The Cars Drive Themselves Now. The Fight Is Over What Comes Next.

The strangest thing about riding in a car with no driver is how quickly it stops being strange. The steering wheel turns on its own, briskly and precisely, the way the keys of a player piano move for no one. A screen offers a greeting. And then the vehicle eases into traffic and drives, for the most part, like a mildly cautious parent ferrying someone else's children home. Within about four minutes, the science fiction has curdled into the ordinary, and the passenger is checking a phone.

That collapse — from wonder to Tuesday — is the essential fact of the driverless car in 2026, and it obscures how far the technology has come and how unsettled its future remains. Waymo, the autonomous-driving company owned by Google's parent, Alphabet, now provides roughly 500,000 paid rides a week in a growing list of American cities, having removed human safety drivers from its front seats entirely in 2020. In February, investors valued the company at $126 billion, a figure larger than Ford and General Motors combined.

And yet Waymo has almost certainly never turned a profit. It loses, by most estimates, well over $5 billion a year. Its ridership has recently plateaued even as it adds cities. It is the clear leader in a field that suddenly includes Tesla, Amazon and several heavily subsidized Chinese firms. Two questions now sit in the back seat of every driverless car: whether any of this can become a durable business, and whether it points toward a world in which almost no one owns a driver's seat at all — or merely a better taxi.

I have taken these rides in both San Francisco and Los Angeles, and the honest verdict is that they are wonderful. Not thrilling — wonderful in the specific way that a well-run train is wonderful. There is no negotiation, no small talk to manage, no wondering whether the route being taken is the route or the scenic version of the route. You get in, the doors lock, the music is whatever you left it as, and the car simply goes. What it removes is not the driving so much as the low-grade social administration of being driven. By the second trip I had stopped watching the wheel entirely, which is either the highest compliment available or the most quietly unnerving thing about it, depending on your temperament.

From a desert race to a city street

The lineage runs back to a bribe. In 2004, frustrated that the auto industry was moving too slowly, the Pentagon's research arm, DARPA, offered a cash prize to any team that could send a driverless vehicle across 150 miles of the Mojave Desert. None of the finalists finished; the best managed about seven miles. The organizers ran it again the following year, and this time a modified Volkswagen built by a Stanford team led by the roboticist Sebastian Thrun completed the entire course alone.

Among the spectators, wearing a hat and sunglasses to avoid being recognized, was Larry Page, a co-founder of Google, who had been preoccupied with the idea of self-driving cars since the 1990s. He hired Dr. Thrun, and in early 2009 a secret project code-named Chauffeur began inside Google's experimental lab. Its founding engineers included Dmitri Dolgov, who is one of Waymo's two chief executives today. The team's ambition was not to cross an empty desert but to survive a city — a far harder problem — and for a year and a half almost no one else at Google knew the effort existed.

The stated purpose was never really about cars. Mr. Page framed it, then and later, around the roughly 1.2 million people killed worldwide each year in traffic collisions, nearly all of it attributable to human error. That argument remains the moral engine of the entire industry, and it cuts two ways. If a machine drives more safely than a person, then every year society declines to hand it the wheel carries a measurable cost in preventable deaths. It is a genuinely powerful case. It is also the case that justifies putting thousands of experimental vehicles onto public streets. Both are true at once.

In 2014, the company built a bubble-shaped prototype with no steering wheel and no pedals, a statement that it meant full autonomy rather than driver assistance. The project was spun out of Google as its own company, Waymo, in 2016. Four years later, in the suburbs of Phoenix, it began letting members of the public ride in genuinely empty cars for money. Everything since has been an attempt to do that in more places, more cheaply, without anything going catastrophically wrong.

How the machine sees

Waymo's technical philosophy rests on a word its chief rival treats as a weakness: redundancy. The company layers three distinct ways of perceiving the world on top of one another, so that where one is fooled, the others can overrule it.

Lidar, the spinning unit atop the roof, fires hundreds of thousands of laser pulses a second and measures how long each takes to return, assembling a precise three-dimensional map of the surroundings that works as well in darkness as in daylight because it supplies its own light. Radar uses radio waves to read the speed of other objects directly and to see through rain, fog and snow that would blind a camera. Cameras, ringing the vehicle, supply what the other two cannot: color and meaning — the difference between a red light and a green one, a stop sign, an officer waving traffic through. The newest generation adds microphones, so the car can hear an emergency siren, and determine its direction, before the ambulance comes into view.

What lidar actually returns. Not a picture but a cloud of measured points, dense enough to resolve individual branches, and just as legible at midnight as at noon. Photographed at Waymo’s exhibit.

