Decoding the Myth of Self-driving Cars and the Competition in U.S.

An assessment of current stage, an analysis of how to win, and a prediction of who will win.

While this video is very impressive, what interests me the most is the first 5 seconds of it — notice that Tesla doesn’t call the brave guy in the car “the driver” nor “the passenger”. He is “the person in the driver’s seat.” He doesn’t drive the car but he’d better be alert to the surroundings so he’s neither a driver nor a passenger who can sleep or browse social media feed in the back seat. In fact, this subtle relationship very much reflects the current stage of self-driving technology: we have made great progress in the past few years and semi-autonomous driving is possible today with certain restrictions — geographical, regulatory, speed limit, driver-required, hands-on-the-wheel, etc. And the technology is currently better at assisting the driver (or, “the person in the driver’s seat”) than taking full control of the car. Cars with the most advanced self-driving system in the market today are still somewhere between Level 2 and Level 3 of the six levels defined by the Society of Automotive Engineers (SAE):

1. It still has a lot of restrictions. To name a few — geographically, it performs better on highways with less variables than a busy city center intersection, as evidenced by the recent news that Uber’s self-driving truck made its first delivery with its autonomous system turned on on the highway. And recently Uber admits its self-driving vehicles in SF have a problem with bike lanes. Legally, in September 2016, the federal government released its first rulebook governing the manufacture and sale of self-driving cars. Now automakers need to pass safety assessment on 15 specific topics to receive approval from the National Highway Traffic Safety Administration (NHTSA) prior to commercial introduction. I’m sure there will be more regulations coming as the technology advances.

2. A self-driving car that can successfully move you from a place to another 999 times out of 1000 is still not a successful innovation and will not be widespread. What matters the most is the one time when it failed and this is particularly true for the self-driving technology. Tesla, one of the pioneers in this technology, for example, involved in multiple fatal accidents in which its Autopilot technology failed to protect the driver in the past few months. Both reputation and stock price took a dive.

Based on the above reasons, I believe what the video shows is a preview of what Level 3 automation looks like for consumers but it shouldn’t make you feel that fully autonomous car is around the corner — even the Level 3 car is not in the market yet. To have a clearer understanding of the current stage and to provide guidance into the future, I’d like to break down self-driving technology into the following 3 components:

Hardware: most technologies already exist — outsource-able

○ Car

○ Sensors (camera, radar, laser, ultrasonic, etc.)

○ Processors (GPU, CPU)

Software: a very very hard piece to crack

○ GPS, Hyper-accurate mapping

○ Decision-making algorithm

○ Routing/Matching algorithm (for the Transportation-as-a-Service model): Let’s call this self-driving-car-as-a-service model TaaS 2.0 and the ride-sharing model with human drivers TaaS 1.0

○ Real-world data to train the software

Market: a very very new market

○ Riders (consumer adoption)

○ Drivers: needed in the transition from TaaS 1.0 to TaaS 2.0

(a hybrid network of human drivers and autonomous cars)

○ Government (regulations, infrastructure: are today’s roads, traffic lights, signs best suitable for autonomous cars? Probably not.)

My assessment of the current stage is: the hardware component is already there but the software and market are far from ready. Whoever first cracks the remaining two pieces will be in the best position to win the competition. This explains why Apple shifts focus from building a car to building the software system; Google spins off Waymo and partners automaker Fiat Chrysler; Uber hurries into testing self-driving vehicle in San Francisco without a permit, etc.

In October, Tesla announced that all Tesla vehicles produced “will have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver.” Looking at the hardware pieces in a self-driving car — camera, radar, ultrasonic sensor, GPU, CPU, car — and you will find none of these are 21st century inventions. Thanks to the pioneers in these industries, those hardwares are accessible, powerful, light-weighted and mostly affordable today. And Moore’s Law tells us that GPUs and CPUs will have better performance and lower costs over time, providing greater computing power in processing all the data captured by car sensors.

