2026 Annual Report to Congress ·
On this page: Executive SummaryKey FindingsRecommendationsRelated Hearings and Research
Executive Summary
Competition between the United States and China for leadership in AI extends across the technologies, infrastructure, and applications that underpin the development and deployment of AI systems. Leadership at one layer can create advantages elsewhere in the stack, but weaknesses at one layer can also constrain the performance of the entire system. This makes control over key AI-related technologies, infrastructure, and supply chains a critical issue for U.S. economic and national security. The United States and its allies and partners currently hold substantial advantages in many of the most technically difficult and concentrated segments of this stack. U.S. firms lead in advanced AI chip design and software; Taiwan dominates leading-edge semiconductor fabrication; the Netherlands maintains a monopoly over extreme ultraviolet lithography systems; and South Korean and U.S. firms currently dominate high-bandwidth memory. Together, these advantages have constrained China’s access to the advanced computing resources required to develop frontier AI systems.
China has pursued technological self-sufficiency for more than two decades, and U.S. technology restrictions since 2019 have increased the challenges—and likely the urgency—of Beijing’s localization campaign. Export controls on advanced AI chips, semiconductor manufacturing equipment, and related technologies have disrupted China’s access to advanced computing capabilities and slowed the development of its indigenous semiconductor ecosystem, even as Chinese firms continue to make progress. China is simultaneously attempting to engineer around its remaining technological chokepoints while leveraging advantages elsewhere in the AI stack, including China’s dominance of critical minerals and materials. These strengths could help China compensate in part for less efficient domestic AI hardware and other technological constraints.
Ultimately, U.S.-China AI competition is unlikely to be determined by dominance at any single layer of the technology stack. Despite constraints on advanced compute, Chinese firms have produced near-frontier models with fewer advanced computing resources, while inexpensive open-weight models have facilitated their rapid international adoption. China could therefore remain dependent on foreign technology at critical portions of the infrastructure layer while gaining influence in global model and application ecosystems. Conversely, U.S. leadership in advanced computing will not automatically translate into broader AI leadership without sufficient energy, networking, and data center infrastructure to deploy these technologies at scale. U.S. and allied export controls can preserve critical technological advantages and slow China’s progress, but their strategic value ultimately lies in buying time for the United States to strengthen or solidify its competitiveness across the entire AI technology stack. Part Two of this research will dive deeper on how China is attempting to overcome its own chokepoints across the AI tech stack, including those associated with software, servers, hyperscalers, and energy constraints.
Key Findings
- The United States and allies hold decisive advantages at the top of the stack, and export controls are the principal tool slowing China. Their continued effectiveness depends on closing loopholes and working closely with allies to prevent their erosion.
- For more than two decades, Beijing has been pursuing self-sufficiency and global competitiveness across the entire semiconductor ecosystem and—more recently—the AI technology stack. U.S. and allied controls did not drive China’s indigenization efforts, but they have added urgency in recent years.
- Although China has made some progress in achieving its goals, it has not yet closed the hardest chokepoints—including the most advanced capabilities in logic fabrication, memory, and semiconductor manufacturing equipment, as well as the Nvidia Compute Unified Device Architecture (CUDA) software moat—and likely will not for some time.
- China is investing heavily to engineer around major chokepoints: alternative lithography, advanced packaging, photonics, and an open-weight, lower-compute model strategy. Success is not assured, but U.S. policymakers need to closely follow developments in these areas.
- China’s dominance in processing and refining several critical minerals and rare earth elements creates upstream vulnerabilities throughout the AI technology stack for the United States and its allies and partners. Beijing has demonstrated its willingness to leverage this position through export restrictions, raising the risk that it could disrupt U.S. access to materials needed for advanced semiconductors, data center hardware, power systems, and other AI infrastructure during a geopolitical or trade dispute.
Recommendations
- The growing proliferation and adoption of Chinese open-weight artificial intelligence (AI) models in the United States carries national security implications for sensitive data, cybersecurity, and dependence on Chinese technology. Uncertainty about how these models behave in sensitive applications, what vulnerabilities they may introduce, and whether safeguards can adequately mitigate those risks warrants greater federal scrutiny. Congress should require enforceable protections for sensitive applications before these models become embedded in critical systems and costly to replace.
