A Roadmap for a World Without Drivers

Recently, a number of analysts have written thoughtful pieces about the future of mobility in a world of self-driving cars. Often, the projections assume that most cars will transition to electric motors over time, replacing the internal combustion engine. Presented below are links to some of the better recent works:

Benedict Evans, of Andreessen Horowitz, offers his analysis here: http://ben-evans.com/benedictevans/2015/7/27/ways-to-think-about-cars

Tory Gattis, a fellow at Center for Opportunity Urbanism, writes here: http://www.newgeography.com/content/005024-preparing-impact-driverless-cars

KPMG offers an analysis of the impact of these changes on the auto insurance market. https://www.kpmg.com/US/en/IssuesAndInsights/ArticlesPublications/Documents/automobile-insurance-in-the-era-of-autonomous-vehicles-survey-results-june-2015.pdf

Brian Johnson, an analyst at Barclays, has a PDF on what he calls “disruptive mobility” here: http://orfe.princeton.edu/~alaink/SmartDrivingCars/PDFs/Brian_Johnson_DisruptiveMobility.072015.pdf

As an investor in public and private markets, I’ve spent a lot of time thinking about the future of transportation. At the moment, my portfolio includes a global auto OEM, two materials suppliers to automotive OEMs, and a semiconductor company that sells to auto OEMs. So this exercise is not theoretical for me — there are risks to be avoided, and opportunities to be captured, from understanding what’s going to happen in the the world of transportation.

We should start with the points about the future of transportation that are nearly “consensus.” I use that word with quotes, because each analyst, consultancy, and forecaster offers opinions that coalesce around these points, though they differ on questions of degree, timeline of adoption, and other details. With that said, the consensus:

• Uber, or someone like them, will offer autonomous shared vehicle services. Widespread adoption of such services would reduce the amount of vehicles needed to accomplish the same amount of transportation by between 75% and 90%, depending on assumptions about utilization rates and consumer preferences for sharing rides.
• Without the need to pay drivers, Uber-like services with autonomous vehicles (AVs) will cost 50% to 90% less than they do today.
• The improved safety of AVs will reduce insurance premiums between 50% and 90%.
• Much of the urban real estate dedicated to parking will get re-purposed.
• There will be much less traffic, even if shared AVs do not take off, from the more efficient driving patterns of AVs.
• Vehicle miles traveled per person will not change much from today.
• The first vehicles on the market will arrive between 2017 and 2020, with rapid changes to the transportation infrastructure following in short order.

Let’s call this the Consensus Model, and let’s stipulate that much of the Consensus Model is correct. Nevertheless, there are profound security risks that will delay and complicate the full deployment of AVs. Once those risks are addressed, consumers will respond to the low cost and high convenience of AVs by increasing their consumption of transportation, most likely to high levels not contemplated by most analysts. And if those two forecasts are correct, most investors and analysts are making fundamental mistakes about the implications of AVs on existing industries, and on industries that will emerge once enabled by AVs. Let’s discuss these issues in order.

Security Risks and Barriers to Adoption

Most analysts are familiar with technology diffusion models. Whether you call them hype cycles, installation and deployment phases, or anything else, the models all distinguish between early hype, the difficulty of implementation, and the eventual success in deployment. Just this week, Gartner placed autonomous vehicles at the pinnacle of hype on their famous adoption model.

As summarized in Fortune, Gartner characterizes the barriers to adoption as technological and legal. The technological barriers to adoption are well known — more capable sensors, cheaper LIDAR, and more powerful software able to adapt to any conditions, etc. The legal barriers are familiar to anyone who has watched the growth of Uber over the last five years.

But there are other barriers to adoption that will be even more difficult to overcome. Most importantly, when autonomous vehicles (AVs) become common, we will all have to get used to to the idea of thousands of cars driving around without passengers. Cars without passengers will present security and safety risks that have never been faced before.

First, whether they are owned by an individual or by a fleet, AVs are the greatest force multiplier to emerge in decades for criminals and terrorists. Whether you’re a school shooter or a religious extremist, the biggest barrier to carrying out your plan is the risk of getting caught or killed by law enforcement. Only the most extreme mental illness, depraved hatred, or religious fervor can motivate someone to take on those risks as part of a plan to harm other people.

