Originally published in November 2017 here

Why incumbent OEMs don’t need to start building their own AI brain for self-driving cars even though autonomous vehicles (AV) will decide their future.

Opinions expressed are solely my own and do not express the views or opinions of my employer.

Not a day goes by without some new announcement in the auto industry, whether it’s the launch of a new initiative or spin-off, a new strategic partnership, or a new technological advancement. As I have written in my previous article Intro to self-driving cars the developments boil down to three broad shifts that are transforming the auto industry as we know it: the shift from human drivers to Autonomous Vehicles (AV); the shift from gas-powered to electric-powered vehicles; and the shift from car ownership to mobility services. From those trends AV is the most important one because AV will eventually enable the breakthrough of the other two trends - not the other way round. AV will make mobility services affordable and it will give electric vehicles the hard advantage required to overtake gasoline vehicles even when they are still only on pair with them economically.

AV is key – so what now?

So if AV is the primary focus, should OEMs now focus on developing & building the “AI brain” necessary for AV themselves to be a leader in this space?

The AI brain is the decision-making hardware and software combination of the car that assesses threats, plans maneuvers, and negotiates the traffic environment. In the next 5-10 years progress will be fast with significant improvements every year. Cars will be continuously upgraded from current Level 2/3 autonomy now to full Level 5 autonomy.

Cars with with better AV capabilities will free up more commute time for the drivers, will feel smoother and have longer range. AV capabilities will therefore be a key source of distinction for OEMs. So do car OEMs have to develop their own “AI brain” to make AV work? Paradoxically the answer is “no” for most OEMs. To understand why it makes sense to look at the following questions:

• Would OEMs be able to develop AV capabilities better & faster by doing it themselves or by working with suppliers?
• Would a technological advantage translate into significant higher market share and/or profits?

Fur the purpose of this article I will differentiate between 3 types of players in the market:

• OEMs: classical car OEMs who have been in the market for more than 10 years
• NEMs (New equipment manufacturers): New players like Tesla
• Suppliers: Mostly tech companies like Google (Waymo), NVIDIA & others

Let’s start with the first question:

Would OEMs be able to develop AV capabilities better & faster by doing it themselves or by working with suppliers?

Replacing the human driver by developing the AI brain for AV is a complex task which has been going on for over 20 years by now. Basically the AI brain needs to be able to sense (by using sensors & cameras), infer (understand) and act accordingly.

AI based technology is unique in the sense that it cannot be solved with “brainpower” alone – you need the right data to train your algorithm as well. Therefore in order to win this race for technological superiority you would need:

• The right talent
• The right data
• The right approach

The right talent

There is a big shortage of deep learning talent and not everybody will be able to get the right people. By far not enough people have been studying the required fields. Incumbent car OEMs are hindered in hiring by a (perceived) traditional culture with a non-software / agile mindset. They are also clearly the junior partners compared to big software companies like Google both in terms of market value and number of software engineers employed. In short: Auto OEMs will have it much harder to attract the right talent compared to “classic” software companies like Google & NVIDIA or the NEMs.

(Interestingly this is another paradox: the war for AI brains will be decided who can get most human brains.)

Ranking:

1. Suppliers
2. NEMs
3. OEMs

The right data

The way machine learning works is that you need data to feed the algorithms to “learn” how to drive. You could argue that OEMs have an advantage here because they have much more cars on the road and therefore also access to much more “driving data”. Unfortunately for the OEMs this is wrong: You need the “right” data to train your algorithm. You basically need to first standardize the hardware you want to use to operate the self-driving car and then get data for exactly this hardware setup. The current AI algorithms are mostly not good enough to transfer the learnings from one hardware-setup to another one. Therefore Tesla has an advantage in this dimension. They at least say that their current cars including the upcoming Tesla 3 will be able to be upgraded to full Level 5 autonomy. This would give them a major advantage in this field if it works as expected.

