- White House OSTP Director Michael Kratsios alleged on July 22 that Moonshot AI acquired Nvidia GB300-equipped servers and accessed GB300 systems in Thailand to train Kimi K3, in violation of U.S. export controls [1] - Kimi K3, released on July 22, is a 2.8-trillion-parameter model with a 1-million-token context window, trained using Blackwell GPUs sourced from at least two Chinese cloud providers [2] - Kratsios also accused Moonshot of large-scale distillation of Anthropic's Fable model, citing a "sophisticated internal platform" designed to evade detection — Anthropic previously filed a complaint alleging 3.4 million fraudulent API exchanges [3] - Treasury Secretary Scott Bessent stated the U.S. has "the ability to sanction" foreign AI models, referencing watermarks of U.S. LLMs found in Chinese models [4] - Moonshot is now seeking additional Blackwell chips to train Kimi K4, per Bloomberg, even as enforcement pressure escalates
[5] The Trump administration publicly accused Chinese AI startup Moonshot AI of acquiring Nvidia's export-controlled GB300 chips and using them to train its flagship Kimi K3 model, marking the most specific White House allegation yet of Blackwell-generation hardware reaching a Chinese AI lab. Michael Kratsios, Director of the White House Office of Science and Technology Policy, posted on July 22 that Moonshot had "acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models" [1].
Kimi K3, released the same day, is a 2.8-trillion-parameter model with a 1-million-token context window. To assemble sufficient compute, Moonshot reportedly stitched together capacity from at least two Chinese cloud providers, as no single supplier could furnish enough Blackwell chips [2]. The GB300 is part of Nvidia's Blackwell generation, explicitly banned from export to Chinese entities under current U.S. semiconductor controls.
The allegations extend beyond hardware. Kratsios accused Moonshot of conducting "large-scale, covert industrial distillation" of Anthropic's Fable model, claiming the startup built a "sophisticated internal platform" that allowed it to "quickly switch between multiple methods of access to avoid detection" [1]. Anthropic had previously filed a complaint in February 2026 alleging Moonshot generated over 3.4 million fraudulent exchanges through distillation attacks on its models
.
[3]## The Allegations Kratsios's statement represented an unusual public disclosure by a senior White House official, naming a specific company and specific chip model. He drew a distinction between legitimate smaller-model distillation — a common practice in AI development — and what he characterized as "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology" [4].
The Thailand connection is particularly significant for the AI compute supply chain. By accessing GB300 systems hosted in a third country outside the direct reach of U.S. export restrictions, Moonshot allegedly exploited a geographic workaround that has long concerned policymakers. The arrangement raises questions about the effectiveness of chip-level export controls when compute can be accessed remotely through cloud infrastructure in intermediary jurisdictions [1].
Enforcement Signals #
Treasury Secretary Scott Bessent escalated the rhetoric, stating the U.S. has "the ability to sanction" foreign AI models and citing the discovery of "watermarks of our U.S. large language models in many of the Chinese models" [4]. The TechCrunch report noted that the administration is evaluating adding Chinese AI labs to the Entity List and considering bans on Chinese open-weight models for American businesses
.
[4]The enforcement trajectory has moved beyond rhetoric. Anthropic's February 2026 complaint alleged that Moonshot generated 3.4 million fraudulent exchanges through systematic distillation attacks, providing a concrete legal foundation for action [3]. If the administration escalates from public accusations to formal enforcement, cloud providers and data center operators in Southeast Asia hosting intermediary compute infrastructure could face substantial legal exposure
.
[6]## Kimi K3 and the Compute Demand Kimi K3's scale — 2.8 trillion parameters — places it among the largest models publicly disclosed by any lab, Chinese or American. Following its release, demand proved so intense that Moonshot stopped accepting new subscriptions within 48 hours because its GPU capacity was stretched thin [2]. The company also relies heavily on Nvidia's H20 chips, a lower-performance chip that remains legal for export to China, for platform inference operations.
Moonshot has committed to releasing Kimi K3's full model weights publicly, with community distribution partners including vLLM, Nvidia, and AMD [6]. The open-weight release makes the model globally available regardless of any future sanctions — a dynamic that underscores the limits of hardware-focused export controls once a model is trained and published.
The Race for More Blackwell #
Even as the White House accusations mount, Moonshot is actively seeking additional Nvidia Blackwell chips to train its successor model, Kimi K4, which is expected to be significantly larger than K3, according to Bloomberg, citing a report from The Information [5]. The pursuit of more restricted compute — a week after being publicly accused of violating export controls — signals either confidence that enforcement will remain limited or desperation to maintain competitive parity with U.S. frontier labs.
Nvidia shares were essentially flat following the reports, edging up 0.1% in Tuesday morning trading [2]. The muted market reaction suggests investors view the diversion as immaterial to Nvidia's revenue, though it highlights a persistent gap between the intent of export controls and their enforcement on the ground.
Implications for the Compute Supply Chain #
The Moonshot case crystallizes three unresolved tensions in the AI export control regime. First, chip-level restrictions become porous when cloud access through third countries provides an alternative pathway to the same compute. Thailand, which has emerged as a growing data center market, is now under scrutiny as a potential conduit for restricted hardware [1].
Second, model distillation — extracting capability from a frontier model through systematic API queries — represents an entirely different vector of technology transfer that hardware controls cannot address. The administration's simultaneous focus on both GB300 acquisition and Fable distillation reflects the multi-front nature of the challenge [4].
Third, the open-weight release of Kimi K3 demonstrates that once a model is trained, the resulting intellectual property proliferates beyond the reach of any single government's control. This dynamic may accelerate policy conversations about restricting not just chips and cloud access, but also the distribution of model weights themselves [6].
Companies mentioned #
Further sources #
[[1] Michael Kratsios (@mkratsios47) on X, July 22, 2026 — White House OSTP statemen… ↗](https://x.com/mkratsios47/status/2079933645888880708)
[[2] Yahoo Finance / Benzinga — China's Moonshot Reportedly Seeks Access To More NVD… ↗](https://finance.yahoo.com/technology/ai/articles/chinas-moonshot-reportedly-seeks-access-160419591.html)
[[3] Glitchwire — Trump Administration Accuses Moonshot AI of Distilling Anthropic's… ↗](https://glitchwire.com/news/trump-administration-accuses-moonshot-ai-of-distilling-anthropics-fable-escalati/)
[[4] TFTC — White House: Moonshot AI Accessed Banned Nvidia GB300 Chips to Build Kim… ↗](https://www.tftc.io/moonshot-ai-banned-nvidia-gb300-chips-kimi-k3-export-controls)
[5] Bloomberg — Moonshot AI Seeks More Nvidia Blackwell Chips for Next Model, Repor… ↗
[6] CryptoBriefing — Moonshot AI hunts for more Nvidia Blackwell chips to train its… ↗ The stories that matter, in one email. Free — unsubscribe anytime.