I’m not undermining the importance of the hardware component of self-driving car. My point is hardware is not the biggest problem in the development of self-driving car. What’s challenging is a) to put all the information captured and processed by all the hardware pieces together to make decisions that are better than those from human drivers in real-time and b) to facilitate consumer adoption and work with the government to co-create regulations and the infrastructure that guide and support the technology. The first piece is software and the second piece is market. Both components are very interesting. First, one should note that the self-driving software is very different from most software we use today. It is very special in that it has near-zero tolerance of errors. In contrast, imagine a productivity app on your mobile phone or an enterprise software a company uses — after releasing the product, developers constantly identify bugs and de-bug and update the software over-the-air while users use the software, report bugs and wait for updates that fix those bugs. For the self-driving software, however, a minor bug will put the passenger’s life at risk. This zero bug-tolerance means much more difficult software development and friction in the market.

Software development: the bar for product release will be much higher. A lot more rounds of validation are required and a different beta testing approach will be needed. It’s almost like playing a game in which a success gives you 1 point and a failure deducts 1000 — you’d better be 100% sure that it will be a success before you give it a shot, right?

Market reactions: regulations will be very strict and even if a self-driving car passes the safety assessment and all other regulatory friction and enters the market, it will be extremely difficult to achieve mainstream market adoption in a short time. I see myself as an early adopter of technology, but I have to admit I feel very nervous for the courageous “man-in-the-driver-seat” in the previous Tesla video. According to the Maslow hierarchy of needs, safety is a fundamental need, so it’s expected that the chasm between early market and mainstream market will be very wide and hard to cross.

Technology Adoption Life Cycle

What this means is although the hardware challenge is mostly resolved, and the government and many companies are investing heavily in research and development of self-driving cars, this technology is still in its infancy and it takes years to mature. I don’t make predictions on how soon it will have a real impact in our lives, because even Tesla, an optimistic and adventurous company (in my opinion) and a pioneer in self-driving technology, admits that it’s not possible to know exactly when its full self-driving capability will be available (note that all the hardware pieces necessary are already built into its cars) because it is “dependent upon extensive software validation and regulatory approval, which may vary widely by jurisdiction.” Again, it shows that the software and market pieces are tougher to crack than many people have thought.


Now let me take a huge leap forward: imagine that technological barriers are gone and companies can manufacture fully autonomous cars (both hardware and software requirements for at least Level 4 are met). What should the go-to-market strategy and business model be? I believe a company has the following 2 options:

1. Sell cars directly to consumers: traditional car manufacturer model

2. Provide services to consumers: self-driving-car-as-a-service (TaaS 2.0)

Which one is a better model for self-driving cars?


I’d like to take the consumer’s perspective and first think about why people buy/own cars. I’ve identified the following motivations:

Control: you know your car is at your command

Reliability: you go to wherever you want whenever you want

Privacy: driving/riding a private car renders more privacy

Economy: lower cost per mile ($0.90 for private cars vs. $1.54 for ride-sharing)

Ownership: it feels good for some to own a car

~ Status symbol: it feels better for some to own a luxury car

Driving is fun!!

Looks like a lot of reasons. Let’s see what’s left assuming there exists a super-efficient TaaS network. Because there will be a ride that takes you to whatever destination you wish 24/7 at your fingertips, control and reliability are no longer valid as motivations of owning a car:

Privacy: driving/riding a private car renders more privacy

~ Economy: it’s debatable which option has lower cost per mile, but the difference certainly shrinks

Ownership: it feels good for some to own a car

~ Status symbol: you can ride luxury cars in TaaS network, but still one can argue it feels great to own a luxury car

Driving is fun!!

Now adding full automation technology into the equation, why would a consumer prefer owning a self-driving car over using TaaS 2.0? Additional motivations are taken out: that you can be the only person in the vehicle (if you desire) invalidates privacy; since cost per mile in TaaS 2.0 will be significantly lower (without a driver taking 80% of what you pay), economy becomes invalid; and obviously you don’t drive either way, so no more “driving is fun.” What’s left:

Ownership: it feels good for some to own a car

~ Status symbol: owning a luxury self-driving car is cool, although you can probably ride a luxury self-driving car in TaaS 2.0 too


Now it’s interesting to see from the consumer’s perspective that the only motivations of owning a self-driving car over TaaS 2.0 are: ownership and status symbol. I believe, however:

1. Without other value propositions, ownership itself as a motivation won’t stand long — all the other benefits of owning a car come together to make ownership meaningful. When all the other motivations are gone, the ownership motivation becomes vulnerable

2. Car as a status symbol simply does not stand for the mass public

It seems that the only market for selling self-driving cars to the consumers is on the higher-end. This implication is reinforced by the fact that the total costs of manufacturing a self-driving car will be quite high, at least in the early years of the self-driving car era before economies of scale kick in. Does this imply a hybrid model for car manufacturers — TaaS 2.0 for the mass public and selling luxury self-driving car to a small group of consumers? Or will they become suppliers for TaaS 2.0 companies? To answer those questions, I’d like to first analyze the viability of the self-driving-car-as-a-service model.


