for access to Blackwells or already obtaining better chips through smuggling, remote access, and overseas purchases (more on all that later).
Either way, U.S. policy has become more open to exporting advanced compute to China, even while insisting that the very leading edge remains off limits.
b. Entity Listings have not kept pace.
Entity List designations are one of the government’s basic tools for stopping harmful technology transfers to Chinese companies, brokers, shell entities, fabs, and procurement networks. But this tool is not keeping pace with the threat. New Entity List additions in 2026 are on track to be the lowest in a decade, even as evidence mounts of Chinese smuggling, diversion, and third-country procurement networks. Commerce has taken some limited steps – including reportedly using quiet “is-informed” letters to restrict transactions involving a specific Chinese chip fab – but that is no substitute for a broader, public, sustained Entity List strategy.
There are obvious candidates for action. These include SMEE, China’s leading domestic lithography-tool maker; its publicly known cutouts; and dozens of fabs reportedly producing restricted chips for Huawei.[7]
The Commerce Department did briefly try to strengthen the Entity List last year through an Affiliates Rule, which would have extended restrictions automatically to foreign entities at least 50 percent owned by a listed party. That made sense, as a restricted Chinese company should not be able to evade U.S. controls simply by buying through an unlisted subsidiary. But the administration shelved the rule after Beijing threatened retaliation, reopening a basic evasion pathway and allowing restricted Chinese firms to keep routing purchases through affiliates that have not yet been individually listed.
c. Controls on semiconductor manufacturing equipment remain loose.
Semiconductor manufacturing equipment, or SME, is the machinery that makes chips. It may be the most important export-control issue of all. Finished AI chips give China compute today, but SME gives China the ability to make advanced chips tomorrow, build its own industrial base, reduce dependence on the United States and our allies, and compete for global market share in chipmaking itself. If the goal is to keep China from catching up in frontier AI compute, cutting off access to the machines that make the chips is essential.
The administration has said some good things about this problem, and has undertaken important negotiations with allies, but these have not yet yielded results. China continues to buy enormous volumes of chipmaking equipment not only from U.S. firms through regulatory loopholes (e.g. exporting from offshore) but from allied suppliers in Japan and the Netherlands, undercutting U.S. firms. That is the core gap: U.S. controls cannot fully work if Dutch and Japanese suppliers keep selling China the tools, components, servicing, and support it needs to expand its manufacturing base.
d. Remote access remains unaddressed.
Another gap in export-control policy is remote access. Export controls traditionally focus on whether a physical item crosses a border (e.g., a chip, a server, a semiconductor tool). But AI compute can be accessed without moving chips at all, as Chinese companies rent access through the cloud to chips physically located in data centers in Southeast Asia, the Middle East, or even Ashburn, Virginia. The chip stays outside China, but the compute and most of the strategic benefit flows to China.
This loophole has been acknowledged for years but never addressed. The Biden administration promulgated a remote-access restriction in its final days, but the Trump administration unwound it and hasn’t yet offered a replacement.
e. Massive smuggling, minimal response.
Independent analysis estimates between 290,000 and 1.6 million H100-equivalents were smuggled to China through 2025, for a median of roughly 660,000, or about a third of China's total compute.[8] The federal indictment in March of a co-founder and board director of server giant Super Micro (one of Nvidia’s largest customers) alleges a $2.5 billion scheme to move servers with restricted Nvidia chips to China via Malaysia, in what would be the largest export-control case in U.S. history. Reuters reported that Super Micro’s buyers included two Chinese universities close to the People’s Liberation Army and previously placed on the export-control Entity List.[9]
Southeast Asia is an apparent hotbed for chip smuggling. Commerce has taken some actions but evidently pursued no new Entity Listings or known major investigations, even after the Super Micro indictment and other detailed, credible reports.[10]
f. The Blackwell loophole – and TSMC too?
Another issue involves BIS’s unusual statement of Sunday, May 31. The statement said that companies need a license to sell advanced AI chips, such as Blackwells, to Chinese companies in third countries such as Malaysia or Singapore. But why did BIS need to say this at all, when President Trump has repeatedly said no Blackwells for Chinese buyers?
The answer is alarming: Chinese companies have apparently been accessing advanced U.S. chips for months through a loophole in U.S. rules created last year.
Beginning in 2022, U.S. controls restricted advanced AI chip exports to Chinese firms, whether in China or abroad. In January 2025, the Biden administration folded those restrictions into the broader AI Diffusion Rule. Then, in May 2025, the Trump administration stopped enforcing the Diffusion Rule (for some good reasons) but didn’t formally replace it. That appears to have created a gap: Chinese firms in China still couldn’t buy Blackwells, but their subsidiaries in Malaysia, Singapore, or elsewhere could buy servers full of them.
