American companies are shifting their AI spending toward Chinese models, with usage rates surpassing 30% since February and peaking near 46% in some weeks, according to OpenRouter data. The main driver is straightforward: Chinese AI providers like DeepSeek and Z.ai deliver comparable performance at a fraction of the cost.
This price gap is too significant for even cautious corporate clients to ignore. CNBC reports that open Chinese AI models cost 60 to 90 percent less than top-tier solutions from OpenAI and Anthropic. For example, the startup Lindy switched all its AI traffic from Anthropic’s Claude to DeepSeek in June and expects to save millions annually. Companies processing large volumes of code, translations, and internal documents find that cost savings often outweigh brand loyalty.
The trend is clear in the numbers. In the first half of 2025, Chinese models accounted for just 4.5% of AI usage among US clients of OpenRouter, but now that figure consistently exceeds 30%. A year ago, it was around 11%. This is not just experimental testing by development teams but a substantial redistribution of workloads.
Quality comparisons have fueled this shift as well. CNBC notes Z.ai’s GLM 5.2 model trails Anthropic’s Claude Opus 4.8 by only 1% on a key AI agent benchmark, while costing roughly five times less. Analysts at Brookings estimate the technological gap between leading Chinese and American models at six to nine months. In AI terms, this is barely a gap-more like the normal cycle between major releases.
Use cases for Chinese AI models among American businesses
American and international companies have not fully migrated away from US-based AI providers. They deploy Chinese models for non-critical tasks such as code generation, translation, summarization, and initial customer request triage. But for applications where industry compliance, audit trails, and data retention guarantees matter, solutions like Claude and ChatGPT remain dominant.
This aligns with how corporate cloud services function. Platforms like AWS Bedrock and Google Vertex AI encourage a multi-model approach, where customers allocate tasks across various providers instead of tying themselves to one. In this setup, cheaper models handle routine workloads, while pricier models are reserved for high-stakes cases where errors have costly consequences.
Trust also plays a key role. In regulated sectors like finance and healthcare, legal teams scrutinize model provenance and data routing as closely as accuracy. This has positioned Chinese models as dependable workhorses, not full replacements for established American AI leaders. Companies are learning to treat AI as a toolkit rather than betting on a single favored vendor.
The situation took a sharp turn in June when Anthropic, under pressure from the US Commerce Department, temporarily restricted access to its two most powerful AI models outside the United States-before lifting those limits. This brief interruption served as a stark reminder that enterprise risks extend beyond price and quality to include export controls. If key AI services can be arbitrarily cut off in certain regions, CFOs scramble to have contingency plans ready.
Potential regulatory changes for Chinese AI models from Beijing
Ironically, Beijing itself may close the window of opportunity for Chinese AI models abroad. Over the past month, China’s Ministry of Commerce has met with major players like Alibaba, ByteDance, and Z.ai to discuss protecting AI tech as a national strategic asset. CNBC reports the government considers restricting foreign access to its advanced AI models.
Such restrictions would create a mirror image of the US export controls currently shaping the market. American businesses use Chinese models precisely because they are accessible, affordable, and sufficiently capable. But once geopolitical barriers start governing access, cost savings will be offset by the need to hedge regulatory risks and insurance.
This dynamic plays into a broader trend. Meta’s LLaMA models have already shown open-source and openly available AI can press down prices even without the Chinese factor. Meanwhile, corporate platforms increasingly offer customers choice among dozens of providers. The AI market is heading toward a future where premium model margins depend primarily on superior security, integrations, and performance in complex tasks-not simply the developer’s nationality.
The real test of this shift will come in the next few quarters as it becomes clear whether China imposes export limits on its leading AI technologies. If it doesn’t, expect mounting price pressure on OpenAI and Anthropic. If it does, enterprise customers will likely accelerate their move toward a multi-model strategy-relying on backup suppliers from the US, Europe, and open-source communities alike.
* Meta, the owner of LLaMA, is designated as an extremist organization in Russia, where its operations are banned.