Moonshot AI, the Chinese lab behind the open-weight Kimi model line, wants a bigger allocation of Nvidia Blackwell chips to train its next flagship, Kimi K4. That is the claim in a new report from The Information, picked up by Wccftech on July 28. The timing matters: the Trump administration is currently weighing whether Chinese-built open-weight models should be barred from the US market altogether. Wccftech’s own headline casts the GPU search as Moonshot “openly defying” Washington, yet the reporting behind that framing does not point to a specific export rule that requesting more chips would break.
Some background explains the appetite for more compute. Moonshot shipped Kimi K3 in mid-July: an open, multimodal system built on roughly 2.8 trillion parameters with a context window near one million tokens. The architecture spreads 896 experts across a wide GPU cluster and trims per-token memory use through a smaller key-value cache. Microsoft’s own estimate puts the potential savings from swapping K3 in for the OpenAI and Anthropic models currently running Copilot at close to $600 million in inference spend, a figure that helps explain why a larger successor model is drawing this much attention.
A separate and more specific set of allegations predates this story. A Trump administration official has claimed Moonshot secretly holds Nvidia GB300 servers and reached additional GB300 capacity through Thailand, while Anthropic has accused the lab of distilling K3 from its own Fable model. Wccftech’s timeline argument cuts against a full distillation claim: Fable came off the ban list in late June, and K3 launched only weeks afterward, arguably too fast for Moonshot to have absorbed an entire competing architecture, though fine-tuning some outputs against Fable remains plausible. Those GB300 and Thailand claims describe conduct that plausibly brushes against export rules. The new reporting about K4 does not; it describes a lab shopping for more legitimate compute, not a confirmed workaround.
Nvidia hardware is not Moonshot’s only supply line. To train K3, its engineers strung together eight-GPU server nodes across several data centers inside China, relying on domestic chips for a meaningful share of the compute. That a lab already running production training on homegrown silicon is still chasing more Nvidia Blackwell capacity says less about defiance and more about the compute gap facing any model larger than 2.8 trillion parameters. Wanting more chips is a routine ambition for a lab racing to ship a bigger successor. Whether the specific channel Moonshot is reportedly pursuing crosses an export-control line is the question this particular report leaves open.
The policy backdrop adds pressure. Dell, Perplexity, and Palantir put their names on a joint letter asking the Trump administration to keep the open-weight ecosystem alive rather than close it to Chinese labs, and Meta, Microsoft, Mistral AI, and chipmaker Nvidia signed on as well. That letter is a direct response to the same possibility driving this story: a blanket prohibition on Chinese open-weight releases. The dispute sits inside the same export-control debate that produced yesterday’s detention of a suspect accused of smuggling advanced chips out of Taiwan, a reminder that Washington’s enforcement net now covers both criminal smuggling cases and corporate compute-acquisition disclosures.
For operators running production workloads on Kimi K3, or evaluating K4 once it ships, the near-term risk is not capability. It is regulatory: if Washington moves toward a ban on Chinese open-weight models, any pipeline built on Kimi today needs a fallback vendor lined up before a rule change forces an emergency migration.
Wccftech (Rohail Saleem) reported this July 28, 2026, citing original reporting from The Information on Moonshot’s pursuit of additional Nvidia Blackwell GPUs.
