Contra Pueyo on Robotaxis

Robotaxis won't be everywhere soon, and it's unclear that Tesla has an edge

As a reminder: Changing Lanes will begin reserving most of its content for paying subscribers in May. Pledge to pay now, at this link, and enjoy Charter Membership rates of $8 per month or $80 per year; that’s a 20% discount on regular rates, which will begin on 6 May 2025.

Earlier this week, on 11 March 2025, Tomás Pueyo published "Robotaxis Are Here," in which he lays out a vision of how robotaxis are going to imminently transform urban transportation. Pueyo argues that self-driving taxis will replace human drivers between 2025 and 2027, and personal car ownership will begin to decline after 2028. In the fullness of time, says Pueyo, the global market in robotaxis will grow to be $3-trillion-to-$5 trillion in value, as robotaxis supplant public transit to become the dominant mode of urban transport.

This kind of techno-optimism captures the imagination, which is good.

But it’s bad that it makes so many questionable claims.

I like Tomás; we’ve spoken a few times, virtually and in person, and I enjoy his newsletter, Uncharted Territories (I encourage you to subscribe here). I share his excitement about the potential long-term impact of driving automation, and I look forward to a world of widespread adoption.

Unfortunately, I think that Pueyo is wrong to think that:

• Robotaxis at scale will arrive quickly; and that

• We know which firm will benefit most.

Huh. I never noticed before, but Tomás actually looks like the Distracted Boyfriend

So today, in lieu of our usual Off-Ramps feature, I’d like to review eight elements of Pueyo's article of which I’m skeptical. The first four have to do with his views on how fast robotaxis will arrive, and what will change in our cities as a consequence. The latter four have to do with his assessment that Tesla, rather than Waymo, is better positioned to dominate the coming robotaxi market.

To be clear, I do think the robotaxi revolution is coming. What’s more, I welcome it. Two things have held it back: the zeal of its enemies, and the uncritical hype of its friends. We need to overcome both, and the way to do that is to have a more nuanced understanding of the technical, regulatory, and human factors that will shape the growth of this technology.

Doing so will not only help us understand the world better, but also help us to shape a world where robotaxis can do us the most good.

Four Missteps Pueyo Makes About How Fast Robotaxis Will Arrive

1. The Long Tail Is Even Longer Than He Thinks

Firstly, Pueyo predicts that robotaxis will replace traditional taxis between 2025–2027, with personal car ownership declining after 2028.

That’s ludicrously soon.

Pueyo ignores the consistent pattern of missed automated vehicle (AV) deployment predictions that has characterized the industry for a decade. As my co-authors and I document in our book, The End of Driving (forthcoming later this year!), the path to widespread automated driving has been marked by repeated timeline slippages. Consistently, as we show in Chapter 2, “Hype, Disillusionment, and a Reset”, the horizons have been stretched, with rollout at scale always two-to-three years away. While Tesla is most notorious for making big claims, the pattern of missing deadlines isn't unique to any one party. It persists because of the genuine difficulty of the technical challenge.

The fundamental issue is the ‘long tail’ problem. Getting an automated system to handle 95% of driving scenarios is challenging but achievable; getting to 99% is significantly harder. Unfortunately, a wholly-automated system that can handle any situation safely and reliably needs to achieve what engineers call ‘six nines’ reliability, i.e., 99.9999%, or even more. This represents a challenge that is at least four orders of magnitude more difficult to solve.

Look at Waymo's actual progress: they began testing without safety drivers in Phoenix in 2017, but it wasn't until late 2023 that they expanded beyond limited areas of Phoenix to parts of San Francisco, and early 2024 for limited operations in Los Angeles. This methodical, city-by-city expansion over more than six years illustrates the realistic pace of deployment, as practiced by the industry leader with unparalleled access to capital.

No one else will be able to do it faster… at least, not without significant blowback (about which more later).

This history should teach us to treat any ambitious timeline we encounter with skepticism. Revolutionary change, outside of the digital realm at least, always takes decades rather than years.

2. Underestimation of Regulatory Difficulty

Pueyo dismisses any policy obstacles to robotaxi adoption, suggesting that safety statistics alone will ensure fast and widespread approval by regulators.

I wish we lived in a world where regulators moved fast, considered the big picture, performed cost-benefit analyses, and welcomed innovation. Regrettably, we don’t.

