Why Are US Startups Routing 30%+ of Tokens to Chinese AI Models? The Inside-China Read

In the eight weeks ending August 2026, five Chinese flagship models launched in succession — and for the first time, the essay claims, the five most-called models on OpenRouter were all Chinese, despite 47% of the platform’s users being US developers [unverified]. US enterprises’ token consumption on Chinese models reportedly rose from 4.5% eighteen months ago to a stable 30%+, peaking at 46% [unverified], as companies like Coinbase and DoorDash shifted routine workloads to cheaper, open-weight Chinese models.

The numbers

• 8 weeks, 5 flagship models
• OpenRouter: top 5 most-called models all Chinese (July 2026) [unverified]
• 47% of OpenRouter users are US developers; 6% Chinese
• US-enterprise token share on Chinese models: 4.5% → 30%+ stable since February 2026, peak 46% [unverified]
• GLM-5.2: mid-June, topped CodeArena; Kimi K3: July 16, 2.8T parameters; DeepSeek-V4-Flash: late July; Qwen3.8-Max: early August, 2.4T-parameter MoE; Seedance 2.5: early August
• Stanford 2026 AI Index: US–China top-model performance gap narrowed to 2.7% [unverified]
• Lindy (25-person startup): API costs down ~90% after migrating from Claude to DeepSeek-V4; bills previously exceeded total payroll
• Coinbase: AI spending nearly halved; internal survey — 91% of engineers don’t need frontier-level performance
• Tenstorrent: costs down 5× after switching from Claude
• Cost comparison: ~$25 (Claude) vs ~$0.18 (DeepSeek) for the same coding workload; some Chinese models cost ~1% of US closed flagships
• Qwen APP update August 7: free research features, scheduled tasks, PC-controlling office assistant

The take inside China

The story Chinese tech media are telling: a release blitz with no precedent. Eight weeks, five flagships — covering text, code, multimodal, and video — while OpenAI and Anthropic update flagships on a yearly cadence.

The essay’s argument: this isn’t brute-force compute. It’s a different technical bet — MoE architecture plus inference-side optimization plus open weights — instead of one giant closed model recouped via high API prices. The payoff shows in adoption: the named switchers are telling (Lindy, Coinbase, DoorDash, Tenstorrent), and the candid counterpoint gets airtime too: US engineers note gaps remain in extreme deep reasoning. The real pattern, the essay argues, is “layered mixing” — Chinese models for the 80% of routine work, US flagships for the hard 20%. Its pointed question: if China eats 80% of workload volume, can the remaining 20% sustain US closed-model valuations?

Meanwhile, the Qwen APP update on August 7 — free research, scheduled tasks, a PC-controlling office assistant — takes the model layer straight to ordinary users; viral Excel-and-PPT demos on Douyin and Bilibili did marketing no benchmark could.

Why it matters outside

If the cost gap (reportedly up to two orders of magnitude) and the open-weight advantage hold, Western AI pricing and business models face real pressure — and enterprises gain leverage via multi-vendor hedging. The trajectory also complicates the assumption that export controls can hold back Chinese frontier AI. English coverage of the same wave: The Register, Reuters, China Daily.

Sources

• The Register: China turns up the heat with open model blitz
• Reuters: Alibaba unveils its most capable AI model
• China Daily: Zhejiang AI models gain ground

Provenance & disclosure. Originally published in Chinese on our WeChat channel on 2026-08-08 (“谁在给硅谷换AI底座？8周5款中国大模型正在改写规则”); drafted with AI assistance under human editorial direction. Translated to English on 2026-08-26 (AI-assisted, human-reviewed). The OpenRouter share figures, named-company quotes, and the Stanford 2.7% figure are the essay’s citations, kept with [unverified] tags; the release wave itself is corroborated by the English sources above. The essay carries a pro-China framing; that framing is preserved as translated opinion, not asserted as fact. This is translated commentary — not a SigPulse measurement. Our first-party measurements live in the dispatches and the /data/ ledger.