Autonomous vehicles neutralize those risks, and they open the door for new types of crime not possible today. A future Timothy McVeigh will not need to drive a truck full of fertilizer to the place he intends to detonate it. A burner email account, a prepaid debit card purchased with cash, and an account, tied to that burner email, with an AV car service will get him a long way to being able to place explosives near crowds, without ever being there himself. How will law enforcement solve physical, violent crimes committed by people who were never at the scene of the crime?

A more recent example is instructive. Dzhokhar and Tamerlan Tsarnaev were identified by an examination of footage from numerous private security cameras that were recording the crowd in downtown Boston during the Marathon. Imagine if they could have dispatched their bombs in the trunk of a car that they were never in themselves? Catching them might have been an order of magnitude more difficult than it was.

Now, a utilitarian would gladly accept an increase in terrorism and crime deaths, in exchange for the vastly greater reduction in accidental traffic deaths that would be prevented by AVs. But we are not utilitarians. We are Americans. And in America, a human life is somehow worth more when it is taken with an explosive than with a bullet.

The reaction to the first car bombing using an AV is going to be massive, and it’s going to be stupid. CNN will go into “missing airplane” mode. There will be calls for the government to issue a stop to all AV operations, much in the same way that the FAA ordered a ground stop after 9/11. But unlike 9/11, which involved a decades-old transportation infrastructure, the first AV bombing will use an infrastructure in its infancy, one that will be much easier to shut down. That shutdown could stretch from temporary to quasi-permanent with ease, as security professionals grapple with the technical challenge of distinguishing between safe, legitimate payloads and payloads that are intended to harm.

The scenario described above — using an AV to commit a violent crime — involves no hacking. Hacking is the second major barrier to adoption that will present unique problems to AVs. Any car with driver controls that are accessible via the Internet presents an amazing target for a hacker. The recent staged disabling of a Jeep Cherokee, which did not involve an autonomous vehicle, is a taste of things to come.

Hacking AVs need not be malicious to present problems. One can easily imagine a teenager sending every AV in his city to his high school at the same time, with or without bewildered passengers stuck inside them. Or, a prankster could simply lock the doors on an entire fleet of vehicles, trapping passengers inside until a solution could be found. The possibilities are endless, mostly because AVs are going to be first Internet of Things product category to be mobile and large enough to transport humans and cargo. Compared to hacking an air conditioner or a remote camera, the average hacker will find AVs an irresistible target. And that’s without even considering hackers with more damaging intentions, like those who hacked Sony Pictures or the federal Office of Personnel Management. Hackers motivated by destruction, rather than thrills, will be even more dangerous when they gain control of AVs.

Of course, raising concerns about how bad actors will use AVs for crime or for hacking is not the same thing as saying that we should ban AVs. The economic and social benefits of AVs will outweigh the risks. But the risks will be novel risks, unlike other risks we’ve encountered before, and they will take time to mitigate. That interval, during which technologists will address many of these issues, will delay deployment of AVs, but it should not stop the deployment.

The third barrier to adoption is scaling and adoption time. It’s hard enough to scale in software — just ask Google, where most Android users are still on versions between two and four years old. In the automotive world, hybrid cars are a good parallel for AVs. As we can see, adoption was not automatic:

Source: http://mnordan.com/2013/02/28/the-very-curious-hybrid-boom/

Now, hybrid cars differ from AVs in material ways. The benefit of adopting hybrid technology — saving fuel— is a much less valuable benefit, with much narrower consumer appeal, than full AV capability. On the other hand, hybrid cars fit seamlessly into the existing use case and industry structure of traditional vehicles, while AVs are far more disruptive. Finally, hybrid cars offer no advantages for the pursuit of mischief or damage relative to traditional gasoline-powered cars.

Given those precedents and caveats, what can we expect the adoption curve to look like for AVs? I believe the adoption will proceed in several stages.

