The Information称月之暗面和阿里训练前沿模型仍依赖英伟达Blackwell芯片

Kratsios claimed Moonshot trained K3 at a data center in Thailand, where Chinese access to Nvidia chips might be allowed under U.S. rules. But Moonshot trained K3 partly in China, according to one of the three people. Chinese rules governing cross-border data transfer make it difficult to move the massive data required for pre-training to overseas data centers, the person said.

Other Chinese firms are also stepping up their open-source AI game—also with help from Nvidia chips. Just days after Moonshot released K3, Alibaba Group unveiled Qwen3.8-Max, its largest-ever model at 2.4 trillion parameters, saying its performance is comparable to that of top U.S. models. Alibaba trained Qwen3.8-Max on Nvidia chips including the Blackwell, according to two people with knowledge of the matter.

In December, The Information [reported](https://www.theinformation.com/articles/deepseek-using-banned-nvidia-chips-race-build-next-model?rc=jnr9wn) that DeepSeek was training its new AI model using Blackwell chips that had been smuggled into China. (Earlier this year, DeepSeek released the new model, V4.)

These episodes highlight the fact that while Chinese startups are disrupting Silicon Valley by developing highly capable open-source AI models, they still depend on U.S. technology to do so. Beijing has poured massive state funds into China’s chip industry for years, and the country is rapidly adopting domestic chips for inference, or the running of AI models. But when it comes to training AI models that aim to compete with Anthropic and OpenAI, Chinese firms still rely on Nvidia.

By using advanced Nvidia chips for training, Chinese AI firms are navigating a risky geopolitical environment. They’re highlighting that U.S. government export restrictions on Blackwell chips aren’t working as intended. That comes at a time, advances of Chinese open-source models and their use by U.S. businesses has raised concerns in the U.S. about the threat China poses in AI. Chinese open-source models are taking share from leading U.S. models, hurting their potential.

Moreover, some in the U.S. claim Chinese firms are training their models using U.S. tech. In his X post about Moonshot, Kratsios also accused the firm of using U.S. AI models’ output to train K3 in a practice known as distillation.

**New Arms Race**

Kimi K3, released on July 16, [sent shock waves](https://www.theinformation.com/newsletters/ai-agenda/new-kimi-k3-model-means-u-s-china-ai-race?rc=jnr9wn) through the global AI industry by demonstrating frontier-level coding capabilities and matching or even surpassing top U.S. models on some benchmarks.

“Kimi K3 has started a new round of arms race” among Chinese AI developers, said a researcher at a major Chinese tech firm who works on model training. This researcher and others working for Chinese AI firms say they need advanced Nvidia chips, which excel in stability and high-speed networking, to train huge powerful models with frontier-level capabilities. Training such models with domestic chips would be very difficult, if not impossible, they say.

To assemble enough computing power to train K3, Moonshot tapped at least two different Chinese cloud firms with Blackwell chips, according to the three people. (It isn’t clear how Chinese cloud firms got the Blackwell chips. Their access is likely in contravention of U.S. export rules.) No single supplier could offer enough chips, forcing Moonshot to connect many eight-chip Blackwell servers together. That’s a much smaller configuration than the biggest U.S. AI companies typically use for training AI models, according to one of the people.

Moonshot’s engineers refined the network design to improve how Blackwell chips communicated with each other and across data centers run by different providers.

**Inference Needs**

Moonshot also relies heavily on Nvidia chips such as H20—which U.S. rules allow Chinese startups to buy—to run K3 on the Kimi platform, according to three people with direct knowledge of the situation. Moonshot recommends cloud firms or other customers that host K3 on their servers to connect at least 64 server chips to run it.

The massive popularity of K3 strained Moonshot’s computing capacity, prompting the startup to stop accepting new subscriptions within 48 hours. “Our GPUs are feeling it,” the company said in an X post last week. Moonshot said it would reopen subscriptions as it adds more computing capacity.

Finding more resources to run the  K3 API for customers will take time. Beijing has discouraged local companies from using Nvidia’s H20 and wants them to use Chinese-designed AI chips. But the large servers needed to run models such as K3, which connect dozens or hundreds of chips designed by firms including Huawei,  are not widely available yet. Customers can wait as long as six months for delivery, according to the three people.

Many Chinese firms use offshore AI data centers with Nvidia chips to train models. Tencent, for example, rented computing power from a data center near Osaka that is equipped with thousands of Blackwell chips owned by Japan’s Datasection, according to [the Financial Times](https://www.ft.com/content/9b47c335-9633-4560-9f57-5736c9d04bef?syn-25a6b1a6=1). Such arrangements are not currently prohibited, but Washington is considering tighter controls on Chinese companies’ remote access to advanced U.S. chips. The Information [reported](https://www.theinformation.com/articles/u-s-investigates-chinese-ai-companies-access-chips-amid-moonshot-accusations?rc=jnr9wn) last week that the Bureau of Industry and Security, an agency within the U.S. Department of Commerce, is formally investigating whether Chinese firms like Moonshot are accessing advanced U.S. AI chips.

While the success of K3 prompted the Trump administration to ramp up its attacks on Chinese models, Nvidia CEO Jensen Huang said in an [interview](https://www.axios.com/2026/07/22/nvidia-jensen-huang-china-open-source-ai) with Axios that U.S. companies should “absolutely” be allowed to use Chinese open-source models like K3. Late last week, Huang made his [first-ever post](https://x.com/JensenHuang/status/2080643682408321103?s=20) on X to share a letter signed by Nvidia and 24 other companies emphasizing the importance of open-source models. “The world needs both frontier closed models and frontier open models,” he said.

So far, [the vast majority of American cloud firms, tech incumbents and startups favor allowing U.S. businesses to use Chinese open-weight models](https://www.theinformation.com/articles/silicon-valley-unites-anthropic-chinese-ai-restrictions?rc=c48ukx). A smaller number of firms, namely Anthropic and OpenAI, have advocated for some restrictions or regulations or otherwise discourage U.S. businesses from using such models over theoretical national security or cybersecurity risks.