OpenRouter Is Running 10 Quadrillion Tokens a Year. The real number is 194 trillion.
A ten quadrillion run rate is one week multiplied by fifty-two. The measured figure is 194 trillion tokens, and the doubling time is the actual finding.
On November 2 Deedy Das posted that OpenRouter had crossed a ten quadrillion annual token run rate. Ten quadrillion is a sixteen-digit number, and it is the least reliable figure in this post.
Start with what was actually measured. In the week of October 26, OpenRouter processed 194 trillion tokens. That is a count. The ten quadrillion is that week multiplied by fifty-two.
194T
Tokens processed in the week of 26 October. The measured number.
10.5 wks
How long token volume takes to double, on Das's own figures.
10Q
The annual run rate. One week, extrapolated across a year.
Why the big number is the weak one
A run rate assumes the next fifty-one weeks look like the one you just had. For a business growing slowly that is a reasonable simplification. For one whose volume doubles every ten and a half weeks it is wrong in both directions at once: it overstates the past, because no earlier week in the year came close to this one, and it understates the future, because by February the weekly figure should be roughly double again.
So the honest version is not "OpenRouter processes ten quadrillion tokens a year." Nobody has processed ten quadrillion tokens. The honest version is that it processed 194 trillion in a week, and that this week was roughly twice the size of a week in mid-August.
Das
"Congratulations to OpenRouter for crossing 10 QUADRILLION annual token run rate today! Token volumes continue to 2x every…"
The curve is the finding
Three checkpoints from the same source, across one year: roughly 1.5 quadrillion a year earlier in 2026, more than 4.5 quadrillion by August, ten quadrillion in the week of October 26. Each one an extrapolation, but they are extrapolations made the same way, so the shape between them holds even where the absolute values are soft.
Why a router is where this shows up first
OpenRouter does not make a model. It sits between applications and the labs and picks which model serves each request. That position is why the number exists at all: no single lab can see this curve, because no single lab sees the traffic it loses to a competitor on price or latency.
How much traffic that is becomes clearer from the earlier checkpoint. Back when the run rate was around 1.5 quadrillion, Das put it at 15 to 30% of Google's API volume, 20 to 40% of OpenAI's, and more than half of Microsoft Azure Foundry's. Those comparisons are from a point when the number was roughly a sixth of today's, and they are his estimates rather than disclosures by any of the three. Even heavily discounted, they describe a router carrying a serious fraction of what the labs serve directly.
It is also, fairly obviously, why Stripe agreed to buy it in August. A company whose business is sitting in the middle of transactions bought the company sitting in the middle of inference. Reported prices run from $7B to $10B and none of them are confirmed, so treat that range as rumour rather than fact.
The strategic point does not depend on the price. Model leadership has changed hands repeatedly this year and will again. Routing volume has only gone one direction, because every one of those handovers is a reason to route rather than commit.
What to take from somebody else's chart
Quote the measured number, not the run rate. If you are putting this in a deck, 194 trillion tokens in a week is defensible and 10 quadrillion a year is an extrapolation somebody will eventually check. The difference costs you nothing and saves the slide.
Assume your own curve is steeper than your budget. If industry volume doubles roughly every ten weeks, a flat annual AI budget set in January is wrong by October in a way no amount of per-call optimisation fixes. Forecast on a doubling, then argue about the multiplier.
Treat routing as the default, not the hedge. The reason this curve lives at a router is that nobody wants to be married to one lab while the leaderboard keeps turning over. If your stack has one model hardcoded in it, you are the counterparty to that bet.
Measure your own doubling time. Most companies cannot say what their token volume did last quarter, let alone what it doubles in. That number is knowable, it is cheap to instrument, and it is the only version of this chart that tells you anything about your own costs.
Somebody else's sixteen-digit number is interesting. Yours is actionable.
Measure your own curve.



