Over the past few weeks, companies around the world have been shifting away from proprietary U.S. solutions toward Chinese open-weight models, driven by significantly lower costs and performance that approaches the market’s high end. The trend is visible in a sharp rise in usage volumes on the AI Gateway of a San Francisco–based cloud platform: open-weight models now account for 29% of token volume—nearly three times their share in April.
Adoption and key figures
Reported metrics show rapid movement. Since mid‑June, daily token volume for Zhipu’s GLM-5.2 on the Vercel gateway has increased fiftyfold, according to the platform. GLM-5.2 runs at roughly one‑fifth the cost of Anthropic’s Claude Opus 4.8. At the same time, DeepSeek’s V4 Flash—a lighter-weight variant of the V4 Pro—has become the single largest model on the gateway, capturing more than 20% of traffic on Wednesday versus about 15% a month earlier.
Why Chinese open-weight models are gaining ground
Two technical and economic forces are driving the shift. First, many proprietary U.S. models require premium cloud subscriptions and per-token billing, which quickly inflate costs for high‑intensity applications. Second, open-weight models can be downloaded freely and run on local hardware, lowering operating expenses for companies able to host instances in‑house or leverage less expensive cloud providers.
Reevaluation of spending and
(The source text ends mid-heading. If you provide the rest of the paragraph or heading, I’ll translate the remainder in the same style.)
