Chinese AI models now hold between 30 and 46 percent of enterprise API token usage flowing through US developer platforms, according to a CNBC investigation published July 7. That same day, Reuters reported that China’s Ministry of Commerce was in talks with Alibaba, ByteDance, and Z.ai about restricting overseas access to their most advanced models. On July 8, US lawmakers confirmed they are investigating Airbnb and Cursor’s parent company Anysphere over their use of Chinese AI. Both governments are now moving to control which AI developers can use — and neither is done moving.
The Numbers Tell the Story #
Eighteen months ago, Chinese models held roughly 2 percent of OpenRouter token volume. Today, they have cleared 30 percent every single week since February 8, 2026, peaking as high as 46 percent in CNBC’s enterprise survey. US models — Google, OpenAI, and Anthropic combined — held around 70 percent of OpenRouter traffic in June 2025. By July 2026, that share sits at roughly 30 percent.
The tipping point was the week of February 9, 2026. Chinese models processed 4.12 trillion tokens on OpenRouter versus 2.94 trillion for US models — the first week Chinese systems outpaced American ones on the platform. DeepSeek V4 Flash now commands 16.3 percent of all OpenRouter token volume, more than any single US provider.
This Is a Cost Story, Not a Performance Story #
Open-source Chinese models are 60 to 90 percent cheaper than US frontier models on output tokens. Here is what that gap actually looks like:
| Model | Output Cost (per 1M tokens) |
|---|---|
| DeepSeek V4 Flash | $0.28 | | Qwen 3.6 Max | $1.20 | | GLM-5.2 (MIT) | $4.40 | | Claude Sonnet 5* | $10.00 | | GPT-5.5 | $30.00 |
That 100x gap between DeepSeek and GPT-5.5 on output tokens is not a rounding error — it is a business model. And in agentic workloads, where a single task might call the model 50 to 200 times, the gap compounds with every loop. GLM-5.2 scores 62.1 percent on SWE-bench Pro, beating GPT-5.5’s 58.6 percent, at one-sixth the cost. The performance excuse for paying a 10x premium is getting harder to make.
Now Both Governments Are Moving #
The political situation just became significantly more complicated — from both directions.
On the US side, the House Select Committee on China and the House Committee on Homeland Security launched a joint investigation in April into companies using Chinese-developed AI. Airbnb’s use of Alibaba’s Qwen and Cursor’s integration of Moonshot AI’s Kimi model are both under scrutiny. The stated concern is data security and ideological influence. The practical challenge: you cannot ban model weights that are already freely distributed on the internet.
On the Chinese side, Reuters reported July 7 that Beijing is considering restricting overseas access to its most advanced AI models — including models not yet released. The proposed framework is tiered: basic open-source models would require a filing, advanced models would go through security review, and frontier models could be restricted to domestic use only. Nothing has been decided, but the direction of travel is clear.
The result is a bifurcated AI market where your production dependencies could be severed by either government. That is a different kind of infrastructure risk than most teams are pricing in.
What to Actually Do #
Your risk profile depends almost entirely on how you deploy, not which model you choose. The guidance breaks down into three tiers.
If you are using hosted Chinese API endpoints (deepseek.com, Qwen direct API, etc.) for any prompt containing proprietary code, customer data, or regulated information: stop. Data processed through those endpoints falls under China’s National Intelligence Law. For EU users, it also violates GDPR Article 46. This is not a hypothetical compliance risk.
If you are routing through a US or EU cloud provider that serves Chinese model weights, your data exposure is lower, but model behavior may still reflect training decisions made in Beijing. Run evals on your specific workloads and document what you tested.
If you are self-hosting open weights on your own infrastructure, the data sovereignty problem largely disappears. Self-hosting breaks even economically at around 2 million tokens per day versus API pricing. Below that threshold, API access through a US-based proxy like OpenRouter or Vercel AI Gateway is typically the better trade.
The teams handling this well are routing per workload: Chinese models for high-volume, low-stakes agentic loops; US models for customer-facing outputs and anything touching regulated data. Airbnb said it uses Chinese models “only through approved US-based service providers” for non-sensitive workloads. That is a reasonable middle position for now — but do not hard-code any single model into your stack while the regulatory picture is still in motion.