Deepseek’s plans to train a new AI model R2 on Huawei’s Ascend chip failed and forced Nvidia to retreat while delaying the launch.
The story, pushed out by Beijing for several months, was one of unstoppable technological advancements and marching to self-sufficiency. However, in reality, there is a habit of chewing. The recent troubles of the Chinese AI Darling Deepshake are examples of textbooks where ambitions meet the hard walls of technical limitations.
After the successful launch of the R1 model in January, Deepseek found himself under pressure from China. The message was clear, according to three people talking about The Financial Times. Use Huawei chips instead of Nvidia chips.
When it comes to actually training a new R2 model, sources say Deepseek encountered a “persistent technical problem” with Huawei’s AI chip. The problem was so basic that the project was stopped. Those who know the situation said this was the main reason the planned launch of the model was scrapped in May, and why they put the company behind them in a market where no one was waiting.
To understand why this is so big, you need to know the difference between AI training and reasoning. Training is a difficult part, such as sending students to university for years of intense learning. It requires a huge amount of power and stability. Reasoning is a relatively “easy” part, like asking questions to graduates.
Deepseek discovered that Huawei’s chips may be ready for the final exam, but they have not yet advanced to rigorous university courses. The company had no choice but to go back to Nvidia’s powerful system to train. Sources say the Deepseek team is trying to work with the R2 model with the Huawei chip in the less demanding inference phase.
The two confirmed that Huawei sent its own team of engineers to the Deepseek office to help the R2 model run on chips. However, even with the experts in the room, they were unable to make the training a success.
Talk to anyone in the industry and they will tell you this is not a big surprise. Huawei CEO Ren Zhengfei said earlier this year that the US is “exposing Huawei’s achievements” and “it’s not that great yet.”
However, Beijing is still actively promoting high-tech giants who prefer to support local hardware, and the Financial Times reports that companies must justify their orders for H20 chips that are compliant with Nvidia’s exports. This is part of the strategy of building domestic champions, but it can force companies to technically choose their choice.
Aside from the issues facing Huawei chips in training, Deepseek founder Liang Wenfeng reportedly told his team he was unhappy with the overall progress towards the R2 model. He is said to aim for them higher and push them to build something that can keep the company up among AI industry leaders.
All top-down directives and national pride still apply the laws of engineering. The Deepseek story reminds us that there are no shortcuts in a global race for AI hegemony. China has played a long game, but for now the Performance Crown remains firmly in Nvidia’s head.
reference: Ren Zhengfei: The Chinese AI Future and the Long Game of Huawei
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expo in Amsterdam, California and London. The comprehensive event will be held in collaboration with other major events, including the Intelligent Automation Conference, Blockx, Digital Transformation Week, and Cyber Security & Cloud Expo.
Check out other upcoming Enterprise Technology events and webinars with TechForge here.
