On January 15, 2026, BIS made a quiet but pointed change to its license-review policy for advanced computing commodities, and the technical detail is in the thresholds, not the politics. The rule shifts the review standard for a defined band of chips, exemplified by the NVIDIA H200 and AMD MI325X, from a presumption of denial to a case-by-case review for exports to China and Macau, but only for parts that sit below specific TPP and DRAM-bandwidth figures and only when the exporter makes a set of certifications. Reading the rule as a classification matter, the news is the precise definition of the band that moved.
The numeric band that moved
The rule targets advanced computing commodities with a TPP less than 21,000, as defined in Technical Note 2 to ECCN 3A090.a and 3A090.b, and a 'total DRAM bandwidth' less than 6,500 GB/s, as defined in the notes to the relevant paragraph of supplement no. 2 to part 748. The rule names the NVIDIA H200 and the AMD MI325X as exemplars of parts in that band. Two things are worth pinning down. First, this is not a blanket loosening of the advanced-computing controls; it is a redefinition of how a sub-threshold slice is reviewed. Parts above 21,000 TPP or above 6,500 GB/s of total DRAM bandwidth are untouched by the new case-by-case policy. Second, the rule introduces 'total DRAM bandwidth' as a controlling parameter alongside TPP, which is notable because it makes the memory feeding the accelerator part of the line-drawing, not just the compute. A part can be inside the eligible band on compute yet outside it on DRAM bandwidth, and both have to be satisfied.
From presumption of denial to case-by-case
Mechanically, the change lives in section 742.6 (Regional Stability) of the EAR, which the rule amends to provide a case-by-case licensing policy for license applications covering the specified advanced computing commodities. 'Case-by-case' is a defined posture in the EAR: it does not mean approval, and it does not mean the prior presumption of denial. It means BIS evaluates each application on its facts against stated criteria rather than starting from a thumb on the scale toward denial. For the covered band, that is a meaningful shift in the default, but it is gated by conditions that keep the agency's discretion intact.
The certifications that gate eligibility
The case-by-case treatment is available only if the chips are commercially available in the United States at the time of publication and the exporter certifies a specific list of facts. Those certifications are unusually substantive. The exporter must certify that there is sufficient supply of the product in the United States; that producing it for export to China will not divert global foundry capacity away from similar or more advanced products destined for U.S. end users; that the recipient has demonstrated sufficient security procedures; and that the item undergoes independent, third-party testing in the United States to verify its performance specifications. Each certification maps to a distinct policy worry. The supply and foundry-capacity certifications guard against export demand cannibalizing domestic and allied access to leading-edge capacity. The security-procedures certification addresses diversion risk at the recipient. The third-party-testing requirement is the most technically interesting: it forces independent verification that a part actually sits in the eligible performance band, closing the door on misrepresented specifications by making the performance claim auditable rather than self-asserted.
Why the testing requirement matters most
The independent-testing condition is the rule's enforcement spine. The entire premise of parameter-based controls is that a part's TPP and DRAM bandwidth determine its treatment, which means the system is only as good as the verification of those figures. By requiring U.S.-based, independent, third-party testing to confirm a covered part's performance specifications, BIS converts the TPP-less-than-21,000 and DRAM-bandwidth-less-than-6,500-GB/s thresholds from datasheet claims into verified facts. That is consistent with the broader trajectory of these rules: as the controls have moved from nanometers to computed performance, the agency has had to build mechanisms to verify the performance, and mandatory third-party testing is the cleanest such mechanism yet. For the covered band of accelerators, the January 2026 rule is therefore less a relaxation than a recalibration, trading a flat presumption of denial for a conditional, verified, certification-gated review keyed to two precise numbers.
Why 'total DRAM bandwidth' joins TPP at the line
The pairing of TPP with 'total DRAM bandwidth' as co-equal eligibility parameters is the rule's most technically telling feature, and it follows directly from how AI accelerators actually perform. A large accelerator's usable throughput on real workloads is frequently gated by how fast it can move data to and from memory, not by its raw arithmetic ceiling. By writing 'total DRAM bandwidth less than 6,500 GB/s' alongside 'TPP less than 21,000,' BIS acknowledged that two parts with identical compute can deliver very different real-world AI performance depending on their memory bandwidth, and it refused to let memory bandwidth ride free. The result is a two-dimensional eligibility box: a part qualifies for case-by-case review only if it sits inside on both axes. This mirrors the logic seen elsewhere in the regime, where HBM and memory bandwidth are treated as first-class objects of control rather than incidental specifications, and it means a vendor cannot move a part into the eligible band by trimming compute while leaving its memory subsystem untouched.
Placing the change in section 742.6, the Regional Stability section, rather than in the part 744 end-use controls, also signals how BIS is reasoning about this band. Regional Stability is the policy basket for items whose concern is the destabilizing accumulation of capability rather than a specific prohibited end use, and routing the H200-class band through a case-by-case Regional Stability review lets the agency weigh each application's facts, the certifications, the testing results, the destination, against that frame. The January 2026 rule thus fits the pattern of the entire advanced-computing regime: precise numeric thresholds (21,000 TPP, 6,500 GB/s), a defined commodity band exemplified by named parts, and an enforcement mechanism, here mandatory independent testing, built to make the numbers verifiable rather than self-asserted.