Hong Kong, October 20, 2023. Word is spreading about a mysterious shipment full of Nvidia A100 processors. Catalog value: around 100 million euros. During the night, teams of buyers from firms like Pinduoduo or Kingsoft flock to get their hands on the treasure, says the Chinese newspaper Meiri Renwu, in a long article spotted by the ChinAI newsletter. This precipitation, in the stifling atmosphere of the peninsula, is explained by the importance and rarity of this type of hardware, GPUs, essential for the manufacture of latest generation AI models, LLMs (for large language models). But also by the fact that these American-designed graphics processors, by far the best on the market, are subject to sales bans by Washington on Chinese soil. We do not know the end of thriller. Who got what, and above all, at what price?
The advent of ChatGPT, in November 2022, was experienced as a “Sputnik” moment by many countries. A sudden awareness of the significant impact that generative AI will have on societies and the global economy. The trade war between the United States and China therefore comes at a particularly bad time for the latter. If it can sometimes count on secret deliveries under the cloak, this lack of equipment specially dedicated to “GenAI” is not enough to bridge the distance. “China is very advanced in autonomous cars and other segments requiring classic AI, but it is behind in generative AI,” observes Anne Bouverot, co-president of the Committee of Experts on Artificial Intelligence set up by Matignon.
“Political tension”
Figures prove this. Until mid-January, China trailed France, the UK, Israel and Canada in money invested by venture capitalists in generative AI start-ups, with slightly more of $500 million, according to data collected by Dealroom.co. If a very recent fundraising from Moonshot AI, of around a billion dollars, puts China more in the race, it nevertheless remains almost 30 times less than the United States. Several experts interviewed by the New York Times thus estimate that Beijing is “at least a year” behind its rival. Early models, some of which are openly built on open source LLMs like Meta’s Llama, do not emerge in performance rankings based on industry standard tests. Most are American, belonging to OpenAI, Anthropic or even Google. Accompanied by a few French (Mistral AI in particular), British (Leeroo) and even Canadians (Cohere). These are the ones used primarily by businesses as well as many applications popular with individuals.
“This situation has increased the level of political tension on the subject in China,” indicates Mathilde Velliet, researcher at the Geopolitical Center for Technologies at Ifri, specializing in Sino-American relations. Prime Minister Li Qiang, visiting this week the giant Baidu – the “Chinese Google” – recalled how important AI was “for the development of new productive forces”. But part of the problem is the diet itself. The latter approached the advent of technology with his characteristic desire for control and censorship. A redoubled mistrust regarding “GenAI”, directly affecting knowledge and language. “It’s one of the fears there: that a program like ChatGPT starts talking politics,” underlines Pierre Sel, a specialist in issues linked to new technologies in China. Therefore, each LLM must be authorized before distribution. The first licenses were only granted slowly last summer, months after the release of ChatGPT. Around forty are now approved, out of more than 230 existing models, according to the South China Morning Post. The quality of the data used for training LLMs can also be questioned: as a reminder, China completely bans Wikipedia on its territory, a royalty-free source that is nevertheless massively used by Western models to improve their reliability.
Second session
All is not lost, however. Moonshot’s massive fundraising is perhaps an illustration of this. Zhipu, 01.AI from eminent researcher Kai-Fu Lee, iFlytek, Qihoo 360, Baichuan are also among these locally emerging firms in generative AI. These access computing power thanks to Baidu, Alibaba or ByteDance, firmly established Tech giants, with significant reserves of GPUs, acquired before the various rounds of sanctions in 2022 and 2023. Well informed, the Meiri Renwu estimates around 200,000 GPUs, half of which are A100 equivalents, in the hands of the owner of ByteDance. Enough to train four to five models of similar scale to GPT-4. The country finally plans to build 20 additional computer centers in two years.
“The quality of Chinese technology”, as well as the level of its “world-class” universities on the subject are also high, recalls Pierre Sel. In particular, as Anne Bouverot pointed out, in the field of non-generative AI. “ByteDance has probably developed the best recommendation algorithm in the world,” explains Pierre Sel. Three other elements argue for Chinese catching up in the short or medium term: the dynamism of open source, allowing easy access to quality technological bricks on generative AI, the trend towards smaller models requiring less processing power. calculation, and the dynamism of its internal market. Baidu already claims 100 million users of its Ernie chatbot.
Real reasons for hope. Even if American sanctions remain a burden for China. “It is likely that the United States will carry out new controls on technology exports, for example on access to the cloud, from which it is always possible to have access to computing power,” points out Mathilde Velliet . And the stock of GPUs will eventually run out, despite the trickery. China is currently investing billions of dollars to boost its own production; forty were added in September. Huawei and the founder SMIC (Semiconductor Manufacturing International Corporation) are leading the remontada. The first marketed 5G smartphones with its own hardware last year for the first time. And it launched a range of training servers specializing in AI called Atlas. A notable progress, although there is still a way to go. “Performance, as far as GPUs are concerned, is still far from Nvidia, and companies are too accustomed to the American software ecosystem provided with its processors,” notes Pierre Sel. Which perhaps explains the choice of the name Atlas, a mythological reference to a titan defeated by Zeus and the gods of Olympus, before being condemned to carry the celestial vault on his shoulders. A good allusion to the monumental challenges facing Huawei, rightly notes the Meiri Renwu. And ultimately, surely, those of China in generative AI.
.