The software, which Waymo describes in three words — sense, solve, go — fuses those inputs into a labeled model of the scene, predicts where every pedestrian, cyclist and vehicle is likely to move next, and computes its own path accordingly, repeating the loop many times a second. Underlying all of it is machine learning trained on nearly 200 million miles of autonomous driving, in which one car's rare encounter becomes every car's lesson.

And what the cameras add. A stitched 360-degree view of a San Francisco street fair, every pedestrian, cyclist and traffic cone boxed and classified. Lidar knows something is there; this is the layer that works out what it is. Photographed at the exhibit.

The contrast with Tesla is the industry's deepest disagreement. Tesla has bet its robotaxi future on cameras alone, reasoning that humans drive with two eyes and that cameras are cheap enough to put the technology in millions of privately owned cars. Waymo's answer is that human eyes come attached to human brains, and that on public roads, at scale, the extra sensors are the difference between a system that is usually right and one that is reliably safe. Which company is correct is not yet a settled question. It is a multibillion-dollar wager on the nature of perception itself.

Three ways of seeing. Lidar supplies shape, radar supplies speed and weather, cameras supply meaning. The bet is that they rarely fail together.

The economics of an empty seat

For all the engineering, the hardest problem Waymo faces may be arithmetic. The company is buried inside Alphabet's "Other Bets" segment, a bucket the parent declines to break out in detail, and what leaks through the quarterly math is stark: operating losses well over a billion dollars every three months, running toward eight billion for the year by some forecasts, against revenue that amounts to a rounding error on Alphabet's balance sheet. In one recent quarter, the segment lost close to five dollars for every dollar it earned.

The $16 billion that investors committed in February — of which roughly $13 billion came from Alphabet itself — is best understood not as a vote of confidence in current results but as a bet that the losses are an investment in a machine that grows cheaper with every mile. The newest hardware costs about half of the previous generation's; sensor prices have fallen roughly 90 percent since 2017. Alphabet's chief executive has suggested profitability could arrive as soon as 2027. A robotaxi that runs sixteen hours a day, drawing no wages and no tips, eventually undercuts any car with a person in it.

The bear case is equally grounded, and one number captures it. Through the first half of 2026, Waymo added cities while its weekly ride count barely moved — meaning each vehicle is now completing fewer trips than it did a year earlier. The company is, for the moment, buying map coverage faster than it is buying riders. Whether that is the awkward middle of a genuine hockey stick or an early sign of a ceiling is the most important open question about the business, and no one, Waymo included, can yet answer it with confidence.

There is a final oddity in the arithmetic, and it is the one most readers will bump into first: you cannot buy any of this. Waymo has no ticker and no public listing. It has raised more than $27 billion from outside institutions across successive private rounds, which means Alphabet controls it but no longer quite owns all of it, and there was no public offering on the calendar as of the middle of this year — though investors who put in that kind of money do not, as a rule, intend to hold forever. The only route in through a brokerage account is Alphabet itself, where the most consequential automotive company of the decade sits as roughly three percent of the market capitalization, attached to a multibillion-dollar annual loss, inside a segment the parent declines to itemize.

Which produces a strange situation for anyone trying to follow the industry by watching share prices. The pure plays that are listed — Tesla, Uber, Baidu, and the Chinese operators Pony.ai and WeRide on American exchanges — are each betting on a different piece of the problem, and none of them is the leader. The suppliers are a cleaner read: the companies making lidar units, chips and driver-assistance stacks get paid whichever architecture wins. None of that is advice, and this is not a stock column. It is simply worth knowing that the balance sheet doing the most to reshape the automobile is one no ordinary investor can hold on purpose.

A drunk driver endangers one intersection. A software defect endangers every intersection, in every city, at once.

A crowded field

For most of its life, Waymo's main competition was a memory of failure. That is no longer true.

The loudest rival is Tesla, which launched a robotaxi service in Austin in mid-2025 and has since expanded to several metropolitan areas — but with only a few dozen vehicles, most still carrying human safety monitors, and without the regulatory approval to operate a driverless commercial service in California. Its appeal to investors is cost: a camera-only Tesla is far cheaper to build than a sensor-laden Waymo, and estimates put its cost per mile at roughly half. If Tesla can make vision-only autonomy work safely and at volume, it could pressure Waymo's economics severely. That "if" remains large, and its crash data is under federal scrutiny.

