The growth in the popularity of and technology contributed to the growth of the market value of NVIDIA above $ 3 trillion. However, its shares fell on Monday by 17%, causing a drop in the market value of the company by almost $ 600 billion, after the announcement of the Chinese startup of Deepseek V3 and R1, capable of competing with the best models of any American company, although they were trained for a small part of the cost of Less advanced chips NVIDIA H800 and A100, writes Fortune.
Also, at the beginning of the week, the AI Assistant app startup Deepseek was first in the ranking of the most popular free applications in the online store in the Apple App Store in the United States, ahead of the Openai Chat-Bot. Moreover, the Deepseek R1 model, designed to challenge the Openai O1 “reasoning” model, can be launched at the workstation, and not in the data center.
Since the powerful NVIDIA accelerators are one of the largest articles in the development of the most advanced AI models, investors began to revise their ideas regarding investments in AI-Business. Yes, Deepseek clearly shocked the AI market, but talking about the collapse of NVIDIA can be premature, as well as the statements that the success of Deepseek means that the United States should abandon the policy of China’s access to the most advanced JI chipes warn Fortune.
Deepseek claims to use 10 thousand NVIDIA A100 accelerators, as well as H800 chips, which is an order of magnitude less than American companies use their most advanced AI models. For example, the XII Ilona Musk built a Colossus computing cluster in tenesess based on 100 thousand NVIDIA H100 accelerators, it plans to expand to 1 million chips.
Image source: Heather Wilde / Unsplash
This gave the reason to some experts to argue that the introduction of US restrictions spurred innovation in China. Fortune consider such conclusions short -sighted and claim that the influence of Deepseek can, paradoxically, sounds at first glance, to increase the demand for the advanced AI chips – both NVIDIA and its competitors. The reason is partly in the phenomenon known as the Jevons Paradox.
The Jevons Paradox, also known as the rebound effect, is named after the 19th century British Economist William Stanley Jevons (William Stanley Jevons), who noticed: when technological progress makes the use of the resource more efficient, the overall consumption of this resource tends to increase. This makes sense if the demand for anything is relatively elastic-the price decreasing due to increase in efficiency creates an even greater demand for the product.
One of the reasons for the weak implementation of AI models in large organizations was their high cost. This especially concerned new “reasoning” models, such as O1 from Openai. DeepSeek models are much cheaper than competitors in operation, so now companies can afford to deploy them for many use scenarios. On the scale of the industry, this can lead to a sharp increase in demand for computing power.
On Monday, Microsoft CEO, Satya Nadella and the former Intel CEO Pat Gelsinger, indicated this in social networks. Valela directly referred to the paradox of Jevons, while Gelsinger said that “calculations are subordinate” to what he called the “law of gas”. “If you make it much cheaper, the market will expand for it … It will make AI much more widespread,” he wrote. “The markets are mistaken.”
Image source: Mark Daynes / Unsplash
Fortune asked the question: “What exactly computing power will it be required?” The top-end accelerators of NVIDIA are optimized for teaching the largest large language models (LLM), such as the GPT-4 from Openai or Claude 3-Opus by Anthropic. For an infority, NVIDIA chips are less suitable than competitors, including AMD and, for example, GROQ, the chips of which allow you to execute AI loads faster and much more efficiently. Google and Amazon also create their own AI chips, some of which are optimized for an infront.
NVIDIA is now occupied by more than 80% of the AI-based market on the basis of the data center (if you exclude custom ASIC cloud providers, its share can be up to 98%) and is unlikely to lose dominance quickly or completely, Fortune noted. Eye accelerators can also be used for an information, and the CUDA software platform has a large and loyal community of developers, which is unlikely to refuse it once. If the general demand for AI chips increases due to the Jevons paradox, the total income of NVIDIA will still be able to grow even when the market share falls due to the increased market.
Another reason why the demand for advanced and chips will probably continue growth, is associated with the peculiarities of the work of reasoning models such as R1. While the abilities of previous LLM types grew as available computing power increased during training, then the models of reasoning depends on computing resources during the infront – the more there are, the better the answers.
Image Source: Kayla Kozlowski / Unsplash
Having launched the R1 on a laptop, you can get a good answer to a difficult mathematical question, say, an hour, while when using accelerators in a cloud for the same answer, a few seconds will go away. For many business applications, a delay or time, the necessary model for response, is of great importance. And in order to reduce the time of the task, advanced AI trackers will still be needed.
In addition, many experts doubt the veracity of DeepSeek’s statements that its V3 model was trained for about 2048 NVIDIA H800 accelerated accelerators or that its R1 model was trained in so small chips. Alexander Wang, CEO of Scale AI, said in an interview with CNBC that, according to him, Deepseek secretly gained access to a cluster of 50 thousand H100 accelerators.
It is also known that the HighFlyer hedge fund, which owns Deepseek, managed to purchase a significant number of NVIDIA less productive accelerators before the sanctions. So it is quite possible that NVIDIA is in a better position than panicing investors suggest, and that the problem with the US export control is not in politics, but in its implementation, Fortune analysts summed up.