New BIS Licensing Policy for H200s: Tough Guidelines, Weak Enforcement
On January 13, 2026, the Trump administration released a new licensing policy for exports of Nvidia H200s and equivalent chips to China under certain conditions. IAPS analysis identifies the following implications:
The policy limits H200 exports to China to less than 50% of total U.S. sales, but that will still dramatically improve China’s capabilities. Exporters will have to certify to BIS that they are selling less than 50% as many chips to China as the number of the same-model chips that have been sold in the U.S since launch. The BIS policy permits exports of all chips below certain performance thresholds, but the criteria for the export tax suggest only Nvidia H200 and AMD MI325X chips will be permitted. Up to 900,000 H200-equivalents of compute¹ could be exported to China under the new policy, before accounting for future U.S. sales. Exporting millions of advanced chips would dramatically increase China's access to AI-relevant compute, accelerate China’s military modernization, and help Chinese AI companies undermine the market share of U.S. products worldwide.
The policy requires exporters to certify they are prioritizing U.S. sales, but the specific constraints are subjective and easily subverted. Exporters would have to certify that sales to China are not delaying U.S. orders or diverting global foundry capacity away from U.S. customers. However, it will be very challenging for BIS to fact-check exporters on claims about complex market conditions, especially if BIS is under political pressure to approve exports, and exporters will be strongly incentivized to certify these things regardless.
The policy requires the Chinese recipient to provide a security plan, but BIS has almost zero leverage to enforce the plan once the chips leave the U.S. The policy requires exporters and ultimate consignees to provide plans for Know Your Customer (KYC) due diligence and physical security, as well as lists of remote access customers in countries of concern and guarantees about non-transfer of model weights or provision of remote access to restricted parties. However, once the chips are in China, it’s up to Chinese companies to enforce the control: BIS has almost zero leverage. BIS is essentially giving China millions of advanced AI chips now, based on highly optimistic promises about future good-faith behavior.
The 25% export fee proposed by the policy may be legally challenged and will not raise a meaningful amount of revenue. The 25% export fee is implemented through a requirement that the chips’ performance be verified at a lab physically in the United States prior to export, and a 25% tariff on chip imports (since all Hopper-generation chips are originally fabricated outside the United States). This arrangement may not withstand legal scrutiny; even if it does, the approximately $6.75 billion this could raise is tiny compared to other government revenue.
What Does the Policy Change?
It allows some exports of H200s or equivalent chips to China if conditions are met.
The policy changes the licensing policy for certain advanced AI chips to China and Hong Kong from a “presumption of denial” to a “case-by-case” licensing policy. The policy does not alter the “presumption of denial” for advanced AI chips to the rest of BIS Country Group D:5, including countries like Russia, Venezuela, and Iran. The administration’s tariff announcement only applies to the H200 and the MI325X, suggesting only those chips will be approved.
Which Chips are Eligible for Export?
H200s or equivalent chips (based on their processing speed and memory bandwidth).
Under the new BIS licensing policy, eligibility is defined by two technical thresholds intended to proxy for advanced AI training capability: “Total Processing Performance” and “total DRAM bandwidth”. To qualify for the new licensing policy, chips must have a total processing performance of less than 21,000, and a total DRAM bandwidth of less than 6,500 GB/s.
“Total Processing Performance (TPP)” measures how fast a chip performs operations on data. Higher TPP allows models to be trained faster (more operations per second) and more cheaply (more operations per unit of power consumption).
“Total DRAM bandwidth” is a measure of how fast a chip can move data back and forth from memory. Higher DRAM bandwidth enables training larger models, using chips more efficiently, and implementing memory-intensive model architectures like “Mixture of Experts” (MoE) that improve capabilities.
50% Quantity Rule
The policy limits H200 exports to China to less than 50% of total U.S. sales, but that will still dramatically improve China’s capabilities.
The new BIS policy restricts exports to China based on total historical sales of that model of chip for use in the U.S. The new BIS licensing policy requires that exports to China of a specific model of chip do not exceed 50% of sales of that chip “for end-use in the United States” from when the chip was first offered for sale to the date of the license application. This means that Nvidia cannot sell more H200s to China than 50% of the total quantity it has ever sold “for end-use in the United States.” The exporter is required to provide BIS with the total historical sales information needed to calculate the limit.
The current limit is estimated to be 900,000 H200-equivalents,³ which we estimate will initially increase by 150,000 chips each quarter as Nvidia and AMD sell additional chips to U.S. customers,⁴ and then plateau as U.S. customers transition to more advanced chips. Although Congress is considering several bills to limit advanced AI chip exports, there is currently no legal prohibition on revision of this policy to allow sales of Blackwell and Rubin chips, and such revision would be consistent with administration actions so far.