1. Car companies (or tech companies entering the auto business) will release AVs in limited quantities at first, perhaps no more than 10,000 units in the first year. These first AVs will be a rounding error in a new car market that moves 17 million units a year in North America, and in a fleet of over 250 million cars on the road today.
2. Among the first 10,000 buyers, or the first 10,000 users of an AV service like Uber, will be people who use the cars in ways that frighten people, create news, and generate demands for regulation. It could be a confused 88-year-old who gets in a car in Kansas and inadvertently drives to Times Square. Or it could be a McVeigh type seeking to create real damage.
3. After the first security event involving an AV, production of new AVs and use of the AVs already on the road will be stopped, reminiscent of 9/11. Every level of government, from your city’s taxi commission to the federal Department of Transportation, will want to get involved. The ground stop might last well past a year. The extreme novelty of the risks and challenges to be addressed will make this policy response more difficult than the local policy response to the rise of Uber and Lyft, and more challenging than the security response to 9/11. America created an entire new Cabinet level bureaucracy, the Department of Homeland Security, to address 9/11. We could expect a response similar in scale to the security threats posed by AVs.
4. Sales will resume at some point. After that, market share will climb at a rate faster than the share gains for hybrid cars. We might reasonably expect 25% market share of new vehicle sales after ten years, and over 75% after 15 or 20 years. An adoption curve this fast would be roughly comparable to recent innovations like cell phones or DVD players, which were much easier to deploy than AVs will be.
5. At that pace, it will take more than 30 years for the full benefits of AVs to diffuse into the economy. A protracted, generation-long deployment phase for an expensive, capital intensive new technology is highly congruent with past examples.

The VMT Rebound

Now that we’ve discussed barriers to adoption and the likely scale of the deployment phase, let’s talk about the second order effects that are likely to develop as AVs take market share. Among the forecasts of the Consensus Model, I believe the prediction of constant Vehicle Miles Traveled (VMT) per person to be the most suspect. Traveling a mile in an AV is going to be deliriously cheap, and when we make something cheaper, we consume a lot more of it.

This concept is not new. In the nineteenth century, the economist William Stanley Jevons observed that as coal got cheaper, consumers used more coal. Economists have subsequently observed the Jevons Paradox in action in other markets, including traffic congestion, energy efficiency, and the consumption of basic materials.

The futurist George Gilder, who was influential during the early years of the Internet, makes a related point. Every economic age, Gilder says, is defined by a key scarcity and a key abundance. When a new technology emerges to change the relationship between scarcity and abundance, old business models break as consumers adopt new behaviors that would have been impossible under the old framework. As the deployment phase matures, the business models that win are ones that would have looked wasteful, frivolous, and even decadent under the ancien régime.

The most important scarcities imposed by transportation are time, attention, cost, and the actions of rival drivers. Driving somewhere takes time. During that time, it takes nearly your full attention (one hopes). Every mile you drive costs money, mostly in the form of fuel and vehicle depreciation. And if you want to travel to the same place as everyone else, at the same time, you will face traffic and pay more in time and attention cost to get there.

So — how cheap will AVs be compared to traditional vehicles? Brian Johnson, in his recent report for Barclays, provides framework for assessing the cost reductions. Johnson suggests that there will be four types of vehicles in the future: traditional cars that still require a human driver, “family autonomous vehicles” which are owned by consumers and used exclusively within a single household, “shared autonomous vehicles” (SAVs) owned and deployed by fleets in a model similar to Uber and Hertz, and “pooled shared autonomous vehicles” (PSAVs) that are like the preceding category, but which carry more than one passenger at time, like Uber Pool or Lyft Line.

Oddly, Johnson’s slides implies that fuel cells will power the SAVs and PSAVs of the future, despite much greater progress by battery-powered cars. For our purpose, we can ignore this technicality and continue to assume that most AVs will also be EVs.

Johnson’s analysis, which assumes 12,000 VMT/year, suggests that traveling one mile in an SAV or PSAV will cost between $0.08 to $0.44. On an annual basis, that works out to $960 to $5280. According to the AAA, an American driving an average sedan will spend $0.58 per mile, for an annual spend of $8700, though their analysis assumes VMT of 15,000 per year. Even accounting for the different VMT assumptions, it’s clear that SAV services will reduce the cost of transportation for consumers by thousands of dollars a year.

