On June 22, 2025, in Austin, Texas, Tesla quietly launched its long-anticipated paid ride trial for the Robotaxi. The rollout wasn’t dramatic. Just around 20 Model Y vehicles appeared on the roads—not the much-talked-about Cybercab revealed last October, the one stripped of pedals and a steering wheel. These were familiar cars. Mass-produced, mostly unmodified. In the back seat, there was a small screen for passengers to interact with. That’s about it.
Tesla chose a few neighborhoods in Austin and invited a limited group of users to try it. Each ride cost $4.20—far below the standard taxi fare. Quickly, videos had started surfacing on social media. The vehicles used in the pilot were just standard Model Ys—no special modifications. There was no driver in the driver’s seat. Instead, a Tesla employee sat in the front passenger seat, serving as a “safety monitor.” At the same time, the cars were being remotely supervised by Tesla’s operations center. But there were no controls on the passenger side—no steering wheel, no pedals, nothing to intervene with. Unless they used the touchscreen in the front row, the safety monitor had no way to take over the vehicle.
It felt ordinary. Deliberately so.
Elon Musk called it the beginning of a new chapter—Tesla’s pivot toward AI and robotics. He says thousands more vehicles will be deployed “in a few months,” spreading to San Antonio, San Francisco, Los Angeles. There was no ceremony. No cinematic reveal. Just 20 cars with staff monitors and a test route. You could call it clunky. Or you could call it the beginning of something massive. Because what looked humble on the surface is where Tesla’s real advantage lies.
These weren’t prototype vehicles. These were off-the-line Model Ys, running a slightly customized version of FSD 13.2.9—not even the speculated 14.0. That’s the point. No unique hardware. In contrast, Waymo’s vehicles are heavily customized. Their rooftops are fitted with complex sensor arrays—multiple LiDAR units, cameras, and radar systems—all visibly integrated into the structure. You can tell at a glance that these are not ordinary cars. And that, in a way, highlights the essence of Tesla’s strategy: by avoiding this kind of visible, high-cost modification, they’ve built the foundation for something that can scale exponentially.
Waymo hasn’t disclosed precise costs, but estimates suggest each vehicle adds between $50,000 and $100,000 in hardware. This locks them into a slower, capital-intensive growth model. Only about 1,500 Waymo vehicles are operational as of May 2025.
In contrast, Tesla is using mass-produced Model Ys as its Robotaxi fleet, relying on its already-proven autonomous driving system. This significantly lowers the barrier to operating a Robotaxi service and ensures a steady supply of vehicles. If, over the next few months, Tesla can use real-world data from multiple cities to show regulators that this unassuming Model Y can provide safe and reliable autonomous rides, the safety monitors in the passenger seat can begin to be phased out—and the profit margin compared to traditional taxis will become clear. More importantly, once that trust is established, Tesla will be in a position to make a logical case to regulators: allow private Model Y owners with FSD to join the Robotaxi network during their downtime. That’s what would make rapid, low-cost expansion truly feasible.
That idea isn’t new. Tesla embedded fleet management code in vehicles as early as the 2022 AI Day. The goal has always been clear: build a self-driving fleet using vehicles already in people’s garages.
2022 AI Day
In 2023, at Investor Day, Musk outlined a vision for a $10 trillion global energy and mobility economy. Robotaxis were a small but crucial footnote in that vision.
2023 Investor Day
If you own a Tesla, the pitch becomes simple: let your car drive others when you’re not using it. It earns money. You earn money. Suddenly, a Model Y is no longer a depreciating asset. It’s a productive one.
Tesla estimates you could make over $30,000 a year this way. At that rate, the car pays for itself in under two years. After that, it’s all profit. That changes everything. Buying a Tesla isn’t just about driving anymore. It’s about investing.
It’s also why the Robotaxi trial in Austin matters. Because if these 20 cars, mostly unmodified and essentially identical to what’s in millions of driveways, can demonstrate commercial viability, then the floodgates open. Tesla has over 3 million vehicles in the U.S., and about half have the hardware needed for full self-driving. Even if just 10% of those owners opt in, that’s a fleet of 100,000 cars. Instantly. That dwarfs anything Waymo has built in a decade.
It’s a shift from cars as consumption to cars as investment. For more than a hundred years, owning a car has been a losing proposition—95% of the time it sits still, costing money in depreciation, insurance, and parking. But with Robotaxi, a new use case emerges: the car that earns its keep.
Of course, the vision is still incomplete. Just a few days into the Austin trial, one Tesla Robotaxi failed to make a left turn properly and drifted briefly into oncoming traffic. The camera-only system Tesla relies on still struggles in some complex real-world scenarios. These edge cases, or “long-tail events,” are where safety systems are truly tested.
The key question now is whether Tesla can use its closed-loop data system to rapidly improve its algorithms, and whether the scale of its data can adequately cover “long-tail” events. That’s what it will take to prove that Tesla’s camera-only approach can match—or even rival—multi-sensor fusion systems, and to build trust in its technology among users.
Just like with large AI models, “scaling law” also applies in the world of autonomous driving. But here, “scale” doesn’t refer so much to model size or parameter count—it’s mostly about the amount of real-world data. One Waymo study suggested that the optimal model size for autonomous driving might be in the tens of millions of parameters. The true ceiling for an autonomous driving system lies in how well it can understand the full diversity of the real world, especially those rare, hard-to-predict “long-tail” events. Without enough data, that understanding remains out of reach. No matter what approach is used, there’s a baseline volume of data you need to cross before performance meaningfully improves.
Long-tail events include things like unusual road users, odd combinations of traffic behavior, or barely visible lane markings during extreme weather. These events might make up less than 0.1% of all driving data, but they’re critical to safety.
Because these edge cases are so rare, they don’t naturally show up often enough in standard model training. Traditional algorithms tend to struggle when such events occur. That’s why Waymo spends enormous effort digging through its data to identify these rare events, and then retraining its models with extra emphasis on them.
Tesla takes a different approach. It uses a “shadow mode” that passively collects data from the entire fleet at minimal cost, creating a vast training dataset very quickly. Any time a driver intervenes or disagrees with what FSD is doing, that instance gets flagged—essentially turning into labeled video data. Over time, this allows Tesla to accumulate enough examples to train the system on rare situations.
But it’s a double-edged sword. If Tesla feeds this unfiltered data directly into the system, FSD might not only learn how people drive—it might also pick up their bad habits. That’s part of why some people describe Tesla’s FSD as smooth and confident, but also prone to cutting corners or bending the rules when it thinks it’s faster. In the long run, Tesla will need to distinguish between human driving that’s useful to learn from, and behavior that’s better ignored. Cleaning up this “dirty data” won’t come cheap.
The next six months will be critical. Will accidents decrease? Will safety monitors be removed? Will regulators approve private vehicles joining the network? Will Musk be distracted by politics again?
If things go right, the effect won’t be subtle. The entire ride-hailing industry—Uber, Lyft, even Waymo—will have to adjust.
For now, it’s just 20 cars in Austin. Quiet. Understated. But the ripples have already begun.