A new model artificial intelligence presented in January 2025, an unknown Chinese startup until then Deepseek. Its model was almost equal to those of the giants, such as Openai and Google, but required significantly fewer computational resources. This move shook the waters and caused embarrassment in the established forces of the industry.
This move initially raised anxiety for investors and disruption to Wall Street, with NVIDIA shares temporarily sinking. However, concern quickly turned into admiration as artificial intelligence research company Deepseek entered the list of Time magazine with the 100 most influential companies of the year.
Six months later, the image is more complex. Rumors of a technological “revolution” may have been premature. Nevertheless, Deepseek’s intervention revealed significant weaknesses in the artificial intelligence ecosystem: dependence on huge data centers, the opacity of the US algorithms, and the threat of China as a rising superpower.
To assess the actual effect of Deepseek, but also to understand why the reaction was so intense, we look at five basic aspects that determine the current momentum in the artificial intelligence industry:
1. Share Prices: Quick recovery
This is the biggest loss of value in Wall Street history: on January 27, the price of Nvidia fell by 17% – given the historically high capitalization of the $ 3.5 trillion market, it was $ 589 billion. Other companies also experienced a black Monday: Cloud providers such as Microsoft, chip manufacturers such as ASML, energy experts such as Siemens Energy.
Nvidia is the “iconic name” of artificial intelligence and, therefore, the dominant issue on the Stock Exchange, as the Morningstar analysis company writes. Graphic processors of the chip manufacturer is central to the development of large linguistic models that work in the background in programs such as Chatgpt. In this respect, the group benefits more than excitement – and suffers more than doubts.
Deepseek challenged a supposed regularity. So far it has been the following: The more data and computing power the developers use in the training of a language model, the more efficient it becomes. The new business broke this vicious circle with some inventive ideas. This made shareholders and analysts asking: need fewer data centers, transformers and processors than we thought?
However, the Deepseek phenomenon was not lasting on the Stock Exchange, as the price of Nvidia recovered and is just below the historic high. However, for a company of this size, the evolution of the price is unusually unstable, Morningstar observes. The chip manufacturer is a pioneer in a technology, the introduction of which is full of uncertainty. Recent proof is that although the company reported strong growth last week, its share price fell.
2. Use of artificial intelligence: long -term growth
There are several reasons for uncertainty. While Openai’s manager Sam Altman is planning the vision of a supernatural, many companies find it difficult to operate with artificial intelligence applications. At the same time, Deepseek and other companies show that progress is feasible with less means.
What does this mean for computing infrastructure providers and their suppliers? According to UBS, they do not need to be worried at the moment. A group of investment bank analysts estimates that the demand for computing power will increase by an average of 115% per year from 2024 to 2027, rather than all uncertainties.
However, there is a shift. Today, it is mainly the developers who overheat processors with the training of their models. In the future, the application of technology – in terminology: Inferencing – will ensure the full use of production capacity. According to UBS, this figure is expected to rise from 20% to 57%, pushed by business applications.
It is worth noting that this is a model of calculation – with dangers, as UBS writes. If, for example, companies need more time to introduce artificial intelligence applications, part of the demand that has been calculated at present will be lost.
3. Price Evolution: Costs decrease, use increasing
At the end of 2024, before the shock of Deepseek, the Guido Appenzeller of the business capital Andreessen Horowitz described that the cost of using artificial intelligence models had been reduced on average ten times in three years.
Deepseek continued this trend. With similar performance, the models cost more than 90% less than the competing models of Openai.
Did the Deepseek thus accelerate the decline in prices? In any case, Openai reduced the costs for the successor of O1, the GPT-O3, by 80% in early August. Also, the new GPT-5 basic model is half cheaper to use.
With each processing unit it processes, the cost continues to decline, with some exceptions such as Claude Opus 4. So far, Deepseek has not managed to undermine US competition. It remains one of the cheapest models, with only a few Chinese providers offering lower use prices than Deepseek.
At the same time, companies are spending more and more money on artificial intelligence. The cheaper a raw material, in this case the artificial intelligence, the more used and the greater its consumption – a phenomenon that is known in the economy as a paradox of Jevons. A recent post on Andreessen Horowitz’s blog states: “Work based on the latest models with advanced conclusions still have a relatively high cost.”
4. Competitive Status: short -lived popularity
Theoretically, Deepseek would be an economical alternative to American providers – and would also be freely accessible. Developers and users can download it for free via the HuggingFace artificial intelligence platform. Through the Website, users can interact freely with Chatbot-or have access to a model with a similar performance to the GPT-4O through a more affordable interface.
But did the Chinese start-up manage to get users from American providers? The One Little Web Analysis Company evaluated the use of Chatbot based on artificial intelligence in browser. According to the results, Deepseek is in second place with only 4%, but with a large difference from Chatgpt, which has a share of 48%. However, Deepseek has been steadily losing popularity since February, while other providers have increased.
It is not possible to accurately determine how many companies use Deepseek models. A survey of 550 IT executives in the US in early 2025 showed that 17% of respondents said they were using Deepseek. Another assessment of the Menlo Ventures business capital, which has invested in the artificial intelligence provider Anthropic, estimates that this figure is just 1%.
In Europe, concerns about data protection may prevent companies from using the model. BSI has warned not to use the model in cases critical for safety. In Italy, Deepseek was withdrawn from the App Store within a few days, also due to security concerns.
According to Thomas Thiele of Counseling Company Arthur D. Little, the critical for businesses is which model is the best and the most expensive. “What was fashionable six months ago, today is old,” Thiele says. According to an evaluation of the Semianalysis online magazine, Deepseek models also need more time to process than competitively. For Thiele, Deepseek is therefore rather symbolic internationally. However, he admits that “he has brought the issue of Chinese open source models.”
5. New race for open source models
Before 2025, publicly available models were considered less efficient than the privately owned models of Openai and Anthropic. Open source models and open weight models are models available for free to download and adapt.
However, according to a Stanford Ai Index assessment, at the beginning of 2025 they gained ground in several performance comparisons – and Deepseek was part of this jump.
So far, Openai prefers a closed approach and maintains its secrets behind the chatgpt. However, in early August, Openai published two Open-Weights models for downloading and adapting: GPT-Thessaloniki and GPT-US-20B.
This is likely to be a reaction to China: the government in Beijing is actively promoting the AI open source, while companies such as Alibaba, Baidu and Moonshot AI publish their models. Advisor Thomas Thiele explains this contrast to the US as follows: “The message is: we do not need to hide and we are open to the rest of the world.”
America has taken into account: the Trump government recently announced in its artificial intelligence action plan that it will focus more on open source models.
If Chinese companies offer the best free models available, they could more and more find their place in companies around the world, as they can be easily downloaded and used. In addition, entire communities of programmers gather around popular open source models, resulting in models from China set the standards.
When Deepseek caused turmoil in the artificial intelligence industry, Meta’s artificial intelligence, Yann Lecunn, wrote on LinkedIn: “In those who see Deepseek performance and think:” China surpasses the US “, you read wrong. The correct interpretation is: “Open source models exceed privately owned models”. After all, in the field of open source, both interpretations seem possible.