The administration’s Sunday guidance has now clarified that this can’t go on. But there may be another critical loophole still unaddressed, this one allowing Chinese firms to order the manufacture of advanced chips by leading-edge allied fabs such as TSMC and Samsung. If so, this too would be in urgent need of a fix.
g. Distillation attacks countered only with name and shame?
Distillation attacks are the mass IP theft of the AI age, and Chinese labs are using them at scale. A distillation attack involves systematically querying a powerful AI model, capturing its outputs, and using those outputs to train or improve a competing model that imitates the original model’s capabilities without permission or compensation.
The administration has recognized the problem, including in a publicly released White House memo in April, but hasn’t taken action to counter it. This comes after some two decades of failing to counter China’s mass cyber-enabled theft of intellectual property, which the U.S. government labeled “the largest transfer of wealth in human history” as early as 2011. Without an effective counter, distillation risks becoming the turbocharged AI-age version of that failure: billions of dollars in American model investment copied, repackaged, commercialized, and exported by Chinese labs.
5. For Argument’s Sake
Before turning to recommendations, it is worth addressing several recurring arguments about AI export controls. These claims come up often, but they should not distract from the central point: America has a real compute advantage, China knows it, and better export controls can help preserve it.
a. Have export controls worked to date?
Yes, especially where they have targeted real chokepoints. The clearest example is EUV lithography, where U.S. coordination with the Netherlands since 2019 has helped keep China from reaching the frontier of advanced chip manufacturing at scale. The same pattern is visible in AI chips: Huawei is not catching Nvidia; it is falling further behind. China’s AI leaders complain openly about compute shortages, China’s government complains about U.S. controls, and Chinese firms and brokers are making major efforts to smuggle controlled chips. Those are not signs of a policy that has failed. They are signs of a policy that is biting.
b. Do export controls only drive successful Chinese indigenization?
No. China has successfully indigenized or dominated many strategic sectors that were never subject to comparable U.S. export controls: high-speed rail, solar, batteries, electric vehicles, shipbuilding, drones, telecom. Beijing’s drive for self-reliance long predates recent semiconductor controls and is rooted in CCP strategy, industrial policy, military-civil fusion, and fear of foreign dependence. More importantly, China is not successfully indigenizing at the frontier in AI chips. It is struggling because controls deny time, tools, know-how, and scale. Reducing controls would not prevent Chinese indigenization – it would subsidize it.
c. Are export controls inconsistent with light-touch regulation at home?
No. A light-touch domestic approach does not mean exporting strategic technology to adversaries. The United States has never understood free enterprise to require selling the most sensitive technologies to hostile powers. We would not have sold rockets to the Soviet Union. Anthropic’s Mythos model is a warning of how powerful AI-enabled cyber capabilities may become, and similar “Mythos-like” moments in biology are likely ahead.
Export controls are not a rejection of innovation but a way to let American companies thrive at home and with trusted partners abroad, while preventing them from helping China build the capabilities that could undermine both American security and their own long-term competitiveness.
d. Should success be measured by U.S. market share in China?
America should want the U.S. AI stack to dominate global markets, but we need to be clear about which markets we mean. It is not enough to ask whether U.S. chips have high market share in China. The relevant stack includes chips, cloud services, frontier models, applications, data centers, and deployment ecosystems.
Maximizing U.S. chip sales into China could actually weaken America’s broader AI position by helping Chinese firms train better models, build stronger cloud platforms, and export Chinese AI infrastructure to third countries. The positive version of export controls is not simply restricting China but helping American cloud, model, chip, and application companies win globally, without arming their strongest strategic competitor.
These arguments point in the same direction: Export controls are not a substitute for American innovation, but they are essential to protecting it. The question for policymakers is: How can policy make them stronger, better enforced, and better matched to the full AI stack?
6. Recommendations
The central task for Congress, I would submit, is to help turn America’s compute advantage into a durable strategy: clear rules, real enforcement, and a full-stack approach that prevents chips, cloud access, semiconductor manufacturing, model security, and domestic AI infrastructure from enabling foreign adversaries to use our own powerful technology against us.
In particular:
a. Enact the historic package of bills now before the Senate.
For example:
- AI Overwatch Act: This would codify limits on exporting the most capable AI chips to foreign adversaries, require Commerce to notify and certify to Congress before approving sensitive AI-chip licenses, and accelerate trusted exports to allies and partners.
- MATCH Act: This targets semiconductor manufacturing equipment – the tools China needs to make advanced chips itself – by pushing alignment with allies such as Japan and the Netherlands and, if necessary, using U.S. authorities over foreign-produced items that rely on American technology.
- Remote Access Security Act (RASA): This would modernize export-control law so adversaries cannot evade chip controls by renting controlled compute remotely through cloud services.
- Chip Security Act (CSA): This counters AI-chip smuggling by requiring advanced chipmakers to implement technical security measures to detect and prevent diversion to unauthorized countries and end users.
- Deterring American AI Model Theft Act: This would create a framework for countering distillation attacks by identifying foreign actors that illicitly extract technical features of American closed-source AI models and sharing threat information with U.S. model owners.
