Tech companies invested in the AI market have seen their stock prices take a hit following claims from Chinese startup DeepSeek relating to cost and performance of a new AI model.
The R1 LLM (large language model) released by the company last week appears to be comparable to those developed by OpenAI and Meta. However, DeepSeek claims to have trained the model using far less advanced chips and for a fraction of the cost, approximately $5.6 million - although there is a lot of caveats to that number (for one, it only counts the single training run, not all the research runs).
DeepSeek and the generative AI perception dilemma
A more efficient AI bursts the bubble of AI gigawatt optimism
In comparison, in July 2024 it was reported that OpenAI’s training and inference costs could reach $7 billion for the year, and the company last week announced 'The Stargate Project,’ a joint venture with MGX, Oracle, SoftBank that is set to invest $500 billion into AI infrastructure over the next four years.
Also last week, Meta CEO Mark Zuckerberg announced the company is planning capital expenditure of $60-65 billion, primarily on data centers and servers, as it seeks to boost its AI capabilities.
Over the weekend, DeepSeek overtook ChatGPT to become the most downloaded app in Apple’s US App Store, with shares in Nvidia, Microsoft, and Meta all falling, seemingly as a consequence of the company’s claims.
Nvidia was down nine percent in pre-market trading, with Meta and Microsoft, which has also made a big bet on AI in recent years, both seeing their shares fall by four percent. Dutch semiconductor equipment manufacturer ASML was also impacted by the slump, down 9.7 percent, while Schneider Electric similarly saw its shares slide by 8.7 percent.
US sanctions on advanced AI semiconductors have, in the main, stopped Chinese companies from having access to the same hardware used by OpenAI and Meta to train their models. Consequently, DeepSeek said it used a stockpile of old Nvidia A100 chips in addition to an unspecified number of H800s, a cut-down AI chip specifically designed by Nvidia for the Chinese market to avoid falling foul of US government export controls.
Although the market has been spooked by DeepSeek’s technology, it should be noted that the company’s claims regarding hardware and training costs have not been verified.
It has been reported that many Chinese companies are finding ways to acquire leading-edge Nvidia chips by buying through secondary markets such as Singapore.