Amazon's Zoox has taken the opposite tack, building purpose-made pods with no steering wheel or pedals and a full complement of cameras, radar and lidar. It began limited commercial service in Las Vegas and San Francisco and is preparing to offer rides through the Uber app. Because its vehicles lack manual controls, its scale depends on federal exemptions, which has slowed it.

Then there is China, where a parallel industry has grown up almost entirely separate from the American one. Baidu's Apollo Go operates across more than ten Chinese cities and has pushed abroad into Abu Dhabi and Dubai; Pony.ai and WeRide, both listed on American exchanges, are scaling fleets at home and in the Persian Gulf, aided by state support and a regulatory climate friendlier to rapid deployment. Chinese operators have begun advertising a metric Western firms rarely cite — how seldom their airbags deploy — as a proxy for safety. Their expansion into Western markets is limited so far, but their manufacturing base is not, a fact that shadows Waymo's own supply chain, as will become clear.

Notably, the company spending the most on autonomy builds no cars and operates no robots at all. Uber, having abandoned its own self-driving effort years ago, now partners with nearly everyone — Waymo, Baidu, and others — and has committed more than a billion dollars to a future fleet of autonomous Rivians. Its argument is that autonomy amplifies what it already owns: tens of millions of daily trips and the data to match them. In that reading, the winner of the driverless race may not be whoever builds the best car, but whoever controls the app that summons it.

The cautionary tale hangs over all of them. Cruise, General Motors' entry, was operating driverless in San Francisco alongside Waymo until one of its cars dragged a pedestrian who had first been struck by a human-driven vehicle. The company's slow and murky handling of the aftermath cost it its license, its founders and ultimately its existence; G.M. wound the effort down. Cruise is the reason every Waymo recall makes news. The industry has already demonstrated that it can forfeit public trust in a single afternoon, and that trust, once lost, is not reissued by regulators.

The car that comes from China

For years, a Waymo was a Jaguar I-PACE — and it is worth pausing on that, because the donor car was not some anonymous appliance. The I-PACE was the most decorated vehicle of its year in the world. In 2019 it took World Car of the Year, World Car Design of the Year and World Green Car in a single evening, the first car ever to win three categories in the history of those awards, and it collected European Car of the Year for good measure. It also happens to explain something passengers notice without knowing why: the ride is genuinely lovely. A low, flat battery floor, a long wheelbase, air suspension on the better-equipped cars and generous tire sidewall give it a settled, unhurried gait that suits a vehicle nobody is trying to hustle. Jaguar has since discontinued it, which is its own small parable — the most award-winning car of its season, alive today mainly as a donor body for somebody else's software — and its death forced a scramble for a successor. The purpose-built replacement is a boxy electric minivan called the Ojai, with sliding doors and a flat floor designed around passengers rather than a driver. It is also Chinese, built by Zeekr, a brand of the Geely group.

That has entangled a technology story in a trade war. Chinese-built electric vehicles face American tariffs of around 102.5 percent; a van that sells for the equivalent of $39,000 in China would arrive priced near $89,000 before Waymo installs a single sensor. To comply with security rules aimed squarely at Chinese "connected cars," the vehicles are shipped without sensors, computers or connectivity and are fitted with Waymo's American-designed hardware at a plant in Mesa, Arizona. Waymo has imported thousands anyway, because even after the tariff the arrangement costs roughly half as much as a Jaguar did. It has also drawn political fire: a United States senator singled out the Geely relationship at a hearing and called for restrictions. As one headline put it, Americans cannot buy Chinese electric vehicles, but Waymo is importing them by the thousand.

The Ojai. Sliding doors, a flat floor, sensor pods at all four corners and no styling gestures aimed at a driver, because there isn’t one. Built by Zeekr in China, shipped deliberately incomplete, finished in Arizona.

The hedge is domestic. Waymo has struck a deal with Hyundai to equip Ioniq 5s built in Georgia — reportedly up to 50,000 of them by 2028. The larger point is that a company that spent a decade insisting it was a software business now needs tens of thousands of physical cars in a hurry, and has discovered that in 2026 a fleet cannot be separated from geopolitics.

The carmaker becomes a supplier

Look at that arrangement from Detroit, or Coventry, or Ulsan, and something uncomfortable comes into focus. Jaguar built the car and then killed it. Zeekr builds a body in China and ships it deliberately incomplete. Hyundai has agreed to build Ioniq 5s in Georgia — reportedly as many as 50,000 by 2028 — to a fleet operator's specification. Uber has committed more than a billion dollars to a future fleet of autonomous Rivians it intends to operate rather than sell. In every one of those deals, the company that knows how to make an automobile is the junior partner.

