Chinese AI Models Lead Reported OpenRouter Token Usage

Chinese models accounted for a majority of token usage in the OpenRouter figures reported by CNBC on September 26. The finding is significant for teams choosing models, but it describes traffic through a gateway rather than the entire AI market. Keep the platform, period and unit attached whenever you cite it.

Written October 1, 2026 from information published by September 26, 2026. Later product developments are outside this article’s scope.

CNBC’s September 26 report gives Chinese models 57–67% of OpenRouter tokens in the week of September 14, compared with 6–13% in February. The figures cover companies in the United States, Europe and OpenRouter’s “Global South” grouping. CNBC separately reports Vercel’s share reaching 55% in August from 11% in January; its geographic breakdown is unspecified. These are platform-supplied figures reported by CNBC, not measurements collected by Digital Applied.

The published range should remain a range. Without the underlying group weights, a midpoint would be an invented aggregate. The two platforms also cannot simply be averaged: the larger traffic population would need a different weight, and the observation windows would need to match.

A token is a unit of text processed by a model. A workload that sends long documents or runs repeated agent steps can consume more tokens than many short conversations. A change in workload mix can therefore move the volume share without a matching change in customer count. That is a reason to label the measure precisely, not a reason to dismiss it.

A hypothetical example shows the distinction. Suppose two models each receive half the tokens, but one costs twice as much per token under an otherwise identical workload. Their volume shares are equal while their spending shares are not. Real billing is more complicated because input, output and cached tokens can have different prices, which strengthens the need to keep the units separate.

Our earlier Chinese-model landscape article addresses the provider context. This report concerns gateway usage; neither should be cited as a census of every enterprise deployment, direct API customer or self-hosted installation.

We have not independently reproduced the September usage shares. A live rankings page observed later cannot establish what it showed during the reported week. To recreate that measurement, a researcher would need a dated export or archived snapshot, the grouping rules, the reporting window and a way to reconcile every included row with the stated denominator.

An honest follow-up measurement would choose its own observation period and state it explicitly. It would record whether the unit is input tokens, output tokens or both, how cached tokens are handled, which requests are included, and how model families are assigned. A label such as “Chinese model” also needs a definition: model developer, serving provider and location of processing are separate properties.

Until those details are available, do not turn platform rankings into a national-market estimate. The useful statement is narrower and still worth knowing: a particular usage report shows a substantial shift within the systems it covers.

For a buyer, adoption can help identify a model worth evaluating. It cannot answer whether that model follows your output contract, handles your documents well or meets the rules for the data you plan to send. Build a small task set from work the team actually performs and define the pass criteria before seeing the outputs.

Keep the task and acceptance criteria fixed when comparing providers. If one configuration gets extra retries or a human correction step, record that as part of its cost. Our model-routing guide explains how to connect those checks to routing decisions.

A practical next step is to identify one lower-risk workload where a different model could compete on accepted-result cost. Route only that workload after evaluation, with a fallback for failures. The permission boundary should remain the same when the model changes.

When presenting this finding internally, put the platform and period in the same sentence as the percentage. Keep the source nearby and distinguish the reported observation from your recommendation. “This candidate deserves a trial” is a defensible purchasing response; “this proves it is the best model” requires evidence this report does not supply.

Teams building a multi-model workflow can use our AI transformation service to define the tests, data boundary and fallback behavior before changing production routing.

A reported range with a named platform and period is a precise statement. Each step away from those three attachments makes it a different, larger claim. The illustrative readings below each take one step. They are patterns of citation, not things a particular writer has done.

Each enlargement is undone by putting the platform, period and unit back in the sentence with the percentage, which is the decision section’s rule and the record in the next section’s purpose.

The four also show why the verification section declines to reconstruct a ranking. Each step away from the report replaces a figure someone measured with a figure someone inferred, and an inferred figure cannot be checked against a dated export, the grouping rules or the stated denominator. The report is quotable because it is attributable; the enlargements are not, because the person who would have to defend them is the writer rather than the platform.

The decision section says to put the platform and period in the same sentence as the percentage and keep the source nearby. A short record per statistic does that once so every later citation inherits it. The platform the figure describes, and whether it is a gateway, a host or a direct provider. The period, as the source states it, so a week in September is never silently compared with a month. The unit: tokens, and whether the source says input, output or both and how cached tokens were handled, which the verification section notes is usually unstated. The population: the geographies or customer groups the figure covers. The source and the date it was read, with the figure as a range where the source gives a range. What the figure does not measure, in a line: revenue, customer count or completed work. And the sentence you may write, drafted once and reused.

An illustrative record for the OpenRouter figure reads: platform, the OpenRouter gateway; period, the week of September 14 against February; unit, tokens as reported by the platform, input-output treatment unstated; population, companies in the United States, Europe and OpenRouter’s Global South grouping; source, CNBC’s September 26 report of platform-supplied data; does not measure, spending, customers or accepted work, and not the whole AI market; sentence, in the week of September 14 Chinese models accounted for 57 to 67 percent of tokens on OpenRouter as reported by CNBC, up from 6 to 13 percent in February.

The record also settles the buying question the fourth section raises. A statistic with its record attached can nominate a candidate for the task set and the acceptance criteria; it cannot stand in for them, because none of its fields describes your documents, your output contract or the serving provider that will handle your data.