U.S. AI Chip Rule Targets Overseas Remote Access | Zeniteq
Loading
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
U.S. AI Chip Rule Targets Overseas Remote Access
Effective November 19, 2026, new U.S. controls tie overseas AI infrastructure to the customers using its compute, with licensing, reporting and ownership checks for affected transactions.
The U.S. Commerce Department’s Bureau of Industry and Security has implemented new export controls targeting remote access to advanced AI computing infrastructure by foreign actors. Effective November 19, 2026, the interim final rule requires data centers and chip suppliers to verify the ownership and headquarters of customers accessing significant compute power.
The regulation mandates end-use licensing for transactions involving companies headquartered in Country Group D:5 or Macau, while also introducing destination-based controls for Indonesia, Malaysia, Singapore, Thailand, and Brazil. Affected data centers must conduct due diligence on customer ownership and workload purposes, with a 90-day window to license existing arrangements for restricted entities. The rule specifically covers hardware classified under ECCNs 3A090 and 4A090, while formally rescinding the previously unenforced AI Diffusion Rule.
The U.S. Commerce Department’s Bureau of Industry and Security has extended advanced-chip controls to address remote computing arrangements through overseas data centers. The interim final rule, announced on November 16, 2026, took effect on November 19, bringing new obligations for chip suppliers, cloud providers and data-center operators. BIS’s announcement describes the target: restricted foreign actors obtaining computing capabilities through infrastructure in third countries.
The consequential change is that a chip’s physical location is no longer enough to assess a covered deployment. The rule creates an end-use licensing requirement for exports, reexports and in-country transfers of advanced computing items to overseas data centers that supply significant compute to certain customers.
For infrastructure teams, that connects hardware procurement with customer ownership and workload information. A data center may sit outside a restricted destination while serving a customer whose headquarters or ultimate parent brings the arrangement within the new control.
What Changed on November 19
The new end-use control targets data centers providing significant compute to, or for, companies headquartered in Country Group D:5 or Macau, including companies whose ultimate parent is headquartered there. D:5 covers U.S. arms-embargoed destinations, including China, according to Morrison Foerster’s analysis of the rule.
The legal mechanism matters. BIS is imposing an end-use license requirement on covered chip and equipment transactions involving those data centers. Describing this simply as a ban on renting GPUs would obscure the supply-chain obligations and defined scope.
The Federal Register summary also calls for customer due diligence by data centers, foreign foundries and chip exporters, including checks for direct or indirect ownership or control by persons in D:5 destinations or Macau.
That means an overseas billing address cannot, by itself, answer the compliance question. Operators need to understand the company behind an account and its ownership structure.
Covered Hardware and Significant Compute
The rule concerns advanced computing items, not every processor used for AI. identifies Export Control Classification Numbers, or ECCNs, including . The firm says the rule leaves the relevant hardware classifications’ technical parameters unchanged.
Two scope questions therefore sit alongside each other:
Does the hardware fall within a covered export-control classification?
Does the data center’s customer and compute arrangement trigger the new end-use control?
The available source summaries do not supply the numerical threshold for significant compute or its measurement conditions. This analysis therefore cannot translate that threshold into a GPU count, GPU-hours or a training-budget cutoff. Operators need the operative regulatory text before making that calculation.
BIS also identifies an exclusion for commercial recommender systems, such as those used by online retailers and social media companies. That makes workload purpose consequential. A recommender deployment and an advanced-model training arrangement require separate assessment, even when both use AI infrastructure.
I’d treat workload classification as an onboarding issue. The exclusion gives providers a concrete reason to document what a customer is buying compute to do.
Two Licensing Checks Now Affect Overseas Deployments
The remote-access control sits alongside a separate expansion of destination-based licensing.
The rule adds license requirements for covered advanced semiconductors going to, or being transferred within, Indonesia, Malaysia, Singapore, Thailand and Brazil. BIS identifies those countries as presenting elevated diversion risks. Additional destinations can be designated later. The Federal Register summary sets out both measures.
Compliance check
What brings it into scope
Relevant qualification
Remote-compute end use
A covered transaction involving a data center supplying significant compute to qualifying D:5/Macau-linked companies
Customer headquarters, ultimate-parent headquarters and workload purpose matter
Five-country destination control
Covered advanced-semiconductor exports, reexports or in-country transfers involving the five named countries
Qualifying U.S.-headquartered recipients can use a reporting-conditioned exclusion
Existing covered services
A data center was already providing qualifying significant compute before publication
A separate 90-day licensing provision requires attention
For the five-country requirement, the exclusion applies to qualifying ultimate consignees or end users headquartered in the United States, or whose ultimate parent is U.S.-headquartered. It comes with reporting obligations under §743.10, according to Greenberg Traurig.
The practical distinction is between where the equipment goes and who benefits from its computing power. Providers need to evaluate both. Satisfying a destination-related condition does not resolve the separate customer and end-use analysis.
For a multinational cloud provider, that also means the corporate structure of a deployment can affect the applicable process. A familiar brand name on a purchase order is less useful than verified information about the consignee, end user and ultimate parent.
Reporting Requires More Than an Account Name
The clearest operational detail in the published analyses is the monthly reporting requirement attached to the qualifying U.S.-headquartered-recipient exclusion.
