The Undercut
Chinese AI models captured nearly half of US enterprise AI traffic in just six months, while Washington is yet to produce a coherent response.
On July 7, CNBC published platform-level data showing that Chinese AI models now handle between 30% and 46% of enterprise API traffic on major US developer platforms. A year ago, that share sat at 4.5%. The next day, President Trump cancelled a signing ceremony for a new AI executive order, the second cancellation in two months, citing concerns about hindering American competitiveness. Chinese models captured the traffic, and nothing in Washington’s response has slowed the shift.
The tokens don’t lie
The CNBC investigation drew on data from OpenRouter, one of the largest API routing platforms. Chinese models have held above 30% of the platform’s enterprise token volume every week since February 8, peaking at 46% in June. US-origin models from Anthropic, OpenAI, and Google have collectively fallen to 35.7%. DeepSeek commands 17.6% of routed tokens as the platform’s single largest vendor, and Alibaba’s Qwen follows at 13.9%.
Lindy, an AI automation company, migrated all of its traffic from Claude to DeepSeek and projects millions in savings. Coinbase runs 1,200 AI agents on Chinese models and halved its AI spending. Uber burned through its annual AI budget in four months as AI coding tool usage surged beyond the company’s ability to tie it to shipped products. Harpreet Arora, Vercel’s head of agentic infrastructure, told CNBC that enterprise teams now route tasks to the cheapest model that clears the quality bar, and Chinese models consistently win that trade.
Good enough and cheaper by far
Open-weight Chinese models run 60 to 90% cheaper than their US counterparts, according to OpenRouter’s Justin Summerville. DeepSeek V4 Flash costs $0.14 per million input tokens, while GPT-5.5 costs $5.00, a differential of roughly 35-to-1. Programming tasks account for more than half of OpenRouter’s total traffic, and Chinese models perform disproportionately well on coding benchmarks at those price points. GLM-5.2, released in June by Zhipu AI under the Z.ai brand, scored within a percentage point of Anthropic’s Opus 4.8 on agentic benchmarks at roughly a fifth of the cost. In its first week on Vercel, GLM-5.2’s daily token volume grew 27-fold, and its customer count grew 80-fold. Xiaomi’s MiMo-V2-Pro recently became the most-used model on OpenRouter by weekly token volume, processing 4.21 trillion tokens per week for a 21.1% platform share. OpenAI holds 7.5%.
The June 12 suspension of Anthropic’s Fable 5 under export controls pulled one of the most capable US models offline for 18 days. Chinese AI usage by US firms surged during the blackout, as enterprises discovered that cheaper alternatives cleared their quality bars. OpenAI’s GPT-5.6 spent its first 13 days in a government-coordinated preview limited to roughly 20 vetted organizations before launching publicly on July 9. During both windows, Chinese open-weight models faced no comparable restrictions on global distribution. Access to Fable 5 returned on July 1. The CNBC data, published six days later, showed that Chinese model share had not retreated since the shift began in February.
Washington dithers
On July 8, Trump cancelled a scheduled signing ceremony for a new AI executive order, the second time in two months that he pulled the same document. The first cancellation, on May 21, followed lobbying from tech industry allies. A draft of the order included up to 90 days of government security review before model release, a requirement that the industry argued would hinder competitiveness. In July, Trump offered the same reasoning himself, telling reporters that he didn’t want to do anything that would “get in the way” of America’s lead over China. The CNBC investigation had published the day before, placing Chinese model share at nearly half of US enterprise API traffic.