Moonshot AI's Kimi K3 Model Trained on Restricted Nvidia Champs
Moonshot AI's Kimi K3 model has caused controversy after it was revealed to have been trained on a massive 20,000-Nvidia-GPU cluster provided by Alibaba. This has sparked concerns that Washington's chip controls are being breached, as the chips used in the cluster are restricted for use in Chinese companies.
The arrangement between Moonshot and Alibaba is significant because Alibaba is not only an investor in Moonshot but also its cloud provider and compute supplier. This has raised questions about how effectively export controls can be enforced when a single company has such deep involvement with both the lab and the infrastructure it uses.
Washington's allegations against Moonshot include claims that the company accessed Nvidia GB300-equipped servers, including those in Thailand, to train its Kimi K3 model. The GB300 is part of Nvidia's Blackwell generation, which US rules are meant to restrict from use by Chinese companies.
The controversy has not stopped investors from circling Moonshot, with reports suggesting that the company is preparing for a final pre-IPO fundraising round that could value it at as much as $50 billion.