Stock Markets July 28, 2026 09:34 AM

Moonshot AI Weighs Larger K4 Model and Seeks More Nvidia Blackwell Chips Amid Export Limits

Beijing startup plans a successor to its 2.8 trillion-parameter K3 but faces U.S. chip restrictions and data transfer constraints

By Jordan Park
Share
Twitter Reddit Facebook LinkedIn
NVDA

Moonshot AI is reportedly exploring development of Kimi K4, a model larger than its current K3, and is seeking additional Nvidia Blackwell-class chips to train it. Sources say K3 was trained using Nvidia hardware, and there is disagreement over whether all training occurred outside China. U.S. export controls and Chinese cross-border data rules complicate access to advanced chips and the movement of large training datasets.

Moonshot AI Weighs Larger K4 Model and Seeks More Nvidia Blackwell Chips Amid Export Limits
NVDA
Summarize with
ChatGPT Perplexity Claude Grok Gemini

Key Points

  • Moonshot AI is reportedly discussing development of Kimi K4, a model larger than its current K3.
  • Kimi K3, an open-source model with 2.8 trillion parameters, was trained using Nvidia chips including Blackwell-class processors, according to multiple sources.
  • The company is seeking additional Blackwell chips to prepare for K4, but U.S. export restrictions and Chinese cross-border data rules complicate access and dataset relocation.

Overview

Moonshot AI is in talks to build Kimi K4, a proposed next-generation model larger than its existing K3, according to people with direct knowledge of the matter. The Beijing-based startup would need more advanced Nvidia AI chips to train the new model, even though U.S. regulations limit the supply of that technology to Chinese firms.

Training of Kimi K3

People familiar with Moonshot's operations said the company trained Kimi K3, the largest open-source model to date with 2.8 trillion parameters, using Nvidia processors, including the most advanced Blackwell series. Three sources provided that information, which aligns in part with a social media statement by Michael Kratsios, a White House senior official.

Plans for K4 and chip access

Two of the sources said Moonshot is actively seeking access to additional Blackwell chips as it prepares for potential future models such as K4. Securing more of that class of hardware would be a prerequisite for training a model larger than K3, according to the people briefed on the company's plans.

Location of training and regulatory frictions

There are differing accounts of where K3 training took place. Kratsios stated that the model was trained at a data center in Thailand, a location where Chinese access to Nvidia chips might comply with U.S. rules. However, one of the three sources said some training occurred partly in China. That person noted that Chinese regulations on cross-border data transfers make it difficult to relocate the massive datasets required for pre-training to an overseas data center.

Constraints and context

Moonshot’s pursuit of more advanced Nvidia chips intersects with U.S. export controls that restrict the flow of certain high-end AI semiconductors to Chinese companies. At the same time, Chinese rules governing the transfer of large training datasets introduce logistical hurdles for moving pre-training data abroad, according to a source.

What remains uncertain

Sources differ on the exact geography of K3’s training and on how Moonshot will resolve both chip access and data transfer challenges if it proceeds with K4. The company’s reported need for additional Blackwell chips and the competing regulatory regimes are the central constraints identified by people briefed on the matter.

Risks

  • Access to advanced Nvidia Blackwell chips may be curtailed by U.S. export controls, affecting AI model development - impacts semiconductor suppliers and AI infrastructure markets.
  • Chinese regulations on cross-border data transfer could hinder relocating the large datasets needed for pre-training to overseas facilities, complicating model training logistics - impacts cloud/data center operators and companies relying on cross-border compute.
  • Discrepancies about where training occurred create uncertainty around compliance and operational strategy, which may affect partnerships and resource planning - impacts AI startups and international data center operators.

More from Stock Markets

X Challenges Australia's Teen Account Ban, Says Proposed Enforcement Powers Clash With International Law Jul 28, 2026 Wall Street Futures Slip Ahead of Fed Decision and Big Tech Earnings as Middle East Tensions Flare Jul 28, 2026 TSMC Gradually Restarts Japan Fab After Kumamoto Quake; Inspections Underway Jul 28, 2026 Takaichi Signals Temporary Cut to Japan's Reduced Food Consumption Tax, Yomiuri Reports Jul 28, 2026 Security Council to Begin Closed-Door Straw Polls on Thursday to Narrow UN Secretary-General Field Jul 28, 2026