There's a marketplace called OpenRouter that pools 400-plus AI models and lets developers pick one per job. Its value is that it shows actual paid token usage from over ten million developers, not benchmark scores but where real money goes, which makes it the closest thing to an honest usage ranking. The June 2026 chart has a surprise at the top: China's DeepSeek is number one at 16.3%, ahead of Anthropic, Alibaba, Google, OpenAI and Tencent. When people objected that Chinese users were padding the numbers, OpenRouter broke out US users only, and it's more striking, not less. The share of tokens US users spend on Chinese models climbed from 4.5% in early 2025 to 30% in 2026, peaking near 46%. Roughly half of America's usage is now Chinese AI. The reason is price: DeepSeek's cheap model costs one thirty-sixth of GPT-5.5. Behind that sit cheaper power, labor and construction, plus an MoE design that fires only part of the brain per question, compression tricks for long documents, and 'distillation', training on the answers of top US models. But here's the twist: heavy usage, little revenue. Chinese models collect under 5% of the fees. The US Congress has already opened an investigation.
Why it matters · The usage crown and the revenue crown have split apart. Chinese models win volume on price while US models keep nearly all the money, a warning for anyone pricing the AI-model business off market share.
Worth asking · Would you use Chinese AI at 1/36th the price if quality matched?