More importantly, the cost in attention associated with transportation will drop to nearly zero. The average American could shift some of the 5.5 hours of television watched per day into the car, and end up with vastly more personal time once freed from the need to pay attention to the road. This possibility has led many people to predict that AVs could enable further suburban sprawl as the costs of transportation fall. A person who moves to a more distant exurb but commutes via a PSAV will pay less money for transportation, have more time for entertainment, and will also pay lower, exurban prices for their housing. It will be an irresistible combination, and it will be just one of many ways that VMT will ratchet upwards once each marginal mile loses its cost in dollars and attention.

Autonomous vehicles will also allow new transportation use cases to emerge. Short haul flights and regional train travel will be hard pressed to compete against trips in FAVs, if the electricity needed to drive an autonomous Tesla from Los Angeles to San Francisco costs less than $10.00. Compared to a $75 fare on Southwest, the cost savings will be high (especially if more than one person is traveling) and the time penalty will be minimal for distances of under 500 miles, when calculated on a door to door basis. Even for those consumers who choose to forego ownership of a vehicle to use a shared service, AVs promise vastly cheaper inter-city travel. For a service like Megabus, the cost reduction when batteries replace gasoline and software replaces drivers is going to be massive. The existing fleet of A320 and 737 aircraft may get redeployed to long haul flights.

It shouldn’t be surprising that planners and analysts look at the prospect for higher VMT enabled by AVs and react with apprehension. After all, when humans are driving internal combustion engine vehicles, increasing VMT increases the associated externalities: traffic, carbon emissions, and the need for parking lots, among many other associated costs. Writing in Slate, Joseph Coughlin and Luke Yoquinto of MIT conclude with a cautionary note:

But still, the high-speed autonomous commute stands as a real possibility. That’s why we should start thinking now about its implications — both positive and negative. We need to make a deliberate decision about how we will live in the future, before the self-driving car makes it for us.

While calls for technocratic planning are understandable, preemptive solutions are unnecessary for adjusting to the increase in VMT that AVs will bring. All the negative externalities that would call for policy solutions are collapsed by the technology of AVs/EVs themselves. Traffic congestion will be reduced by the better driving, closer spacing and platooning capabilities of software. Pollution and carbon emissions would be irrelevant if the AVs were also EVs, and in fact the carbon advantage of EVs would increase over time as the grid deploys more renewables to meet renewable portfolio standards. And of course there would be far less need for parking.

Also, demand aggregation, currently in its infancy, has the potential to further reduce the cost of PSAV services. Just this week, Uber began testing “smart routes” that incentivize users with lower fares to catch an UberPool on a major street nearby, making the overall system more efficient. There will be more innovation along these lines. Some of the ideas that have been discussed include neighborhood commuter jitneys that depart for downtown every 15 minutes, AVs as demand aggregation for public transport, and AV versions of company-specific transportation options, like the Google bus service in the Bay Area.

Finally, new users of transportation, like children and seniors, are likely to travel many multiples of the distance they do today once AVs lower the costs and barriers to their safe transportation.

Every one of these AV use cases will share a common characteristic: they will be outrageously cheap (in dollars, but also in time, attention, and non-excludability) compared with today’s transportation options. If there is a rebound effect, the question then becomes — how big will it be? How elastic is the demand curve for transportation? A few other cases of rebound effect are instructive.

Among household appliances, none have gotten more efficient, more quickly, than refrigerators and air conditioners. Through 2010, “the average refrigerator sold in the United States…uses three-quarters less energy than the 1975 average, even though it is 20% larger and costs 60% less.” But consumers have spent that savings on more refrigeration. A suburban house is now likely to have a fridge in the kitchen, one in the basement, a wine chiller somewhere in the house, and a mini fridge in a home office, in addition to the 20% larger refrigerator in the kitchen. The story is similar with air conditioning. Between 1993 and 2005, “the energy efficiency of residential air-conditioning equipment improved twenty-eight per cent, but energy consumption for A.C. by the average air-conditioned household rose thirty-seven per cent.”

