Six months ago, Waymo said it was handing out 250,000 robotaxi rides a week. Now the company says that number is 450,000. Read those two figures back to back and you get a better sense of where autonomous driving actually is than from any conference keynote.
I want to sit with that jump for a second, because I think most people reading about robotaxis are still filing them under “someday.” Someday doesn’t usually double in half a year.
What the numbers actually describe
A ride is a small, boring, verifiable thing. Someone opened an app, a car showed up, and it took them somewhere. Multiply that by 450,000 and repeat it every week. On top of that, Waymo says it plans to expand into 12 new cities by 2026, with Dallas and San Diego among them.
That’s the whole verified picture, and I’m going to stick to it rather than dress it up. But the shape of it tells you something: this is no longer a demo. It’s an operation with a service footprint, and it’s growing city by city rather than all at once.
Why I keep pointing at this on an AI explainer site
Most of what we cover here is software agents. Things that read your email, book your travel, sift your spreadsheets, file your tickets. Useful, occasionally uncanny, and almost entirely confined to a screen.
A robotaxi is the same basic idea wearing a much heavier coat. It perceives, it decides, it acts, and it does all of it without a person approving each step. The difference is that when a chat agent gets something wrong, you get a bad paragraph. When a two-ton agent gets something wrong, you get a headline.
So I treat robotaxi ride counts as the most honest AI benchmark we have. Nobody grades a robotaxi on a leaderboard. It’s graded by whether a stranger in Dallas gets home.
Three things this teaches non-technical readers
- Real autonomy is measured in repetitions, not capabilities. The impressive part isn’t that a car can drive itself once. It’s that it does it hundreds of thousands of times a week with the same outcome.
- Geography is a feature. An agent that works in one city doesn’t automatically work in the next. Each new market means new streets, new weather patterns, new local driving habits, new rules. Expanding to 12 cities is a technical claim as much as a business one.
- Scale is the hard part, not the idea. Plenty of teams have built agents that work in a controlled setting. Very few have built ones that hold up when the setting stops being controlled.
The part that isn’t about technology
The framing around Waymo lately has less to do with sensors and more to do with rules. Who gets a permit. Which city says yes. What happens after an incident. Who’s liable when the driver is a piece of software.
I don’t have verified spending figures to share on that front, so I won’t pretend to. What I can say is that regulatory attention is the natural next chapter for any AI system that stops being a product and starts being infrastructure. Once a service is moving hundreds of thousands of people a week, city councils and state agencies become part of the engineering problem. Not metaphorically. Literally, in the sense that a rule can shut down a fleet faster than a bug can.
This is the thing I’d flag for anyone trying to understand AI without a technical background: the ceiling on agent deployment is increasingly legal and political, not computational. The models are the easy conversation now. Permission is the hard one.
How I’d watch this from here
If you want a simple way to track whether autonomous agents are actually working in the physical world, ignore the announcements and watch two numbers. Rides per week, and cities served. Both are difficult to fake and both are easy to check.
Right now those numbers say 450,000 and growing, with 12 more cities on the plan for 2026. That’s not a promise about the future. It’s a description of a service running today, in specific places, for specific people, with an empty driver’s seat.
Whatever you think of self-driving cars, that’s the clearest signal we’ve gotten that AI agents have started doing work that leaves the screen. The rest of the industry is still arguing about benchmarks. This one is arguing about parking.
đź•’ Published: