"Access to advanced GPUs is key for us," a source close to Moonshot AI said, underscoring the firm’s push for Nvidia Blackwell chips to train its next-generation AI model. The startup plans to build Kimi K4, a much larger successor to its current 2.8 trillion-parameter Kimi K3, which already holds the record as the largest open-weight AI model globally.
The challenge: Nvidia’s most powerful Blackwell GPUs, including the GB300 version, are barred from export to China under strict U.S. Department of Commerce rules. Moonshot reportedly circumvented these limits before by training K3 partially at a data center in Thailand a tactic aligned with U.S. policies allowing Chinese companies to lease hardware located abroad. However, industry insiders reveal that some training may have still occurred inside China, a risky move considering tight data transfer regulations that restrict cross-border dataset movements necessary for AI development.
Further controversy surrounds Moonshot’s use of "model distillation" techniques allegedly involving Anthropic’s Fable, an established AI model. This process, where a smaller model learns from a more complex one, has drawn scrutiny from U.S. officials. Treasury Secretary Scott Bessent indicated that the U.S. might impose sanctions if evidence confirms that open-weight Chinese AI models like Moonshot’s are built leveraging American models without proper authorization.
Funding-wise, Moonshot is reportedly eyeing a fundraising round valuing the company at around $50 billion ahead of a potential public listing, signaling strong confidence in their tech despite geopolitical hurdles. Other Chinese firms, such as Alibaba with its Qwen3.8-Max model, have also been connected to Blackwell chip usage, reflecting the growing appetite for cutting-edge AI hardware amid ongoing export tensions.
This content is for informational purposes and not investment advice.
