Moonshot Seeks More Nvidia Blackwell Chips to Train Kimi K4

The interesting thing about Moonshot's next move is the context around it. Alongside a live White House push to prove the Beijing-based startup used banned Nvidia hardware to train its current flagship, The Information reported that Moonshot is already lining up more Blackwell chips for a Kimi K4 that would be significantly larger than K3, according to sources cited in the report.

For scale, K3 is a 2.8-trillion-parameter open-weight model that Moonshot released last week, and the company has positioned it against the leading closed frontier systems. Going 'significantly larger' from there is not a minor step. It implies a training run that needs meaningfully more high-end accelerator time, which is exactly the resource US export controls are designed to keep out of Chinese hands.

The accusation half of this story is where you should hedge. Michael Kratsios, Director of the White House Office of Science and Technology Policy, said Moonshot acquired servers equipped with Nvidia GB300 chips through Thailand and covertly distilled Anthropic's Claude to boost K3, per CNBC. The Commerce Department's Bureau of Industry and Security has opened an investigation into whether Chinese firms are circumventing restrictions by routing workloads through third-country cloud providers. Moonshot's plans for K4 land against that live enforcement backdrop, and the reporting explicitly frames the K4 chip push as continuing despite the restrictions.

What the reporting doesn't give you is how Moonshot expects to actually get the chips this time, whether the alleged Thailand pathway is still open, or what a bigger-than-2.8T model would look like in compute budget, inference cost, or release timing. Treat the K4 plan as stated intent, not a training run already in progress, and take Kratsios's specifics as an allegation until the BIS probe produces more.

The upside for the rest of the field is straightforward. Each time a Chinese lab pushes an open-weight model to a new size class, the community running models locally gets more capable weights to work with. If K4 ships in a form similar to K3, developers get another frontier-adjacent model outside a hosted-only paywall, and the pressure on Western closed labs to justify their pricing climbs another notch.