The escalating technology war between the United States and China has found a new flashpoint in the cloud. Bloomberg reported that the White House Office of Science and Technology Policy (OSTP) has formally accused Chinese AI startup Moonshot of bypassing US export controls to train its flagship Kimi K3 model, alleging the company utilized a massive cluster of 20,000 restricted Nvidia chips hosted in Thailand through Alibaba Cloud.
The allegations highlight the growing difficulty of enforcing hardware embargoes in an era of globalized cloud computing. According to the OSTP, Moonshot accessed the GPU cluster remotely, allowing its engineers in Beijing to train Kimi K3, a 2.8-trillion-parameter model that has benchmarked competitively against top-tier US systems, without ever importing the chips into China. The case has sent shockwaves through Washington’s policy community, raising fundamental questions about whether export controls can remain an effective tool in an age of borderless cloud computing.
The Thailand Cloud Loophole
The core of the US accusation centers on how Moonshot accessed the compute. US export controls strictly prohibit the sale of advanced Nvidia chips, such as the H100 and the newer GB300 Blackwell architecture, to entities within China. However, the rules governing cloud access to these chips located outside of China have historically been murky.
Moonshot reportedly exploited this gray area by renting compute time on servers physically located in Thailand. Because the chips were never imported into China, the arrangement technically sidestepped direct import bans, while allowing the company’s engineers to train their models remotely. The entire Kimi infrastructure runs on this 20,000-chip cluster provisioned through Alibaba, which holds a 36% stake in Moonshot — a vertical integration that makes enforcement exceptionally challenging for US regulators.
This arrangement shifts the regulatory target from a simple hardware transaction to a complex web of corporate relationships and cross-border data flows. Alibaba Cloud’s role as both a major investor and the infrastructure provider for Moonshot creates a situation where the US government would need to act against one of China’s largest tech conglomerates to effectively shut down the pipeline.
The Distillation Accusation
The hardware allegations are only half of the story. The OSTP also accused Moonshot of conducting a “large-scale, covert industrial distillation” campaign against Anthropic’s Fable model. Distillation involves using outputs from a more advanced, proprietary model to train a smaller or competing model, a practice that violates the terms of service of most major US AI providers.
The White House claims Moonshot developed an internal platform to query Fable at scale, harvesting the data to accelerate the development of Kimi K3. Treasury Secretary Scott Bessent amplified the warning, stating that “when PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table.”
The timeline of these events has sparked debate within the AI research community. Anthropic briefly pulled Fable 5 from public access on July 1 due to export control concerns, while Kimi K3 launched on July 17. Some analysts question whether a 16-day window is sufficient to conduct industrial-scale distillation and integrate the results into a 2.8-trillion-parameter model, though the OSTP maintains its assessment.
Moonshot’s Defiant Trajectory
Despite the severe allegations from Washington, Moonshot’s financial trajectory remains unhindered. The company recently closed a $3.5 billion funding round at a $35 billion valuation, as covered in our earlier report on Moonshot’s record-breaking fundraise, and is reportedly already seeking new backers at a $50 billion pre-money valuation ahead of a planned Hong Kong IPO.
The startup’s commercial traction is equally robust, with annual recurring revenue reportedly hitting $300 million in June. Furthermore, Moonshot is actively seeking to acquire more Nvidia Blackwell chips for its next-generation models, signaling a clear intent to maintain its compute pipeline regardless of US regulatory pressure.
For Nvidia, the situation presents a profound dilemma. The company must navigate strict compliance with US export controls while watching Chinese firms build massive, purpose-built AI training facilities using its hardware via third-country cloud providers. As the US government weighs potential sanctions or Entity List designations, the Moonshot case serves as a stark reminder that in the global AI race, compute finds a way.
The case also raises broader questions about the enforceability of the entire US chip control regime. If a leading Chinese AI lab can train a frontier model on 20,000 restricted Nvidia chips simply by routing the compute through a cloud provider in a third country, the strategic value of export controls as a tool for maintaining US AI dominance is fundamentally called into question.
What the Moonshot case ultimately illustrates is the fundamental tension at the heart of the US export control strategy: hardware embargoes are a blunt instrument in a world of borderless cloud computing and deeply integrated corporate relationships. The US government is now faced with a choice between significantly tightening the rules around third-country cloud access, a move that would affect US cloud providers like AWS and Microsoft Azure operating in Asia, or accepting that the current framework has a critical loophole that China’s most ambitious AI companies are already exploiting at scale.