A cost-efficient approach to AI heralded by a Chinese startup is challenging widely held assumptions about the amount of computational power needed for commercial deployment of the technology.
The launch of an open-source Chinese AI chatbot is sending shockwaves throughout the tech industry. Said to rival OpenAI’s o1 model, Deepseek was developed with a fraction of the resources that leading Western outfits employ, casting doubt on the premise that progress in AI requires massive amounts of cutting-edge chips. As a result, semiconductor stocks, including Dutch equipment makers ASML and ASM International, took a tumble on Monday.
Deepseek, a venture set up only last year by hedge fund billionaire Liang Wenfeng, claims it used a little over 2,000 Nvidia H800 GPUs to train its language model. The H800 is based on Nvidia’s H100, modified to comply with US export restrictions. Originally launched in 2022, it’s no slouch, but being two generations behind Nvidia’s soon-to-be-launched Blackwell generation, neither can it be considered top-of-the-line. Other than on hardware, Deepseek is said to have spent less than 6 million dollars on training, compared to routinely 100+ million dollars by Western firms.
This article is exclusively available to premium members of Bits&Chips. Already a premium member? Please log in. Not yet a premium member? Become one for just €20 for the first year and enjoy all the benefits.
