Tech stocks plummet as investors react nervously to DeepSeek
Investors ask why Chinese company can seemingly deliver comparable product for fraction of cost of US rivals
US and Asian tech stocks plummeted in value yesterday as investors reacted to the newly launched Chinese chatbot DeepSeek topping the Apple app store charts over the weekend after apparently having been trained with a fraction of the resources of its US competitors.
The Nasdaq closed down 3.1%. The drop at one point wiped more than $1tn off the index from its closing value of $32.5 trillion on Friday,
Nvidia had the worst day, registering a record one day loss for a company. The GPU manufacturer suffered a fall in share price of 17%, reducing its market value by nearly $600b. Apple overtook Nvidia as the most valuable listed company in the US because of the latter’s collapse in value.
Other tech stocks that took a hammering included Alphabet, which lost $100b, and Microsoft, which saw a relatively modest fall of $7b.
The fall in value has continued for Japanese tech firms linked to the AI sector. Nvidia supplier Advantest fell by more than 9%, while tech investor SoftBank - a key partner in Trump’s Stargate AI project - fell by 5% having already 8% of its value on Monday.
Early indicators this morning suggests that the panic is subsiding, with Nvidia shares recovering 6% in Frankfurt, while Oracle and Palantir also registered gains.
The attention generated by DeepSeek led to it becoming the victim of what the company said was a “large-scale malicious attack”. In the event, the attack hasn’t proved to be too damaging; new users are experiencing limited registrations but service for existing users was unaffected.
DeepSeek, founded in 2023 by former hedge fund chief Liang Wenfeng, claims to have developed its V3 model for slightly less than $6 million. The lower cost and innovation was necessary because exports of the highest power GPU chips to China have been banned since 2022, forcing those developing LLMs to innovate faster than their well-funded US rivals.
DeepSeek developed bespoke algorithms to build its models using reduced-capability Nvidia H800 chips, according to a research paper published in December. The company’s models are also open source and incorporate a reasoning feature that articulates its thinking before providing responses.
Whether V3 is truly comparable to the most recent iteration of ChatGPT is a moot point. Some investors dispute whether DeepSeek really has used the lower powered chips, and whether the cost it has given is accurate. There are also concerns about the censorship of responses in certain areas, such as when asking about Tianmen Square.
Nonetheless, given that OpenAI spent approximately $100 million to train GPT-4o, DeepSeek’s apparent success leaves Sam Altman with some explaining to do to investors. He said yesterday on social media that it was “invigorating to have a new competitor”.
He also said: "DeepSeek's r1 is an impressive model, particularly around what they're able to deliver for the price."
Computing says:
Several of our interviewees recently have emphasised the extent to which 2025 has to be the year where generative AI starts to deliver tangible financial returns for businesses as opposed to being full of promises but very few use cases.
DeepSeek has exposed the fragility of the AI bubble by delivering a product which is broadly comparable to Chat-GPT for a fraction of the compute and a fraction of the cost.
Even President Trump called it a wake-up call.
Investors have poured hundreds of billions of dollars into companies like OpenAI but Sam Altman’s response yesterday that “we will obviously deliver much better models” only served to emphasise the ‘jam tomorrow’ message coming from the sector.
With investors warning that the AI bubble is beginning to look dangerously unsustainable, especially coming as it does at a time of increasing nervousness about borrowing costs, pressure is mounting on those who have delivered relatively little despite astronomical investment.