Self-Driving Cars: Pros and Cons at 270M Miles

A self-driving car drives itself with no person responsible for the driving task. Whether that is good or bad stopped being a thought experiment once one company published enough miles to argue about with data.

The pro is measurable. Across more than 270 million fully autonomous miles, Waymo reports its driver was involved in 95 percent fewer serious-injury-or-worse crashes than human drivers in the same areas. The con is measurable too. The safety analysis behind it covers five of Waymo's metro service areas, the record took seventeen years to produce, and many of the companies that set out to build it alongside Waymo are gone.

Dmitri Dolgov, Waymo's co-CEO and one of the founders of the Google self-driving car project that began in 2009 and became Waymo in 2016, went through both halves on stage with Sequoia partner Konstantine Buhler at Sequoia Capital's AI Ascent 2026 in May. What follows is that ledger, checked against Waymo's published data, and then the part that transfers to anyone shipping an autonomous product of any kind.

The short version

The pros, on the record:

• Fewer serious crashes, measured and published, on the roads where the service runs.
• The largest safety gains go to pedestrians, cyclists and motorcyclists, the road users who absorb the most harm today.
• Perception that does not blink, get tired, or look away, including in situations a person cannot physically see.
• Once the driver is validated, expansion is fast and getting faster.

The cons, on the record:

• Availability is local. Every new city is work, not a software update.
• The sensor envelope has hard edges, and the system is honest about them.
• The hardware is still expensive enough that cutting its cost is a stated engineering goal.
• The record took seventeen years and the survival rate among the companies that tried is brutal.
• The safety evidence is the operator's own analysis of its own operating areas.

The pro that counts: the harm being displaced is enormous

The case for autonomy starts with the baseline it replaces, and the baseline is grim. The World Health Organization puts global road traffic deaths at approximately 1.16 million a year in its July 2026 update, revised down from the 1.19 million reported in its 2023 status report, and road injury remains the leading cause of death for people aged 5 to 29. More than half of those deaths are vulnerable road users: pedestrians, cyclists and motorcyclists.

Dolgov's version of the number is the one that lands: "somebody loses their life to a crash on our roads every 26 seconds." He treats it as the company's founding provocation rather than a statistic. "the status quo is not okay," he said, and called challenging it something that matters to everyone at the company.

That framing matters for reading the safety data, because the comparison is not against a perfect driver. It is against the actual one.

What the published data says, with the scope attached

Waymo releases updated safety analysis on a rolling basis, and the three most recent releases make the trend legible:

• March 2026: over 170 million fully autonomous miles, 92 percent fewer crashes causing serious or fatal injuries, 82 percent fewer crashes involving any reported injury.
• June 2026: more than 220 million miles through the end of March, 94 percent fewer serious or fatal injury crashes, and Atlanta reaching statistical significance on its own with 5.4 million miles and zero serious-or-fatal injury crashes where a human benchmark predicted roughly 1.2.
• September 2026: more than 270 million miles through the end of June, 95 percent fewer serious-injury-or-worse crashes, 82 percent fewer injury crashes overall, and 841 injury-causing crashes that did not happen.

At the AI Ascent conversation in May, Dolgov was working from the 170 million mile release. His summary was that the Waymo driver is "more than 13 times safer than a human driver when it comes to serious injury causing collisions" in the cities where it operates, and he converted it into the form that makes it real: "that 13x reduction means that we are preventing a serious injury every 8 days."

Two honest caveats belong next to those numbers, and neither of them cancels the result.

The first is scope. This is Waymo's analysis of five of its own metro service areas, compared against human crash benchmarks for surface streets in those same areas over the same period. It is a strong claim about Atlanta, Austin, Los Angeles, Phoenix and San Francisco, not a claim about driving in general. The second is that the comparison is made regardless of fault, which cuts in Waymo's favor rather than against it: a vehicle can be blameless in a crash a better driver would have avoided entirely, and counting those anyway is the harder standard.

What makes the whole thing usable is that the methodology is published. A reliability number published without its denominator and its date cannot be checked by anyone. If you are building anything where failure has a cost, how you define and publish your own reliability is the decision underneath every claim you will later want to make.

The second pro: perception that sees what a person cannot

Two moments from Dolgov's account show the difference between a careful driver and a system with different senses.

Dolgov describes a rider on an electric scooter who lost control and fell directly in front of a Waymo. The car was able to "swerve and brake and everybody walked away." A person might have managed the same. Many would not have.

