Moonshot AI cancels $150 million GPU rental after US closes cloud loophole
Chinese AI companies had been getting around the US export ban on Nvidia chips by renting them in data centers in Singapore, Malaysia and Indonesia. Now the US Commerce Department requires a license from roughly 300 chips, while AI labs typically need thousands. As a result, Moonshot AI cancelled a $150 million contract to rent a B300 cluster for training Kimi K4. If you rent GPUs in Asia, check with your provider about licensing.
AI-processed from The Decoder; edited by Hamidun News
Chinese companies could not buy Nvidia's most advanced chips directly because of US export restrictions, so they rented them in data centers in third countries: Singapore, Malaysia and Indonesia. This is what people call the cloud loophole: on paper, the buyer purchases a service, not hardware. The US has now closed it.
A license from the Bureau of Industry and Security (BIS) at the Commerce Department is required starting at a threshold of roughly 300 advanced AI chips, while Chinese labs typically need several thousand to train a model. As a result, the Beijing startup Moonshot AI cancelled a $150 million contract with an Indonesian cloud provider. It had planned to rent an Nvidia Blackwell Ultra B300 cluster there to train its Kimi K4 model.
If you rent GPUs from cloud providers in Asia yourself, this story bears directly on your risk: a deal can fall through over licensing and customer screening.
What exactly was closed
The old scheme was simple. A Chinese company rented compute in a data center abroad and trained its models there. Under US rules, that counted as a cloud service, not a hardware shipment.
Now such data centers must obtain a BIS license. The agency controls exports of sensitive technologies. Existing operators got 90 days to transition, and throughout that period they must carefully screen customers for ties to China.
The license threshold is very low. According to the law firm Sidley Austin, it corresponds to roughly 300 advanced AI chips. For comparison, Chinese labs typically need several thousand processors to train large models. In other words, practically any serious cluster falls under the rule.
Who has already been hit
According to The Information, Moonshot AI cancelled the $150 million contract and paused talks on other deals. It had planned to rent the B300 cluster starting next year to train Kimi K4. Because of its enormous compute needs, the company had agreed to prepay about 50%, instead of the 15% that is customary in the industry.
Access to Blackwell chips abroad also mattered for the previous model, K3. According to the same report, K3 matched the best American models in July and became the trigger for a BIS investigation.
The pressure is coming from the money side too. Bernstein analysts led by Qingyuan Lin warn that training costs will rise. Bloomberg reports that five banks are considering pulling out of a $480 million loan to the Singapore-based cloud provider Aolani. The money was meant to fund deliveries of Nvidia chips in Malaysia for ByteDance, the owner of TikTok.
Physical shipments have stalled too
According to the Financial Times, TSMC and Samsung have halted some processor shipments to Chinese customers in recent days while they file license applications. The new rules also close a workaround through independent chip packaging and testing companies. Chinese developers used that route to order powerful processors without direct US approval.
Samsung customers reportedly affected include ByteDance and Kunlun, Baidu's AI data center unit. The exemptions for non-AI chips have also been scrapped. Licenses are now required even for advanced smartphone processors for Xiaomi, Oppo and Vivo. Industry insiders expect serious production delays.
Chinese companies already have it hard: memory prices are historically high, and domestic fabs such as SMIC lack sufficient capacity. Huawei recently advised customers to source memory themselves for its new AI chip.
Why Washington is pushing so hard
The tightening is attributed to a New York Times investigation into a scandal involving Alice Huang, the former girlfriend of ex-Nvidia chief Jensen Huang. According to the newspaper, Alibaba was the hidden main customer and financier of her Singapore company Megaspeed, which bought more than $4 billion worth of Nvidia chips. These claims rest on media reports and analysis by the firm Infrawatch, and Alibaba officially denies that its services were used for military purposes.
The newspaper also reports that the chips were used by Zhejiang Lab, a laboratory linked to military research. According to the NYT, researchers there worked on AI-based cyberattacks and automated vulnerability discovery.
What this means in practice
For Chinese labs, training large models will become more expensive and slower. They will have to rely on domestic chips, which have their own bottlenecks.
Other market players should take three things into account. First, cloud providers in Southeast Asia will screen customers more strictly. Second, licensing queues may delay hardware deliveries. Third, when planning large GPU contracts, you need to build in legal risk and a fallback option up front.
It is still unclear how quickly BIS will issue licenses and how many deals will survive the 90-day transition period.
- Nov 19, 2026Two pods share one A10G GPU on Amazon EKS: NVIDIA Run:ai gave each 8 GiB
- Nov 19, 2026Anthropic goes public and raises $81.3 billion, beating SpaceX's $75 billion record
- Nov 19, 2026Open DaisyRose-27B solves 12% of Terminal-Bench tasks, up from 5.6%
- Nov 19, 2026Reson8 Open-Sources a Speech Model 60 Times Smaller Than NVIDIA's: 7.86% Word Error Rate vs. 7.78%
Key Questions Answered
What exactly was closed?
The old scheme was simple. A Chinese company rented compute in a data center abroad and trained its models there. Under US rules, that counted as a cloud service, not a hardware shipment. Now such data centers must obtain a BIS license. The agency controls exports of sensitive technologies. Existing operators got 90 days to transition, and throughout that period they must carefully screen customers for ties to China. The license threshold is very low.
Who has already been hit?
According to The Information, Moonshot AI cancelled the $150 million contract and paused talks on other deals. It had planned to rent the B300 cluster starting next year to train Kimi K4. Because of its enormous compute needs, the company had agreed to prepay about 50%, instead of the 15% that is customary in the industry. Access to Blackwell chips abroad also mattered for the previous model, K3.
What are the key details on Physical shipments have stalled too?
According to the Financial Times, TSMC and Samsung have halted some processor shipments to Chinese customers in recent days while they file license applications. The new rules also close a workaround through independent chip packaging and testing companies. Chinese developers used that route to order powerful processors without direct US approval. Samsung customers reportedly affected include ByteDance and Kunlun, Baidu's AI data center unit. The exemptions for non-AI chips have also been scrapped.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.