Waymo Has Driven 50 Million Miles. Here Is What That Data Is Actually Worth.

Every mile a Waymo vehicle drives generates roughly one terabyte of raw sensor data. That number comes from the combination of LiDAR point clouds captured 10 times per second, camera frames from multiple high-resolution cameras running at 30 frames per second, radar returns, GPS coordinates, IMU readings, and the vehicle’s own decisions about steering, braking, and acceleration. One terabyte per mile is an approximation, but it is a useful one.

Waymo has driven 50 million autonomous miles.

Do the math and you arrive at a staggering figure. Waymo has generated something in the neighborhood of 50 exabytes of raw driving data. That is 50 million terabytes. For context, the entire Library of Congress contains roughly 20 terabytes of text data. Waymo’s dataset is 2.5 million times larger. Netflix’s entire streaming library fits in about 30 petabytes. Waymo’s dataset is more than 1,600 times that.

This is not a vanity metric. This is the foundation of the most advanced autonomous driving system in the world, and it represents a competitive moat that may be functionally impossible for any other company to replicate.

What one mile of Waymo data actually contains

To understand the value of 50 million miles, you first need to understand what a single mile contains. A Waymo vehicle is not just a car with a camera strapped to the roof. It is a rolling data collection platform that perceives the world through multiple overlapping sensor systems, each generating its own stream of structured information.

LiDAR point clouds. Waymo’s fifth-generation Driver system uses multiple LiDAR units that fire laser pulses in every direction, measuring the precise distance to every surface within range. Each LiDAR rotation produces a point cloud containing hundreds of thousands of individual distance measurements. At 10 rotations per second, one minute of driving produces roughly 600 separate point clouds, each one a detailed three-dimensional snapshot of the vehicle’s surroundings. These point clouds capture geometry that cameras cannot: the exact shape and distance of every object, surface, and obstruction around the vehicle, regardless of lighting conditions.

Camera frames. The vehicle carries multiple high-resolution cameras covering a full 360-degree field of view. Each camera produces 30 frames per second. Across all cameras, a single mile of driving generates tens of thousands of individual image frames. These frames capture color, texture, signage, lane markings, traffic signals, and the appearance of other road users. Unlike LiDAR, which excels at geometry, cameras excel at classification: distinguishing a cyclist from a pedestrian, reading a stop sign, or recognizing a construction zone.

Radar returns. Radar sensors provide velocity information that neither LiDAR nor cameras can match. Radar can measure the speed and direction of every moving object in the vehicle’s surroundings, even in conditions where LiDAR and cameras are degraded, such as heavy rain or fog. Each mile generates a continuous stream of radar measurements that help the system understand not just where objects are but where they are going.

Vehicle telemetry. Every steering input, brake application, acceleration command, and route decision is logged with millisecond precision. This telemetry creates a complete record of what the vehicle decided to do at every moment and, critically, why it made that decision. When the vehicle encounters a pedestrian stepping into the road, the telemetry captures the exact sequence: detection, prediction, planning, execution. This decision data is arguably more valuable than the sensor data itself because it represents the accumulated judgment of the system.

Edge cases and near-misses. Buried within those 50 million miles are millions of rare and unusual situations. A mattress falling off a truck on a Phoenix highway. A dog running across a San Francisco intersection. A construction crane swinging a load over a road in Austin. Emergency vehicles approaching from multiple directions. These edge cases are disproportionately valuable because they represent the scenarios that separate a functional autonomous system from a dangerous one. You cannot manufacture edge cases in simulation. You find them by driving real miles on real roads.

The data flywheel that money cannot buy

Waymo’s data advantage is not static. It compounds.

Every mile the fleet drives adds new data to the training set. That data is used to improve the Waymo Driver software. The improved software drives more safely and handles more scenarios. Safer driving enables expansion to new cities and new operating conditions. Expansion means more miles. More miles mean more data. The cycle continues.

This is the data flywheel, and it is the core reason why Waymo’s lead is so difficult to challenge. A competitor starting from zero today would need to drive 50 million miles of their own to accumulate an equivalent dataset. At Waymo’s current pace of roughly 5 to 7 million new miles per year, that would take a decade of continuous operation with a comparable fleet, and by the time the competitor reached 50 million miles, Waymo would be at 100 million.