The second is stranger. A Waymo was stopped at a red light. A bus crossed the intersection and halted, partially blocking it. The light turned green, the car began to move, and then it started behaving defensively toward a pedestrian it had no line of sight to. A pedestrian then emerged from behind the bus, and the car nudged around them.

Dolgov says the behavior surprised him until the team worked out what had happened. "it doesn't see through solid objects," he said. The lidar had bounced a signal underneath the bus and returned a sparse read of a person's feet moving, and that was enough for the system to infer a pedestrian and predict where they were going.

Read that story twice, because it is both a pro and a con. The pro is real: no human driver infers a pedestrian from the motion of feet under a bus. The con is in the same sentence. The system does not see through the bus. It got lucky with geometry, and the team only understood the capability after the fact. An autonomy envelope has edges, and the people building it are often finding those edges by observation rather than design.

The con on the map: autonomy is local

An under-reported limitation of self-driving cars is that they are not a product you can download. They are a deployment.

At the time of the talk Waymo was operating in eleven cities, and Dolgov's description of adding one is a list of unglamorous work: collect the data, characterize the environment, validate the driver, and start the conversation with the local community. "it's on us to earn the trust of" the people there, he said, because the product is new to them regardless of how many miles it has elsewhere.

The technical half of that is improving fast. "the driver is generalizing incredibly well" now, he said, so a new city is increasingly a validation exercise rather than a re-engineering one. That shift is visible in the pace. In February 2026 Waymo opened to the public in Dallas, Houston, San Antonio and Orlando on the same day, the first time it had launched multiple cities simultaneously, taking it to ten commercial metro areas. By September 2026 it had opened public service in Denver, San Diego and Tampa, then Las Vegas two weeks later, reported as its fifteenth market. London and Tokyo are announced.

Still, the con stands. If you do not live in one of those metro areas, none of the safety record is available to you, and the reason is not software.

The con in the bill of materials

Dolgov described the sixth generation of the Waymo Driver, and the way he described its goals is the most direct thing said on stage about cost. The focus, he said, has been on "performance but also on simplification, drastic cost reduction, and high scale volume production."

Companies do not set drastic cost reduction as a headline engineering goal for something that is already cheap. The hardware stack is a real constraint on how fast this can reach the roads where most road deaths happen, and 92 percent of those deaths are in low- and middle-income countries rather than in wealthy metro areas with a dense sensor supply chain. This is the same economic shape that governs every physical AI product: the intelligence gets cheaper on a predictable curve, and the thing it has to move through the world does not.

The con that killed the field: the long tail is the product

Waymo's timeline is the clearest argument here. The project set itself two goals at the start in 2009: drive 100,000 miles in full autonomy, and complete ten routes of 100 miles each around the Bay Area, each one end to end with no intervention. About a dozen people, working days on hardware and nights on testing. It took roughly 18 months.

Seventeen years after that start, the same organization is at 270 million miles. The distance between those two facts is the entire problem.

Dolgov's explanation of why the field kept producing hype cycles is the most transferable thing he said. Autonomous driving, he explained, has always had the property that "it's very easy to get started, but it's very difficult to take it all the way to a real product, full autonomy and superhuman performance." So every genuine breakthrough, convolutional networks, transformers, large language models, triggers the same reaction. It "reshapes the early part of the curve", but it "doesn't change the long tail of it."

The companies that read the reshaped early curve as a shortened total curve are mostly not here. Argo AI, funded by Ford and Volkswagen, shut down in October 2022 when no further investor appeared, and both parents moved their attention to driver assistance. Uber sold its self-driving unit to Aurora in December 2020 and put $400 million into the buyer, which is what offloading a unit you have decided not to finish funding looks like.

General Motors went further and stopped funding Cruise's robotaxi business in December 2024, redirecting the work to driver assistance for personal vehicles after years of spending on it.

Dolgov's survival strategy was not technical. It was believing the mission mattered while "not looking for a kind of easy wins or quick solutions or silver bullets", and staffing the team's stamina to the real length of the problem.

The part that transfers: de-risk in sequence, then scale in parallel

Here is the line from the conversation that is worth more to a founder than any of the safety statistics. Asked where the business goes next, Dolgov said Waymo had "transitioned from an intentional sequential de-risking of the driver and key parts of the business to rapid parallel global commercialization."

That sentence describes a deliberate two-phase shape, and both words in each phase are doing work.

Sequential de-risking is the slow phase, and the sequencing is intentional. Waymo took eight years from starting fully autonomous operations to serving the public in four cities. It did not scale geographically during that period, because each new city multiplied an unvalidated risk.