This is a genuine inversion. For a century the automaker owned the thing that mattered: the brand, the dealer network, the financing arm, the badge on the hood, and above all the relationship with the person who paid. In the robotaxi economy the automaker supplies a body and sells it wholesale, and the software company keeps the rider, the app, the route data, the pricing and the margin. It is roughly the position a contract manufacturer occupies relative to a consumer-electronics brand — competent, essential, and interchangeable. Nobody standing on a curb in Santa Monica orders a Jaguar. They order a Waymo, and a Jaguar turns up.

A Jaguar, working. The most decorated car of 2019 — three World Car titles in one evening — now discontinued by its maker and earning its living as somebody else’s hardware.

Nobody orders a Jaguar. They order a Waymo, and a Jaguar turns up.

There is a design consequence too, and it is quietly delightful. Take the driver out of the brief entirely — no sightlines to protect, no wheel to package, no pedal box, no reason for a long hood or a low cowl — and ask engineers to optimize purely for the people being carried. What they arrive at, every single time, is a box. Amazon's Zoox pod is a box with sliding doors and a flat floor. The Ojai is a box with sliding doors and a flat floor. Strip away a hundred years of styling convention built around a person holding a steering wheel, and the shape that emerges is the humble minivan, which is a conclusion this publication reached some time ago by a different route. The most advanced vehicles on earth are converging on the school-run van. The engineers did not set out to prove the point. They simply ran out of reasons not to.

The safety argument, and its critics

Everything above is the promise. The reckoning is the safety record, and the truth of it lives in the gap between two sets of real numbers.

Waymo's case is strong and largely supported by outside parties. Across nearly 200 million driverless miles, the company reports roughly 91 percent fewer crashes causing serious injury and about 92 percent fewer crashes involving pedestrians than human drivers over the same roads. Reinsurers who actually pay claims have examined its liability record and found it markedly cleaner than the human baseline. On the broadest measure, the machine does appear meaningfully safer than the average driver — and the average driver is a low bar, prone to texting, drinking, speeding and fatigue.

The counterargument, advanced in a 2026 study from researchers at Hunter College, is that the average driver is the wrong comparison. Measured against professional for-hire drivers — the cabbies and veteran ride-hail drivers who do this for a living — the study found that Waymo's rate of serious-injury crashes ran about 1.5 times higher per mile, even as its overall crash rate remained lower. Waymo disputes the framing. But the finding lands a fair blow: the claim that the technology is "safer than humans" depends heavily on which humans.

The company's rapid expansion has also exposed a pattern of failure that its scale makes newly visible. Since early 2024 Waymo has filed a series of voluntary software recalls — for a car that struck a utility pole; for vehicles that repeatedly and illegally passed stopped school buses in one city; for robotaxis that drove into flooded roads; and, in mid-2026, for nearly 4,000 cars that misjudged freeway construction zones. Federal investigators at two agencies have an open inquiry into a January 2026 collision in which a Waymo struck a child near a Santa Monica elementary school. No one was killed in these episodes, and a company filing its own recalls is arguably the safety system working as intended. But the nature of the failures is the point: they are not random human lapses but systematic blind spots, shipped simultaneously to thousands of identical cars and, too often, discovered in traffic.

Nor has the public been uniformly grateful. During immigration protests in Los Angeles in 2025, crowds set Waymos on fire — the driverless car proving a nearly perfect target, a wealthy and watchful machine with no human inside to reason with. Waymo suspended service in Los Angeles and San Francisco. The episode is a useful corrective to the corporate brochure: not everyone experiences the empty driver's seat as progress.

What it does to the car you own

The mechanism to watch is not cars but miles. A privately owned vehicle sits motionless more than ninety percent of its life; a robotaxi works sixteen hours a day. So the first thing a fleet takes is not your car, it is your errands — the airport run, the commute, the trip to dinner where somebody wants a second glass of wine. Those miles migrate first because they are the miles nobody enjoys.

Follow that and the downstream effects are reasonably predictable, even if their scale is not. New-vehicle demand softens first in dense metropolitan areas, and softens at the cheap end — the second car, the commuter appliance, the runabout bought out of logistics rather than desire. Franchised dealers get squeezed twice over, because the profit in a dealership has never really been the sale; it is the service bay, and electric fleets both need less maintenance and do it themselves at their own depots. Insurance is stranger than it first appears: if autonomous fleets absorb a large share of the highest-risk urban miles, the human risk pool left behind is not automatically cheaper to cover, and there is a plausible world in which insuring a person to drive in a city becomes conspicuously more expensive than it is today. And then there is the real estate. Robotaxis do not park; they redeploy. Every American city is carrying an enormous inventory of structures built for a fleet that idles, owned by people who have not yet had to think about what happens when it stops idling.

