Chinese AI firms gain traction as Anthropic and OpenAI restrict models

Chinese AI firms gain traction as Anthropic and OpenAI restrict models

US export controls on flagship AI models are backfiring spectacularly, pushing American companies straight into the arms of cheaper Chinese alternatives

The US government tried to build a wall around its best AI models. Companies responded by walking through the open door next to it.

Since the US enacted export controls on Anthropic’s Mythos 5 and Fable 5 models in June 2026, Chinese AI companies have surged from bit players to serious contenders in the American enterprise market. Data from OpenRouter shows Chinese models now account for over 30% of AI token usage among US firms on a weekly basis, peaking at 46%. Before February 2026, that number averaged just 11%.

The great migration

The math here is brutally simple. Chinese AI models from companies like DeepSeek, Z.ai, and Alibaba’s Qwen are priced 60-90% lower than their top US equivalents. Some are priced at roughly 5% of what comparable Anthropic models charge.

Advertisement

DeepSeek has led individual model usage share during the adoption surge. But the most dramatic story belongs to Z.ai’s GLM-5.2 model, which saw a 27-fold increase in daily token volume within its first full week after launching in late June 2026.

The shift isn’t limited to experimentation, either. AI startup Lindy moved its entire operation from Anthropic’s Claude to DeepSeek, citing substantial cost savings.

These aren’t inferior knock-off models being adopted purely on price. Chinese open-weight and open-source models have become genuinely competitive on global benchmarks and usage metrics, including on platforms like Hugging Face where developers vote with their downloads.

How US policy created its own problem

The export controls on Anthropic’s flagship models were designed to protect American AI advantages. When the US government temporarily suspended access to Mythos 5 and Fable 5, it sent a clear signal to every CTO in America: your access to the best US models can be revoked at any time, for reasons that have nothing to do with your business.

Chinese open-source models, by contrast, can be downloaded, self-hosted, and run without depending on anyone’s permission.

What this means for investors

Companies that are adopting Chinese models stand to improve their margins significantly. If your AI inference costs drop by 60-90%, that flows almost directly to the bottom line. Investors should watch for earnings reports from AI-dependent companies that quietly switched providers, because those margin improvements will show up.

Interest from Alibaba, ByteDance, and Z.ai has already sparked discussions about potential export regulations from China. If Beijing decides to impose its own restrictions on model access, the companies that migrated to Chinese AI could find themselves in exactly the same position they were trying to escape, just with a different government holding the switch.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.

Chinese AI firms gain traction as Anthropic and OpenAI restrict models

Chinese AI firms gain traction as Anthropic and OpenAI restrict models

US export controls on flagship AI models are backfiring spectacularly, pushing American companies straight into the arms of cheaper Chinese alternatives

The US government tried to build a wall around its best AI models. Companies responded by walking through the open door next to it.

Since the US enacted export controls on Anthropic’s Mythos 5 and Fable 5 models in June 2026, Chinese AI companies have surged from bit players to serious contenders in the American enterprise market. Data from OpenRouter shows Chinese models now account for over 30% of AI token usage among US firms on a weekly basis, peaking at 46%. Before February 2026, that number averaged just 11%.

The great migration

The math here is brutally simple. Chinese AI models from companies like DeepSeek, Z.ai, and Alibaba’s Qwen are priced 60-90% lower than their top US equivalents. Some are priced at roughly 5% of what comparable Anthropic models charge.

Advertisement

DeepSeek has led individual model usage share during the adoption surge. But the most dramatic story belongs to Z.ai’s GLM-5.2 model, which saw a 27-fold increase in daily token volume within its first full week after launching in late June 2026.

The shift isn’t limited to experimentation, either. AI startup Lindy moved its entire operation from Anthropic’s Claude to DeepSeek, citing substantial cost savings.

These aren’t inferior knock-off models being adopted purely on price. Chinese open-weight and open-source models have become genuinely competitive on global benchmarks and usage metrics, including on platforms like Hugging Face where developers vote with their downloads.

How US policy created its own problem

The export controls on Anthropic’s flagship models were designed to protect American AI advantages. When the US government temporarily suspended access to Mythos 5 and Fable 5, it sent a clear signal to every CTO in America: your access to the best US models can be revoked at any time, for reasons that have nothing to do with your business.

Chinese open-source models, by contrast, can be downloaded, self-hosted, and run without depending on anyone’s permission.

What this means for investors

Companies that are adopting Chinese models stand to improve their margins significantly. If your AI inference costs drop by 60-90%, that flows almost directly to the bottom line. Investors should watch for earnings reports from AI-dependent companies that quietly switched providers, because those margin improvements will show up.

Interest from Alibaba, ByteDance, and Z.ai has already sparked discussions about potential export regulations from China. If Beijing decides to impose its own restrictions on model access, the companies that migrated to Chinese AI could find themselves in exactly the same position they were trying to escape, just with a different government holding the switch.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.