What you need to know
- An X post by AI analyst claimed that the new DeepSeek AI is being powered by Huawei’s Ascend 910C chip.
- Although the model was first trained on Nvidia's H100, the company is now shifting gears to get a more domestic product.
- The X post also claims that DeepSeek might train its next AI model using 32,000 Huawei 910C chips.
DeepSeek AI is the new bot on the block these days, and this Chinese AI model is going head-to-head with other U.S.-based AI companies.
While understanding how this large language model is gaining its skills, in a recent X post, AI analyst Alexander Doria confirmed the chip that powers DeepSeek. Doria stated that DeepSeek’s R1 large-language model (LLM) was first trained using Nvidia’s H100 but now it fully functions on Huawei’s Ascend 910C chip to generate responses. (via Tech Advisor.)
Even though the 910C chip isn't as dominant as the Nvidia's H100, the Chinese company wants to keep the chip domestic instead of going with U.S. based cores, reducing their reliability on expensive chips. Presumably, DeepSeek is testing its LLM to operate as good as it did on Nvidia chips.
"910C are (slightly) less performant and, even more importantly, does not come yet with a good interconnect which is critical for training," Doria added.
This could potentially bridge the gap between companies needing expensive chips that could power their AI models, functioning on par with big tech AI. Additionally, Doria stated that DeepSeek might train its next AI model (V4) using 32,000 Huawei 910C chips. However, it remains to be seen how the Chinese chip maker will meet this demand, however 'chip independence' is something that the Chinese companies are working on, to help them become autonomous.
The DeepSeek R1 is a reasoning model that's built on the V3 large language model and is claimed to be developed at a fraction of the cost— under $6 million to train its model. When compared to U.S. brands like ChatGPT that shelled out millions of dollars to create almost the same AI experience.
"The next chip, the 920c, is aiming for B200 performance (the current Nvidia flagship)."
I feel this should be a much bigger story: DeepSeek has trained on Nvidia H800 but is running inference on the new home Chinese chips made by Huawei, the 910C. pic.twitter.com/6IAgQlQ3ouJanuary 28, 2025
Despite using a moderately powerful chip, DeepSeek's AI is outperforming U.S AI rivals like chat GPT AI model. For instance, DeepSeek V3, has become extremely efficient at complex tasks like coding and essay writing.
DeepSeek's impact has already started to take effect as Nvidia, took a massive hit on Monday, losing $593 billion in market value as tech stocks tanked, marking the biggest one-day loss any company has ever seen on Wall Street.
