Open Source · JULY 14, 2026
Chinese open weights take 41% of Hugging Face downloads and all six top OpenRouter slots
Real production-traffic data from OpenRouter and Vercel's AI Gateway shows US model share collapsing from 70% to 30% in twelve months, with DeepSeek, Z.ai, Tencent, Xiaomi, and MiniMax now processing three times more tokens per week than American labs.
Chinese open-weight models took 41% of Hugging Face downloads this spring and now occupy all six top slots on OpenRouter, with Anthropic's Claude Opus 4.7 sitting at seventh. The frontier hasn't moved. The volume tier has.
Across the two production-traffic datasets that matter, OpenRouter and Vercel's AI Gateway, the trajectory is the same. Chinese models process roughly 18 trillion tokens per week to the US labs' 5.5 trillion, a three-to-one gap. Twelve months ago it ran the other way: US models held about 70% of OpenRouter token share in June 2025, and 30% in June 2026. Chinese open-weight share climbed from 1.2% to nearly 30% inside that same year, according to an OpenRouter and Andreessen Horowitz analysis covering more than 100 trillion tokens of anonymised usage. DeepSeek alone is now the single largest provider on OpenRouter at 16.3% of token volume; DeepSeek, Z.ai, Tencent, Xiaomi, MiniMax and Qwen together account for roughly 44% of top-ten volume.
The Vercel data tells the same story on a shorter clock. Open-weight models jumped from 11% of all AI Gateway tokens in April 2026 to 29% in June. DeepSeek went from under 1% of gateway tokens in April to 17% in May, driven by V4 Flash, which launched at $0.14 per million input tokens and $0.28 per million output, pricing that runs roughly 20–50x below comparable Anthropic models and 8–12x below Qwen 3.6 Plus and Kimi K2.6. Z.ai's GLM 5.2 grew its daily token volume about 27x and its customer count about 80x in a single week.
And yet Anthropic still commands 65% of AI Gateway spend, and 70–80% on high-stakes workloads like back-office agents. Claude Fable 5 and GPT-5.6 aren't losing customers so much as losing the low-stakes half of each customer's traffic. Chinese open-source models, as CNBC reported, run 60% to 90% cheaper than the leading Anthropic and OpenAI offerings, and buyers have noticed.
This is the split the market is quietly pricing in. Frontier labs keep the margin-rich workloads where reliability, safety-tuning and enterprise contracts justify the premium. Everything else, the routing, the summarisation, the drafting, the retries, gets handed to whichever open-weights provider is cheapest that week. Harpreet Arora, Vercel's head of agentic infrastructure, and Justin Summerville at OpenRouter have both narrated the same behaviour in different vocabularies: production teams are stratifying their stacks by task value.
The parallel isn't the 2017 ICO cycle or any of the other Chinese-tech-catching-up scripts. It's closer to the memory-chip market of the 1990s, when Korean and Taiwanese entrants pushed US incumbents up the value chain into design and specialty parts, leaving commodity DRAM to whoever could produce it at cost. Frontier stays American. Inference, increasingly, doesn't.
Sources
- The real AI race may no longer be at the frontier (TechCrunch)
- Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge (CNBC)
- DeepSeek enters the fight for token volume, Anthropic continues to dominate spend (Vercel)
- What's next for Chinese open-source AI (MIT Technology Review)
- Open-Weight Models from China Are Capturing a Growing Share of AI Usage (Trending Topics)