VFF - The signal in the noise
NewsTrending

Chinese AI Startup Moonshot Trained K3 on Restricted Nvidia Chips

Read original
Share
Chinese AI Startup Moonshot Trained K3 on Restricted Nvidia Chips

Beijing-based AI startup Moonshot has trained its Kimi K3 model, the world's largest open-source model with 2.8 trillion parameters, using Nvidia's advanced Blackwell chips despite U.S. export restrictions on such technology to Chinese firms. The company is now seeking additional Blackwell chips to develop Kimi K4, a significantly larger successor model. The situation highlights the tension between U.S. chip export controls and Chinese AI development capabilities.

  • Moonshot AI trained Kimi K3 (2.8 trillion parameters) on Nvidia Blackwell chips, circumventing U.S. export restrictions
  • The startup is actively seeking more Blackwell chips to build Kimi K4, a larger next-generation model
  • Kimi K3 is the world's largest open-source model, confirming statements by White House official Michael Kratsios
  • U.S. rules prohibit advanced Nvidia chips from being sold to Chinese firms, yet Moonshot obtained them

This case demonstrates that U.S. export controls on advanced AI chips may not be effectively preventing Chinese companies from accessing cutting-edge hardware needed for frontier model development. Moonshot's success in obtaining and deploying Blackwell chips suggests enforcement gaps or workarounds in the restrictions, raising questions about the efficacy of current policy tools designed to maintain U.S. technological advantage in AI.

For AI infrastructure and chip vendors, this signals ongoing demand from Chinese AI labs for advanced processors despite regulatory barriers. For investors and competitors, it indicates that Chinese AI startups can achieve world-class model scale and performance, potentially reshaping the competitive landscape in generative AI development.

  • U.S. chip export restrictions may not be achieving their intended effect of limiting Chinese AI development capabilities
  • Moonshot's ability to train a 2.8 trillion parameter model suggests Chinese firms have found ways to access restricted hardware
  • The demand for Blackwell chips from Chinese startups will likely persist, creating ongoing pressure on enforcement mechanisms
  • Kimi K3's scale and open-source release could accelerate AI capability development globally, including in regions subject to U.S. restrictions

Monitor whether U.S. authorities take enforcement action against Moonshot or investigate how the company obtained Blackwell chips. Track whether additional Chinese AI startups publicly acknowledge using restricted Nvidia hardware, and watch for any policy responses or tightening of export control mechanisms. Also observe Kimi K4's development timeline and whether Moonshot can secure the additional chips needed for training.

Share

Subscribe to the newsletter

The latest stories and analysis, delivered to your inbox.

Free. No spam. Unsubscribe any time.

Related stories

Snowflake launches agent governance layer to control enterprise AI costs
Model Release

Snowflake launches agent governance layer to control enterprise AI costs

Snowflake launched Cortex AI Gateway, a centralized control layer for governing how AI agents access enterprise data and tools, alongside security integrations with 1Password, Aembit, Linx Security, SailPoint, and Saviynt. The platform addresses a fundamental security gap: traditional enterprise security assumes humans are the actors, but AI agents operating at machine speed can exploit permission gaps and amplify existing risks. Snowflake positions itself as the control plane that decides what agents can do with enterprise data, rather than allowing each vendor to build closed ecosystems.

by michael.nunez@venturebeat.com (Michael Nuñez)· VentureBeat AI
Microsoft Bets on Cheap, Specialized AI Over Frontier Models
TrendingNews

Microsoft Bets on Cheap, Specialized AI Over Frontier Models

Microsoft unveiled MAI-Cyber-1-Flash, a custom-built AI security model, and Project Perception, an agentic defense platform designed to automate vulnerability detection and remediation. The system scores 96% on the CyberGym benchmark while cutting costs roughly in half compared to Microsoft's current production setup. The architecture routes 90% of security tasks to the smaller, cheaper model and escalates the remaining 10% to OpenAI's GPT-5.4, reflecting Microsoft's strategy to compete on cost efficiency rather than raw model size.

by michael.nunez@venturebeat.com (Michael Nuñez)· VentureBeat AI
Enterprise AI Agents Need Context, Not Just Models

Enterprise AI Agents Need Context, Not Just Models

At VB Transform 2026, SAP's Max McPhee outlined how enterprises can move beyond chatbots to autonomous AI agents by grounding them in company-specific context through knowledge graphs and governance controls. The key difference between assistants and true agents lies in providing enterprise context rather than relying on general knowledge, combined with identity and permission controls that prevent agents from circumventing access restrictions. SAP's recent acquisitions of LeanIX and Signavio, plus investment in n8n, are designed to help agents navigate complex, multi-system enterprise landscapes where SAP represents only a portion of the technology stack.

· VentureBeat AI
AI Industry Faces New Threat: Autonomous Agent Cyberattacks
TrendingNews

AI Industry Faces New Threat: Autonomous Agent Cyberattacks

Hugging Face CEO Clement Delangue called for 'radical transparency' in response to what he described as the first autonomous agent cyberattack targeting OpenAI, which he characterized as an 'unprecedented event' requiring an 'unprecedented response.' The statement signals growing concern within the AI industry about security vulnerabilities as autonomous systems become more capable. Details about the nature of the attack, its scope, and OpenAI's response remain limited in available reporting.

by Anthony Ha· TechCrunch AI