In the actual world, regulators bear strong mandates to avoid incurring liability, often with what Joseph Heath calls “a near-complete disregard for questions of cost and efficiency”. And their unofficial mandate is to avoid attracting headlines, which means avoiding the assumption of risk. So it largely doesn’t matter how wonderful robotaxis might be; that quality, on their own, will not swing any weight. The administrative state will allow them to operate, but only at the most measured pace.

Worse, there is not one administrative state, but many.

AVs face a patchwork of regulations that vary not just between countries but between states and even municipalities. In the United States, each state has its own regulatory framework for AV testing and deployment. California's regulations differ from Arizona's, which differ from Florida's. (The same fragmentation applies in Canada.) This fragmentation means that approval in one jurisdiction doesn't translate to approval in another. Expanding markets will require battle after battle, one city or state at a time.

And those battles will have high stakes. As we’ve seen, a single incident that Cruise mishandled led to California regulators revoking the company's permit to operate driverless vehicles statewide, ultimately killing the company. The aftermath of this incident has prompted other AV companies like Zoox to adopt slow, conservative deployment plans.

Moreover, safety is only one factor in the regulatory equation. Policymakers also consider labour displacement concerns, equity issues, and the effects on both sustainability and liability. Regulatory approval for AVs will continue to be incremental, for years to come, as companies build relationships with regulators in each jurisdiction they enter, and inevitably face hostility or indifference in some places. This means not only that adoption will take a long time, but also that it will be uneven.

3. Weather Will Be Challenging for a While

Speaking of uneven deployment, Pueyo ignores the question of weather.

As my interview with Prof. Waslander at the University of Toronto shows us, heavy rain and snow continue to baffle AVs. Cameras see almost nothing in a blizzard, lidar is overwhelmed with signal noise from raindrops and snowflakes, and radar has limited range. Worse, the neural networks that rely on this sensor data are confused by winter conditions, even as those conditions demand faster, more decisive decision-making than is usual.

Winter weather, and heavy rain, are not ‘edge cases’ but regular conditions in many parts of the world. The fact that AVs have trouble seeing in them or thinking about them is a profound difficulty.

As I concluded in my winter-weather newsletter, I think these problems will be solved. But they aren’t being solved quickly, meaning that even if we get faster robotaxi deployment than I expect, we will certainly get it in the sunbelt only.

4. There Are Infrastructure and Support-System Requirements

Finally, Pueyo is thinking hard about robotaxis. That’s good, but it leads him to think less about the infrastructure needed to support them. That’s a critical omission.

For a firm to introduce a large-scale robotaxi fleet into a city, it will also need to introduce:

• Specialized charging infrastructure. Electric robotaxis operating continuously would need high-capacity, strategically located charging stations

• Maintenance facilities. Robotaxis operating 24/7 will require more frequent maintenance than privately owned vehicles. The industry would need to develop new maintenance protocols and facilities designed specifically for high-utilization automated vehicles

• Pick-up-and-drop-off zones. As my co-authors and I discuss at length in Chapter 8, “Matters of Scale”, in The End of Driving, robotaxis at scale will make city transportation networks ungovernable, unless cities also transform large stretches of street parking spaces into pick-up and drop-off (PUDO) zones. As some cities get to high levels of robotaxi uptake, other cities—seeing the chaos that is unleashed without PUDO zones—will insist that such zones be introduced first before robotaxis can reach high levels of market share

These infrastructure requirements represent massive capital investments that will take vast sums to finance and years to implement. The chicken-and-egg problem is clear: building up sufficient infrastructure in advance of the vehicles being in revenue service is prohibitively expensive, but widespread vehicle deployment will be, to some extent, constrained without the supporting infrastructure.

The reasonable thing to expect is slow build-ups: small fleets with small infrastructure footprints that grow as the market does. In other words, not fast, sudden, exponential growth.

In a sense, what Waymo has done in San Francisco is the easy part. It will become more challenging as Waymo, or any other robotaxi firm, claims more and more of the total addressable market of rides. Then, new problems will emerge: safety edge cases that encourage regulators to slow market penetration, weather edge cases that slow deployment, and need for infrastructure improvements in the world that will take time and money to produce.

Taken together, this means that a future of ubiquitous robotaxis, for most of the world, is not going to emerge in the next three years.

Four Missteps Pueyo Makes About Tesla Dominating the Robotaxi Space

Pueyo makes a bold claim about the competitive landscape in robotaxi:

Waymo has a big advantage in that it already has full self-driving cars. However, every other advantage is on Tesla's side: Its vision-only bet sounds reasonable, it has a massive fleet already, it's vertically integrated, it has huge manufacturing expertise, it has much more data, and its AI is probably the best.

His analysis overlooks several crucial matters.

