What's going on with Waymo? It turns out: A lot!
It's been a year since we last wrote about Waymo, and a lot has happened. Back in April, Waymo hit 250,000 paid robotaxi rides per week, up from just 50,000 a year ago. We've seen quite a few headlines recently and wanted to check in on how the old robotaxi was faring. It turns out, quite well! In particular, the safety numbers explain why they're pretty much crushing everyone else. Let's dive in and take a look, then check out their latest expansion news, and finally wrap with where all this robotaxi stuff is headed.
The Safety Stats
When we covered Waymo in September 2024, some shockingly impressive safety stats had just been revealed (by Timothy B. Lee, of the fantastic AI blog Understanding AI). In the time that's passed since, that data has more or less stayed consistent.
In fact, a massive, recent peer-reviewed study just analyzed Waymo's performance over 56.7 million rider-only miles (that's with zero human behind the wheel). The results?
- 85% fewer suspected serious injuries compared to human drivers (39% to 99% confidence interval).
- 79% fewer overall injuries.
- 91% fewer airbag deployment crashes.
But here's where it gets wild. When they broke it down by crash type:
- 96% reduction in intersection crashes (where 27% of fatal crashes happen).
- 92% fewer pedestrian crashes.
- 82% fewer cyclist crashes.
- 82% fewer motorcycle crashes.
- 93% fewer single vehicle crashes.
As Kevin Roose tweeted, "I don't know how people can look at Waymo's safety data (or just ride in one) and not feel like we live in an age of miracles. The car! Drives! Itself!"
Now, these stats comes with one big caveat mind you: this data sample size is probably too small to accurately predict Waymo’s impact on “fatal crashes.” Waymo itself says, “as fatal crashes are the most rare type of crashes, there is not yet sufficient mileage to make statistical conclusions about fatal crashes alone.”
Right now, the fatality rate in the U.S. is ~1.3 per 100 million miles driven (which includes highway driving, so likely lower on regular streets, where Waymos are currently driving). So it remains to be seen if Waymo’s fatality rate is lower or higher than that. They’ve already driven over 100M miles, but they’ve only begun testing on highways recently, so the data can’t be perfectly mapped yet. This is the trend to watch, but even still: the fact that they produce significantly less crashes in general is massive progress.
Here's how Waymo ensures these safety numbers are meaningful and sustained: As Waymo published in a paper back in June, they know their driver is ready for the road by assessing 12 distinct criteria for safety determination. This includes...
- Collision avoidance testing against behavioral benchmarks.
- Predicted collision risk per million miles compared to human benchmarks.
- Rules of the road compliance.
- Vulnerable road user interactions.
- High severity scenario assessment.
Most interesting is how Waymo's approval process, which involves actively controlling risk through:
- Gradual rollout windows with phased deployment.
- Scale limits (maximum allowed mileage).
- Geographic distribution of mileage based on local risk factors.
- Real-time suspension capability for unexpected behavior.
Why do they do all this? Because the "absence of unreasonable risk" is the current industry standard definition of safety for autonomous vehicles. Rather than declaring they've achieved this standard, Waymo uses the 12-criteria framework for how the industry should begin to evaluate it. They admit their proposed criteria "are not intended to be sufficient for the determination of absence of unreasonable risk" and that "minimum benchmarks will need to be agreed upon broadly by industry and regulators" before anyone can make such a determination.
In other words, the autonomous vehicle industry is still debating not just whether current systems are safe enough, but what "safe enough" even means—a fundamental question that remains unanswered. In the meantime...
The Expansion Blitz
Armed with these stats, Waymo is moving fast...
Coming Soon:
- Nashville 2026: Partnering with Lyft, with Lyft's Flexdrive managing fleet operations.
- SFO Airport: Just got approval for three-phase rollout (safety driver → staff testing → commercial ops).
- Arizona Public Transit: Partnering with Via to integrate robotaxis into Chandler's public transit this fall.
- Planned launches in Atlanta, Miami, Washington D.C.: All launching in 2026.
And in case you need a refresher, where is Waymo operating today?
Already Operating:
- San Francisco, Los Angeles, Phoenix, Austin.
- Over 4 million rides in 2024 alone.
- Over 95M miles driven in these four cities as of June 2025.
- Now expanding from downtown cores to suburbs.
Now, let's talk growth trajectory:
- May 2023: 10,000 weekly rides.
- May 2024: 50,000 weekly rides (5x growth).
- April 2025: 250,000 weekly rides (another 5x).
As of today’s date (September 21, 2025), we know that Waymo has surpassed 250,000 weekly rides, but not by how much. We also know that they’ve reached over 10M paid rides in total.
If this rate continues, they'll hit 1 million weekly rides by early 2026. For context, that's still tiny compared to Uber's 250 million weekly rides globally, but in the cities where Waymo operates? They're already taking meaningful market share.