There are 2 major requirements for a company to enter the TaaS 2.0 business:

A. A fleet of self-driving cars: unlike today’s Uber or Lyft which is essentially a marketplace that matches drivers and riders, a self-driving car as a service company needs to either own a fleet of cars or partner with a third-party supplier that provides a fleet of cars. Since building a fleet either way will require a lot of capital expenditure, a company needs to have excellent ability to raise capital or a huge amount of free cash flow to enter the market. To give a quick estimate of capital expenditure required to achieve supply side liquidity with a fleet of self-driving cars — In 2014, Uber claimed it does 1 million rides per day globally. I believe the number has grown rapidly in the past two years but for simplicity let’s say Uber now has 1 million rides per day in the U.S. market. If we assume a self-driving car can complete 50 rides per day, Uber needs 1M/50 = 20,000 cars to achieve supply side liquidity in today’s U.S. market. Assuming each self-driving car costs $100,000, those 20,000 cars add up to a tremendous $2 billion investment.

B. Matching and routing capabilities: owning a fleet by itself means nothing if the company doesn’t have the ability to dispatch those autonomous cars efficiently to riders in real-time. This new version of on-demand economy — replacing human drivers with intelligent machines on the supply side — requires next level precision in matching and routing algorithm to ensure a smooth pick-up, drop-off, and all the on-trip variations (change in destination, carpooling, backtracking for another pickup, etc.). An autonomous car is trained to do a lot of things, but making the right pickup decision is probably not one of them. There has to be a system that centrally manages and dispatches the fleet: just like hyper-accurate mapping is required for any self-driving cars to operate on the road, hyper-accurate matching and routing is required for any self-driving car as a service companies that replace human-human interactions with machine-human interactions during pick-up, on-trip, and drop-off. A TaaS 2.0 company adds value by providing an exceptionally powerful matching and routing algorithm on top of a fleet of self-driving cars. In other words, if a company doesn’t have such capability, it has no chance to compete in this market no matter how many self-driving cars it has or how good those cars are.

A company must have both A and B to compete in this space. Self-driving car manufacturers have A but lack B whereas ride-sharing companies like Uber or Lyft have B but need A. Unless car manufacturers can catch up with ride-sharing companies in matching and routing capabilities, which is unlikely because ride-sharing companies have years of advantages with the technology and the market, the better business model for the former is to sell standard self-driving cars to TaaS 2.0 companies/partner with TaaS 2.0 companies and sell luxury self-driving cars to high-end consumers.

For ride-sharing companies, they can either develop self-driving technology in-house or purchase autonomous cars from manufacturers that can make them to become TaaS 2.0 companies. Remember — the software component is the key to self-driving cars, so whoever crack the software piece will have a huge advantage over other players in the industry. It’s not surprising that ride-sharing companies are actively exploring the possibility of developing self-driving software in-house (Uber-Volvo, Lyft-GM partnerships), since doing so will give them more control over service quality than outsourcing to third-party manufacturers, and more importantly the self-driving software is a moat against all other players in the self-driving car industry (other ride-sharing companies, car manufacturers, software companies, etc.). Routing and matching capability is a moat against car manufacturers and other non-TaaS companies, and a ride-sharing company would love to create a moat against other ride-sharing companies that have similar routing and matching capability.

