US Clamps Down on Cloud GPU Rentals: 12M TPP Threshold, Linked Data Centers, 5 New License Countries
The US Bureau of Industry and Security issued an interim final rule extending advanced AI chip export controls to remote cloud access, setting a 12 million aggregate TPP threshold, merging linked data center capacity, adding license requirements for five countries including Singapore and Malaysia, and scrutinizing indirect use by restricted entities.
Background: Huang's Departure and Reuters Report
On November 12, 2026, NVIDIA announced that Jensen Huang resigned as CEO, president, and director, with CFO Colette Kress becoming interim CEO. The same day, Reuters reported that Chinese AI firms — including a large cloud provider and smaller labs — were contacting overseas data centers to pre‑book NVIDIA GPU capacity, fearing that Huang‑related controversy could trigger tighter US remote‑compute restrictions. The article notes this reflects corporate risk assessment, not a confirmed causal link between Huang's exit and subsequent rulemaking.
Why Overseas GPU Rental Became a Workaround
US rules already blocked direct export of high‑end NVIDIA GPUs to China. Chinese firms therefore turned to foreign data centers: the operator buys and hosts the hardware abroad, and the Chinese customer submits training jobs remotely. The chips never enter China, and the computation occurs offshore. However, existing US regulations already covered cases where a restricted entity becomes the sole or primary remote user of a data center, potentially requiring a license. The new rule expands this scope.
New Interim Final Rule: "Addressing Remote Access to Advanced Semiconductors"
On November 16, BIS published the interim final rule (IFR) drafted by the Treasury and National Security Council, effective November 19 via the Federal Register. It adds §744.24 to the Export Administration Regulations (EAR), targeting large‑scale remote compute supply to entities headquartered or ultimately owned in Country Group D:5 (which includes mainland China) or Macau.
12 Million Aggregate TPP Threshold
The rule triggers when an exporter, reexporter, or transferor knows the controlled items will be used in a data center and provides more than 12 million aggregate TPP (Total Processing Performance) to a qualifying entity. TPP is defined per ECCN 3A090 as theoretical multiply‑accumulate performance multiplied by operand bit‑width ; aggregate TPP sums the TPP of all controlled chips/devices in the transaction. It is not a simple GPU count and cannot be equated with raw FLOPS.
Figure 3: GB200 NVL72 rack cross‑section — a single cabinet as a TPP aggregate unit
Linked Data Centers: Merged TPP Calculation
Modern GPU clouds distribute capacity across multiple facilities connected by high‑speed networks. The IFR introduces Linked Data Centers : two or more data centers linked physically or digitally that share resources, balance workloads, replicate data, and operate as a unified whole. For §744.24 applicability, the TPP of all linked centers must be aggregated . Mere network connectivity does not automatically create linkage; actual resource sharing and joint scheduling are required.
Figure 4: Dual‑plane network topology enabling cross‑facility resource pooling
Indirect Use and End‑Use Changes
The rule also captures indirect use : if a non‑D:5 entity rents GPUs but the data center knows the compute is actually training models for, or delivering models to, a D:5 entity, the transaction may require a license. Moreover, if already‑deployed GPUs are repurposed to serve qualifying entities above the threshold, the in‑country transfer and end‑use change can trigger new authorization requirements even though the hardware hasn't moved.
90‑Day Transition for Existing Services
Data centers already providing large‑scale compute to covered entities before the rule's effective date have 90 days to obtain a BIS license; otherwise they risk subsequent export restrictions.
Five New License Destinations: Singapore, Malaysia, Thailand, Indonesia, Brazil
Separately, BIS added five countries to the license‑required destination list for advanced computing chips and related equipment. The stated rationale is diversion risk: controlled items might transit these jurisdictions to restricted end users. The rule includes a conditional exclusion : if the ultimate consignee or end user is headquartered in the US, or its ultimate parent is US‑based, the specific destination license requirement is waived provided the exporter files monthly reports with BIS (due the first business day of the following month) detailing parties, destination, chip model, ECCN, TPP, quantity, and export date. All other cases face case‑by‑case review considering the destination's export control regime and enforcement cooperation.
