“Two years ago the main concern was China, today, in Europe, the biggest concern is the United States.” This is how Per Roman, founder of the venture capital fund Bullhound Capital, summarizes the change sweeping the artificial intelligence market.
While more and more Western companies are choosing models developed in China to contain costs that are now sharply rising and to diversify dependence on US suppliers, Beijing is considering limiting access beyond its borders, turning the most advanced AI into a strategic asset to protect.
COSTS AND PERFORMANCE DRIVE COMPANIES TOWARD CHINA
From Silicon Valley to Europe, companies are progressively supplementing or replacing models developed by OpenAI and Anthropic with Chinese alternatives, attracted by significantly lower costs and capabilities now considered competitive. Supporting this trend, CNBC states, was also the shift of several OpenAI and Anthropic enterprise services from fixed-rate subscriptions to consumption-based billing systems, which have increased expenses for many businesses.
According to OpenRouter, the share of tokens processed by US companies via Chinese models has exceeded 30% weekly since February 8 and has reached peaks of 46%. The average for the previous twelve months was 11%, while in the first half of 2025 it had dropped to 4.5%.
Kyle Chan, a researcher at the Brookings Institution, notes that “Chinese AI models are particularly interesting for American companies now that AI costs are exploding,” emphasizing how businesses today are much more attentive to the economic sustainability of their implementations than in the past.
THE CASES OF DOORDASH, SIEMENS, AIRBNB, AND LINDY
According to the Financial Times, among the companies that have already integrated models developed in China is DoorDash. Co-founder Andy Fang explained that the group entrusts the “lower-level work” to the Kimi K2.6 model from the Chinese startup Moonshot AI, reserving Anthropic’s Fable only “for the more difficult tasks.” According to Fang, this combination “offers far superior performance at a lower cost” compared to the previous solution based solely on Anthropic’s models.
Siemens has also chosen to diversify its AI tools. The German group uses models developed by DeepSeek and Z.ai alongside those from US labs, Nvidia, and the French company Mistral, explaining that it wants to maintain “flexibility” in choosing technologies.
Airbnb has confirmed using “a limited number of models of Chinese origin,” specifying, however, that these are run exclusively through approved US-based service providers to ensure data protection and business operations security.
The Californian startup Lindy has instead made a complete migration, transferring 100% of traffic from Anthropic’s Claude models to DeepSeek. CEO Flo Crivello called the change “transformative,” explaining that it will allow the company to save millions of dollars in a few months and that switching to the DeepSeek V4 model has improved performance in “many key use cases.”
MODELS THAT HAVE LESS AND LESS TO ENVY FROM US ONES
One of the factors favoring the spread of Chinese models is their open source or open-weight nature, which allows companies to download them, run them on their own servers, and adapt them to internal data, maintaining much greater control compared to proprietary systems like ChatGPT or Claude.
For Vipul Ved Prakash, CEO of Together AI, the best open-weight models cost between 10 and 60 times less than corresponding proprietary models. Justin Summerville, head of data and analytics at OpenRouter, estimates instead that Chinese open source models can be 60% to 90% cheaper than the main Anthropic and OpenAI solutions.
The technological gap, CNBC writes, is also continuing to narrow. Z.ai’s GLM-5.2 model, presented in June, was enthusiastically received in Silicon Valley. According to Harpreet Arora of Vercel, it recorded the fastest adoption rate among all models monitored by the platform in 2026: in the first full week, the daily token volume increased about 27 times and the number of customers grew about 80 times.
And Marc Andreessen, co-founder of Andreessen Horowitz, wrote that “many competent people and industry insiders argue that GLM-5.2 is the first Chinese model capable of matching and often surpassing the public models of the major American labs without compromises.”
GEOPOLITICAL UNCERTAINTY CHANGES STRATEGIES
Alongside economic reasons, the weight of geopolitical considerations is also growing. The temporary US export ban on Anthropic’s Mythos and Fable models has pushed many companies to reduce dependence on American suppliers.
As reported by the FT, Ben Grinnell, head of AI at the British company Newton, believes that even after the Fable ban was lifted, “market perception has changed forever.”
For this reason, companies like Timebutler have started shifting part of their workloads from Anthropic’s Claude model to Alibaba’s Qwen models to diversify suppliers, while, according to RTP Global, for many European startups a Chinese model hosted internally now represents the safest choice.
BEIJING CONSIDERS LIMITS ON ACCESS TO THE MOST ADVANCED MODELS
However, while the international spread of its models grows, China is considering limiting foreign access. According to Reuters, the Ministry of Commerce has discussed with Alibaba, ByteDance, and Z.ai possible restrictions on the most advanced AI models, along with stricter rules against technology theft and new controls on startup funding. Among the hypotheses is also a system that would subject the most advanced models to security checks or limit their use to the domestic market only.
AI AS A STRATEGIC RESOURCE
The possibility of limiting model exports reflects the growing strategic importance attributed to artificial intelligence by both Beijing and Washington. Chinese authorities, according to Reuters, particularly fear that Anthropic’s Mythos model could be used to identify software vulnerabilities and target China’s interests.
At the same time, the United States has already introduced restrictions on foreign citizens’ access to Anthropic’s most advanced models, while several American experts are calling for regulation of the use of Chinese technologies as well. In this context, AI models no longer represent just technological tools but assume an increasingly central role in economic competition and national security.