OEMs have two options to counter this approach from Tesla: They can start implementing the required hardware in their upmarket cars like AUDI is doing it with the new AUDI A8. In those cars the additional hardware costs are easier to swallow. Also they are very likely to be used a lot so they will give the OEMs lots of (data) miles on the road. OEMs can then further share this data with their suppliers who also get the data from other OEMs. With this pooling approach they are very likely to get enough data to develop their own AVs.

In the supplier space Google (Waymo) has probably gathered the most data so far on their own. At least in California they have gathered more than 10 times more Miles than any other company so far. They have just recently expanded these tests to a live taxi service.

Ranking:

• NEMs (Tesla currently leading)
• Suppliers (Google/Waymo)
• OEMs

The right approach:

The best people & data can mean nothing if the whole team walks into the wrong direction. Choosing the right approach to solve the “AV problem” can make or break the whole endeavour.

So what are the main decisions you need to take to develop an AV?

• Hardware: Which hardware to use?
• Vertical integration: Which components do you build yourself?
• Algorithms: Which algorithms do you use for your data?

Hardware

One of the biggest questions here is if LIDAR is required to observe the environment or “normal” optical cameras are enough. LIDAR is using Lasers to detect the environment so it is independent of fog, night or other other environmental factors. The downside of LIDAR is that it is relatively expensive. Tesla is betting on Level 5 autonomy without LIDAR while Google / Waymo is putting a lot of work into making LIDAR cheaper to use. It is currently unclear which approach will win in the end. However I do expect that this will be resolved in the next 1-3 years and then everybody will basically work against a similar and more or less standardized hardware configuration.

If LIDAR wins the current TESLA approach becomes obsolete and the first AV will probably be roled out in mobility services like robo taxis where the additional LIDAR cost is not really a problem because AV still offsets the costs of a even more expensive human driver.

Vertical Integration

This is basically the old iPhone vs. Android questions. Apple has chosen to be fully vertically integrated by having control over all parts of their products. They may choose to get suppliers for certain parts but in the end they have full control over all parts. If the market is not mature this normally allows for faster progress as you can make sure all parts work well together even when no standards exist (yet). However when the market matures Android was able to pick up and overtake it. In software the weird thing is that the mass market stuff is most often better than the specialized software because the best people want to work on the most-used software.

This also explains the current state of the market. Currently vertically integrated players like TESLA are leading however suppliers with the “Android approach” are picking up speed and will very likely eventually take over the market.

Algorithms

There are several algorithms & approaches to get the AI brain to drive the car based on the data used. For the purpose of this paper let’s just talk about 2 of the main options. You can either use an “End to End learning approach” or splice up the whole problem in several smaller ones with different approaches for each of the problems. Similarly to the hardware discussion currently it is not clear which approach will win.

Using a end-to-end learning approach will more likely result in a more direct jump from Level 2 to Level 5 autonomy. Splicing up the problem into several smaller issues will result in the cars getting smarter one step at a time. My personal guess is that there is no “one size fits all solution”. The key point here is that either way my guess is that you can only get 1-3 years of “algorithm advantage”. Software engineers share knowledge openly (and change jobs) and therefore any advantage will not be limited to one company for long.

So let’s now come back to the initial question:

Would OEMs be able to develop AV capabilities better & faster by doing it themselves or by working with suppliers?

Do they have the right people? They can get lucky in hiring but no unfair advantage compared to Suppliers like Waymo. Do they have the right data? Other players in the market have better data. Do they have the right approach? They can get lucky again but again no unfair advantage here. Also an advantage in algorithms will probably not last long.

So to answer the question: OEMs are very likely faster by using suppliers to develop the AV capabilities than by trying to build it themselves. Some bigger OEMs like Daimler may still be able to pull it off but in general they should rely to work with suppliers.

So let’s now look at the second question:

Will a technological advantage translate into significant higher market share and/or profits?