Used values split along the same seam. A vehicle whose whole case was utility — cheap, sensible, gets you there — is competing directly with a service that is also cheap, sensible and gets you there, and does not require insurance, parking or a down payment. A vehicle whose case was never utility is not competing with that at all. Nobody has ever bought a Miata to solve a transportation problem.

Which is where this stops being a eulogy. The reflexive assumption is that autonomy is the end of the enthusiast's car, and it is at least as reasonable to argue the opposite. If the machine takes the joyless miles, what remains in the garage is the portion of driving somebody actually chose. Nobody buys a manual gearbox in order to sit in traffic on the 405. Nobody specifies a limited-slip differential for the school run. For a long time the private car has had to be two things at once — the thing you wanted and the thing you needed — and the compromise has usually been resolved in favor of the thing you needed. A world that hands the errands to a computer is a world where the car in your garage no longer has to justify itself on usefulness. It only has to be worth driving.

The honest counterweight is that this is a comfortable story and comfortable stories about the future have a poor record. Driving is a learned skill, and skills that stop being necessary tend to stop being taught; licensing rates among young people have been drifting downward for two decades already, for reasons that predate any of this. It is not difficult to imagine a city where operating a car yourself becomes expensive, restricted and faintly antisocial rather than simply optional. Both futures are live. What is not in question is that the private automobile is about to stop being the default answer to the question of how a person gets somewhere, for the first time in roughly a hundred years, and that everything about the business was built on the assumption that it always would be.

Will every car drive itself?

Which leaves the question that matters most, and that no company can yet answer: is the driverless car the future of all driving, or only of the trip you do not want to make yourself?

The economics point, at least for now, toward the taxi. A robotaxi justifies its expensive sensors because it works nearly all day; a privately owned car, which sits idle more than 90 percent of the time, cannot easily amortize tens of thousands of dollars of lidar and computing it uses two hours a day. The current systems also depend on high-definition maps of the exact streets they serve, painstakingly built and maintained, which is why Waymo expands one geofenced city at a time rather than switching on a nation. And a fleet can be watched, updated and pulled over remotely in a way a million individually owned cars cannot. Those are the reasons autonomy is arriving first as a service you summon, not a feature you buy.

The opposing vision belongs largely to Tesla, whose entire strategy assumes that autonomy can be made cheap enough to sell to ordinary owners, and that those owners' cars might one day earn money as robotaxis while their drivers sleep. If it works, it collapses the distinction between the private car and the fleet. If it does not, the two worlds stay separate: robotaxis in mapped cities, and privately owned cars that grow steadily more capable without ever quite taking full control.

That second, quieter transformation is already underway, and it is easy to miss because it wears a duller name: driver assistance. Systems that keep a car in its lane and a safe distance from traffic, and that now permit genuinely hands-free driving on mapped highways, make up the largest and fastest-growing slice of vehicle automation — a partial autonomy that millions of people already use without calling it that. The likeliest near future is not a single leap to a world without drivers but a bifurcation: fully driverless robotaxis scaling city by city, and personally owned cars that handle more and more of the highway while still expecting a human to be ready for everything else. Full self-driving in a car you own, everywhere, in all weather, remains the hardest and most distant prize — and may, for reasons of cost alone, never be the norm.

There are limits that money and machine learning have not dissolved. The systems remain concentrated in temperate, well-mapped, densely traveled cities, where each vehicle can be kept busy; the economics look far worse on a rural road driven twice a day. Snow, heavy rain and the endless long tail of the genuinely unexpected still give engineers trouble. Regulation is a patchwork, and public trust, as Cruise learned and Waymo is relearning with every recall, is fragile. And there is the harder-to-model matter of whether people even want it — whether the car, for a meaningful number of drivers, was ever only transportation, or was also something they chose to do with their own hands.

What is no longer in doubt is the engineering. A machine that crossed a desert on a dare now threads rush-hour San Francisco more reliably than most of the humans around it, hundreds of thousands of times a week, and gives a blind rider in Phoenix a way to reach a doctor's appointment without asking anyone for help. That much is real, and it is remarkable, and it deserves to be seen as remarkable for longer than the four minutes it takes to become ordinary. The rest — the profits, the rivals, the reach into everyday ownership — is still being decided, out on the road, one empty driver's seat at a time.