- Статьи
- Economy
- Neck and neck: China has almost caught up with the United States in terms of AI metrics
Neck and neck: China has almost caught up with the United States in terms of AI metrics
The gap between the American and Chinese language models is rapidly narrowing: the "Chinese", who at one time lagged behind in the field of artificial intelligence (AI) for several years, are now literally two or three points behind the leaders according to LiveBench estimates, and in some indicators surpass them. Considering that Asian models are much cheaper and offer open scales, their attractiveness is leveled even more noticeably. Izvestia investigated the current state of market development and what could happen when and if China takes the lead.
We rolled to the finish line
Since 2023, the position in the AI race has changed dramatically. Three years ago, the release of GPT-4 from OpenAI seemed like an unattainable peak. The first Chinese products, such as Ernie Bot from Baidu, caused condescending reviews in the West due to weak logic, which worsened the coherence of responses, and political censorship did not contribute to optimism - America unconditionally dominated.
Now everything has changed. Models from OpenAI, Anthropic, and Google still hold the top spots in global rankings. They are better able to handle complex programming and complex logic chains. However, Chinese developments have firmly occupied positions immediately behind the leaders. Products from the DeepSeek startup, as well as Alibaba's Qwen and Moonshot AI's Kimi models, are consistently among the top 10 or 15 of the world's best systems. The distance between the first echelon (USA) and the second (China) has decreased from a year and a half to several months.
According to the latest LiveBench data, DeepSeek v4.1 Flash ranked sixth, Moonshot's Kimi ranked 11th, and Alibaba's Qwen ranked 13th. By the way, they were all ahead of the best model from SpaceX. The Chinese have proven their ability to create algorithms that meet 95% of the requests of ordinary users and corporate clients in terms of response quality, even if they are slightly inferior to the leaders in some subtleties.
At the moment, the main problem of Americans is not even in the models themselves, but in the way they are distributed. American giants rely on closed proprietary systems, access to which is sold through expensive subscriptions and APIs. Chinese developers choose open source, putting the weights of their advanced models in the public domain. This allows developers around the world to download them and deploy them on their own servers.
Where Chinese companies do sell API access, they use dumping. The cost of processing tokens from DeepSeek or Alibaba is ten times lower than that of OpenAI. The reasons are clear. First, US export restrictions on the supply of advanced Nvidia chips forced Chinese engineers to look for non-standard ways. Unable to build data centers with hundreds of thousands of processors, they focused on architectural optimization, so their models require fewer computing resources to operate. Secondly, you need to make your product attractive in a highly competitive environment. Beijing is keen for startups in India, Brazil and the Middle East to build their applications based on Chinese platforms. Actually, this is how competition works inside China, and not only in the AI industry.
The illusion of profitability
This creates problems for the American financial model. In 2024-2026, Bigtech invested huge sums in the purchase of iron and the construction of infrastructure, with trillions of dollars in capital expenditures. Investors are not demanding to show the return of these funds yet (the main thing is expansion and progress), but this moment must come sometime.
The official accounts of a number of American AI leaders are trying to show optimism, but a detailed audit raises questions. The recent history of the financial performance of the company Anthropic is noteworthy. The startup has reported on reaching operational profitability. But in the second approximation, accounting tricks are immediately visible. The company's management has extended the amortization period for server equipment. This made it possible to drastically reduce costs in quarterly reports, eventually making paper profits, although the physical outflow of cash remained at the same level. In addition, payments to partners, such as Amazon, have disappeared from operating expenses.
The problem is already there. And if corporate customers realize that they can use the free Chinese model to sort mail, analyze documents, and operate call centers, they will begin to massively abandon more expensive subscriptions to American services.
In a scenario where Chinese neural networks capture a significant share of the global market, American companies will not be able to discourage investments in data centers. Yes, Chinese developers will also incur losses — no one has figured out how to train models cheaply. But the financing structure in China is different: Alibaba, Baidu, and independent startups have government funds and direct lending behind them, and they can "absorb" losses for quite a long time. Beijing can allow its companies to operate in deep negative conditions until 2030 or longer, considering this as an investment in global geopolitical dominance. American companies depend on the mood of their shareholders.
The Kings' Last Argument
The historical norm is that when a technological advantage disappears, government coercion comes into play. As soon as Chinese open-source models finally catch up with their American counterparts in all benchmarks, we will almost certainly see Washington move to non-market measures. The ban on the export of chips is clearly not limited to the case. The American administration may use mechanisms for a direct embargo on the use of Chinese software code. Interestingly, even high-tech companies from the United States are against such pressure, but it all depends on how to present the situation to the public.
Companies using Chinese neural networks will be subject to checks for threats to national security. It is likely that protective tariffs will be imposed on digital services, a ban on placing Chinese models on the servers of American cloud providers (AWS, Azure), and the threat of secondary sanctions for those international corporations that try to build their businesses on Chinese algorithms.
Obviously, the United States will not allow market mechanisms to destroy its multibillion-dollar (if not trillion-dollar) investments in artificial intelligence. The competition of algorithms will end where the jurisdiction of the sales representative and the Ministry of Finance begins. The global Internet may be divided into isolated technological blocks, and the choice of a neural network for work or entertainment will become a matter of policy.
Переведено сервисом «Яндекс Переводчик»