Or to frame it the other way round: If you are late will you lose significant market share? Will there be an “iPhone moment” for the traditional OEMs similar to what Nokia experienced in the smartphone world? In order to answer this I would like to differentiate between the private car market and the market for mobility services.

Private car market

Let’s first look at the private cars market: Is it like the smartphone market where a 3 year lead in technology leads to a total disruption of the market with big incumbents like Nokia failing? Or is it more like the TV market where a lead in technology rather results in slight changes in market share and where other factors such as marketing, availability, production capabilities are replacement cycles are more important?

Lets compare those 3 markets:

The main takeaway here is that the car market has really long replacement cycles. People don’t buy a new car every 1-2 years like they buy a new mobile phone. They buy a new car maybe once every 5-10 years on average. Therefore even if you manage to achieve full autonomous driving a few years before the rest of the market you will not be able to create an “iPhone moment”. Also if you are the first player you will probably have to prove over years that your cars are safe market by market so people are allowed to use the full AV mode. So you may win up to 10% or so of market share but this won’t be enough to kill off most of your competition.

So even if a vertically integrated player like Tesla manages to go to full AV earlier, it is very unlikey that Tesla will be able to create an “iPhone moment”. Companies like Daimler who have centered their brand around technical excellence will take a hit on their brand but they will eventually catch up based on their production capabilities & brand value.

So in the private cars market a technological advantage of up to 3 years will very likely not lead to significant shift in market share.

Mobility Service market

The market dynamics are different in the mobility services market. Here the replacement ratio is much shorter - probably more like 2-4 years than 5-10 years because the cars are used much more often. Also the production capabilities problem is (initially) much less of a problem. At least in order to build enough cars for a city you need more like 1000 cars than millions. Also the mobility services market is expected to generate the biggest share of the profits in 20 years So it is important to be successful in this space. Also players like Google could be real competitors to the OEMs in this space as you can see from the recent announcement of Google testing Level 4 capable robo taxis. Traditional suppliers like Bosch are not far behind with the promise to do similar tests 2018 in Germany . The obvious player UBER allegedly tried to steel technology from Google but is now facing legal issues and does not seem to have a good team in place.

So yes the market for mobility services is much less “secure” than the B2C market but it is still unlikely that this will cause an “iPhone moment” for the industry:

While mobility services will quickly become very profitable they will only represent 10-20% of the revenue in the next 20 years. People won’t give up their cars so quickly (see above, cars have a replacement ration of 5-10 years). As long as they have their cars they will continue to use them.

Furthermore regulatory issues will still take time and needs to be taken on market by market. So even if you manage to operate your fleet of robo taxis a few years earlier than the competition in some selected markets it will take years to scale this globally.

So will a technological advantage translate into significant higher market share and or profits?

As we have learned in the previous section the tech advantage in autonomous driving will be very likely only 1-3 years. This is not enough to create an “iPhone moment” for the traditional OEMs because they are protected by:

• long replacement cycles of cars
• production capability issues of new entrants
• regulatory issues which will slow down the adoption of this new technology

There will be increased profits for the players who manage to get mobility services “on the road” at scale first. However this will only be temporary as long as the incumbent players are able to catch up in max 1-3 years. However If they take much longer OEMs are in indeed danger of going extinct.

So what now?

So despite their fate being determined by their success in the AV space, most OEMs will be able to and should get the necessary technology from suppliers. Developing this technology themselves will very likely take longer and will be more costly than by working with suppliers. There will be always exceptions but for most OEMs instead of building the AI brain, these are the 3 issues OEMs should focus on to succeed in the new AV world:

• Fast follow on AV by working with suppliers (while keeping access to the data)
• Use their influence on the regulatory scheme to delay AV introduction
• Leverage their unfair advantage on production capabilities once they have AV capabilities

Fast follow on AV by working with suppliers (while keeping access to the data)

As discussed above it is important to keep the technology advantage for first mover to max 1-3 years. Therefore it is important to work with the right suppliers to get access to the technology as soon as possible. If you have the chance put the required sensors in your upmarket cars to get enough data to co-develop AV with your suppliers. Consider working with several suppliers at once to not get too dependent on one. This is basically the same as with working with all other suppliers.

