Self-driving cars could generate trillions of dollars of revenue beginning in the 2020s. GM (GM) is one of the companies pushing most aggressively to seize this opportunity. GM aims to deploy fully autonomous cars for a ride-hailing service in 2019.

GM is vertically integrating the three layers of self-driving cars: vehicle hardware, the self-driving software, and the ride-hailing service. This differentiates it from its competitor Waymo (GOOG, GOOGL), which has no announced plans for vertical integration.

However, GM is not using production fleet learning: the collection of neural network training data from its fleet of production cars. This could be the disadvantage that causes GM to deploy self-driving cars less quickly, less widely, and at a lower level of capability than Tesla (TSLA), which is using production fleet learning.

Why vertically integrate?

Vertical integration makes sense for two reasons. The first reason is technical: it allows for a tighter coupling between vehicle hardware and self-driving software. The second reason is financial: controlling all three layers means that a company will capture all of the revenue from an Uber-like autonomous ride-hailing service.

In the above graph, ARK Invest estimates the breakdown of revenue between the manufacturer of the vehicle hardware, the developer of the self-driving software platform, and the operator of the ride-hailing service. By vertically integrating these three layers, a company can capture all the revenue.

Software/service integration

My observation is that self-driving software developers are strongly incentivized to operate their own ride-hailing services in order to capture that much larger share of revenue. Operating a ride-hailing service is by far the easiest of the three layers. Why give away that revenue to some company with an app? Just develop your own app. If you can get self-driving cars to work, you can certainly do that.

So, I don’t think there will be any distinction between the self-driving software developer and the ride-hailing service operator. Waymo, a self-driving software developer, is testing an autonomous ride-hailing service. Uber, a non-autonomous ride-hailing service operator, is attempting to develop its own self-driving software in-house. The self-driving software and the ride-hailing service will be vertically integrated.

Hardware/software integration

Vehicle manufacturers are also disincentivized to partner with a company like Waymo — since that means helping their competition — unless the manufacturer is cut in on the revenue that comes from operating a ride-hailing service. Even so, any deal will have a finite duration. It’s in the interests of a vehicle manufacturer to acquire one of the many startups working on self-driving car software. Otherwise, when the deal ends, Waymo may simply play that manufacturer out of the game.

Conversely, Waymo risks getting played out of the game if manufacturers refuse to supply it with custom-built cars that can be integrated with Waymo’s sensors and software. This analysis also applies to Uber and other non-manufacturers working on self-driving cars.

This is the advantage that GM has over Waymo. As a car manufacturer with an in-house software development effort — its own internal version of Waymo — it can capture all the revenue associated with self-driving cars. GM can also avoid the strategic complexity of self-driving partnerships in a zero-sum game where the risk of non-cooperation is high. That’s in addition to the technical advantage of tight hardware/software coupling.

Lead generation

A brief side note: in the graph, ARK Invest also includes payments to third-parties for lead generation. My hunch is that services like the Tesla Network or GM’s Cruise Anywhere won’t need any exogenous lead generation. Self-driving cars automatically catch people’s attention. I think the technological marvel will get people to try an autonomous ride-hailing service, and the usefulness and cost savings will make them long-term customers. Customers may get a small discount when their friends use a referral code, but that’s about all I can foresee. Perhaps ARK Invest sees something I don’t.

The advantage of production fleet learning

I believe GM has most likely chosen the right strategy by vertically integrating. However, I believe the company will continue to be at a technological disadvantage if it does not implement production fleet learning.

At last count, GM’s Cruise subsidiary is testing 180 self-driving cars in two locales: San Francisco (including downtown) and suburbs in the Phoenix metro area. Cruise CEO Kyle Vogt said in November that the goal is reach a rate of 1 million miles driven per month in early 2018.

Tesla, by comparison, has over 125,000 cars equipped with self-driving hardware being driven by customers, sometimes assisted by Enhanced Autopilot, all around the world. That’s in addition to Tesla’s undisclosed number of test vehicles.

Simply multiplying the number of vehicles by the average miles driven in the U.S., we can estimate that Tesla’s fleet of Hardware 2 vehicles (as they are called) are driving somewhere around 124 million miles per month. This rate will accelerate as more Model S and X vehicles are sold, and especially as Model 3 production ramps up.

As of yet, GM has not announced plans to equip its production vehicles with self-driving hardware. Instead, it has announced plans to build a car with no steering wheel or pedals that is intended for autonomous ride-hailing, not for sale to customers.

