A single data point is describing where enterprise AI spending is actually going better than most of this quarter's earnings calls. According to CNBC, citing OpenRouter data, the share of tokens US companies route to Chinese AI models has stayed above 30% every week since 8 February 2026, peaking as high as 46% — up from an average of 11% over the prior twelve months, and just 4.5% in the first half of 2025. For any technology leader budgeting AI spend, that curve demands a decision.
Why it's happening: price, mostly
The driver isn't ideology — it's economics. OpenRouter's data lead told CNBC that open-weight Chinese models can run 60% to 90% cheaper than leading Anthropic and OpenAI systems. Vercel reported that Z.ai's GLM-5.2 saw the fastest model adoption it tracked in 2026 — daily token volume up roughly 27x and customer count up roughly 80x in its first week. And the capability tax that used to come with the discount has narrowed sharply: GLM-5.2 reportedly landed within a percentage point of Anthropic's Opus 4.8 on one closely watched agentic benchmark, at about a fifth of the cost.
The reason is structural. As token prices for frontier models rise and usage explodes, costs have become unpredictable — CNBC reported that Uber burned through its entire annual AI budget in four months as coding-tool usage outran anything it could tie to a shipped product. As Vercel's head of agentic infrastructure put it to CNBC, when a task doesn't need the best model, teams route it to the cheapest one that's good enough — and Chinese models are winning that trade.
The honest caveats
Before you migrate anything, three things the reporting itself is careful about:
- OpenRouter is one venue, not the whole market. The 30–46% figure reflects a developer-heavy routing platform. It does not cleanly separate production enterprise traffic from experimentation, and the broader market still leans on Claude and GPT for the hardest work.
- "Good enough" is task-specific. The pattern isn't wholesale replacement; it's routing. Teams keep frontier models for complex reasoning and send commodity tasks to cheaper ones.
- Geopolitics cuts both ways. The US is weighing tighter rules on model releases, and — per Reuters reporting — China has discussed restricting exports of its most advanced models. A model you standardise on today could face access friction from either government tomorrow.
A decision framework
Rather than a yes/no on "Chinese models," decide per workload across four axes.
1. Data sensitivity. What data enters the prompt? Regulated personal data, customer records, or proprietary code demand extra scrutiny of where inference runs and what's retained. For EU firms this is a GDPR question before it's a cost question. The strongest argument for open-weight models is also the most overlooked: you can run them on your own infrastructure, keeping data entirely in your control — often a better sovereignty story than any US-hosted API.
2. Task criticality. A model drafting internal summaries and a model making credit or safety decisions sit in different risk classes. Route accordingly, and keep a human in the loop where the cost of being wrong is high.
3. Total cost, not sticker price. Factor migration effort, evaluation, prompt re-tuning, and the operational cost of self-hosting. A 90% cheaper token is not 90% cheaper in production — but at scale the difference is still often decisive.
4. Portability. The winning architecture isn't a single provider; it's a routing layer that lets you switch models per task and swap providers when price, performance, or policy shifts. Build the abstraction and today's geopolitical uncertainty becomes a config change rather than a re-platforming.
The takeaway
The rise of capable, cheap, open-weight models is genuinely good news for enterprises — it restores leverage that had drifted entirely to a few frontier labs. But "cheaper" is a workload decision, not a company-wide flag. Classify your use cases, keep sensitive data on infrastructure you control, and build for portability so no single lab or government owns your roadmap.
Where Educatifu fits
We help companies design multi-model AI architectures — routing layers, self-hosted open-weight deployments, and the governance to use them safely under EU rules. If your AI bill is climbing and you want options without betting the roadmap on one provider, get in touch.