White House accuses Moonshot AI of evading Nvidia chip export controls following the Kimi K3 release, escalating enforcement action.

A model hailed as a breakthrough in efficient AI training just days after its launch has become the focal point of an escalating U.S.-China technology dispute.

Michael Kratsios, Director of the White House Office of Science and Technology Policy, publicly accused Chinese AI company Moonshot AI of accessing banned Nvidia chips to train Kimi K3, its 2.8-trillion-parameter open-weight model. According to the allegation, Moonshot avoided directly importing restricted Nvidia GB300 chips into China. Instead, the company reportedly accessed GB300-equipped servers through infrastructure in Thailand, training on the hardware remotely without ever bringing the physical chips across a restricted border.

Kratsios further alleged that Moonshot operated an industrial-scale covert distillation campaign against Anthropic's Fable model, using methods designed to rotate access patterns and evade detection.

The disclosure sent Nvidia shares lower as investors assessed implications for export-control enforcement.

Kimi K3's July 16 release as an open-weight model—meaning its trained parameters are already downloaded, mirrored, and running globally—creates a policy complication. U.S. regulators cannot recall model weights already in the public domain, even if they pursue sanctions against Moonshot. That fact constrains how effective any policy response can be.

This marks the second major disruption from a low-cost Chinese model in 18 months. DeepSeek's R1 did similar damage in January 2025, erasing roughly $600 billion from Nvidia's market value in a single day. The critical difference here is the explicit allegation of sanctions evasion rather than merely efficient training—a distinction that raises stakes for export-control policy and for any company depending on current enforcement to protect its position.

The moment is sharpened by Nvidia CEO Jensen Huang's recent public arguments against restricting open-weight models, exposing tension between Nvidia's commercial interest in maximizing chip demand and Washington's national-security priorities.

Analysts debate how much of Kimi K3's capability gains stem from distillation versus fundamental efficiency research. Moonshot has previously stated that export controls pushed it toward training efficiency over raw compute scaling—a narrative predating current allegations that complicates easy conclusions about the model's origins.

Some in the industry have also highlighted the political dimension: U.S. companies openly considering Chinese open-weight models is itself noteworthy, and reflects how competitive Kimi K3's pricing and performance have become.