NVIDIA shares dropped 12% in pre-market trading as the world was captivated by AI models from the Chinese startup DeepSeek.
The DeepSeek V3 model, introduced last year, has already surpassed ChatGPT in the American AppStore, while the new R1 reasoning model, released last week, became the most popular AI model on HuggingFace. The interest in this case was sparked by the fact that both developments are open-source models, publicly available without any subscriptions, and duplicate (or even surpass) the latest AI versions from companies such as Meta or OpenAI.
At the same time, the main advantage (as well as the reason for the drop in shares of leading chip manufacturers) lies precisely in their training method, which does not require billions of dollars and a lot of time.
In a December research article, DeepSeek noted that its V3 was developed in just 2 months, used a cluster of 2000 specialized chips from Nvidia, and cost only $6 million. For comparison: modern models usually require 16,000+ chips and just GPT-4, according to OpenAI CEO Sam Altman, cost more than $100 million.
“DeepSeek clearly does not have access to as many computational resources as American companies, and somehow they managed to develop a model that looks very competitive,” says semiconductor analyst from Raymond James, Srini Padjuri, hinting at export restrictions imposed by the USA in 2021.
CNBC reports that the extensive model from DeepSeek “prompted investors to inquire about the cost of training” and question the intensive computational approach favored by leading global AI companies. As a result: NVIDIA shares, which had ensured a meteoric rise last year by developing AI chips, fell 12% in pre-market, joining the European chip manufacturer ASML, which reported single-digit percentage losses in stock prices. Similarly, a drop (-8.32% on Japanese markets) was recorded by the Japanese investment conglomerate Softbank, which recently made headlines due to President Trump’s initiative, involving $500 billion in investments for building AI data centers in the USA.
It is worth noting that the amount of $6 million, announced by DeepSeek, has so far hardly been confirmed, and, according to experts, “does not include costs associated with prior research and experiments with architectures, algorithms, or data.”
Spelling error report
The following text will be sent to our editors: