Efficiency of AI tool Deepseek sparks debate

Analysts estimate that AI upstart's hardware expenditures exceed $500m

Life

The efficiency of DeepSeek as an AI tool has generated debate within the industry. Although the Chinese start-up claims that its large language model DeepSeek-V3 was trained using only 2.8 million GPU (graphics processor unit) hours at a cost of $5.6 million, this figure stands in stark contrast to the billions spent by US tech giants on similar projects.

DeepSeek’s success has fueled speculation about the true size of their investment. Documents reveal that its parent company, hedge fund High-Flyer Quant, has built a significant computing infrastructure. In 2019, Liang Wenfeng, founder of both High-Flyer and DeepSeek, invested heavily in GPUs for algorithmic trading. High-Flyer’s website highlights the development of Fire-Flyer 2, a supercomputer cluster capable of reaching 1,550 petaflops (unit that measures computer processing speed) – similar to some of the world’s most powerful supercomputers.

Analysts at SemiAnalysis estimate that DeepSeek’s hardware expenditures exceed $500 million, when considering research and development costs in addition to total cost of ownership. Their projections even suggest a potential server expenditure of $1.6 billion, including $944 million for operational costs associated with compute clusters.

Despite DeepSeek’s varying figures on spending, the software innovations are undeniable. The company’s models show that performance is not solely dependent on hardware investment. Analysts at Morgan Stanley recognise this point and argue that DeepSeek’s output serves as a game changer, supporting AI applications built on their models. Analyst Tilly Zhang of Gavekal emphasized that the race for leadership in AI now involves more than just access to advanced chips, it requires the ability to use them effectively.

Emerce

Read More: AI Artificial Intelligence DeepSeek