Moonshot AI has started planning K4 just after open-sourcing its 2.8-trillion-parameter Kimi K3 on July 27, with the Beijing startup now seeking more Nvidia Blackwell chips to train the larger next-generation model, according to BlockTempo.
The company released K3 model weights, a technical report, and key infrastructure at the same time. It also said K4 is already on the roadmap and will be "significantly larger" than K3.
No single provider can supply enough Blackwell capacity
BlockTempo reported that no Chinese cloud provider can currently offer enough Blackwell compute on its own. Moonshot’s solution has been to combine compute resources from at least two domestic cloud companies and have engineers optimize networking across data centers.
That setup is intended to keep training jobs running across different server rooms and even different cities without latency dragging down the entire process.
The report says this patchwork approach shows how export controls, rather than removing demand, have pushed companies toward more complicated engineering workarounds when a single supplier cannot meet requirements.
Inference demand is also straining available hardware
The same pressure is visible on the inference side. Under current controls, customers using H20 chips that can still be legally obtained need to connect at least 64 server chips to run Kimi K3, according to the report.
Getting 64 chips linked up and working together already carries meaningful hardware and networking costs. BlockTempo said this helps explain why Moonshot wants to assemble Blackwell capacity despite the compliance risks involved. The issue is not only electricity use or data center space, but whether the company can absorb the latest wave of demand and retain the users coming in.
U.S. official made allegations, but no technical evidence has been released
One week earlier, Michael Kratsios, director of the White House Office of Science and Technology Policy, wrote on X that the U.S. government had information indicating Moonshot distilled Claude Fable 5 to develop K3.
Kratsios also alleged that Moonshot had built a sophisticated internal platform that could quickly switch between multiple access methods to avoid detection, and that it had obtained GB300-equipped servers in Thailand.
BlockTempo noted that Kratsios has not published technical evidence supporting those claims. As presented in the report, the allegations remain a one-sided government claim rather than a confirmed investigative finding. The report added that such accusations could leave cloud providers and data center operators in Southeast Asia exposed to unexpected legal risk.
Experts have also questioned the distillation claim
The report said some experts dispute the timeline behind the allegation. Claude Fable 5 was only made public on July 1, leaving a three-week window. On that basis, they argue K3 could not have been built primarily by distilling Claude Fable 5 because the schedule would have been too tight from a technical standpoint.
Nvidia’s stance remains unchanged
As for Nvidia, the report said its position has not changed: the stronger the model, the greater the inference demand, and chip sales would rise rather than fall.
In BlockTempo’s framing, export controls may slow chip flows, but they do not stop demand itself.