Moonshot reportedly seeks more NVIDIA Blackwell GPUs for Kimi K4 despite U.S. export pressure
Moonshot AI appears to be pushing deeper into the global AI race, even as its strategy risks drawing fresh scrutiny from Washington. After the success of its newly released Kimi K3 model, the Chinese AI company is reportedly trying to secure access to a larger supply of NVIDIA Blackwell GPUs to train its next-generation Kimi K4 model.
The move could escalate tensions between China and the United States, especially at a time when the Trump administration is weighing tougher restrictions on advanced AI hardware and open-weight AI models from Chinese companies.
Moonshot recently introduced Kimi K3, an open-source AI model with 2.8 trillion parameters designed for frontier-scale intelligence. The model supports multimodal capabilities, features a context window of roughly 1 million tokens, and is built to deliver fast, competitive performance against leading AI systems.
One of Kimi K3’s most notable strengths is its efficiency. The model uses a sharply reduced KV cache and memory-saving optimizations that help lower the cost of inference on large-scale data center hardware. Its architecture distributes 896 experts across many GPUs, allowing it to handle demanding workloads more efficiently than many comparable models.
That efficiency has attracted significant attention. Some estimates suggest major AI customers could potentially save hundreds of millions of dollars in inference costs by shifting certain workloads away from more expensive closed models and toward Kimi K3.
However, Moonshot’s rapid rise has also placed it directly in the middle of the intensifying U.S.-China AI rivalry.
Moonshot has faced allegations that it used outputs from a leading Western AI model to help refine Kimi K3. Some U.S. officials have also raised concerns that Chinese AI developers may be using advanced NVIDIA GPUs outside China, including in countries with looser access restrictions, to train or improve large models while avoiding American export controls.
According to recent claims, Moonshot may have accessed NVIDIA GB300-class systems, including through overseas channels. These allegations remain politically sensitive because NVIDIA’s most advanced AI chips are central to U.S. efforts to limit China’s access to frontier computing power.
Still, the timeline raises questions. The model that Moonshot was accused of using for distillation reportedly became available again near the end of June, while Kimi K3 was released around mid-July. That short window would make it difficult to fully distill a model as large and architecturally distinctive as Kimi K3. A more plausible scenario is that Moonshot may have used external model outputs to fine-tune certain responses, rather than build the entire system from another model.
Now, Moonshot’s ambitions appear to be growing even larger.
Fresh reporting suggests the company is seeking access to a much bigger pool of NVIDIA Blackwell GPUs for Kimi K4, its next major AI model. If Kimi K4 surpasses Kimi K3’s 2.8 trillion parameters, it would require an enormous amount of compute power for training, testing, and optimization.
Moonshot is not relying solely on foreign chips. Its engineers reportedly used domestic AI hardware and connected multiple eight-chip servers across separate data centers in China to train Kimi K3. That demonstrates a growing ability among Chinese AI firms to work around hardware constraints through distributed infrastructure and system-level engineering.
Even so, domestic alternatives are still widely believed to trail NVIDIA’s most advanced GPUs in raw performance, software maturity, and training efficiency. For a model as ambitious as Kimi K4, access to Blackwell-class hardware could provide a major advantage.
This is why the situation is so sensitive. If Moonshot obtains high-end NVIDIA GPUs despite export controls, it could trigger a stronger response from the Trump administration. Washington has already been tightening restrictions on advanced semiconductors, and AI model access is becoming the next major front in the technology conflict.
The administration is reportedly considering whether to restrict or ban open-weight AI models from China. Open-weight models are AI systems that publicly release the trained numerical parameters, or weights, that determine how the model processes information and generates responses. These models are easier for developers, researchers, and companies to inspect, modify, and deploy.
Closed-weight models, by contrast, keep those parameters private. Many leading commercial AI systems use this approach to protect intellectual property, reduce misuse risks, and maintain centralized control.
Supporters of open-weight AI argue that the approach fuels innovation, lowers barriers for startups, improves transparency, and helps researchers build safer systems. Critics worry that powerful open-weight models could be adapted for harmful uses or strengthen geopolitical rivals.
Several major technology and AI companies have urged the U.S. government not to damage the open-weight AI ecosystem. Their position is that open AI development is critical for competitiveness, research, and long-term technological leadership.
Moonshot’s reported pursuit of more NVIDIA Blackwell GPUs for Kimi K4 highlights the central dilemma in the current AI race. Advanced models require vast computing resources, and the companies that can secure those resources are more likely to define the next generation of artificial intelligence.
For Moonshot, Kimi K4 could be a chance to prove that Chinese AI companies can compete at the frontier of large-scale model development. For the United States, it may become another test of whether export controls can meaningfully slow China’s AI progress.
Either way, the battle over GPUs, open-weight models, and AI infrastructure is becoming one of the most important technology stories of the year.
