The Signal
OpenRouter's State of AI report found that Chinese open-weight models averaged about 13% of weekly token volume and approached 30% in some weeks in its dataset. That is strong evidence of meaningful adoption on one routing platform. It is not evidence that Chinese models process 60% of United States business AI traffic.
What OpenRouter measured
OpenRouter is a marketplace and routing layer that lets developers access models from multiple providers through a common interface. Its State of AI report analyzes activity observed on that platform, primarily from November 2024 through November 2025. The report covers a large and useful sample, but it is not a census of global enterprise AI use.
Within that sample, Chinese open-weight models averaged about 13% of weekly token volume. Their share rose sharply at several points and reached nearly 30% in some weeks. That pattern shows that developers will move meaningful workloads when a model offers a compelling mix of capability, price, openness, and availability.
Why open weights change routing behavior
Model choice is becoming a portfolio decision. Teams can reserve frontier proprietary models for tasks where marginal reasoning quality matters, then route extraction, classification, drafting, coding support, and other repeatable work to lower-cost options. Open-weight models also give sophisticated teams more deployment control and a path to private hosting.
This weakens the idea that one laboratory automatically captures every layer of AI economics. A strong base model can be valuable while routing platforms, inference providers, fine-tuning systems, evaluation tools, and application owners compete for the rest of the margin. Lower switching costs make measured reliability and total cost more important than brand alone.
The limits of the dataset
OpenRouter describes its analysis as observational. The platform serves a particular developer population, more than half of usage is outside the United States, and geography is inferred from billing information. The report does not publish the cross-tab needed to claim a specific share of United States business tokens for Chinese models.
Token volume also is not the same as revenue, number of customers, production criticality, or economic value. A high-volume, low-cost model may process many tokens while a premium model earns more revenue from smaller but more valuable tasks. Usage share should therefore be read as adoption evidence, not a complete market-share calculation.
What operators should do
The practical signal is to design for model substitution. Maintain evaluations that reflect real tasks, track latency and failure rates, and compare costs at the workflow level. A routing policy should include data-handling rules, provider jurisdiction, incident response, and a fallback when a model or endpoint changes.
For investors, the durable question is where switching remains difficult. Proprietary data, workflow integration, distribution, trust, and verified outcomes can retain value even as inference becomes cheaper. Open models increase pressure on undifferentiated model access while expanding the number of economically viable AI applications.
Security and governance belong in the benchmark
A model that wins on price and benchmark scores may still be unsuitable for a regulated or sensitive workflow. Operators need to know where prompts are processed, what is retained, which subprocessors are involved, how abuse monitoring works, and whether contractual controls match the data classification. Geographic origin alone is an incomplete proxy for those questions, but ignoring jurisdiction and supply-chain dependency is equally incomplete.
The mature buying process combines task-level evaluation with governance. Test accuracy on representative work, red-team failure modes, document data flows, define human review, and maintain an exit path. Open weights can offer greater control when a team has the infrastructure and expertise to run them. A hosted endpoint can offer faster deployment. The right decision depends on the full operating system around the model, not a leaderboard snapshot.
Benchmarks also decay as models and workloads change. Save a versioned evaluation set, record the model and provider version, and rerun it when routing rules change. Cost should be measured per successful task, including retries and human correction, rather than per input token alone. That process converts a volatile model market into an operational advantage that can be reviewed and improved.