Back to AVs. The Consensus Model holds that vehicle sales will fall, and the total automotive fleet will shrink, in response to the deployment of SAV & PSAV services. Barclays forecasts that the fleet will be 60% smaller and sales will be down 40%.

Their assumptions, presented in some of the previous slides shown above, are that an SAV will displace 9 traditional vehicles and travel 64,000 miles per year. They further assume that a PSAV will displace up to 18 traditional vehicles, but that they will also travel only 64,000 miles per year. That’s not very different from a New York City taxi, which travels 70,000 miles per year.

These load factors are strangely low. A car traveling 64,000 miles per year at an average of 20 miles per hour will only be in service for 8.7 hours per day. There are reasons to believe AVs will travel much faster than 20 miles per hour, and that they will be in use more than 9 hours per day.

First, platooning will allow AVs on freeways to travel at much higher speed and density than traditional cars. Second, at full deployment, waiting at red lights will be a thing of the past, as this video shared by Benedict Evans sugests.

Third, charge times are likely to be only an hour or two per day, using modern high power charging infrastructure. Fleet owners will have every incentive to charge quickly, getting their cars back on the road to earn revenue.

So, how much more transportation will people consume? With costs likely to fall between 50% to 90%, a consumer could increase VMT by anywhere from 2X to more than 5X while spending no more on transportation than today. Relative to the Consensus Model, I believe we will have more people, in more AVs, traveling more miles every year, requiring a larger fleet than is assumed, and higher annual vehicle production. Further, once the security risks have been addressed early in the installation phase, during the deployment phase this increase in VMT will be accompanied by none of the costs and externalities associated with higher VMT today.

Finally, we must consider the possibility, which may be an inevitability, that VMT with a person in the car will come to be the minority of miles traveled by AVs on the highways of the future. In his post, Benedict Evans hints at the kinds of on-demand services that might emerge:

They have the potential greatly to expand the adoption of on-demand, and so to transform who buys cars and why.

Removing the drivers from an on-demand car service cuts the cost, since you don’t have to pay them and also since lower accident rates mean cheaper insurance (though this applies to your own car too). But in addition, autonomous cars expand supply for on-demand services, since many more cars are available to be used for on-demand when their owners aren’t using them. This will creates all sorts of second-order effects and feedback loops.

Any service priced at a premium today for its on-demand convenience, and available only to the subset of the population able to pay for convenience, could be ten to a hundred times larger once AVs lower the price of convenience nearly to zero. Delivery services are just the easiest type of service to imagine; there will be others.

We can begin to glimpse the future demand response to AVs. People won’t travel just 12,000 to 15,000 miles per year. They might travel 30,000 to 50,000 miles per year. And they might generate non-occupied AV journeys bringing them goods and services on demand that would create another 50,000 to 100,000 VMT per year. Total transportation demand, in the face of per mile cost reductions of 50% to 90%, might logically respond by rebounding to a new equilibrium where consumption is 10X the prior demand.

Implications for Investors

To recap, my thesis for the installation and deployment of AVs is as follows:

1. We are not even in the installation phase yet, though we can glimpse it.
2. During the installation phase, we will face significant security risks inherent to AVs that will be difficult to address.
3. The government is likely to overreact to these risks.
4. Once past the security risks, the low cost and high benefits of AVs will lead to much higher adoption, and many more miles travelled (by cars with humans in them, and without humans in them) than anyone is assuming today.
5. The increase in VMT will have very few externalities associated with it, because the externalities are exclusively associated with human drivers and internal combustion engines.

What, then, does this scenario imply as an investor?

First, very few companies founded during the installation phase of a technology survive to be the leaders of the deployment phase. Among leading US Internet companies, only Amazon, eBay, Google, Netflix, PayPal, and Priceline were founded in the 1990s. It’s equally likely that over the next few years, among the dozens or hundreds of startups founded to take on the AV market, only one or two a year will be thriving a decade hence.