900,000 H200-equivalents is about 2 years of current Chinese chip production. The Institute for Progress estimates that China can produce only between 60,000 and 160,000 Blackwell-equivalents, or between 227,400 and 606,400 Hopper-equivalents, each year. Selling 900,000 H200-equivalents to China would catapult it two years into the future in its efforts to accumulate compute, and sales under this policy will likely be even higher over time.
Exporter Certifications: Sufficient Supply and Foundry Capacity
The policy requires exporters to certify they are prioritizing U.S. sales, but the specific constraints are subjective and easily subverted.
Exporters are required to certify that exports are not delaying U.S. orders, but BIS has no way to verify this. The new policy mandates that exporters certify that exports to China will not result in delays in fulfilling orders for any U.S. customers. However, BIS has no way to verify this certification, unless companies whose orders are delayed choose to complain to BIS and can demonstrate an export linkage. BIS would be relying on the assertion of an exporter who is heavily incentivized to certify there will be no delays, without a reliable way to check this claim.
Exporters are required to certify that exports are not diverting global foundry capacity, but BIS has no way to verify this either. The new policy also mandates that exporters certify that exports to China will not divert global foundry capacity away from end-users in the United States. However, again, without access to foundries’ order books and internal communications, there is no way for BIS to meaningfully fact-check this. The situation is worse than for the sufficient supply criteria, because U.S. companies whose orders have been delayed have an incentive to complain to BIS, but foundries have no incentive to report on capacity allocation.
Recipient Certifications: KYC, Physical Security, and End-Users
The policy requires the Chinese recipient to provide a security plan, but BIS has zero leverage to enforce the plan once the chips leave the U.S.
Recipients are required to provide detailed Know Your Customer (KYC) due diligence and physical security plans, but BIS has no way to enforce compliance with these plans. The licensing policy requires exporters to obtain detailed information on how chip recipients will carry out Know Your Customer (KYC) due diligence and maintain the physical security of the chips. However, BIS has no ability to enforce these conditions once the chips are in China. There is no “kill-switch” on an Nvidia H200, and the threat of denying future chip sales is not very useful: if an applicant can purchase hundreds of thousands of H200s legally in one transaction, it may not need to buy more, and threatening to cut off chips won’t deter it.
Recipients are required to provide lists of proposed remote users, commit to not providing remote access to unauthorized users, and commit to not transfer model weights without authorization, but BIS may not even be able to detect these violations. The licensing policy requires exporters to certify that remote access will not be provided to prohibited end-users (for example, China’s military, intelligence services, or WMD programs), and to provide a list of proposed Infrastructure-as-a-Service (IaaS) remote access end users in countries of concern.⁵ If the recipient provides IaaS services, they must also commit to not transferring model weights without BIS authorization or providing remote access to models trained using the chips to prohibited end-users.
However, BIS is unlikely to be able to even detect these violations, let alone successfully deter them. BIS enforcement has been described as running on “Google searches and Microsoft Excel” due to the bureau’s outdated technology and minimal resources, and it is unclear how BIS will detect a digital transfer of model weights on a network in China, or monitor who is actually receiving remote access. The requirements of the policy are sound, but for BIS to enforce them it would need to aggressively monitor the digital activities of chip recipients in ways the policy does not legally provide for and BIS may not be able to implement.
25% Export Fee
The 25% export fee proposed by the policy may be legally challenged. The licensing policy only applies to direct exports from the U.S., and mandates that chips be tested at a lab physically in the United States prior to export. This will force the chips to be imported to the U.S. before being exported to China, allowing the export fee to be framed as an import tariff. While Nvidia and AMD, who benefit from this arrangement, are unlikely to sue to end it, federal courts have previously frowned on novel interpretations of tariff powers.
The export fee will not raise a meaningful amount of revenue. We estimate, at maximum, the export fee would raise around $6.75 billion. By comparison, the U.S. federal tax take in 2025 was about $5.23 trillion. In exchange for doubling China’s access to AI-relevant compute, the U.S. would increase its annual tax revenue by 0.13%, or around half of the cost of a single Gerald R. Ford-class aircraft carrier.
Endnotes
Based on "Unpacking the H200 Export Policy" by Janet Egan and James Sanders, CNAS
Full calculation by the Institute for Progress (IfP) of TPPs of various chips with sources
Based on "Unpacking the H200 Export Policy" by Janet Egan and James Sanders, CNAS
Per IfP analysis of Nvidia’s 2024 production of Hopper-generation chips
Belarus, China, Cuba, Iran, Macau, North Korea, Russia, and Venezuela