P.S: Did you know that if you order an Uber in the cities and regions where Waymo operates, you’re likely to get offered one? It costs about the same as a human driver (except no tip, obviously), probably to stay competitive and not deflate the market. Also, Waymos aren’t cheap to make… around $300,000 a car by some estimates (although these numbers might be completely made up; it seems the hard math ranges from $160K to $300K depending on the labor involved). But, let's look at the ROI potential on these bad bots, even on the high end of $300K, with a little back of the napkin math:
- Waymo fleet size: 2,000 vehicles as of Aug. 2025.
- Total weekly rides: ~250,000 rides per week.
- Average Uber ride cost: ~$30 (based on current market data). Could be as high as $50-60 in Los Angeles, depending on the time of day, or as low as $17.
Now, based on the total rides per week, each Waymo currently completes 125 rides per week (250,000 ÷ 2,000).
To recover the ~$300,000 manufacturing cost at $30 per ride: A single Waymo needs to complete ~10,000 rides to break even.
- At the current rate of 125 rides/week, it would take 1.5 years to break even.
- To break even in just 1 year, each Waymo would need to do 192 rides per week (1.5x current rate).
Now, this calculation only covers the initial manufacturing cost and assumes Waymo charges similar prices to Uber for a steady period of time. It doesn't include:
- Ongoing operational costs (maintenance, electricity, insurance, depot operations).
- Software development and updates.
- Fleet management costs.
- Profit margins.
- Or an entire vehicle getting fire-bombed during a protest, like what's happened before!
... So the actual profitability timeline would likely be longer when including these operational expenses (or, conversely, could be shorter if the actual per car cost is much lower than estimated). That said, the ~10,000 rides figure gives you a solid baseline for the manufacturing cost recovery alone.
Oh, and if you thought Uber was out of the self-driving game, guess again: Uber just invested $300M in Lucid Motors to launch a new premium robotaxi service "in a major US city" next year.
The Competition Is Getting Real (Sort Of)
So in case you haven't noticed, there's A LOT of companies working on self-driving cars. Not as many as there was back in 2021, or 2019, but check out a few of the recent headlines we've seen lately:
- Amazon's Zoox finally joined the party in September, launching free robotaxi rides in Las Vegas. Their approach is... different. No steering wheel, no pedals, seats face each other. It looks like a toaster on wheels, and honestly? That might be genius. If you're in the strip, Zoox is currently offering free rides between Top Golf, Area15, and various Strip hotels.
- Chinese Players Are Going Global: WeRide and Pony AI are partnering with Singapore's Grab and ComfortDelGro to launch robotaxi services in Singapore in 2026. They're deploying 11 vehicles for testing this month on 12km routes in the Punggol neighborhood. Baidu's Apollo Go is eyeing Malaysia and Singapore launches too.
- And let's not forget about Tesla: After launching in Austin on June 22 with much fanfare, Tesla's robotaxi has been... messy. They had 3 crashes in their first month with just 12 vehicles on the road, all while safety monitors were still in the cars. But, the app got 2 million downloads on day one when it went public September 4th, so the demand is currently there (FYI the app is still just a waitlist last we checked). And despite Musk promising to remove safety monitors by end of 2025, they're still riding shotgun in every vehicle. Also, y'know what's not really cool? Tesla's trying to hide crash details from regulators. We get it, crashes happen, but let's not hide them. Though, let's be fair: Waymo is also getting stricter about what data it shares with regulators, too.
Now, credit where credit is due to Tesla: They've accumulated over 3.6 billion miles of FSD data, with users now driving about 8 million miles per day. That's an absolutely massive dataset—more than any other company has collected with consumer vehicles. Morgan Stanley's Adam Jonas recently tested FSD himself, and said it drove 99% of his recent 1,400-mile road trip, calling it "a game changer."
The latest FSD v13 is showing about 489 miles between necessary interventions, according to crowdsourced data. That's a significant improvement from earlier versions, and it shows Tesla is making progress on its shaky record. But here's the reality check: Tesla's own head of FSD says they need to hit 700,000 miles between critical interventions to match human safety levels. They're currently at about 1,680 miles between critical interventions for frequent users. That's a big gap.
How does Waymo stack up? Sources report Waymo has ~17,000 miles between critical disengagements. Keep in mind though: since Waymo has no human safety driver, "disengagements" occur either when the system recognizes a flaw and pulls over/stops, or when a remote human operator takes over. Even still, if Waymo has ~17,000 miles per intervention, that's ~10x better than Tesla, but Waymo still needs a 41x improvement to reach human-level safety (17,000 → 700,000). Tesla, meanwhile, would need a 417x improvement to reach human-level safety (1,680 → 700,000).
Now, to be fair to Waymo, ARK Research estimates Waymo averages ~500,000 miles between police-reported collisions as of Q4 2024, which is much closer to the human benchmark when measuring actual accidents rather than interventions. And as of data from last year, over 22 million rider-only miles, Waymo achieved 84% fewer airbag crashes, 73% fewer injury crashes, and 48% fewer police-reported crashes compared to human drivers.