Parallel commercialization is what the slow phase buys. Four cities in one day. Over 20 million fully autonomous rides by May 2026, half of them in the preceding seven months. 100 million additional autonomous miles between the March and September safety releases alone. None of that speed was available earlier, and all of it was available at once.

The mistake this framing catches is the common one: running both phases at half speed simultaneously. Scaling a product whose core risk is unresolved, and de-risking it slowly because you are also busy scaling. That produces a company that is neither safe nor fast.

Two of Dolgov's architectural points explain why the first phase cannot be skipped and then retrofitted.

The first is about where safety lives. "safety has to be the non-negotiable foundation", he said, and it has to be built into the model architecture, the training and evaluation recipe, and the mindset of the team from day one. His reason is precise rather than moral: "how you go about the first 90% is totally different problem" from how you reach the next nines. The techniques are not on a continuum. Getting quickly to competent and getting slowly to reliable are different engineering programs, and a team optimized for the first is not one step away from the second.

The second is about architecture. Waymo's foundation model is end to end, powering three related tasks, the driver, the simulator and the critic. But Dolgov is blunt that a plain end-to-end system is not sufficient for the destination: "there's a massive difference between using end-to-end versus purely relying on it." Waymo augmented the learned representation with structured intermediate representations, and the reason is entirely about the last mile. Those structures are what make runtime validation of the agent possible, and what make closed-loop evaluation, closed-loop training and richer reinforcement learning reward functions practical. He is explicit that you might not need any of it for a driver-assist system, a prototype, a demo or a small deployment, and that it is unavoidable if you intend to go all the way.

That is the lesson with the widest reach. The machinery that lets you validate, evaluate and constrain an autonomous system has to be designed into it, because it is structural rather than additive. A team that ships the capable version first and plans to add the safety layer later is not behind schedule. It has built something with no place to put one. If you are giving agents real authority in your own product, raising autonomy one workflow at a time is the same discipline at a smaller scale, and the sequencing argument is the same one.

What to do this week

• Write down the single risk that would make your product unshippable if it went wrong at scale, in one sentence. If you cannot name it, that is this week's work and nothing below matters yet.
• Decide explicitly which phase you are in. Sequential de-risking or parallel scaling. Pick one and tell the team which, because running both at half speed is the failure mode.
• Check whether your architecture can be validated at runtime, not just evaluated offline. If nothing inside the system can inspect what the model is about to do, you have a demo architecture.
• Take your most impressive capability demo and count how many of its runs you selected. The unselected failure rate is your actual number.
• For the workflow with the highest consequence, write down the reliability level you need before you choose the architecture rather than after.
• Publish one reliability number you are prepared to stand behind, with the denominator and the date attached. Being first in your category to publish a real one is a position competitors cannot copy with a landing page.

If you want the operating system around this, from sequencing autonomy to pricing the work your agents do, that is what AI Operating System for Startups is built to teach.

Sources

• Waymo's Dmitri Dolgov: 20 Million Rides and the Road to Full Autonomy, Sequoia Capital's AI Ascent 2026, published 1 May 2026. The conversation this article distills, and the source for the hype-cycle argument, the perception stories, the city-launch playbook, the sixth-generation hardware goals and the sequencing framing. Profiles: Dmitri Dolgov and Konstantine Buhler.
• Waymo's published safety analyses, the source for every percentage and mileage figure in this article: September 2026 (270 million miles, 95 percent, 841 crashes), June 2026 (220 million miles, Atlanta) and March 2026 (170 million miles, one serious-injury crash prevented every 8 days), with the full methodology at Waymo Safety Impact.
• World Health Organization, for the global road death baseline: the road traffic injuries fact sheet updated 20 July 2026 (approximately 1.16 million deaths a year, 92 percent of them in low- and middle-income countries) and the earlier Global status report on road safety 2023, whose 1.19 million figure is the 2021 estimate Dolgov's 26-second line tracks.
• Waymo's simultaneous four-city launch: Ready to Ride: Dallas, Houston, San Antonio, and Orlando, February 2026. The September expansion into Denver, San Diego and Tampa: Waymo, TechCrunch and Reuters, followed by Las Vegas on 14 September 2026.
• The consolidation of the field: Argo AI's shutdown reported by TechCrunch and Reuters, Uber's sale of its self-driving unit to Aurora reported by TechCrunch and The New York Times, and General Motors ending its funding of Cruise's robotaxi business in December 2024, reported by TechCrunch and Reuters.
• Crash reporting obligations for automated driving systems: NHTSA's Standing General Order on Crash Reporting.