The flywheel also improves in quality over time, not just quantity. Waymo’s simulation infrastructure runs billions of additional miles of virtual testing each year. But those simulated miles are seeded with real-world scenarios extracted from on-road data. A real near-miss captured in San Francisco becomes a template for thousands of simulated variations. A real traffic pattern observed in Phoenix becomes the basis for synthetic training data. The simulation system is powerful precisely because it is grounded in 50 million miles of actual driving experience.

What it cost to build this dataset

Alphabet has invested more than $5 billion in Waymo since the Google Self-Driving Car Project launched in 2009. That figure includes research and development, vehicle procurement, sensor manufacturing, computing infrastructure, personnel costs, and the operating expenses of running a fleet of autonomous vehicles across multiple cities for years.

To put that number in perspective, $5 billion is more than the GDP of some small countries. It is roughly the annual revenue of Airbnb. It is three times what Figure AI has raised in total funding to build humanoid robots. And most of that money was spent before Waymo generated a single dollar of commercial revenue.

The investment breaks down across several categories, though Alphabet does not disclose a granular breakdown.

Vehicle costs. Each Waymo vehicle is a heavily modified Jaguar I-PACE equipped with a custom sensor suite worth hundreds of thousands of dollars. The fleet currently numbers roughly 700 vehicles. At an estimated $200,000 to $300,000 per fully equipped vehicle, the fleet alone represents an investment of $140 million to $210 million. Earlier generations of vehicles using custom Chrysler Pacifica minivans added additional hundreds of millions in vehicle costs over the years.

Sensor development. Waymo designs and manufactures much of its own sensor hardware, including its fifth-generation LiDAR system. Developing custom LiDAR, camera, and radar systems from scratch, and achieving the reliability needed for commercial deployment without a safety driver, required years of engineering effort and hundreds of millions in R&D.

Computing infrastructure. Processing 50 exabytes of raw data requires enormous computing resources. Waymo maintains its own data centers and cloud infrastructure for training its neural networks, running simulations, and managing fleet operations. The annual compute cost alone likely runs into the hundreds of millions of dollars.

Personnel. Waymo employs between 2,000 and 3,000 people, many of them highly specialized engineers in robotics, machine learning, and autonomous systems. At Silicon Valley salary levels, the annual personnel cost likely exceeds $500 million.

The point of this accounting exercise is simple. Waymo’s dataset was not cheap to build. The $5 billion investment represents a barrier to entry that is measured not just in money but in time. Even a company willing to spend $5 billion today would need years of on-road driving to accumulate comparable data. The money buys the vehicles and the sensors and the engineers. It does not buy the miles. The miles have to be driven.

Waymo vs. Tesla: two philosophies of perception

The most common comparison in autonomous driving is Waymo versus Tesla, and it centers on a fundamental disagreement about how autonomous vehicles should perceive the world.

Waymo uses a multi-sensor approach. LiDAR provides precise 3D geometry. Cameras provide color and classification. Radar provides velocity. The three sensor types complement each other’s strengths and compensate for each other’s weaknesses. LiDAR works at night when cameras struggle. Cameras read text that LiDAR cannot. Radar sees through rain and fog that degrade both LiDAR and cameras.

Tesla uses a camera-only approach for its Full Self-Driving (FSD) system. Elon Musk has argued that humans drive with two eyes and a brain, so vehicles should be able to drive with cameras and a neural network. Tesla removed radar from its vehicles in 2021 and has never used LiDAR in production. The FSD system relies entirely on eight cameras and a custom neural network running on Tesla’s proprietary AI chips.

Tesla’s FSD system has accumulated approximately 35 million miles of autonomous driving. That figure is smaller than Waymo’s 50 million, but the comparison is not straightforward. Tesla’s miles come from a fleet of consumer vehicles driven on diverse roads across the entire United States, while Waymo’s miles come from a dedicated fleet operating in specific geofenced cities. Tesla’s data has more geographic diversity. Waymo’s data has more sensor richness.

The critical difference is not the quantity of miles but the quality of data per mile. A Waymo mile includes LiDAR point clouds that give the system ground truth about the three-dimensional geometry of every scene. When a Waymo vehicle detects an object, it knows the object’s exact distance, size, and shape with centimeter-level precision. A Tesla mile includes camera images that the neural network must interpret to infer depth and distance, a fundamentally harder problem.

