Update: White House says Moonshot used banned GB300s for Kimi K3

An export-control charge now stacks on the distillation claim, raising the odds Kimi K3's maker gets fenced off — and the case to keep your stack swappable.

Nowline JUL 23 8:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The new charge: banned GB300s, routed through Thailand

    White House OSTP director Michael Kratsios (Jul 22) says Moonshot trained Kimi K3 on Nvidia GB300 servers accessed in Thailand — an export-control end-run stacked on the earlier claim it distilled Anthropic's Fable. That's two independent legal exposures now, not one.

  • What flips for you: your access to K3

    If this lands Moonshot on the Entity List, US builders on Kimi K3's hosted API inherit compliance risk overnight, and the Jul 27 open-weights drop turns into mirror-it-now-or-never. If you depend on K3 (2.8T params, frontier-tier coding), pull the weights the hour they post.

  • How they're proving it: the model keeps calling itself Claude

    Redwood Research's Ryan Greenblatt published a cross-entropy analysis showing K3 self-identifies as 'Claude' far more than chance — suggestive of distillation, though web-scraped data is an innocent explanation. No access logs or watermarks released yet; the case is statistical, not forensic.

  • The hedge: keep provenance out of your architecture

    Sanctions are still 'on the table,' not signed — nothing's blocked today. But the move is model-agnostic plumbing: an abstraction layer that lets you swap K3 for DeepSeek V4 (MIT, open, ships Jul 24) without a rewrite. Treat provenance as a legal question, not a code dependency.