The capitalization of the largest US technology corporations continues to break records, supporting the entire American stock market. At the same time, the only driver of this growth is investors’ belief that multibillion-dollar investments in artificial intelligence will yield enormous profits in the near future. Meanwhile, Chinese developers are releasing one neural network model after another, at much lower costs than American ones, and they are also open-source. This highlights a key problem of the US AI sector—its enormous expense, both in terms of costs and stock prices. Could Chinese breakthroughs lead to a new crisis in stock markets, Izvestia investigated.
What’s the price of a token for the people?
In the AI economy, the basic unit of calculation is the token—a fragment of text or code processed by the model. Nvidia CEO Jensen Huang called the token a new commodity, as it is through it that all AI services are priced for developers and corporate clients, while ordinary users typically pay a fixed subscription fee.
Until recently, the main part of the market consisted of chatbots operating on a “question-answer” principle. Token consumption was relatively low. But now the industry is transitioning to agentic artificial intelligence. AI agents do not just generate texts; they independently perform complex multi-step tasks: write program code, conduct market research, manage databases, and interact with external applications without human intervention.
This transition is accompanied by rapid growth in computing resource consumption. According to joint calculations by researchers from MIT, Stanford, and Google DeepMind, completing one complex agentic task requires up to 3,500 times more tokens than a simple logical response in a chatbot. A typical dialogue fits within a couple of thousand tokens, while the work of an autonomous agent writing software or analyzing the market instantly consumes millions. Developers of AI agents already spend hundreds of millions of tokens daily.
The cost of 1 million tokens is becoming a decisive factor for any technology business. And it is in this field that Chinese developers offer conditions that make American models less competitive. AI labs from China strive to keep up with American ones in performance while significantly surpassing them in cost and openness. DeepSeek once shook the markets; now, the successes of the Chinese have become routine.
Parade of models
In June 2026, the Beijing lab Zhipu (also known as Z.ai) introduced its newest system, GLM-5.2, announcing an intention to make advanced intelligence publicly accessible. Almost simultaneously, Moonshot AI released the Kimi-K3 model, and tech giant Alibaba previewed Qwen 3.8.
All these releases are quite impressive. In terms of the number of parameters—a measure of model complexity obtained during training—the Chinese newcomers surpass most known analogs. Kimi-K3 has 2.8 trillion parameters, making it the world’s largest open-source AI system. Alibaba’s Qwen 3.8 operates with 2.4 trillion parameters. American industry leaders—OpenAI and Anthropic—prefer to keep this data about their systems secret.
But the main difference lies in pricing. 1 million output tokens for Z.ai’s GLM-5.2 model costs $1.92, while Anthropic charges $50 for a similar volume for its flagship model, Claude Fable 5. The difference exceeds 25 times. An even more aggressive policy is pursued by the DeepSeek lab, which announced a 75% price reduction for API tokens. Its DeepSeek-V4 Flash model offers rates where using the Chinese system costs 1/34 of the price of American competitors.
The research company Artificial Analysis conducted stress testing, during which leading models performed 657 complex office and administrative tasks requiring huge token consumption. The results were notable. Completing these tasks on Anthropic’s Opus 4.8 model cost $1,000. Using Z.ai’s GLM-5.2 cost $270. The DeepSeek-V4 Flash model handled the entire volume of work for just $14.
Moreover, the newest Chinese model, Kimi-K3, not only demonstrated outstanding cheapness but also surpassed Anthropic’s Fable 5 in the number of successfully completed tasks without errors, costing the customer three times less.
A shot in the foot
The spread of Chinese software is supported by the actions of the American administration itself. Seeking to protect its technological superiority, the White House on June 12, 2026, banned non-US residents from using Anthropic’s advanced Claude Fable 5 model. For the first time, access to commercial AI was blocked by a government decision. The reaction was the opposite. Companies worldwide realized the danger of complete dependence on American proprietary systems, access to which could be revoked at any moment. Against this backdrop, open Chinese models, whose parameters are made publicly available for download and local operation, look like a safe haven.
Even American corporations are beginning to reconsider their strategies. According to data from the payment platform Ramp, a sharp increase is being recorded in the number of US companies paying for DeepSeek services. Microsoft is already exploring the possibility of integrating Chinese models into its flagship office assistant, Copilot, due to the exorbitant costs of using American systems.
Beijing skillfully uses the price advantage of its models for geopolitical influence. At the World Artificial Intelligence Conference in Shanghai, PRC Chairman Xi Jinping announced the creation of cooperation centers for AI application within six regional organizations covering almost the entire Global South. For developing countries seeking modernization but lacking billion-dollar budgets, cheap, open-source Chinese AI is becoming the only solution.
This trend is also confirmed by the scientific community. A publication in Nature indicates that researchers worldwide are massively switching to Chinese open-source models (Qwen, GLM, DeepSeek). They are powerful, free or very cheap, and can be run on local equipment, completely removing scientific groups from the scope of US export controls.
For Russia, under unprecedented sanctions pressure and cut off from American cloud services, Chinese open-source has become a key factor in the survival of the domestic AI industry. Unable to legally use closed models from OpenAI or Google, Russian developers and large corporations actively download the open weights of Chinese systems, fine-tuning them on their own data to create sovereign corporate platforms. This allows maintaining the pace of economic digitalization under severe hardware constraints.
A trillion-dollar collapse
However, Chinese models still have vulnerabilities. According to research, they are less effective at processing complex logical tasks requiring multi-day autonomous analysis and sometimes consume more tokens for “thinking” through an answer than American analogs. Nevertheless, for 90% of routine office tasks, their capabilities are already more than sufficient.
For Wall Street, this situation poses a very serious threat. The market capitalization of American tech giants is inflated by expectations that they can sell AI at high tariffs in an attempt to recoup their trillion-dollar investments in Nvidia chips and data center construction. If China makes basic intelligence a publicly available and very cheap commodity, the profitability of American AI business will go into the red, and Silicon Valley’s dominance could be undermined. Given that the entire US stock market is now supported by technology corporations, the scale of a collapse could be extreme. In January 2025, when the release of the DeepSeek R1 model wiped $1 trillion from US stock exchanges in one day, this could have been just a rehearsal for such a scenario.
Despite the seriousness of the threat, the likelihood of a total catastrophe for the American AI industry, and therefore (in our time) the stock market, is somewhat exaggerated. There are a number of barriers that in many cases will be nearly impossible to overcome.
The first is corporate security and regulation. Large Western businesses (banks, medical corporations, the military-industrial complex) will never integrate models created in China into their closed circuits. The risks of hidden backdoors, data leaks, and direct bans from US and EU regulators completely cut off Chinese open-source from the premium corporate segment of developed countries. If anyone tries to do this, they might get a reprimand from Washington. Even the most powerful corporations are not omnipotent and fear political leaders, especially when national security issues are at stake.
The second limitation is the ecosystem. American giants offer clients not just a neural network but a deeply integrated environment: cloud storage, cybersecurity systems, office suites (like Copilot integration into Microsoft products). Clients pay for a seamless experience and technical support, not just for raw token generation.
At the same time, interesting conflicts may arise outside the US. Given that relations between the Americans and a number of European countries are far from ideal, many of them will at least try to ignore the “strong recommendations” of the US. Who would refuse obvious benefits? And this will only increase transatlantic economic tensions.
About the Author
