Moonshot AI Reportedly Secures 20,000 NVIDIA Blackwell GPUs via Alibaba to Scale Kimi K3.

Moonshot AI Secures 20,000 NVIDIA Blackwell Chips via Alibaba to Power Kimi K3 Architecture

According to undisclosed industry sources, Chinese AI startup Moonshot AI the developer behind Kimi K3, currently recognized as one of the most powerful open-weights AI models has secured an agreement to acquire an additional 20,000 NVIDIA Blackwell GPUs mediated through e-commerce and cloud giant Alibaba.

Previous reports indicate that Moonshot AI relied on a cluster of approximately 20,000 NVIDIA Hopper-architecture chips sourced through entities in Southeast Asia (a region U.S. regulatory scrutiny previously identified as including Thailand). However, regarding past hardware access, Alibaba explicitly issued a public denial stating it had never provided Moonshot AI with access to NVIDIA H200 accelerators.

The relationship between Moonshot AI and Alibaba represents a classic tech industry dynamic of simultaneous competition and partnership (coopetition). While Alibaba serves as a major financial backer of Moonshot AI, its internal Qwen AI engineering team actively competes for model dominance. Despite Alibaba's recent launch of Qwen3.8-Max, benchmark evaluations show it still trails behind Moonshot's flagship Kimi K3 in specialized reasoning and coding capabilities.

Why is upgrading to NVIDIA's Blackwell architecture so critical for heavy-duty open models? The transition from Hopper (H100/H200) to Blackwell delivers a 3-5x increase in AI processing performance and significantly higher high-bandwidth memory (HBM) capacity. For high-parameter models like the Kimi K3, this hardware leap will alleviate multi-node network bottlenecks, enabling large-scale integrated expert architectures (MoE) to operate with drastically lower latency.

Due to stricter BIS (US Bureau of Industry and Security) export restrictions on direct GPU sales, leading Chinese AI labs are increasingly relying on third-party cloud data centers and intermediary channels in Southeast Asia to access high-end computing power for training new generations of LLMs.

The collaborative competition between Alibaba and large cloud provider Moonshot AI mitigates risk by funding cutting-edge AI startups while simultaneously building its own internal foundational models. Facilitating access to GPU computing for Moonshot AI allows Alibaba to leverage its massive cloud infrastructure and market share, even as its internal Qwen team races to bridge the performance gap.

 

Source: Yahoo! Finance 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

Moonshot AI Secures 20,000 NVIDIA Blackwell Chips via Alibaba to Power Kimi K3 Architecture

According to undisclosed industry sources, Chinese AI startup Moonshot AI the developer behind Kimi K3, currently recognized as one of the most powerful open-weights AI models has secured an agreement to acquire an additional 20,000 NVIDIA Blackwell GPUs mediated through e-commerce and cloud giant Alibaba.

Previous reports indicate that Moonshot AI relied on a cluster of approximately 20,000 NVIDIA Hopper-architecture chips sourced through entities in Southeast Asia (a region U.S. regulatory scrutiny previously identified as including Thailand). However, regarding past hardware access, Alibaba explicitly issued a public denial stating it had never provided Moonshot AI with access to NVIDIA H200 accelerators.

The relationship between Moonshot AI and Alibaba represents a classic tech industry dynamic of simultaneous competition and partnership (coopetition). While Alibaba serves as a major financial backer of Moonshot AI, its internal Qwen AI engineering team actively competes for model dominance. Despite Alibaba's recent launch of Qwen3.8-Max, benchmark evaluations show it still trails behind Moonshot's flagship Kimi K3 in specialized reasoning and coding capabilities.

Why is upgrading to NVIDIA's Blackwell architecture so critical for heavy-duty open models? The transition from Hopper (H100/H200) to Blackwell delivers a 3-5x increase in AI processing performance and significantly higher high-bandwidth memory (HBM) capacity. For high-parameter models like the Kimi K3, this hardware leap will alleviate multi-node network bottlenecks, enabling large-scale integrated expert architectures (MoE) to operate with drastically lower latency.

Due to stricter BIS (US Bureau of Industry and Security) export restrictions on direct GPU sales, leading Chinese AI labs are increasingly relying on third-party cloud data centers and intermediary channels in Southeast Asia to access high-end computing power for training new generations of LLMs.

The collaborative competition between Alibaba and large cloud provider Moonshot AI mitigates risk by funding cutting-edge AI startups while simultaneously building its own internal foundational models. Facilitating access to GPU computing for Moonshot AI allows Alibaba to leverage its massive cloud infrastructure and market share, even as its internal Qwen team races to bridge the performance gap.

 

Source: Yahoo! Finance