Just in case you weren’t online the past couple of days, DeepSeek, a Chinese AI startup made waves in the international tech scene for having capabilities of top American AI chatbots such as ChatGPT, while being powered by much cheaper Nvidia H800 chips. This triggered a “panic” reaction among Nvidia investors, resulting in many of them pulling out due to concerns about Nvidia’s future.
What must have sent even more chills down the American AI companies’ spines is that DeepSeek is using Huawei’s Ascend 910C chipsets as its “brains” for its R1 large-language model (LLM), according to a leak by AI analyst, Alexander Doria. The Nvidia chips were used to train the AI in its initial development phase.
Huawei’s 910C chip is slower but better built for AI
So, why did DeepSeek opt for Huawei’s chips rather than Nvidia’s? Is it because the former offers more processing power than the latter? Well, not at all. Looking at their raw performance figures, the Huawei 910C chips are weaker than Nvidia’s H800, and more importantly, lack a good interconnect too, which is essential for training.
However, what makes the Huawei 910C shine is that it’s made with inference in mind. This theoretically allows DeepSeek to achieve similar results with other AI chatbot competitors with fewer chipsets, and more importantly, at lower costs.
DeepSeek’s choice to opt for Huawei’s chipsets rather than Nvidia’s might also be proof that Chinese AI companies like itself can operate without depending on Western-supplied components.
Huawei’s 910C chip can’t entirely replace Nvidia yet, but its successor might
At the end of the day, Huawei’s 910C chipsets still can’t fully replace Nvidia’s H100 in training DeepSeek’s AI models just yet. Still, DeepSeek is considering training its next-gen AI model, the DeepSeek V4 using 32,000 units of the Huawei 910C.
Huawei is also looking to compete squarely against Nvidia’s current top dog, the B200 Performance GPU, with its upcoming Ascend 920C chipset. It will be a game changer if DeepSeek is to utilise this chipset to train its future AI models at much lower costs.
