Deepseek, a little-known Chinese start-up whose affordable technology corresponds to the tastes of the Titans of the most formidable American technology such as Openai and Meta Platforms, has raised investor concerns concerning the need to pay billions of dollars in the research and development.
The Chinese Artificial Intelligence Society (AI) has published its open source model, R1, earlier this month. The model’s capacities correspond roughly to those of the advanced models of Openai, Anthropic and Google while having significantly lower training costs.
“The impressive performance / cost dynamic has raised concerns on investors concerning the need for billions of dollars in capital spending by large American technological companies and the billions more than they plan to spend on a genetive in years to come, “said Malik Ahmed Khan, an analyst in actions from Morningstar.
Do you have any questions about the biggest subjects and trends around the world? Get the answers with SCMP knowledgeOur new content platform organized with explanators, FAQs, analyzes and infographics that are given to you by our award-winning team.
Large American technological companies can follow the plunge and reproduce some of the AI training techniques that Deepseek fits to lead to R1 cost on Tuesday, he said in a research note on Tuesday.
When Deepseek was launched, it quickly became the most downloaded application in Apple American AppStore. The better than expected performance of Deepseek led to a sale of Nvidia and technological peers at Wall Street on Monday. Its shares plunged 17% to US $ 118.58, erase nearly $ 600 billion in market value.
The NVIDIA logo and a decreasing stock graph are visible in this illustration taken on January 27, 2025. Photo: Reuters alt = Nvidia’s Logo and a decreasing stock graph are visible in this illustration taken on January 27, 2025. Photo: Reuters>
“It should have happened, and it could have come from anywhere,” said Alicia Garcia-Herrero, chief economist of Asia-Pacific to the French investment bank Natixis. “He came from China, all the more, in a way, to see a correction.”
Wei Sun, the main analyst of the AI at Counterpoint Research, said that the NVIDIA sale reflects changing perceptions of AI development.
“The success of Deepseek questions the conviction that larger and more computer power models lead to better performance, constituting a threat to the GPU growth strategy of NVIDIA,” said Sun. “He also questions the need for massive investments such as the Stargate project of American $ 500 billion. By obtaining advanced results with fewer resources, Deepseek highlights the potential of an effective innovation on a pure scale, marking a turning point in AI’s race. “
Although the correction could have been an initial reaction to the media threw around Deepseek, Garcia-Herrero said that it did not think it would be a permanent correction.
“The United States will double on investment in AI,” she said. “There will be more money in the United States for the AI industry, if not more than before. Because before, it was somehow understanding that China was delaying, especially in The AI generation and now we see that this was not the case.
Although Deepseek did not have access to AI’s main accelerators, such as Best Hopper and Blackwell products from Nvidia, its large language model (LLM) would be equal with some of the best models published by Openai and others.
“Deepseek raises concerns that future LLM will be developed with less GPU of Nvidia, lower capital expenses of hyperscalers, and perhaps less need for massive energy construction in the United States,” said Brian Colallo, Stratege of Morningstar.
Nvidia on Monday replied to the bracing generated by the Chinese IA company and declared that its advances show the usefulness of its tokens for the Chinese market and that more of its tokens would be necessary for the future to meet the demand for services Deepseek.
Deepseek claims to have formed his model on only 2,048 GPU NVIDIA H800 over a period of two months, which, at US $ 2, explains the number of US dollars announced.
This stimulates a reassessment of AI investment strategies, attention to the sustainability of AI capital expenses, the IA competitive landscape and the monetization of AI, according to Richard CLODE, Director of Janus Henderson portfolio.
Deepseek invoices much less than Openai to use its models – about 20 to 40 times lower, according to Janus Henderson. This adds to the concern about the monetization of AI, given the extraordinary quantities of capital expenditure deployed in the West.
Last year, OPENAI declared that it has closed a long -awaited financing cycle for an evaluation of US $ 157 billion, including 6.6 billion US dollars that the company has lifted from a large investment list and large technological companies, in particular Microsoft and Nvidia.
Analysts say that Deepseek shows that a Chinese company can compete with the highest high IA companies in the United States, despite current Chinese access to advanced American semiconductor technology.
“The breakthrough could be that Deepseek can reach a cutting -edge model without advanced equipment,” according to ING analysts.
According to ING, the market must be price in a slower adoption rate of high performance IT and business income selling AI models.
