The Kimi model is not China's victory in the AI race: it's a reminder that without Nvidia, even the strongest Chinese model cannot function.
The internet was filled with celebrations around Moonshot's Kimi after the model showed impressive performance against American competitors.
Many were quick to declare that China no longer needs the US in the AI race.
Sorry to spoil the celebration for the Chinese,
but we have some updates on the matter.
According to reports, the Kimi model was trained using a cluster of approximately 20,000 Nvidia chips, obtained through Alibaba.
Sources we obtained claim these are H200 chips, among Nvidia's most advanced chips for the current generation.
Alibaba, of course, denied providing H200-based computing power, but did not deny providing Moonshot access to approximately 20,000 Nvidia chips,
This means the current debate is not whether there were Nvidia chips, but exactly which model was used and where they are physically located.
And here the story gets really interesting:
According to information we obtained, Kimi's parent company, Moonshot, has already managed to open an access channel to Nvidia's Blackwell processors through Southeast Asia.
Here the big question arises:
If this involves renting computing power through cloud services, there might not be a violation of US export restrictions.
If this involves a direct purchase of the chips, the story could become a significant regulatory problem for the United States.
A US government official noted that one of the transit points in the process is Thailand, and estimated that those chips were used to train the Kimi K3 model.
Additionally, Moonshot is already looking to further increase Blackwell-based computing power to train the next generation of its model.
The implication for investors is clear: despite the impressive progress of AI companies in China, Nvidia is still at the center of the value chain.
US export restrictions may make access to chips difficult, but according to reports, they are not succeeding in blocking it completely.
If alternative supply routes indeed exist through third countries or via cloud services, the United States may further tighten enforcement and oversight of the supply chain.
For Nvidia, this is another reminder that demand for its advanced chips continues to be enormous, even from customers who cannot purchase them directly.
The battle has long ceased to be between AI models,
the real battle is over who holds the chips that enable their training.
And whoever holds the pickaxe during the gold rush usually doesn't run out of work.