Under §743.10, the exporter, reexporter or transferor must report transaction information. Greenberg Traurig’s summary lists the required categories:
Transaction parties and roles, including relevant parent-company headquarters and the ultimate parents of the consignee and end user.
Ultimate destination and end users, including end users that differ from the ultimate consignee.
Product details, including the paragraph-level ECCN, product name, stock-keeping unit or SKU, total processing performance and quantity.
Transaction dates for each export, reexport or in-country transfer.
Total processing performance, abbreviated TPP, is a hardware-performance reporting field here. Teams should preserve the applicable product classification and performance information alongside quantities, rather than expecting a customer-management system to hold everything required.
These reports concern covered hardware transactions. They should be kept distinct from the separate data-center reporting route referenced by BIS.
Reports under Note 1 to §742.6(a)(6)(iii)(A)(1) and §743.10.
Reports under §744.24(b)(2).
The page directs filers to designated email channels and says attached reports must be machine-readable. A scanned PDF needs optical character recognition so its text can be captured. BIS also warns that the instructions may change.
The useful takeaway is to organize reporting by its legal basis. An exclusion-related submission, a monthly transaction report and a data-center notification should each have an identified owner and the correct submission route.
My view is that the difficult work will be connecting existing records. Hardware classifications often sit with procurement or export-compliance teams; ownership information sits with legal or customer onboarding; workload details sit with technical account teams. The rule makes those records relevant to the same assessment.
Existing Customers Face a 90-Day Licensing Provision
The November 19 effective date includes a specific provision for data centers already serving affected customers.
Morrison Foerster’s analysis describes §744.24(b)(3) as requiring data centers that provided qualifying significant compute to D:5/Macau-headquartered companies before publication to obtain a license within 90 days of the effective date to continue those services.
The distinction between obtaining a license and submitting an application is important. Operators should not plan around an assumption that filing alone satisfies the provision.
The law firm also flags questions about the provision’s scope and its interaction with the commercial recommender exclusion. That makes it risky to treat the 90-day period as a blanket grace period for every existing customer or deployment.
Existing contracts now need a focused review: which customer receives the compute, who ultimately owns that customer, what workload the service supports and which covered items underpin the deployment. A contract signed before November 19 is not enough information to settle those questions.
The Developer Impact Is Availability Before Price
Developers comparing AI pricing may first encounter the rule through capacity approval, scheduling or additional customer checks. A low quoted GPU rate has limited value if the provider cannot confirm that it can supply the planned training run.
The rule is intended to constrain access to large-scale computing power for covered customers. Its actual effect on model-training schedules will depend on licensing decisions, available infrastructure and enforcement. The published evidence does not establish a measurable change in GPU rental prices.
A practical response for cloud and infrastructure teams is to make three reviews work together:
Customer review: Verify headquarters, ultimate-parent headquarters and relevant ownership or control.
Workload review: Establish the intended use and whether an exclusion may apply.
Infrastructure review: Match the service to the covered hardware, destination and transaction records.
This is the rule’s most consequential operational implication: export-compliance information now has a stronger role in capacity planning. Sales teams considering a large compute commitment need a process for resolving those questions before promising availability.
Technical teams can help by ensuring that customer, workload and infrastructure records can be reconciled. That is a practical inference from the rule’s information requirements, not a claim that BIS mandates a particular cloud architecture or logging product.
The AI Diffusion Rule Is Formally Rescinded
The same interim final rule formally rescinds the earlier AI Diffusion Rule, whose non-enforcement BIS announced in May 2025. The Federal Register summary says that earlier framework was never enforced.
Morrison Foerster says the changes remove most of that framework, including the 4E091 classification for AI model weights, while retaining License Exception Advanced Compute Manufacturing and the worldwide destination scope of the advanced computing Foreign Direct Product Rule.
For compliance teams, this requires an update to the underlying rule set. Procedures built around the old diffusion framework need to be reconciled with the new destination and end-use controls.
Comments Are Open, but the Rule Is Effective
BIS is accepting comments until January 19, 2027. The rulemaking identifier is RIN 0694-AK50, and the Regulations.gov ID is BIS-2026-0628, according to the Federal Register notice.
Operators assessing implementation problems can use that process to address definitions, reporting mechanics and the treatment of existing arrangements. The comment period does not postpone the November 19 effective date or replace the separate licensing assessment for existing services.
Frequently Asked Questions
3 questions
1
Does the Rule Ban All Overseas GPU Rentals?
No. It targets defined advanced-computing transactions and qualifying remote-compute arrangements involving D:5/Macau-linked customers. Hardware classification, customer ownership, compute scale and exclusions affect coverage.
2
Which Five Countries Face New Chip License Requirements?
Indonesia, Malaysia, Singapore, Thailand and Brazil. Qualifying U.S.-headquartered recipients, including those with a U.S.-headquartered ultimate parent, can fall within a reporting-conditioned exclusion.
3
Are Commercial Recommender Systems Excluded?
The Federal Register summary identifies an exclusion from the relevant end-use license requirement for commercial recommender systems. Operators still need to assess the workload and any separate requirements affecting the transaction.