Cheap and surly: the Chinese have tipped the global AI market with DeepSeek

Something of a panic has set in in Silicon Valley. Last week, a little-known Chinese company released a new large language model of artificial intelligence, which turned out to be as good as its analogs made in the United States. That would be half the trouble: the main problem is that the "Chinese model" turned out to be dozens of times cheaper and required incomparably less computing resources. In essence, we are talking about a real revolution that could spill over far beyond the artificial intelligence market. It is quite possible that now we are witnessing the beginning of a new "dot-com crash". The details are in the Izvestia piece.

"Sputnik Moment."

At this year's Davos forum, executives of American technology corporations admitted that the gap from China in the development of large language models (LLM) is very small - a year at best. But it turns out that this estimate is an overestimate as well. On January 20, the Chinese offered the DeepSeek R1 model, which was an instant hit. By the end of last weekend, it had already climbed to the first place in the ranking of Apple and Google Store applications. And no wonder: independent tests showed that the program is at least as good as the latest OpenAI, Anthropic and Meta (banned in Russia), and in some respects even better.

The key difference with all these models is that R1, like previous DeepSeek variants, is free for private use. Anyone can download the virtual assistant to their smartphone, talk to it on the company's website or install it locally on a computer (provided the hardware is powerful enough). Almost all American competitors have a proprietary distribution system, it requires payment by subscription.

The popularity of the model has gone so far that it was forced to temporarily restrict access to registration. Although the company justified that the site was overwhelmed by malicious attacks, most likely it simply could not withstand the global demand for the product.

All this has had a bombshell effect in circles close to hitech and AI development in particular. American investor and engineer Mark Andriessen called the situation the "Sputnik moment," comparing the DeepSeek R1 launch to the Soviet Union's launch of the first artificial satellite in 1957, which turned out to be a cold shower for the Americans, who were confident in their superiority in rocket and space technology. Microsoft CEO Satya Nadella acknowledged that the emergence of a new player in the industry will dramatically increase competition, which will ultimately be good for the development of artificial intelligence. OpenAI CEO Sam Altman welcomed the rival and promised to develop improved models to survive the competitive race. For the most part, the stance of the CEO of American Internet companies was a good mine in a bad game: few expected such an effect.

What DeepSeek is

One might expect the competition for the American giants to come from their Chinese counterparts - corporations like Alibaba or Tencent. But DeepSeek's models were developed by a modest Hangzhou-based startup formed in 2023 by High Flyer hedge fund manager Liang Wenfeng, who founded the company with just 10 million yuan ($1.4 million) of fixed capital. In a matter of months, the firm released several models, the penultimate of which, the V3, had already attracted national attention. However, the R1 has shown truly colossal progress, reaching the world's tops.

At the same time, the funds, which the company operates, are quite limited in comparison with both American and Chinese competitors. The exact amount invested in the development of the latest model is not officially announced, but according to indirect calculations we are talking about a seven-digit (in U.S. dollars) number. The most frequently mentioned figure is $5.5 million, which looks ridiculous compared to the $75 million allocated for training the latest ChatGPT model.

More importantly, DeepSeek spent significantly less computing power. The company admitted that it only took about 2,000 Nvidia H800 chips to train the model. These aren't the newest or most powerful chips, and it didn't have to circumvent the sanctions imposed on Chinese consumers to buy them. This is again pennies compared to the processor systems used by American competitors costing tens of millions of dollars.

The minuses of the new model are not very many yet. There is talk of restrictions in terms of censorship, but scandals with various filters have also been heard regarding American designs, as well as other Chinese offerings. For most purposes in which the program can be used, this is not particularly critical. Others, on the contrary, warn that the restrictive barriers in R1 are easily bypassed and the model can be used for illegal activity. Finally, as mentioned above, the company is not yet very ready for its success and cannot provide access to everyone. But on the horizon of even a couple weeks, this problem is not too significant.

Minus $600 billion in a day

The implications for the artificial intelligence industry are already huge - as well as for the broader economy. News of DeepSeek's success triggered an epic sell-off in the U.S. stock market. The Nasdaq index plummeted by 3% in one day, with the world's richest people, who co-own powerful technology companies, losing a combined total of more than $108 billion of their wealth over the course of those hours. The situation looks disturbingly similar to the collapse of Internet companies in 2000, when the market suddenly realized that their valuation was grossly overvalued. That collapse led to nearly 10 years of stagnation in the fastest growing sector. Will it happen again now? It's too early to tell, but the risks are non-zero, given the record share price to earnings ratios of high-tech corporations.

Foremost among them is Nvidia, which lost 17% of its capitalization, or about $600 billion, on Jan. 27 - by a huge margin the record daily drop in U.S. stock market history, if we're talking about absolute numbers of market value. The graphics processor company has seemed to win the lottery time after time over the past 10 years, confirming that shovel makers make the most money during the gold rush. First there was the cryptocurrency boom, then the demand for video cards during the pandemic and now AI. The corporation's securities rose 239% in 2023 and 171% in 2024. Due to this jump Nvidia in the shortest possible time surpassed Microsoft and Apple, becoming the most expensive company in the world (about $3.45 trillion before the fall).

But now it turns out that you don't necessarily need the most "tricked-out hardware" to produce high-quality and up-to-date AI models. This automatically means that chips in such volumes probably won't be needed either, and Nvidia's competitors may well get into the game. And the question of whether hundreds of billions of investment in the pursuit of ever more powerful chips is justified becomes a question. And there are implications not just for Jensen Huang's company. But more broadly, regarding U.S. actions in this field.

Starting in the late 2010s, the Americans began to embargo the supply of advanced chips to China, gradually tightening the screws - all to prevent the creation of competitive Chinese AI. But DeepSeek R1 shows that these efforts have largely gone to waste, and the bans have only stimulated innovation, confirming the adages "you can't be too clever" and "you can't be too smart" (the latter in relation to American tech giants spending large sums on technical support instead of finding elegant solutions). It is quite possible that something similar will soon happen in the production of chips themselves - if, of course, the semiconductor race will still be relevant in a few years.

For Russian companies, what is happening may also be a good lesson. The domestic AI industry, although among the world's top, is still inferior in its products to the Americans and the Chinese. This case, however, shows that the lag in budgets is not critical and can be compensated for in non-standard ways.