A September 26 check of a saved Vercel AI Gateway export uncovered a decisive gap: the text-lab data ends partway through July 29, before the month at issue begins. From that snapshot, the desk cannot calculate what share of August traffic went to models from Chinese labs—whether counted by tokens, requests or estimated spend. The finding limits what this data can verify; it does not establish that Chinese models fell short of any reported share.
The question was narrow: for August 2026, how much Vercel gateway traffic went to models made by identified Chinese labs? The desk compared the saved export with Vercel’s account of its classifications and OpenRouter’s description of its rankings. The evidence cutoff was September 26 at 15:30 UTC. This was a review of public documents and the saved data, not a fresh download of August results or a test of either gateway.
Three shares, not one
Vercel’s lab export separates requests, tokens and spend. A request is a call sent to a model; tokens measure text processed. Vercel calculates spend at published market-rate prices, including for customers who bring their own API keys. It is therefore a normalized estimate, not a record of what every customer paid. Each measure answers a different question about use of the gateway, and a lab’s position can change sharply between them.
DeepSeek’s July 28 entry makes the distinction concrete. Its token share was much larger than its request share, and its estimated-spend share was smaller still. Those numbers describe one lab on one day, not an August total for models from Chinese labs. But they show why a claim about a majority of tokens cannot simply be retold as a claim about a majority of calls or dollars.
DeepSeek’s share of text-lab tokens on July 28 in the saved Vercel export.
DeepSeek’s share of text-lab requests on the same day.
DeepSeek’s share of text-lab spend on the same day.
There is also a monthly arithmetic problem. Each Vercel export row is a lab’s percentage of one measure on one day. To get an August share weighted by traffic, the calculation would need the corresponding daily totals. Adding the daily percentages would not work, and an ordinary average could give a quiet day the same influence as a busy one. The checked export supplies percentages rather than those totals; tokens, requests and spend would each require their own denominator.
Who made the model?
Before calculating a Chinese-origin share, a researcher needs a rule for identifying the models in that group. The maker of a model and the company serving it through a gateway need not be the same. Open-weight status—whether a model’s weights are available—is a third distinction, not a nationality. The desk’s review had no finalized Chinese-origin model roster. Substituting a list of serving providers would change the question from who made the models to who delivered them.
Vercel’s earlier Production Index illustrates the difference. For open-weight token and spend shares, it grouped four labs serving their own models: DeepSeek, MiniMax, Moonshot and Z.ai. For enterprise adoption, it counted open-weight models regardless of provider, a broader rule. Vercel called the serving-lab measure conservative and put it at 29% of gateway tokens but under 4% of spend. Those figures cover June traffic. Neither that month nor that open-weight classification answers the August Chinese-origin question.
OpenRouter has a different denominator
OpenRouter documents a rankings API that accepts an August date range and returns daily token counts for its top 50 public models. Every model outside each day’s top 50 goes into one “other” row. A defined origin rule could classify the named models, but it could not assign every token in that combined row. The rankings also exclude private models, private endpoints and zero-data-retention traffic. The desk examined the API documentation, not an August response.
That model-ranking response provides token counts, not matching model-level request or spend series. OpenRouter’s separate app rankings do include request counts, but apps cannot be treated as model-origin groups. Its token counts also come from upstream providers’ own tokenizers. OpenRouter cautions that a token from one provider is not directly comparable with a token from another. Those limits matter even for a comparison confined to OpenRouter.
Matching the calendars would not match the populations: Vercel’s export concerns traffic routed through Vercel AI Gateway, while OpenRouter ranks eligible public traffic on its own service. An OpenRouter share cannot fill in Vercel’s missing August result, and combining the gateways would not produce a whole-market share. A checkable Vercel answer needs a defined origin group, August observations and the traffic totals needed to weight each of the three measures.
Reader comments
Newest comments first. Replies stay oldest first.