Tesla’s massive amount of driving data may not turn out to be an advantage — to know for sure would require information nobody has yet. Here’s one way it could be an advantage: There may be rare driving circumstances that occur only once in 100 million miles. At a rate of 1 million miles per month, it would take GM over eight years to encounter such a case. At a rate of 124 million miles per month, Tesla will encounter it in a little over three weeks.

Tesla has not shared much detail about how its production fleet learning system works, but this is my best guess based on what we know. The software running in a customer’s Hardware 2 car can flag a situation as unusual based on sensor data or flag that the human driver took a different action than the self-driving software, running in “shadow mode”, would have taken. Then the car will upload all the associated data, including video from the cameras, to Tesla’s engineers for review.

If there are enough situations to confound a self-driving car that occur only once in 100 million miles, and these sometimes cause a crash, it is possible that a self-driving car trained on an insufficiently large data set will exceed the human rate of 190 crashes per 100 million miles. This is hypothetical. As I said, we do not yet have that information. However, if this problem exists, it is a problem that GM can’t solve with its current approach, and a problem Tesla can solve.

Production fleet learning could also, theoretically, simply train Tesla’s neural networks on routine driving tasks at a much higher rate than GM’s, leading to a higher success rate sooner. This could allow Tesla to launch earlier than GM with software that has a higher level of capability.

Finally, Tesla can use the sensors on its fleet of production vehicles to collate up-to-date HD maps of everywhere they drive regularly. Whereas GM plans to launch its autonomous ride-hailing service in a few select cities to start, Tesla could launch in a much larger geographic area.

Revenue and market cap implications

ARK Invest's model forecasts that autonomous ride-hailing services will generate $2 trillion in annual revenue in 2025. GM currently has an 8.2% global market share in auto sales. The same market share in autonomous ride-hailing would generate $164 billion in annual revenue for the company.

It might be tempting to use GM's 0.38 price/sales ratio or the auto industry average 0.52 price/sales ratio to calculate market cap contribution. However, there is an opportunity for higher margins in the autonomous ride-hailing market than in auto sales. This is because the revenue per vehicle is so much higher.

Here's a rough estimate. The projected lifetime of an electric vehicle is 500,000 miles. One taxi driver has gone 250,000 miles in a Tesla Model S with minimal wear. ARK Invest forecasts a price of 35 cents per mile for autonomous ride-hailing services. A study on Uber found that 64% of miles are travelled with a passenger in the car. This means an autonomous electric vehicle can expect 320,000 paid miles in its ride-hailing career. At 35 cents per mile, that's $112,000 in revenue per car. A Chevy Bolt's base price is $37,500. An autonomous Chevy Bolt would generate almost 3x as much revenue. The expenses would simply be the costs of running the vehicle.

The key is bringing down the cost of the autonomy hardware, particularly LIDAR, to a point where the gross margins of autonomous ride-hailing exceed that of auto sales. This is why GM acquired a LIDAR manufacturer in an effort to bring the cost of LIDAR units down from $20,000 today to $500. It should be noted as a risk for GM that it is not yet producing the sensors it plans to use for autonomous ride-hailing at the desired cost.

The autonomous ride-hailing business is more like the taxi business than the auto sales business. The price per mile will be much lower, but expenses per mile will also be much lower because companies won't need to hire drivers. According to data from the government of Canada, the average net margin for the 97% of taxi and limousine services that are profitable is 31%. Assuming the same net margin for autonomous ride-hailing, GM would generate $50.8 billion in net profit on $164 billion of revenue.

Using the S&P 500's historical average price/earnings ratio of 15.7, $50.8 billion in profit would be valued at $798 billion. That's a 1,270% increase from GM's current market cap of around $63 billion.

Conclusion

Out of all the auto manufacturers, GM and Tesla are pushing ahead the hardest on autonomous driving. Both plan to launch fully autonomous cars in 2019. Both are taking a vertically integrated approach, which I believe is the right decision. This gives them an advantage over non-vertically integrated companies like Waymo.

Tesla, however, is pushing ahead harder than GM. By equipping 120,000 vehicles and counting with self-driving hardware, it is setting itself up to collect billions of miles of driving data per year. This accords Tesla a known advantage in HD maps and plausible advantages in both routine driving and dealing with rare events.

Autonomous ride-hailing will be a much larger market than these companies’ core auto sales business. The growth potential is immense. However, it’s also a market that may admit fewer major players than auto sales. Since an autonomous ride-hailing service will improve over time, two or three companies with a head start in a particular country could leave the rest permanently behind. Auto manufacturers that are not aggressively pursuing autonomous driving right now should be considered high risk long-term investments.