This distinction matters enormously for training. Waymo can train its perception system using LiDAR as a teacher signal. The system knows exactly where every object is because the LiDAR measured it directly. Camera perception can then be trained against that ground truth. Tesla, lacking LiDAR, must rely on other methods to establish ground truth, including fleet-sourced data where human drivers provide implicit labels through their driving behavior.

Neither approach is definitively superior. Tesla’s camera-only system benefits from the massive scale of its consumer fleet and the low marginal cost of adding another vehicle to the data collection network. Waymo’s multi-sensor system benefits from richer data per mile and more reliable ground truth for training. The debate will ultimately be settled by real-world safety outcomes over millions of miles of commercial operation.

How Waymo turns miles into money

For most of its existence, Waymo was a pure research project that generated costs but no revenue. That changed with the launch of Waymo One, the company’s commercial robotaxi service.

Waymo One currently operates in four US cities: Phoenix, San Francisco, Los Angeles, and Austin. The service expanded to Austin and Atlanta in 2025, with additional cities planned. Riders hail a Waymo through the Waymo One app, much like calling an Uber or Lyft. The vehicle arrives, the rider gets in, and the car drives itself to the destination without a human driver.

The scale of the service is substantial and growing. Waymo consistently completes over 150,000 paid trips per week across its service areas. At an estimated average fare of $15 to $25 per ride, that translates to roughly $2.25 million to $3.75 million in weekly ride revenue, or approximately $120 million to $195 million in annualized ride revenue.

Those revenue estimates are rough, and Alphabet does not break out Waymo’s financials in detail. But the direction is clear. Waymo is generating real commercial revenue from autonomous rides, making it the only company in the world with a large-scale commercial robotaxi operation.

The revenue itself is important, but it is not the main point. The main point is that every paid ride generates more data. Every trip through downtown San Francisco at rush hour, every late-night pickup in Phoenix, every rain-soaked ride in Austin adds to the dataset. Waymo has built a business model where the act of generating revenue simultaneously generates the data needed to improve the product that generates the revenue. The commercial operation and the data flywheel are the same thing.

The expansion strategy: Austin, Atlanta, and beyond

Waymo’s expansion from its original Phoenix test ground to San Francisco, Los Angeles, Austin, and Atlanta follows a deliberate pattern. Each new city adds not just more miles but different kinds of miles.

Phoenix offered wide roads, consistent weather, and relatively predictable traffic patterns. It was the ideal starting ground for a commercial service. San Francisco added steep hills, dense traffic, aggressive cyclists, complex intersections, and fog. Los Angeles brought highway driving at scale, sprawling suburban roads, and some of the most congested freeways in the world. Austin added a fast-growing Sun Belt city with rapidly changing construction zones and a different driving culture. Atlanta will add a Southeast US city with distinct road designs, weather patterns, and traffic behavior.

Each city is not just a new market. It is a new training environment. The edge cases that a vehicle encounters in San Francisco’s Chinatown are fundamentally different from those it encounters on a Phoenix highway or an Austin residential street. Geographic diversity in the training data makes the system more robust because it forces the model to generalize across conditions rather than overfit to a single environment.

The safety case: 85 percent fewer injuries

Data is only as valuable as the outcomes it produces. Waymo’s safety record provides the most direct evidence that 50 million miles of data translate into real-world performance.

A 2023 study conducted in partnership with Swiss Re, one of the world’s largest reinsurance companies, examined over 7 million miles of rider-only driving (no safety driver) in Phoenix and San Francisco. The study found that Waymo vehicles were involved in 85 percent fewer injury-causing crashes per mile compared to the human driving benchmark. Not 85 percent fewer total incidents. 85 percent fewer crashes that caused injuries to people.

That figure deserves emphasis because it addresses the central question of autonomous driving: is the technology actually safer than human drivers? For Waymo, after 50 million miles of data collection and continuous improvement, the answer appears to be yes, by a substantial margin.

The safety advantage is a direct product of the data. More training data means the system has seen more scenarios. More scenarios mean better predictions. Better predictions mean fewer situations where the vehicle is surprised. And surprise is the root cause of most crashes, whether human or autonomous.