1. Tesla Going Vision-Only Is Risky

Pueyo presents Tesla's vision-only approach as a reasonable bet.

I am not so sure. I’m currently writing a major piece about this, to be published later this year, so I won’t go into great detail now. So instead, I’ll put it simply and quickly: Tesla's choice to use cameras and AI is not obviously going to win over Waymo’s choice of sensor redundancy and sensor fusion. As Steven Waslander and I discussed, optical systems will always struggle with or in low light, glare, precipitation, and circumstances of extreme visual contrast. Tesla’s bet is that sophisticated neural-network AI will be able to overcome the limitations of its sensor set, and it will, in many cases.

But not all. Teslas with automated driving capability will continue to struggle in some conditions, even as Waymo and Zoox (and other entrants) perform admirably, because those conditions are the ones where redundant sensing modalities like LiDAR and radar provide critical advantages.

And there, for Tesla, is the rub. Regulators will be able to say, with justice, that since some self-driving vehicles can master these conditions and drive safely, all must be able to, or the law will not permit them to operate. ‘This is the standard’, they might say; ‘we know some products can meet it; if yours can’t, well, that’s your problem, not anyone else’s. It’s the product, not the standard, that has to change’.

And if I were Waymo, I would be making exactly this case to regulators. Indeed, Waymo is already doing it.

At a public webinar earlier this year (for which footage is available on YouTube), Waymo outlined its 2025 regulatory strategy. The firm’s head of public policy indicated the firm plans to lobby for a national public disengagement database: that is, a database of every safety incident involving a self-driving car. Waymo is comfortable pushing for that because they know it’s a standard they can meet, and that some of their competitors can’t.

Given the current environment in Washington, I don’t expect the ploy to work, not this year. But I do expect much more of this approach in the years to come. Tesla’s insistence that camera-only is sufficient makes it vulnerable in ways some observers don’t appreciate.

2. Tesla’s AI Lead Isn’t Obvious

Pueyo claims that Tesla's AI is "probably the best," but I don’t know why he thinks this.

Is it because Tesla has Grok, showing that it knows how to build neural-network AI? Sure, but Waymo, through Alphabet, has Gemini, which shows the same thing.

Is it because Tesla has access to data on which to train its neural network, courtesy of Tesla’s years of selling cars studded with cameras? Sure, but Waymo has been operating cars with cameras on roads for years as well. That dataset will certainly be smaller, but higher-quality, as they’ve been watching for precisely those situations they want to solve for.

Is it because Tesla has been offering ‘Full Self Driving’ for years and has built a good AI to operate an automated vehicle? Sure, but so has Waymo, which has been optimizing for automated driving, not driver assistance.

On this claim, Pueyo may be right, but he might also be wrong. I’m not sure why he’s so confident in the ground he’s taken. I think it’s more reasonable to believe that each company has specific strengths in different domains… which means it’s premature to think that one is probably going to win.

3. Tesla has Quantity, Waymo has Quality

Let’s drill down on this claim some more.

Pueyo emphasizes Tesla's data advantage, noting they have “100x more data than Waymo”. He’s referring to raw miles driven, and it is an advantage: quantity has a quality all its own, as the proverb has it. And it’s certainly true for training neural networks, where the bitter lesson says more data wins every time.

But I don’t think that’s the end of the story. Sure, Tesla has lots of data. But what kind of data is it?

Tesla's data comes from its customer fleet, driving under human supervision, in conditions chosen by those customers. While vast in quantity, this data has limitations:

• It represents routine driving conditions, and not edge cases

• It lacks consistent annotation and validation

• It comes from systems that still require human oversight

That last bullet may need unpacking. Tesla's fleet data comes primarily from human-supervised driving. That wouldn’t be an issue, but for the fact that the company aims to create systems that operate without human supervision. In other words, there is a domain shift here: the (distribution of the) training data doesn’t predict conditions in which the system will operate, and as such doesn’t help build a system that can flourish in those conditions.

As an example, consider

Timothy B. Lee

’s story of an AI that learned to play Mario Kart by watching footage of good human players. Human players, being good at the game, will keep their kart in the centre of the track as they proceed, and so the AI will learn to do the same. Unfortunately, good human players will rarely ride the edges of the track, much less leave it. And so the AI will not see (enough) data on how to recover from such situations. Consequently, when the AI plays Mario Kart, things will go well… at first. But as soon as it drifts away from the centreline of the track, it will not know how to recover, and falter. And then its errors will begin to compound, leaving it hopelessly unable to proceed.