- Chip EQUIP Act: This would help ensure that taxpayer-supported American fabs do not rely on semiconductor manufacturing equipment from foreign entities of concern. That is a basic supply-chain integrity measure for the domestic chip industrial base.
These bills are interlocking and complementary: Overwatch addresses chips flowing directly to China. CSA addresses smuggling and diversion. RASA addresses remote access. MATCH addresses semiconductor manufacturing equipment. Chip EQUIP protects American fabs. The distillation bill addresses model theft.
Each solves one key piece of the puzzle.
b. Strengthen oversight of Commerce and BIS.
This should include a full accounting of the apparent Blackwell loophole: what shipped, to whom, with what guidance from BIS, and what is being done now. It should also include the foundry due-diligence question, the status of the Affiliates Rule, the pace of Entity List designations, and whether BIS has the staffing, technical expertise, intelligence support, and enforcement capacity needed for this mission.
BIS is now a frontline agency in the U.S.-China technology competition. Lawmakers should treat it that way and ensure that it is fulfilling the mission that Congress and the President have set out.
c. Address BIS tools for AI policy beyond export controls.
Some key AI competition issues relate to Commerce authorities distinct from export controls. One is ICTS, for “Information and Communications Technology and Services.” This committee could codify ICTS authority and direct Commerce to use it to secure America’s domestic AI data-center buildout.
AI data centers are becoming critical infrastructure powerful enough to affect the national electric grid. Their security depends on the integrity of hardware that often comes from China, including power inverters, electrical transformers, optical transceivers, AI servers and other systems that could be compromised or remotely manipulated. We should not build the foundation of our AI future on insecure foreign-adversary technology.
In short: America should innovate aggressively at home, export confidently to allies and trusted partners, and deny China the chips, cloud access, manufacturing tools, and stolen model capabilities it needs to overtake us.
I look forward to your questions.
Endnotes
- Vladimir Putin, 2017: “Artificial intelligence is the future, not only for Russia, but for all humankind. It comes with colossal opportunities, but also threats that are difficult to predict. Whoever becomes the leader in this sphere will become the ruler of the world.” ↑
- Chris McGuire, “China’s AI Chip Deficit: Why Huawei Can’t Catch Nvidia and U.S. Export Controls Should Remain,” Council on Foreign Relations, December 15, 2025, https://www.cfr.org/articles/chinas-ai-chip-deficit-why-huawei-cant-catch-nvidia-and-us-export-controls-should-remain#chapter-forecasting-the-quality-of-us-and-chinese-ai-chips. ↑
- Hannah Petrovic and Azeem Azhar, “Inside the Chinese AI Labs Where America’s AI Controls Created Its Toughest Competition,” Exponential View, Substack, May 13, 2026, https://www.exponentialview.co/p/inside-chinese-ai-labs-efficiency-moat. ↑
- Georgia Adamson, Saif Khan, Tao Burga, and Tim Fist, “Should the U.S. Sell Blackwell Chips to China? Assessing the Impacts of Exporting the NVIDIA B30A AI Chip,” Institute for Progress, October 25, 2025, https://ifp.org/the-b30a-decision/. ↑
- “What Are Chinese AI Companies Saying About the Impacts of Export Controls?,” Konstantin F. Pilz, June 7, 2026, https://www.konstantinpilz.com/posts-publications/chinese-statements-on-compute. ↑
- Emily Wilkins, “Lutnick Gets Grilling on Nvidia Chip Sales to China in Letter from Sen. Chris Coons,” CNBC, May 1, 2026, https://www.cnbc.com/2026/05/01/nvidia-chips-jensen-huang-h200-lutnick-chris-coons.html. ↑
- Ryan Bacic, “The U.S. Restricted China's Chipmaking-Tech Champion. Its New Spinoff Is Another Story,” Kharon, March 5, 2026, https://www.kharon.com/brief/semiconductors-lithography-smee-amies-technology-entity-list-export-controls; Isabel Juniewicz, “Diversion and Resale: Estimating Compute Smuggling to China,” Epoch.ai, April 29, 2026, https://epoch.ai/publications/chip-smuggling. ↑
- Juniewicz, “Diversion and Resale.” ↑
- Eduardo Baptista, “Chinese Universities with Military Links Bought Super Micro Servers with Restricted AI Chips,” Reuters, March 27, 2026, https://www.reuters.com/world/china/chinese-universities-with-military-links-bought-super-micro-servers-with-2026-03-27/. ↑
- Mackenzie Hawkins, Andy Lin, and Kari Soo Lindberg, “Nvidia’s Biggest Southeast Asian Partner Dogged by China Chip Smuggling Questions,” Bloomberg, December 22, 2025, https://www.bloomberg.com/news/features/2025-12-22/nvidia-partner-megaspeed-draws-china-chip-smuggling-concerns-in-us. ↑