Since then, the data has remained impressive (feel free to download it and check it out yourself). So if you look at it from the perspective of measuring actual collision rates rather than interventions, Waymo appears to be much closer to 700,000 mile human benchmark
And as Waymo wrote in their "absence of unreasonable risk" paper, "no methodology, on its own, is capable of providing appropriate coverage with adequate confidence" and so they use"signal-based evaluation" where "signal from multiple methodologies are overlaid" to get a comprehensive risk picture.
No matter what the ACTUAL gap is, what's needed to fill it? More training data, probably. While Tesla collects real-world data, companies like Waabi are using generative AI and simulation to train autonomous trucks almost entirely in virtual environments. Their "Waabi World" simulator has achieved 99.7% accuracy "in matching simulated scenarios to real-world outcomes", and they're planning fully driverless trucks by end of 2025... all while spending a fraction of what competitors have on real-world testing.
As Waabi CEO Raquel Urtasun notes, "No one has yet built the Matrix for self-driving cars," but simulation-first approaches may be the future as they're safer, cheaper, and can test millions of edge cases that would take decades to encounter on real roads.
Where's All This Going? Drivers Becoming Fleet Managers
We all know about Elon Musk's plans to allow anyone to become a fleet manager for their Tesla robotaxi. Here's what nobody's talking about for Waymo: During Alphabet's Q1 earnings call, CEO Sundar Pichai dropped this gem: Waymo is considering "future optionality around personal ownership as well."
Translation: You might eventually buy your own Waymo that drives itself (similar to Elon's plans for becoming your own robotaxi fleet manager). Imagine owning a car that can earn money for you while you sleep by giving rides (or multiple cars giving multiple rides). Or never worrying about parking because your car just... goes home.
There's Just One Problem: A Policy Vacuum
As Wharton Professor Ethan Mollick pointed out, "It seems like there is not enough of a policy response to the existence of self-driving cars." With 2.4 million injured and 40,000 killed in US accidents annually, an 85% reduction in serious injuries should be front-page news. But TBH, it's kind of... not?
Cities are starting to notice, though. NYC just approved Waymo testing last month (with safety drivers for now). Singapore's government is also fast-tracking approvals. And even traditionally car-centric cities are asking "what's our robotaxi strategy?"
As a result of this, the US NHTSA unveiled a new framework that judges self-driving cars on performance (like collision avoidance) rather than hardware (lidar, radar, etc.). This is huge news for Tesla's camera-only approach, which now gets equal regulatory treatment to Waymo's sensor-heavy setup.
Key changes of the framework include:
- Removing requirements for manual controls (steering wheels, pedals, windshield wipers).
- Streamlining approvals for vehicles without human driver features.
- Creating a single national standard vs. state-by-state rules.
The goal: "unleash American ingenuity" while maintaining safety standards. Now, critics worry the standards might not be strict enough, but the changes could definitely accelerate deployments across the industry. Either way, the long term goal of this framework is to create a national standard for robotaxi regulation, instead of 50 regulatory frameworks from all 50 states (let alone independent rules in every city).
There are still kinks to be worked out, such as preventing robotaxis from entering active emergency zones. And if we're not careful in how we design autonomous networks of robotaxis, it could lead to a lot more traffic in major cities, like what's happened with ride-hailing (though as with any complicated system, the increased congestion impact from ride-hailing depends on lots of human factors, so we can assume the same of robotaxis).
We're guessing the traffic problem will be more complicated if multiple thousands or millions of fleet managers are all trying to optimize to make the most money possible... that seems like "Carmaggedeon" 2.0 in the making!
What Happens Next
We're going to make a declaration, at least for this leg of the race: the robotaxi wars are over. Waymo won. At this point, everyone else is fighting for second place.
That said, the real battle is just beginning: Who gets to define what "car ownership" means in 2030? If Waymo can reduce crashes by 90%+ AND it and Tesla can potentially sell you a car that pays for itself, the entire auto industry is about to get disrupted harder than taxis got disrupted by Uber. Not overnight, of course; it'll probably look similar to the transition to electric vehicles, which is still happening (good recent data on this) but incredibly... bumpily.
The safety data isn't just good PR... it's regulatory gold. Every city official seeing these numbers has to ask: If we can prevent 85% of serious injuries, how can we NOT approve robotaxis?
It might end up being that the general public's low opinion of robotaxis (and AI in general, culminating in a new Luddite movement) will become the largest barrier to mass adoption. Right now, residents in Santa Monica are annoyed at the beeping sounds Waymos make (no joke), nor do LA residents like that Waymos just idle around a lot in neighborhoods and even sometimes block driveways. As expected, current Uber and Lyft drivers have staged protests against their newborn robo-brethren, and most interesting of all, a new study from HBS found that people tend to blame robot drivers more than human ones because the technology is still "unsettled" and new.
So you can expect three barriers to full-scale robo-taxi adoption to persist in the near term: the need for even more data (for better simulation, and for safe highway usage), the need for regulatory clarity (for use across more cities, as well as for dealing with fleet management), and the need to improve "the vibe."
We'll see which of these three will get solved first...