This is the principal problem in building neural networks generally. For his part, Lee tells that story to illustrate his claim that Tesla's near-infinite training data isn't a silver bullet.

In contrast, Waymo's data collection is purposefully designed for automated-driving development. Waymo systematically tests edge cases, unusual scenarios, and challenging conditions, creating a dataset that may be smaller in raw miles but will be more valuable for training robust driving-automation systems. Their approach includes synthetic data generation, structured testing protocols, and comprehensive annotation, creating data specifically designed to address the challenging ‘long tail’ they will need to get to six-nines reliability.

So Tesla’s data does not make it as unbeatable as Pueyo seems to think.

4. Company Culture Matters

To me, the most glaring gap in Pueyo's analysis is the significant difference in regulatory relationships and company culture between Tesla and Waymo. These differences are highly relevant in any assessment of where the two companies are in the race to wide-scale robotaxi deployment.

Waymo has pursued a methodical, safety-first approach to regulatory approval, working closely with authorities at each step of deployment. To quote myself, in the inaugural post of Changing Lanes: "Waymo has moved slow and steady, offering only limited service areas, growing them incrementally, and slowing things down when issues arise. It's clear that this is the right strategy. The proof is that when Cruise abandoned this approach, it paid a steep price”.

In contrast, Tesla has often taken a more-confrontational stance toward regulators. Take, for instance, its ongoing struggle against requirements that motor vehicles be sold through dealerships (a case where I think the firm is in the right). At the moment, of course—meaning mid-March 2025—the CEO of Tesla, Elon Musk, has unparalleled influence on federal regulators in the USA. The cost of that has been incredible brand damage with regulators abroad: recently the UK and China raised significant concerns with Tesla’s driver-assist systems and imposed limitations on their use. Meanwhile, animus against Musk as a person is so high in certain quarters that the Canadian government, and perhaps some European countries, are contemplating tariffs specifically targeting Tesla vehicles; I would not be surprised if, when Tesla offers unsupervised Full Self Driving in these places, they are held to a higher standard than their competitors.

To quote myself again: “None of this inspires confidence that Tesla will be able to sustain good relations with any regulator long enough for its robotaxi business to succeed”.

I would only add that this cultural difference between firms extends beyond its leadership to its overall organizational philosophy. Waymo has demonstrated willingness to delay deployments when safety concerns arise, while Tesla has sometimes pushed forward with feature releases it should have known better than to deploy. In the highly regulated domain of automated driving, where a single incident can trigger widespread regulatory backlash (as seen with Cruise), these cultural differences have a high likelihood of affecting Tesla’s ability to scale its robotaxi business.

A Truly Uncharted Territory

I’m on record as saying that the robotaxi revolution is coming, and that it’s not coming as fast as I would like.

It’s certainly not coming as fast as Pueyo suggests. His ambitious timeline—robotaxis replacing human drivers by 2027 and personal cars starting to decline by 2028—significantly underestimates the technical, regulatory, infrastructure, and human factors that stand in its way.

The path forward for robotaxis resembles instead the incremental, city-by-city expansion demonstrated by Waymo over the past six years. Full autonomy will first establish itself in geographically limited areas with favorable weather, regulatory environments, and infrastructure before expanding outward. Trying to do it faster will only prompt backlash.

As that rollout happens, I would be cautious in assuming that Tesla will win the market. While Tesla possesses impressive manufacturing capabilities and advantages in data collection, Waymo's methodical approach to regulatory approval, system safety, and purpose-built data collection shouldn't be underestimated.

What does this mean?

• For investors, recognize that the robotaxi revolution is a marathon, not a sprint; patient capital will be rewarded, and bets on overnight transformations will not

• For policymakers, understand that you can maximize the public benefits of robotaxi, and minimize its disruption, by preparing the right regulations now; it will reach your city faster if you prepare the groundwork for it

• For the average person, don’t delay your next car purchase in anticipation of robotaxis making ownership obsolete next year

Pueyo is correct that robotaxis are here, and that they are going to become common. But he’s incorrect to assume that this change will happen quickly. Like many technological revolutions before it—from electrification to the internet—the transformation will follow an S-curve: slower than optimists predict at first, but accelerating as technical, regulatory, and infrastructure hurdles are cleared. The companies that succeed will be those that recognize this reality and plan accordingly, building sustainable businesses that can weather the inevitable challenges of bringing revolutionary technology to market.

In the meantime, I’ll continue to enjoy Pueyo’s dispatches from Uncharted Territories, but encourage him to temper his excitement with patience.