Megaspeed Case Illustrates Prior Enforcement
A joint DOJ‑Commerce statement (Nov 17) revealed an investigation into Singapore‑based Megaspeed beginning in 2024. In May 2025, Commerce requested NVIDIA's sales records related to Megaspeed; in June, NVIDIA placed Megaspeed on its stop‑ship list, and US investigators conducted on‑site end‑use checks with Singapore authorities. Megaspeed denied illegal diversion and said it cooperated. The investigation predates the Huang controversy and remains ongoing; no public finding of violation exists yet.
Impact on Chinese AI Firms and GPU Cloud Operators
The article stresses that the new rule does not amount to a blanket ban on overseas GPU rental. §744.24 sets specific item, capacity, end‑user, and end‑use conditions; licenses are granted case by case. Transactions outside those conditions, or those that secure authorization, may proceed. However, firms needing sustained, large‑scale GPU clusters must now factor in:
Whether the data center needs a license for the specific engagement.
Whether the end‑use complies with the rule.
Whether contract execution could be disrupted by licensing delays or denials.
Large‑model training demands continuity — long‑running jobs, cross‑node communication, checkpointing, fault recovery, and scheduling all require stable environments. A license‑driven interruption could force retraining plan changes. So far, no public evidence shows specific Chinese AI training jobs halted or price shifts attributable to the rule.
Data Center Operator Challenges
Operators build GPU clusters expecting long‑term leases to maintain utilization. If some customers require extra licenses, or future end‑use changes jeopardize contracts, operators must reassess business arrangements. This pressure may propagate to NVIDIA and its supply chain. Huang previously expressed desire to serve China; the new management now faces a regulatory regime that extends beyond where chips are sold to how they are used in third‑country data centers.
Broader Regulatory Shift: AI Diffusion Rule Repealed, FDP Rule Retained
The same IFR formally rescinds the Biden‑era AI Diffusion Rule (issued Jan 2025, never implemented, withdrawal announced May 2025). It removes the country‑tier framework, certain data center authorization mechanisms, and ECCN 4E091's model‑weight controls. However, the Foreign Direct Product (FDP) Rule for advanced computing chips retains global reach: even when controlled chips are integrated into third‑country‑manufactured servers or equipment, the transaction may remain subject to US export controls.
Conclusion
US oversight of advanced compute has shifted toward destination, end‑use, and customer relationships . For Chinese AI firms, overseas GPU rental remains a potential compute source, but supplier compliance and sustained delivery assurance require deeper due diligence. Domestic AI chips and software ecosystems will enter more procurement evaluations, contingent on actual performance, software compatibility, supply volume, and total cost of ownership. Huang's departure did not stop advanced compute supply, but the new rule adds licensing hurdles to compute services previously routed through third‑country data centers. The most immediate question for firms planning long‑term overseas GPU contracts is no longer just whether hardware can be deployed and jobs submitted, but whether the contracted compute can be delivered continuously for the required duration.
References
[1] NVIDIA: "NVIDIA Announces CEO Transition", Nov 12, 2026.
[2] Reuters: "Chinese AI firms rush to secure Nvidia chip capacity as spy scandal stirs fear of US curbs", Nov 12, 2026.
[3] US Commerce BIS: "BIS Issues Rule Addressing Remote Access to American Super Intelligence Technology", Nov 16, 2026. https://www.bis.gov/press-
[4] Federal Register / BIS: "Addressing Remote Access to Advanced Semiconductors", Doc. 2026-23529, Nov 16, 2026 preview.
[5] US DOJ & Commerce: "Joint Statement on Megaspeed Investigation", Nov 17, 2026.
[6] Reuters: "US releases details of Megaspeed probe, says investigation began in 2024", Nov 17, 2026.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