Waymo has also published its safety methodology in unusual detail for the industry. The company shares quarterly safety reports that include data on crashes, contacts, and incidents. This transparency is itself a competitive strategy. By establishing a public safety track record, Waymo makes it easier for regulators to approve expansions and harder for competitors to argue that their less-validated systems deserve equivalent treatment.

The regulatory moat

Data and safety create a third advantage that is often overlooked: regulatory access.

Waymo’s safety record and transparency have earned it regulatory permissions that no other autonomous vehicle company currently holds at comparable scale. The company operates fully driverless commercial vehicles without a safety driver in multiple US cities. Getting those permissions required years of engagement with state regulators, city governments, and federal agencies. Each permission was granted on the basis of demonstrated safety performance backed by millions of miles of data.

A competitor entering a new city must go through the same regulatory process. But without a comparable safety record, the path to approval is longer and less certain. Regulators are inherently cautious about putting driverless vehicles on public roads. A company that can point to 50 million miles and an 85 percent reduction in injury-causing crashes has a fundamentally different conversation with regulators than a company with 1 million miles and no published safety studies.

This regulatory advantage compounds in the same way the data advantage does. More cities mean more data. More data means better safety. Better safety means easier regulatory approval for more cities. The cycle reinforces itself.

What Waymo’s data means for humanoid robots

This article appears on a site that primarily tracks humanoid robots, so it is worth asking: what does Waymo’s data advantage mean for the broader field of autonomous machines?

The answer is more direct than you might expect. The core challenge in both autonomous driving and humanoid robotics is the same: building a system that can perceive an unstructured environment, predict what will happen next, and act safely in real time. The sensor modalities differ, the actuators differ, and the operating environments differ. But the fundamental AI problem of perception, prediction, and planning is shared.

Waymo has demonstrated that real-world data is the critical ingredient for solving this problem. Simulation helps. Clever algorithms help. But nothing replaces the diversity and complexity of millions of hours of real-world operation. The humanoid robotics industry is learning this same lesson. Companies like AgiBot that are deploying thousands of robots into real factories are accumulating real-world manipulation data that will prove more valuable than any amount of lab testing.

The data flywheel model, where commercial deployment generates the data that improves the product that enables more deployment, applies to humanoid robots just as it applies to autonomous vehicles. Waymo proved the concept. The humanoid industry is now following the same playbook, just a few years behind.

The $175 billion question

How much is Waymo’s data actually worth? This is not a question with a precise answer, but we can triangulate.

Alphabet reportedly valued Waymo at approximately $175 billion in internal assessments during 2025. That valuation implies that investors believe Waymo’s technology, market position, and growth potential are worth more than companies like Goldman Sachs, Starbucks, or Nike. A substantial portion of that valuation rests on the data asset, because the data is what makes the technology work, and the technology is what makes the market position defensible.

Another way to approach the question is replacement cost. What would it cost a competitor to replicate Waymo’s dataset from scratch? At a minimum, they would need to spend $5 billion on vehicles, sensors, computing infrastructure, and personnel. They would need to drive for 15+ years to accumulate equivalent miles. They would need to navigate regulatory processes in multiple cities. And by the time they finished, Waymo would be 15 years further ahead.

The data is worth what it would cost to replace, and for any practical competitor, the replacement cost approaches infinity. That is the definition of a moat.

What comes next

Waymo is not standing still. The company has partnered with Geely’s Zeekr brand to develop a next-generation purpose-built autonomous vehicle. Unlike the current modified Jaguar I-PACE fleet, the next-generation vehicle will be designed from the ground up as an autonomous platform, with the sensor suite and computing hardware integrated into the vehicle’s architecture rather than bolted on afterward. This should reduce per-vehicle costs and improve sensor placement.

The expansion pipeline includes additional US cities beyond the current four, with Atlanta already in the launch process. International expansion remains a longer-term goal, with regulatory environments in Europe and Asia presenting different challenges and opportunities.

And every day, the fleet drives more miles. Every mile adds more data. Every data point makes the system slightly better, slightly safer, slightly harder to catch.

Fifty million miles is a number. But it represents something much larger: a dataset that took 17 years and $5 billion to build, that no competitor can replicate on any reasonable timeline, and that improves automatically every time a Waymo vehicle pulls out of a parking lot and onto a public road.

In the race for autonomous machines, whether they have four wheels or two legs, data is the ultimate currency. And by that measure, Waymo is the richest company in the field.