Chinese AI models have gone from a rounding error to nearly half of US enterprise token traffic in about a year, and OpenAI and Anthropic are watching the shift happen in real time.
Every week since February 8, Chinese-origin models have carried at least 30% of the enterprise token volume moving through OpenRouter, the API routing platform used by thousands of AI startups and, increasingly, large companies choosing between providers. According to a CNBC investigation published July 7, that share hit a weekly peak of 46%. A year ago it averaged 11%. In the first half of 2025 it was just 4.5%.
DeepSeek alone accounts for 17.6% of all tokens routed through the platform, roughly 5.13 trillion tokens a week, making it the single largest vendor on OpenRouter, ahead of every US lab. Alibaba's Qwen ranks second among the Chinese providers at 13.9%, or 2.77 trillion tokens weekly. The whole platform has exploded too. OpenRouter carried about 5 trillion tokens a week in April 2025. By April 2026 that figure topped 20 trillion.
Why Enterprises Are Switching
The reason isn't mysterious. If you've priced out an enterprise AI contract this year, you already know it: DeepSeek's V4 Flash model costs $0.14 per million input tokens through its own API. OpenRouter lists some providers even lower. OpenAI's GPT-5.5 costs $5.00 for the same million tokens. That's not a discount. That's a different business.
Moonshot AI didn't help matters for the incumbents. On July 16, the company released Kimi K3, a 2.8 trillion parameter open-weight model with a million-token context window, and it plans to publish the full model weights on July 27. Once that happens, enterprises can fine-tune and self-host a frontier-class model without sending a single API call to San Francisco.
This isn't hypothetical. Airbnb chief executive Brian Chesky has said the company runs Alibaba's Qwen model for customer service work, calling it fast and cheap. DoorDash executives have pointed to similar savings, routing lower-complexity tasks to cheaper, Chinese-built models instead of paying premium API rates to Anthropic or OpenAI for every query.
Washington's Pushback, and What Comes Next
House committee investigators questioned Airbnb and Cursor in April about their ties to Chinese AI developers, including Moonshot AI, DeepSeek, MiniMax, Alibaba, Zhipu AI, ByteDance, Tencent and Baidu, requesting internal security reviews and details on whether customer data touched those systems. The concern isn't paranoia. South Korea's Personal Information Protection Commission found that DeepSeek had transferred user prompts to Beijing-based Volcano Engine Technology without consent. The Pentagon isn't waiting for a similar finding at home. The fiscal 2026 defense authorization law directed intelligence officials to create guidelines requiring DeepSeek's removal from intelligence community systems, and New York, Texas and Virginia have already barred the app from state government devices or networks.
Here's the problem for anyone hoping regulation solves this fast. Banning a hosted service is straightforward. Banning an open-weight model a company already downloaded and is running on its own servers is not. Commerce still hasn't added DeepSeek to its Entity List. The delay looks less like an oversight than Washington weighing how hard it wants to escalate a fight its own companies are already settling with their wallets.
For OpenAI and Anthropic, the number that matters isn't the ideological one. It's the token count. Every workload that shifts to a $0.14 model instead of a $5 model is revenue that never shows up on a renewal. Both companies have cut prices this year, but neither has closed the gap to DeepSeek's or Qwen's rates, and every enterprise buyer who notices becomes a data point in next year's procurement budget.
The AI moat argument used to be about which lab built the smartest model. Increasingly it's about which model is cheap enough to run at scale, and for a growing share of corporate America, that answer isn't coming from San Francisco anymore. If you manage a model budget, that's the number worth watching, not whichever lab wins the next benchmark.
Also read: China Publicly Rejects Anthropic's Claim That Alibaba Stole Claude's Data, Hugging Face Says an Autonomous AI Agent Swarm Breached Its Systems Over a Weekend, Qwen3.8 Teases a 2.4 Trillion Parameter Open Model as Alibaba Chases Kimi K3
