Does China finally hold its own OpenAi? Despite the attempts of the giants Tencent or Alibaba to compete with the American leader of the generative AI, the surprise came this winter from a start-up completely unknown, Deepseek. His language models – “V3” published on December 26 and “R1” on January 20 – are measured according to the sector standards at “GPT -4O” and “O1” of Openai, who feed the conversation agent Chatgpt.
But it is not the most surprising. Deepseek achieved these prowess at infinitely lower costs than that of his rival. “V3”, for example, claimed only $ 5.57 million for his training. Against more than $ 100 million for “GPT-4O” from Openai to which it is measured. “V3” also required 11 times less calculation time than Meta Llama-3, another major actor of the AI. The whole thing, with a lower number of specialized chips (GPU), and less sophisticated, due to American restrictions imposed on China in this area for several years. Like balls on the feet, in the idea of slowing down the progress of the Empire in the middle in AI.
It is an understatement to say that this Washington strategy, driven by Donald Trump during his first mandate and continued by Joe Biden, has lead in the wing. China, especially via Huawei, learned to manage this lack of fleas, and invests massively in its local production. Deepseek, a subsidiary of the High-Flyer hedge fund, had managed to get a few thousand chips before the valves close. Well took him.
Price war
The ascent of the Chinese start-up from Hangzhou, not far from Shanghai, has other implications for the United States. Deepseek’s forced sobriety questions massive investments currently operated by the Americans. Several experts believe they feed a “speculative bubble”. No later than Wednesday, Openai, Softbank and Oracle announced the “Stargate” project, endowed with $ 500 billion. An unprecedented envelope in order to enrich their infrastructure dedicated to artificial intelligence, from data centers (data centers) to power plants. And to water, in the long term, the whole world of their solutions.
Black point of this approach: pricing pressure is already growing. “The cost of the IA solutions is a major concern of the companies that are buying them at the moment,” confirms Ghislain de Pierrefeu, consultant specializing in artificial intelligence within the Wavestone cabinet. This, while the expected productivity gains thanks to the generative AI are always vague. The general public, too, is ordered to go to the checkout. Microsoft recently increased the price of its office consequences by around 30 % due to the AI. Openai has launched a “pro” subscription to Chatgpt at 200 dollars per month, which he does not even consider profitable.
Deepseek calls into question American gigantism, and its economic consequences. Perhaps for a long time, since the company led by Liang Wenfeng, 40, also shared all of its models and its open source methodology. A clear difference here with Openai. Not only can anyone download and access the best in the generative AI, at the start of 2025. But scientists from around the world now have access to the Deepseek “miracle method”: A mix of technical tips, innovations and optimizations in learning data that made it possible to make your models more efficient and less energy -consuming. So, less expensive for use. “A very impressive demonstration of research and engineering under resource constraints,” said on X Andrej Karpathy, the former AI manager at Tesla.
Geopolitical asset
This method is already emulated. First in China, at Bytedance, Tiktok’s parent company, which suffers from the same flea restrictions as Deepseek. Its “Doubao 1.5” model, presented Wednesday, January 22, also boasts of high performance, comparable to OpenAi models, associated with a very low cost to use – about 50 times less than GPT -4O. There is no doubt that Beijing found a good vein. Liang Wenfeng, the boss of Deepseek, was also invited during the week to an important meeting on AI, organized by the authorities with a small group of sector experts. The local press has seen it as a symbol of its influence, which should not be limited to Asia. European companies, which do not suffer from the same restrictions but are late on the Americans, could also be inspired by its approach.
Could the “Deepseek method” encourage Openai to question his? Surely not. Control of the United States on the china chips to China is only going to increase, under Trump. While Openai will continue to access advanced tools, thanks to Nvidia. Chatgpt’s father is able to keep his leadership and dictate the investment tempo. China is likely to have trouble exporting its models, even at low prices. “There is a blockage in the West to the use of Chinese solutions,” said Ghislain de Pierrefeu. For data confidentiality issues. Of culture, too. Recall that the censorship of the Communist Party has passed on Deepseek as on the other models of AI. Impossible to question him about the sovereignty of Taiwan or the events of the Tiananmen square in 1989.
But not everyone will be so uncompromising, prognostic the consultant of Wavestone: “countries that do not recognize themselves in Western values will look closely at what Deepseek realizes …”. A quota that is growing, from year to year. Despite its provenance, the start-up arouses as much distrust as admiration. Including in the country of Uncle Sam. “We live in a time when a non -American company keeps OpenAi’s original mission alive: a truly open and innovative search that gives power to all,” wrote on X L ‘Jim Fan engineer, eminent personality at Nvidia. A stretch noticed against the one who had placed himself in his early days as a future giant of the Open Source, before giving it up for commercial reasons. And a filigree victory for the soft power Chinese.
.