OpenRouter, the marketplace developers use to route requests across AI models, lists 97 providers. In the seven days to October 30, ten of them carried 54% of the tokens that flowed through it. The top 25 carried 83%. The remaining 72 split the last 17%, which works out to about 66 billion tokens a day each. That is real traffic, but it is a small slice of a market where a few names take most of the volume.

What this is and isn't. The figures below come from Jimmy Research, an independent tracker that sums OpenRouter's own per-provider “tokens processed” series over a trailing seven days. It lists the top 25 providers individually and reports the other 72 as one group, so we can state the group's total and average but not each provider's number. The daily figures are the weekly totals divided by seven. OpenRouter is one marketplace, not the whole inference market, and its traffic skews toward developers who shop across models. It says nothing about what these companies earn elsewhere.

The numbers

Total volume for the week was about 192.9 trillion tokens, or roughly 27.6 trillion a day. OpenAI led at about 2.7 trillion a day, followed by DeepSeek (2.1T), Tencent (1.7T), Together (1.5T), InferenceNet (1.4T), Amazon Bedrock (1.3T), Wafer (1.2T), DeepInfra (1.1T), StepFun (1.0T) and Nvidia (0.9T). Those ten account for about 104 trillion tokens a week, or 54%. The 25th provider, Google AI Studio, still handles about 311 billion a day. Below that line sit 72 providers with a combined 33 trillion tokens a week. That is about 4.7 trillion a day in total, about 66 billion a day on average, and every one of them is below the 25th-ranked provider.

Money is even more concentrated

Jimmy Research also estimates revenue, and says the figures are illustrative and not checked against OpenRouter's books. On its estimate, the top ten providers earn about 81% of the roughly $91.5 million of weekly spend, and Amazon Bedrock alone earns about 26%. The bottom 72 share about 6%, roughly $5.4 million a week across all of them. Tokens are not revenue: the providers at the top are often serving pricier models, and the tail is often serving cheap or free ones. Treat the revenue split as a direction, not a measurement.

Why a long tail is hard to run

Several of the big names are model makers serving their own models, not hosting companies, so “provider” covers very different businesses. For a company that rents GPUs and resells capacity, volume matters: idle hardware still has to be paid for, and the debt used to buy it comes due on a schedule. A provider handling tens of billions of tokens a day on a marketplace where price is the main thing buyers compare has little room to charge more.

Our read. This is the pattern we expect in a market financed ahead of demand. A few providers collect most of the volume, and many others compete for what is left. We think a lot of the tail will merge, specialize or close, and we think it's likelier to happen in 2027 than in 2026. This is our view, not something the data shows. OpenRouter is a snapshot of one marketplace, and a small provider here may do most of its business somewhere else.

What this means, in plain English

Behind the long list of AI providers, a handful handle most of the work. If you build on AI, the question to ask any smaller provider is how long it can keep prices low and the lights on. If you invest in or follow this market, expect fewer names to still be around once the debt-funded buildout has to start paying for itself.

What we're watching

Whether the top-10 share keeps rising in later weeks. Whether any large provider in the tail announces a merger, a pivot or a shutdown. And whether OpenRouter itself publishes provider-level volume for all 97, which would let anyone check the tail one name at a time.

Sources: Jimmy Research, OpenRouter provider data (7 days to Oct 30, 2026), OpenRouter rankings.