U.S. Commerce Department Investigates Chinese Firms Leasing Computing Power in Southeast Asia
0xBroomberg
The U.S. Bureau of Industry and Security has for the first time placed cross-border remote compute leasing under export-control scrutiny, triggered by allegations that Moonshot AI used Nvidia GPU clusters in Thailand to train its models; the move signals a systematic tightening of the grey channel Chinese AI firms have relied on.
What exactly is this probe targeting?
BIS enforcement has launched a dedicated review focused not on physical chip smuggling but on Chinese AI firms remotely renting high-end Nvidia compute from Southeast Asian data centers.
This means → the regulatory logic has shifted from "control hardware leaving the country" to "control compute crossing borders" — even if the chips never leave the local server room, submitting training jobs remotely may now trigger liability.
The direct trigger: White House science and technology policy chief Michael Kratsios publicly named Moonshot AI on July 22, alleging it tapped Nvidia GB300 clusters deployed in Thailand to train its AI models and conducted "large-scale model distillation" of Anthropic's top models.
What did Moonshot AI do, and why was it singled out?
Moonshot AI released Kimi K3, a large model with 2.8 trillion parameters that matched or partially surpassed top closed-source models from OpenAI and Anthropic across multiple benchmarks.
In plain terms = a Chinese company used U.S.-controlled chips in a third country's data center to produce results rivaling America's strongest models — that hit a nerve in Washington.
This reflects how fast Chinese AI breakthroughs are arriving — fast enough for the U.S. to conclude its current export controls have a gap.
Is renting remote compute actually illegal?
Under current Export Administration Regulations, a Chinese firm remotely renting locally procured compute in Southeast Asia is not directly illegal — the rules are built around physical goods crossing borders, and remote job submission falls outside their explicit scope.
Three actions do trigger administrative or criminal liability: ① an entity already on the Entity List renting controlled compute without a license; ② using shell companies to conceal the real beneficiary or submitting false documentation; ③ physically moving controlled chips back to China.
In plain terms = most of the time it is legal, but the moment identity fraud or hardware movement is involved, the grey zone becomes a criminal red line. That is why "train offshore, run inference onshore" has become the pragmatic compromise for many Chinese AI firms.
How fast is the Southeast Asian compute market growing?
Contracted data-center capacity in Malaysia's Johor state surged from under 400 MW at end-2023 to over 2.3 GW — nearly a six-fold jump in under three years.
Nvidia's SEC filings confirm the trend: Singapore-billed revenue rose from roughly 15% of Nvidia's total (about $2.7 billion) at end-2023 to a steady 22%.
Nvidia noted that over 99% of chips settled through its Singapore entity did not stay locally but were distributed by large cloud providers across the region. This means → Singapore serves as a billing hub, not a final compute destination — Nvidia uses this distinction to distance itself from any allegation of irregular chip diversion to China.
How is the U.S. planning to close this loophole?
On the legislative track, Congress is advancing the Secure Remote Access Act (H.R. 2683), which would legally equate remote cloud-compute access with physical chip exports, granting BIS full regulatory authority. The bill passed the House 369–22 and is now before the Senate, not yet enacted.
On the executive track, the Commerce Department is requiring overseas data centers running U.S.-made chips to implement strict KYC (Know Your Customer) checks, verifying the actual beneficial owner of each tenant.
This means → once the Senate passes the bill, the grey-area logic of "if the chip didn't leave the country it's not an export" will be closed entirely — remote leasing and physical shipment will carry the same legal weight.
Does China have enough compute of its own?
Ministry of Industry and Information Technology data from July 20: China's intelligent-compute capacity reached 2,185 EFLOPS as of end-June 2026.
But industry experts flag a structural imbalance — mid- and low-end compute is abundant, while the high-end capacity needed to train the largest models remains in chronic shortage.
In plain terms = the headline number looks strong, but the top-tier GPUs that can handle trillion-parameter training runs are still not enough. That is exactly why some firms turn to offshore compute in the short term — not because they prefer foreign chips, but because domestic high-end capacity cannot yet support the most cutting-edge training workloads. As the Secure Remote Access Act advances through the Senate, how long this offshore channel stays open is the pivotal variable for China's AI compute strategy.
Content is for reference only, not financial advice.