US companies now route more than 30% of their AI tokens through Chinese open-weight models every single week, and have since February 8, 2026 — with the share peaking as high as 46%. That is up from an 11% average across the previous twelve months, and as low as 4.5% in the first half of 2025. The finding, from a CNBC investigation of usage data on the developer platform OpenRouter, marks the clearest sign yet that the cost advantage of open Chinese models has moved from a curiosity to the default for a large slice of American engineering teams.
Key facts
- Chinese models' share of US enterprise tokens on OpenRouter has stayed above 30% every week since Feb 8, 2026, peaking near 46% — versus an 11% average the prior year.
- The startup Lindy moved 100% of its traffic from Anthropic's Claude to DeepSeek, saving millions within months.
- GLM-5.2 (from Zhipu AI) reported roughly 80x customer growth and 27x daily token volume on Vercel in its first week.
- Reported by CNBC; usage rankings visible on OpenRouter.
The background a non-expert needs: "tokens" are the unit AI models bill by — chunks of text going in and coming out. Enterprises spend real money per token, and at scale that bill runs into the millions. For two years the assumption was that you paid a premium to the top Western labs because their models were simply better. That assumption is breaking. Chinese open-weight models — meaning anyone can download and run them — now score comparably on real coding benchmarks while costing 60 to 90 percent less. GLM-5.2, from Beijing's Zhipu AI, ships under a permissive MIT license and posts coding scores in the same range as GPT-5.5, at a fraction of the price.
The most concrete illustration is the AI startup Lindy. Its CEO Flo Crivello described the economics bluntly, saying the cost curve had "crashed to the ground," and moved the company's entire workload off Claude and onto DeepSeek — saving millions of dollars within months. What is emerging alongside this is an "advisor model" pattern: teams default to a cheap Chinese model for the bulk of requests and escalate to an expensive Western frontier model only for the hardest cases. Think of it like hiring a competent generalist for everyday work and calling in the specialist consultant only when the problem genuinely demands it.
Why it matters: this is the structural market story of the quarter. It reframes every frontier launch — including Grok 4.5's aggressive pricing — as a defensive move against a floor that Chinese open-weight models have already set. It also ties directly to the ongoing debate about whether closed frontier models are overpriced for the median enterprise task.
The honest caveat: token share is not revenue share, and OpenRouter is one platform, not the whole market. Many of these tokens are experimental or cost-optimized workloads that were never going to pay frontier prices anyway, and regulated industries face data-governance questions about routing traffic through Chinese-origin models regardless of price. The 46% peak is a striking number, but it measures where the cheap, high-volume tokens go — not where the highest-stakes work lands.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)