According to the Tencent boss, AI doesn’t need as many GPUs

0
Overclocking.com video

If Nvidia is at the top of its game, it’s largely thanks to the market and industry buzz around AI. However, over the past few months, some people have expressed the view that the current hype is looking more and more like the kind of bubble we saw a few years ago. Deepseek’s thunderous arrival on the scene was the first blow to the current euphoria. Recently, at Nvidia’s GDC, it was Intel’s former boss who claimed that Nvidia was, in a way, pushing the consumption of AI. In the last few hours, James Mitchell, Tencent’s Chief Operating Officer, has said as much… and even more.

Tencent servers

Tencent is not just anyone, including in the field of AI. The company is already making massive use of these technologies in its key activities. In gaming, it trains new players and optimises online matches, while in advertising, it refines the analysis of consumer behaviour for more targeted campaigns. However, the company has no plans to accelerate its investment in dedicated hardware.

Indeed, while Tencent is massively introducing AI for its own needs, the Chinese giant is also penalised by the sanctions and embargoes affecting its country. A situation that has forced a change of model… But it has also been a revelation:

Last year, it was thought that each new generation of LLM required a greater number of GPUs. This period has come to an end with the advances demonstrated by DeepSeek.

And now the industry, and us within it, are achieving much greater productivity in LLM training with existing GPUs, without needing to add more GPUs at the rate previously expected […] Chinese technology companies are spending less […] as a percentage of turnover than some of their Western competitors. But we believe that this has been the case for some time because Chinese companies generally focus on efficiency and the use of GPU servers, which does not necessarily compromise the ultimate efficiency of the technology developed. I think the success of DeepSeek has really symbolised and confirmed this reality.

Tencent has its Hunyuan Turbo S AI model and has shown that in certain areas it is even ahead of DeepSeek.

Deepseek, on the other hand, announced very low training costs for its AI from the outset. These revelations temporarily disrupted the market. It should be noted, however, that the opacity of Chinese communications on these subjects makes it impossible to make a totally rational and objective judgement about what lies behind Deepseek.