For three years the defining feature of U.S. advanced-computing export policy toward China has been the presumption of denial: a license application to ship the most capable AI accelerators was assumed to be rejected unless an exporter could overcome that posture. A rule from the Bureau of Industry and Security (BIS) changes that posture for a specific class of chips, moving from presumption of denial to a case-by-case review. The headline beneficiary is the Nvidia H200 and its equivalents — but the conditions attached are the part that determines whether the door is actually open.
The change is narrow and deliberate. It does not deregulate; it reclassifies how applications are weighed. Under a presumption of denial, the burden is effectively insurmountable for most applicants. Under case-by-case review, BIS evaluates each application on its merits — which is a meaningfully more permissive stance while still keeping every shipment under license. The covered scope is the H200 and equivalents, plus less advanced chips, and the rule conditions eligibility on the items being commercially available in the United States at the time of publication.
"The Bureau of Industry and Security (BIS) is revising its license review policy for exports of certain semiconductors to China and Macau--changing it from a presumption of denial to a case-by-case review."— Federal Register, BIS rule 2026-00789, source
The single most consequential word in the rule is "China and Macau," because it names exactly the destinations where the presumption of denial had been the operative posture. Shifting that posture for the H200 — a high-bandwidth-memory-rich data-center accelerator that sits a tier below Nvidia's most capable parts — is a calibrated move, not a blanket loosening. The structure of the policy lives in the four certifications the exporter must make.
The four certifications are the actual policy
An export is not waved through simply because review is now case-by-case. The rule requires the exporter to certify four distinct things. First, that there is sufficient supply of the product in the United States — a guard against the export draining domestic availability. Second, and most revealing, that production of the product for exports to China "will not divert global foundry capacity for similar or more advanced products for end users in the United States." Third, that the recipient has demonstrated sufficient security procedures. Fourth, that the item undergoes independent, third-party testing in the United States to verify its performance specifications.
The second certification is where the policy's logic is clearest. The binding constraint in advanced AI silicon is not design but manufacturing capacity — specifically leading-edge foundry output and the advanced-packaging and high-bandwidth-memory supply that feeds it, all of which are sold out well ahead. By requiring exporters to certify that China-bound production will not divert that scarce capacity away from U.S. end users, BIS is treating foundry and packaging capacity as a strategic resource to be rationed, not merely a commercial input. It is an unusually explicit acknowledgment that the bottleneck is upstream of the chip itself.
The fourth certification — independent third-party testing in the United States to verify performance specifications — is an enforcement mechanism against a known evasion risk. Performance thresholds have historically defined the line between controlled and uncontrolled chips, and verifying that an exported part actually meets its stated specification closes a gap that paper attestations alone would leave open. The recipient-security certification, meanwhile, addresses diversion and end-use concerns once the chip is in country.
What it signals, and what it doesn't
For the supply chain, the practical reading is that a conditional, monitored channel for H200-class exports now exists where one effectively did not. That has commercial significance for the chipmaker and for the customers on both ends, but it is not a return to open trade. Every shipment remains licensed, every application is reviewed individually, and the certifications create an audit trail and a set of representations that can be checked and enforced. A company modeling its addressable China demand off this rule should anchor to the certifications, not the headline shift — particularly the foundry-capacity-diversion test, which a manufacturer with a tight capacity allocation may find genuinely hard to satisfy.
It is also worth keeping the scope honest. The rule covers the H200 and equivalents and less advanced chips; it is not a statement about the most advanced parts, and "equivalents" will be the subject of careful interpretation. The commercial-availability-in-the-U.S. condition further ties eligibility to a snapshot in time. As with any controls action, the live questions are how BIS applies the case-by-case standard in practice, how strictly the foundry-diversion certification is policed, and whether the policy proves stable or becomes another waypoint in an evolving control regime.
The mechanics of why the H200 sits at this particular fault line are worth spelling out. The H200 is a data-center accelerator distinguished largely by its generous high-bandwidth-memory configuration, which is what makes it valuable for serving and training large models — memory capacity and bandwidth, not just raw compute, increasingly determine how large a model a single accelerator can handle. That same HBM is among the most supply-constrained components in the entire industry, sold out well in advance and dependent on a handful of memory makers and on advanced-packaging capacity to integrate it. So a rule that conditions H200 exports on not diverting foundry and packaging capacity is targeting the exact resource that makes the chip both desirable and scarce. The certification is not a generic supply-chain platitude; it is aimed squarely at the physical bottleneck that defines this class of part. That precision is what distinguishes a thoughtfully drawn control from a blunt one, and it is why reading the conditions closely matters more than reading the headline.
The throughline for anyone tracking this beat is that the controls are getting more conditional and more instrumented rather than simply tighter or looser. Rather than a binary allow-or-deny, BIS is building a graduated structure where access is contingent on verifiable conditions tied to the real bottleneck — foundry and packaging capacity. Controls reshape the addressable market, and this rule reshapes it not by opening the gate but by attaching a meter to it. Follow the certifications, because that is where the actual policy lives.