Congress should require the Secretary of Commerce, through the Office of Information and Communications Technology and Services (OICTS), in coordination with the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) and relevant federal agencies, to:
- Restrict deployment in sensitive applications unless safeguards are demonstrated. Within 180 days, identify sensitive uses—including systems controlling critical infrastructure, performing national security functions, or processing bulk sensitive personal data—and prohibit new deployment of covered models in those uses absent a time-limited authorization based on independently evaluated safeguards. Authorizations should address the model, its deployment environment, and its access to data, networks, and tools.
- Require identification and mitigation of existing exposure. Direct federal agencies and covered operators in designated sensitive sectors to inventory existing deployments and implement enforceable mitigation or replacement plans within deadlines established by the Secretary.
- Establish continuing security obligations. Require independent security testing, reporting of significant security incidents, reassessment following material modifications, and verification of compliance. Direct the Secretary to restrict or prohibit the commercial provision or deployment of designated models where identified risks cannot be adequately mitigated.
- Reduce strategic dependence. Require providers serving designated sensitive applications to maintain documented capabilities to replace covered models and migrate affected systems and data to alternatives meeting federal security requirements.
Within 180 days, the Secretary should also complete a consolidated assessment, building on existing federal evaluations, the National Telecommunications and Information Administration’s (NTIA) open-model assessment, and federal supply-chain restrictions. The assessment should distinguish risks inherent in model weights from those arising from hosting, software, updates, and system integration; identify gaps in testing and verification; and inform subsequent requirements without delaying initial implementation.
Congress should authorize and appropriate funding, information-gathering and enforcement authorities, and penalties for noncompliance. The legislation should provide narrowly tailored exceptions for security research and authorized national security activities and require annual reporting to Congress on deployment exposure, compliance, mitigation effectiveness, and remaining dependencies.
- Americans should know when the government services they receive and the commercial products they use rely on AI models developed by Chinese companies. Clear disclosure would help consumers, businesses, and investors assess potential risks to data security, privacy, service reliability, and national security. Congress should make it U.S. policy that Americans know when governments and businesses rely on AI models developed by Chinese companies. Congress should:
- Require federal agencies, and any state or local government or other entity that receives federal funds or federal tax credits, to disclose any use of AI models developed by an entity headquartered in, or organized under the laws of China, or owned or controlled by such an entity, including models fine-tuned on or derived from them, wherever they are hosted. Disclosure should appear in clear language on their public-facing platforms and in their filings with the federal government.
- Require businesses that use such models in any part of a product or service offered in the United States to consumers or other businesses to disclose that use clearly on their websites and in product documentation and treat failure to disclose as an unfair or deceptive act or practice enforceable by the Federal Trade Commission.
- Require reporting companies, as part of quarterly and annual disclosure filings with the U.S. Securities and Exchange Commission, to disclose as a material business risk the extent of their use of such models.
- Sustaining U.S. leadership in AI requires reliable access to advanced chips and the manufacturing capacity needed to produce them. Exports to China should not come at the expense of capacity needed to support U.S. AI development and deployment. Requiring auditable exporter certifications would help ensure that export decisions account for these constraints and protect U.S. access to critical capacity.
Congress should ensure that AI chip exports to China do not divert capacity from the U.S. market, including by amending the Export Control Reform Act of 2018 (ECRA) to condition exports, reexports, or in-country transfers of advanced AI chips to China on exporter certification that they will not divert advanced semiconductor manufacturing or AI capacity from the U.S. market. The legislation should:
- Cover upstream capacity, including advanced-node wafer fabrication, high-bandwidth memory, and advanced packaging—not finished-chip inventory alone;
- Require auditable capacity-allocation data submitted to the Bureau of Industry and Security, subject to ECRA recordkeeping and enforcement provisions, including criminal penalties for false certification;
- Require the Bureau of Industry and Security to report annually to the committees of jurisdiction on certifications, validation methods, and exports coinciding with unmet U.S. or allied demand for equivalent capacity; and
- Prohibit regulatory waivers or narrowing of the requirement without an affirmative national security determination transmitted to Congress for a 30-day review period.