Second, incumbent producers won’t face any risk of obsolescence until well into the deployment phase. And the more physical infrastructure in the supply chain that must be retooled, the longer into the deployment phase it will take. That’s why music companies suffered from the Internet before TV broadcasters and cable networks, who are only now facing the structural changes to their business models.

With the deployment phase for AVs likely to be more than a decade away, investors who are worried about the impact of AVs on traditional car companies are premature. It would be like selling Disney in 1998 because you were worried about Broadcast.com. Like Disney in 1998, automotive OEMs have plenty of time to prepare for an autonomous future. One or two might even surprise us and succeed with the transition.

Third, if the rebound effect of lower prices emerges as I have forecast, then the size of the fleet, and the annual number of vehicles produced, will be a lot closer to current numbers than most analysts have modeled. Today, American consumers buy over 17 million vehicles that they drive around 12,000 miles per year. In the future, American consumers and fleets might still be buying around 17 million vehicles per year, but those vehicles could travel 120,000 miles per year rather than 12,000. Some of those miles might have a human passenger, while others won’t. Even allowing for the greater simplicity of EVs, with fewer moving parts than internal combustion engine vehicles, they will wear out quickly if traveling over 100,000 miles per year.

We can make educated guesses about the frequency of the replacement cycle. As mentioned earlier, the average taxi in New York travels 70,000 miles per year. The same document notes that the average New York taxi is just over three years old.

With these data points, we can begin to estimate the replacement cycle for AVs. We can assume that an AV is driven at either half, equal, or double the 70,000 miles of a typical New York taxi. Similarly, we can estimate that fleets will turn over their vehicles after they have traveled between 100,000 and 500,000 miles. With those parameters, we can estimate service life in years.

So, AV fleet owners are likely to replace their vehicles every few years. If we can estimate the size of the fleet, we can then use the replacement cycle in years to estimate annual AV unit sales, similar to the widely reported SAAR for traditional vehicles today.

Johnson’s report for Barclays suggests that without a rebound effect in VMT/person, transportation demand in the US could be met with 70 million traditional vehicles and 33 million autonomous vehicles, split between FAVS, SAVs, and PSAVs. If we assume, simplistically, that the rebound effect generates demand for a fleet 1.5 to 2.0 times Johnson’s estimates, we can arrive at a guess.

In the following table, the fleet replacement cycle in years is from the middle column in the previous table showing vehicle service life, and the fleet size row is Johnson’s 33 million estimate in the first column, scaled by 1.5 and 2.0 in the following two columns.

It should be apparent that this model is acutely sensitive to inputs — AV sales per year might 4.62 million, or they might be 46.2 million! But the middle cases — in which vehicles are replaced after three to four years (corresponding to 200k to 300k lifetime miles) and the fleet size between 1.o and 2.0 times Johnson’s estimates — yields sales estimates that are not too different from the current 17 million units per year. All that’s required to keep sales roughly flat is a fleet 50% bigger than Johnson’s assumption, replaced every three years. So AVs do not portend apocalypse for Detroit.

Fourth, while the automotive OEMs face the risk of disruption during the deployment phase more than a decade from now, most auto suppliers will not be affected by the transition to AVs and EVs. The cars of the future will still need tires, aluminum structural components and body panels, and infotainment systems. The analogy with television is strong — we may watch television differently than we did in 1995, but we still need actors, directors, and sceenwriters just as much today as we did in the past. That said, auto insurance companies, body shops, and suppliers of parts exclusive to ICE engines, face a very real risk of permanent obsolescence.

Fifth, as on-demand transportation becomes abundant and absurdly cheap, entrepreneurs will develop wonderful new services. If those services existed with today’s cost of transportation, they would feel luxurious, sinful even. But at the new, lower cost of transportation enabled by AVs, these new services will have mass market appeal. Just because it’s hard to envision any specific service that might emerge, doesn’t make it any less inevitable that many such services will be developed.

Finally, even with the security risks that will emerge, we should welcome this transition. The consumer surplus will be worth thousands of dollars per year per capita. AVs will be better than human-driven cars in nearly every way, and while it will take a while to get to the AV future, it will be worth it.

*Thanks to Conor Sen, who reviewed an early draft of this post and provided valuable feedback.