NewsMacroMoonshot AI Seeks Nvidia Blackwell Chips for New Kimi K4 Model as U.S. Export Rules Tighten

Moonshot AI Seeks Nvidia Blackwell Chips for New Kimi K4 Model as U.S. Export Rules Tighten

Author: Coincentral·

Key Takeaways

  • Moonshot AI is discussing Kimi K4, which would be significantly larger than its 2.8 trillion-parameter Kimi K3.
  • The company is seeking more Nvidia Blackwell GPUs even though U.S. rules bar exporting those chips to China.
  • Kimi K3 was reported as trained at least in part in Thailand, while one source said some training also occurred inside China.
  • U.S. officials alleged Moonshot distilled Anthropic’s Fable model, and Scott Bessent said sanctions could follow if that is confirmed.
  • Moonshot has paused new Kimi K3 subscriptions because demand exceeded capacity and is reportedly seeking new funding at a potential $50 billion valuation.
Moonshot AI Seeks Nvidia Blackwell Chips for New Kimi K4 Model as U.S. Export Rules Tighten

Chinese AI startup Moonshot AI is planning a new, larger model called Kimi K4, according to a report from The Information. The company would need more Nvidia chips to train it — chips that U.S. rules say Chinese firms cannot have.

MOONSHOT SEEKS MORE $NVDA BLACKWELL CHIPS FOR LARGER KIMI K4 The Information reports Moonshot AI is discussing Kimi K4, which would be significantly larger than its 2.8 trillion-parameter Kimi K3, for which they are seeking access to more Nvidia Blackwell GPUs despite U.S.… pic.twitter.com/XA6mQYcy4M — Wall St Engine (@wallstengine) July 28, 2026

MOONSHOT SEEKS MORE $NVDA BLACKWELL CHIPS FOR LARGER KIMI K4

The Information reports Moonshot AI is discussing Kimi K4, which would be significantly larger than its 2.8 trillion-parameter Kimi K3, for which they are seeking access to more Nvidia Blackwell GPUs despite U.S.… pic.twitter.com/XA6mQYcy4M

— Wall St Engine (@wallstengine) July 28, 2026

Moonshot recently released Kimi K3, an open-source AI model with 2.8 trillion parameters, making it the largest open-weight model in the world at present.

To train K3, Moonshot used Nvidia Blackwell GPUs, including the advanced GB300 version. Those chips are not permitted to be exported to China under U.S. Department of Commerce rules, which has made access to advanced training hardware a central constraint for Chinese AI developers.

Michael Kratsios, the science adviser to President Donald Trump, said Moonshot trained K3 at a data center in Thailand. Under U.S. rules, Chinese companies may be allowed to access Nvidia chips through servers located in other countries.

However, one source told The Information that K3 was partly trained inside China. Chinese data transfer regulations make it difficult to move the large datasets required for pre-training to servers overseas.

Distillation allegations

Kratsios also accused Moonshot of distilling Anthropic’s Fable model to help develop K3. Distillation is a method in which a smaller or weaker model is trained using outputs from a more powerful one.

U.S. Treasury Secretary Scott Bessent said the Trump administration could sanction open-weight Chinese AI models if distillation from American models is confirmed.

Moonshot and Nvidia did not respond to requests for comment.

Other Chinese AI companies have also reportedly used Blackwell chips. Alibaba’s Qwen3.8-Max model is among those named in reports, underscoring how export restrictions have become part of a broader scramble among Chinese labs to secure access to frontier hardware.

Demand and funding

Moonshot recently had to pause new subscriptions to Kimi K3 after demand exceeded its capacity.

The company is also reportedly seeking to raise more money before going public. Reports place its potential valuation at $50 billion.

For Kimi K4, Moonshot is said to be seeking more Blackwell GPUs, and two sources told The Information that the company is actively working toward that goal. That effort comes as U.S. chip restrictions remain in place, even as Chinese AI firms continue to look for access to advanced Nvidia hardware through overseas infrastructure and other channels.

Nvidia CEO Jensen Huang recently signed a letter to the U.S. Congress supporting open AI models, writing that open models “strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.”

The chip restrictions remain in place, but reports suggest Chinese AI firms continue to find ways to access advanced Nvidia hardware.