What export controls do
Export controls regulate transactions involving specified goods, software, technology, destinations, end users, and end uses. In AI, U.S. controls have targeted advanced computing chips, high-bandwidth memory, semiconductor-manufacturing equipment, software and technology, support by U.S. persons in defined circumstances, and entities associated with restricted programs. Sanctions are a different legal tool, generally restricting dealings with designated countries, entities, or people.
Neither is a universal AI “off switch.” A chip control can constrain access to hardware subject to U.S. jurisdiction; it does not delete algorithms, stop all inference, or prevent every domestic system. The Export Administration Regulations are technically detailed and change frequently. Organizations need current, qualified compliance advice rather than relying on a general article.
The policy timeline matters
In October 2022, the U.S. Bureau of Industry and Security imposed major controls on advanced computing and semiconductor manufacturing related to China. It revised and expanded them in 2023 and 2024, including performance-density parameters, high-bandwidth memory, manufacturing tools, software, end-use rules, and Entity List additions (BIS advanced-computing overview).
In January 2025, the outgoing administration issued the Framework for Artificial Intelligence Diffusion, which would have created a wider country-tier and data-center authorization system. Its compliance requirements were scheduled for May 15, 2025. On May 13, the Commerce Department announced it would not enforce the rule and would rescind it, while issuing chip-diversion and supply-chain guidance (BIS rescission announcement).
Therefore, the Diffusion Rule should not be described as the present global framework. It did not take effect as planned. Later rules and current EAR provisions must be checked directly.
In January 2026, BIS changed the license-review policy for Nvidia H200, AMD MI325X, and comparable chips to case-by-case review for approved Chinese customers that meet specified security, supply, testing, and compliance conditions (Commerce announcement). Current EAR pages also contain destination, end-user, license-exception, and model-weight provisions; the official electronic rules, licenses, and entity lists—not old summaries—control (BIS EAR).
This account is current through September 13, 2026 and will age quickly.
The control surfaces are different
Chips and high-bandwidth memory affect training and high-volume inference, but capability is not determined by one chip specification. Clusters combine accelerators, networking, memory, power, software, and time.
Manufacturing equipment and technology aim farther upstream. Advanced chips depend on lithography, deposition, etching, metrology, design software, and specialized knowledge across several countries. Multilateral alignment matters because substitution can reduce unilateral leverage.
Cloud access can give a customer compute without shipping hardware. Due diligence and end-use controls may address remote access, but they raise privacy, jurisdiction, and implementation questions.
Model weights are software artifacts that can be copied and used for inference. Certain export rules can cover advanced-model weights, but defining scope, nationality, publication, and foreign-produced items is difficult. Restricting weights also affects open research and competition.
Services and people may fall under restrictions on technical assistance, transactions, or designated entities. These measures affect collaboration and mobility and therefore require narrow definitions and due process.
Lumping these into “chip sanctions” makes it impossible to evaluate effectiveness.
What success can mean
Controls may seek to delay a specific military or surveillance program, increase its cost, reduce scale, preserve a technological lead, gain bargaining leverage, or prevent a particular transaction. These goals require different evidence. “The country still built a capable model” does not prove a control had no effect; “shipments fell” does not prove the intended security outcome occurred.
Assessment can examine availability and price of controlled inputs, cluster construction, domestic substitution, manufacturing yields, model-training scale, diversion cases, and downstream capability. Much of the relevant evidence is classified, commercially sensitive, or confounded by other policies. Public claims should acknowledge that uncertainty.
Export controls can buy time, but time has value only if used for diplomacy, defense, evaluation, or resilience. Delay alone is not a safety strategy.
Evasion and adaptation
Restricted actors can use intermediaries, mislabel shipments, rent overseas compute, repair old equipment, stockpile, redesign chips, improve algorithms, or invest in domestic supply. Enforcement therefore needs end-user checks, beneficial-ownership analysis, serial and shipment records, intelligence sharing, penalties, and cooperation with allies.
Overbroad rules can increase incentives for smuggling, fragment markets, and accelerate indigenous alternatives. They can also deny benign users, raise prices, or push technical standards away from U.S. firms. Narrow rules can be easier to administer but leave substitutions. This is a design tradeoff, not evidence that controls are inherently effective or futile.
Algorithmic efficiency further complicates thresholds. Better training and inference can produce more capability per chip. Rules require review as hardware, packaging, and system design change. Performance thresholds also invite products designed just below the line.
Strategic and safety tradeoffs
Controls can slow the accumulation of concentrated capability by a rival state, including military uses. They can also intensify a geopolitical race, reduce scientific exchange, or convince each side that dependence is intolerable. If allies view extraterritorial rules as unfair, coordination may weaken.
From an AI-safety perspective, limiting access by a high-risk actor may reduce one threat while concentrating frontier development in fewer companies or countries. Concentration can improve oversight but increases market power and common-mode failure. Wider diffusion can support defensive research and resilience while making recall and consistent safeguards harder.
Human rights matter. Controls aimed at military or mass-surveillance capability may be defensible; restrictions based only on nationality can damage researchers and communities. Licensing should include clear criteria, timely decisions, appeal, and humanitarian or civil-society considerations where law allows.
Sanctions require special care
Sanctions may freeze assets, prohibit services, or isolate designated actors. Their impact can extend beyond the target through banks, cloud providers, universities, and supply chains. Secondary effects should be monitored, and exemptions should be workable rather than symbolic.
Public advocates should not attempt informal enforcement through harassment or interference with shipments. Suspected diversion belongs with corporate compliance teams and authorized government channels. Evidence should be preserved and reported lawfully.
A more coherent policy package
Export controls work best with:
- multilateral alignment among major equipment and chip-producing jurisdictions;
- current technical definitions and regular review;
- resources for licensing and enforcement;
- secure, privacy-respecting cloud due diligence;
- support for defensive research and supply-chain resilience;
- diplomatic communication about objectives and red lines;
- measurement of costs, evasion, substitution, and actual security outcomes;
- sunset or review processes where evidence can change the rule.
Domestic safeguards remain necessary. Preventing an overseas actor from obtaining chips does not ensure that a domestic developer evaluates models, secures weights, or responds to incidents.
The practical conclusion
Export controls and sanctions can constrain specific AI inputs and actors. They cannot control “AI” as one object, and their effects are neither zero nor fully measurable from public model releases. The January 2025 Diffusion Rule is a particularly important caution against stale descriptions: it was rescinded before its scheduled compliance date.
As of September 2026, the responsible approach is to cite the current EAR, specify the hardware, technology, weight, service, destination, end user, or end use at issue, and state the policy objective. Controls should be lawful, targeted, enforceable, coordinated, and assessed against alternatives—not treated as a permanent technological blockade.