← Back to feed

The US Bureau of Industry and Security is reviewing whether Chinese AI companies can legally reach restricted Nvidia hardware by renting it remotely from overseas data centres, Bloomberg reported on August 7. The investigation is running in parallel to existing enforcement against physical chip smuggling, which is already covered by US export controls.

The legal distinction matters. Export controls are written around hardware moving across borders. When an H100 or B300 server sits in a Singapore colocation facility and a Chinese AI lab rents API-based access to it over the internet, the chip never enters China. Under current rules, that transaction may not fall within BIS’s jurisdiction.

The Two Paths BIS Is Mapping

Physical smuggling: Restricted Nvidia GPUs transported into China, either directly or through third countries. This is already a violation — the hardware itself crosses into a prohibited destination. BIS and allied enforcement agencies have been tracking this route since the first major export controls in 2022. The Nvidia B300 grey-market pricing that emerged in early 2026 — servers trading at 82% premium inside China — reflects how real this channel is.

Remote cloud access: Chinese AI firms rent compute from hyperscalers, neoclouds, or private operators in countries without GPU restrictions. The chips stay in those countries. The Chinese firm accesses them over HTTPS the same way any customer would. Current US export rules do not clearly address this pattern.

Bloomberg’s reporting indicates BIS enforcement is now specifically mapping which overseas providers are serving Chinese AI customers at scale, and what the regulatory basis would be to restrict that access.

Why This Matters Now

The timing follows a string of disclosures about Chinese AI capability advances that have surprised US analysts. DeepSeek R2 ran a 32B dense model capable of 92.7% AIME on a single RTX 4090. DeepSeek V4-Flash cleared 82.7% on Terminal-Bench 2.1, within 6 points of OpenAI’s best public score, using Huawei Ascend hardware domestically. If Chinese labs are supplementing domestic Ascend compute with offshore Nvidia access, the export controls may be preserving the appearance of a capability gap rather than its reality.

The ATOM report published earlier this year put China ahead of the US in open-weight AI downloads — 1.15 billion to 723 million — with Qwen models driving most of the gap. Chinese AI is not being held back by the absence of frontier capability. The question BIS is now asking is whether restricted Nvidia hardware is reaching Chinese AI workloads through a path the original controls did not contemplate.

What a New Rule Would Look Like

Blocking remote cloud access to restricted hardware would be a significant extension of US export control jurisdiction. It would require either rules that restrict where GPU-equipped servers can accept connections from, or secondary sanctions on cloud providers that serve Chinese AI customers without verification. Both approaches would be technically complex and diplomatically contested.

The US has already tested one version of this logic with the AI Diffusion Rule, which stratified country-level access to advanced chips and AI model weights. Extending that framework to include remote compute rental would represent its next major iteration.

BIS has not published formal findings. The review is ongoing.

Key Data Points

• Rule gap: US export controls restrict hardware movement, not remote API access to hardware overseas
• Hardware in question: Nvidia H100, H200, B200, B300 — restricted under EAR
• DeepSeek V4-Flash Terminal-Bench 2.1: 82.7% (August 2026), Huawei Ascend-native
• Prior precedent: AI Diffusion Rule stratified country access to chips and weights; cloud compute not yet addressed
• Bloomberg report date: August 7, 2026