When working with suppliers you need to make sure you own the data as well. Autonomous driving requires the vehicle to sense and respond to myriad conditions—weather, other vehicles, pedestrians, structures, and countless unforeseen obstacles. The brain doesn’t get smarter through traditional programming methods —smart algorithms. Instead, it depends on machine learning, in which it is constantly fed and trained on streams of new data. For this reason, the OEM, and not just the brain supplier, must have full access to the data.

In addition, by owning the data, the OEM can retain the option to develop its own proprietary brain at some future point, or to switch brain suppliers when necessary. Brain suppliers like Google need to be kept in check because with advances in robotic manufacturing, they could very well start building their own cars at some point and threaten OEMs’ existence then.

Furthermore the vehicle’s digital platform, or operating system, should enable updating every three to six months, vastly more frequently than the traditional three- to six-year model upgrade cycle. Because of this, OEM designers and engineers must think very differently than before. OEMs need to foster a new culture, one closer to that of the software industry than to a traditional hardware manufacturer. That’s not to suggest that OEMs should release products to the marketplace as “works in progress”. Instead, I mean that the new data and insights that are crucial to the optimal operation of the AV must be worked into upgrades in a far shorter time period. “Building your own platform” also does not mean they need to code every line of code themselves. They can build on what is available as much as possible, e.g. consider using genivi.org used by BMW & other.

Use their influence on the regulatory scheme to delay AV introduction

The space for self-driving cars is regulated for good reasons: For example a hacking or hijacking of an AV on the road with passengers inside would be disastrous. Still, OEMs must seek a good balance between security and practicality. Risk-tolerant competitors like Tesla are already moving ahead and gaining market share.

Tesla for example does not always use physically separate processors to run the different components of the car’s brain or infotainment systems. Tesla uses modern software architecture to ensure the AV brain always has sufficient computing resources available. Some OEMs find this design approach too risky, instead advocating completely separate hardware for each of those components. This is a delicate balance - OEMs need to avoid wanting to build the perfect solution; in the first iteration they just need to be statistically better than humans.

These kinds of discussions can & will also be used by the incumbent OEMs to actively delay the approval of fully Autonomous Vehicles. OEMs should have teams in place for all major markets to make sure AVs are only approved when they are also ready to launch.

Leverage their unfair advantage on production capabilities once they have AV ability

Once they have AV capabilities OEMs should use their “unfair advantage” in production capabilities to offer lower cost options and/or higher quality to consumers. While Tesla is proud to be able to offer a 30k+ Tesla 3, companies like Volkswagen or Toyota will be able to mass-produce AV capable cars for 20k EUR or less on scale. This will significantly drive market share.

This in turn will enable those companies to also win the mobility services market. It is very likely that all OEMs will offer to their B2C customers the service to “use” their private cars while they are not using it themselves. This will be more cost effective than companies owning cars only for B2B services since they are splitting the capital costs of these costs with their customers. If you are a premium company like Daimler you can still compete with Tesla and other NEMs on brand & production quality since Teslas are currently not very well built.

So the AV paradox is that AV will decide their future but still most OEMs are not well advised to develop their own AI brain to make self-driving cars work. There will be no “iPhone moment” which will totally shake up the industry.

Interestingly Elon Musk seems to be aware of this. He did not start Tesla to create it the “new Apple” of the car industry. He openly says that he started Tesla to accelerate the change from gasoline cars to electric cars by forcing the other OEMs to invest heavily in that area. By releasing the Model 3 he has almost done this - given that they can figure out the production problems in the next year or so. I personally expect him to step down as CEO from Tesla in the next 1-3 years to fully concentrate on his real mission to bring people to Mars with SpaceX then.

Or maybe he comes back to the car world because he discovers his love for (self) flying cars :)

‍