According to a report by The Information, Moonshot is planning a significantly larger successor to its recently launched Kimi K3 model.
- The report said Moonshot trained its 2.8 trillion-parameter Kimi K3 model using Nvidia chips, including advanced Blackwell GPUs, breaching U.S. export restrictions.
- Moonshot is also said to have stitched together computing capacity from at least two Chinese cloud providers because no single supplier could provide enough Blackwell chips.
- White House official Michael Kratsios alleged last week that Moonshot accessed Nvidia GB300 systems in Thailand despite U.S. export controls.
Chinese artificial intelligence startup Moonshot AI is seeking access to more of Nvidia Corp.’s (NVDA) advanced Blackwell chips to train its next-generation Kimi K4 model.
According to a report by The Information citing people familiar with the matter, Moonshot is planning a significantly larger successor to its recently launched Kimi K3 model.
Nvidia shares edged up by 0.1% in Tuesday morning’s trade.
How Moonshot Trained K3
The report said Moonshot trained its 2.8 trillion-parameter Kimi K3 model using Nvidia chips, including advanced Blackwell GPUs, despite U.S. export restrictions that are intended to prevent Chinese companies from accessing the technology.
According to the report, the startup is now looking to secure additional Blackwell chips to support development of K4, underscoring its continued reliance on Nvidia’s cutting-edge hardware for training frontier AI models.
The report added that Moonshot stitched together computing capacity from at least two Chinese cloud providers because no single supplier could provide enough Blackwell chips. Engineers also optimized networking across the separate data centers to improve communication between the GPUs during training.
Moonshot Faces GPU Constraints
According to the report, Moonshot also relies heavily on Nvidia’s H20 chips to run Kimi K3 on its platform. The startup reportedly recommends that customers hosting the model connect at least 64 server chips, while surging demand forced it to stop accepting new subscriptions within 48 hours of the model’s launch because its GPU capacity was stretched.
The report added that Beijing has discouraged local companies from using Nvidia’s H20 chips in favor of domestically designed alternatives. However, the large AI servers needed to run frontier models using Chinese chips remain in limited supply, with customers reportedly facing delivery wait times of as long as six months.
Trump Official Previously Said Moonshot Breached NVDA Export Controls
The report comes a week after Michael Kratsios, Assistant to the President and the thirteenth Director of the White House Office of Science and Technology Policy, alleged in a post on X that Moonshot had acquired Nvidia GB300-equipped servers and accessed GB300 systems in Thailand to train its AI models despite U.S. export controls.
He also claimed the company developed a sophisticated platform to access U.S. AI models while avoiding detection. Kratsios further alleged that Moonshot used large-scale distillation of Anthropic’s models to develop Kimi K3.
While he said the U.S. supports legitimate AI distillation to create smaller and more efficient models, he argued that “large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.”
NVDA stock is up 5% year-to-date and 11% over the past 12 months. The S&P 500 ETF (SPY) is up 16% over the past 12 months, while the Invesco QQQ Trust (QQQ) is up 19%.
The Global X Artificial Intelligence & Technology ETF (AIQ) is up 27% during this period, while the iShares U.S. Technology ETF (IYW) is up 28%.
For updates and corrections, email newsroom[at]stocktwits[dot]com.<