The next question is, what’s a better go-to-market strategy for TaaS 2.0? Instead of launching a fleet of autonomous cars, the better strategy will be providing a hybrid network of human drivers and autonomous cars (TaaS 1.5) and transiting to an entirely autonomous fleet (TaaS 2.0) over time. Since self-driving car is so disruptive an innovation, routing and matching technology needs real-world data, expected and unexpected variations on the road, to adjust to the complexity of dispatching those autonomous cars, and the market needs time to build up trust and adopt this TaaS 2.0. Unlike self-driving software that has to be 100% ready before going to the market, routing/matching software needs a test-troubleshoot-and-iterate process, and much of the process can’t happen in the lab or on the test track: learnings must be from the real road. The hybrid network is a critical intermediate stage that does the following things:

1. Market: facilitate consumer adoption

2. Market: push for regulations and road infrastructure to adapt to this new technology

3. Software: collect massive amount of real-world data to improve routing and matching software

4. Software/Market/Hardware: test all aspects of user experience and human-machine interactions, reveal problems and troubleshoot

That being said, whoever first penetrates the market with a hybrid network of human drivers and autonomous cars to provide transportation as a service will be in the best position to win the self-driving car market. This looks easy, but my previous analysis points out several underlying assumptions for launching this hybrid network: a) the company has validated the viability and reliability of its self-driving software so the car is ready to enter the market; b) the company has the necessary routing and matching capability to dispatch a TaaS network and has enough drivers and autonomous cars in the first place to achieve supply liquidity; c) frictions in the market (regulations, infrastructure, consumer adoption, etc.) are low enough for this TaaS 1.5 to launch. Until those minimum requirements are met can a company compete in this market.


So who’s in the best position to win?

I created the following table to visualize the strengths and weaknesses of some of the most active players. For each criterion, companies are evaluated on a scale from 1 to 5, 5 being the best among all five companies. Please note that these are relative scores, not absolute scores; for example, that Google (now Waymo) gets a 5 in decision-making software doesn’t mean it has a perfect decision-making software — it only means Google (now Waymo) is leading in this category, no matter how far it is from really cracking the secret code of decision-making for self-driving cars.

Google Waymo and Uber are leading the race.

Explanations on scores in each category:

Capital: both Google and Apple have a lot of free cash flow and healthy core business to support their self-driving car projects. Uber has an incredible ability to raise capital (15 funding rounds with $11.5B raised) but obviously it has to spend money elsewhere too. Lyft has 9 funding rounds with $2B raised and spends aggressively to compete with Uber in the ride-sharing business. Tesla needs huge amount of capital on the development and production of Model 3 and construction of the Gigafactory so it’s a little cash-strapped.

Software — GPS, mapping: Google has over 10 years of experience in mapping and the acquisition of Waze in 2013 gives it additional data and expertise. Apple released Apple Maps in 2012 and mapping vehicles were spotted and associated with Apple’s self-driving car project. In 2016, Uber announced the plan to invest $500 million into creating its own maps and has been very active in acquisition of and partnership with mapping company since 2015. Tesla also has the plan to build high definition maps and has been using Tesla drivers to collect data for the development of maps since 2015. Lyft, on the other hand, relies on third-party maps including Google Maps, Waze and Apple Maps and has yet to announce a plan to develop its own maps.

Software — Decision-making: Google (now Waymo) started road-testing its self-driving technology in 2009 and has accumulated over 2M miles with fully self-driving cars. It has the most experience and data that are so important in training the software. In April 2016, Tesla claimed it has over 47M miles driven on Autopilot. Although Autopilot is not fully autonomous, it’s impressive that Tesla owners become the trainers for Tesla’s self-driving software, as the software learns as Tesla owners drive. Leveraging Autopilot mileage from hundred thousands of Tesla owners enables Tesla to learn fast. In 2015, Uber announced strategic partnership with CMU and launched Advanced Technologies Center in Pittsburgh to focus on research and development of autonomous driving technology. In 2016, Uber made another big push by launching self-driving rides in Pittsburgh and acquiring Otto, a self-driving truck company. Recently, Uber expands its self-driving vehicle testing to San Francisco. Partnership, acquisition and road test will give Uber invaluable expertise and data in developing self-driving cars. Both Lyft and Apple are behind in the race. Although Lyft has partnered with GM, it seems to me that GM provides the self-driving technology while Lyft provides the ride-sharing platform in this partnership. Apple struggles with its Project Titan and recently shifts focus from building a car to developing self-driving software.

Software — Routing/matching: Both Uber and Lyft are obviously leaders in this category. Since Uber has over 80% market share, which means more rides processed and more data collected, I’m giving Uber a point higher than Lyft. Google launched its ride-sharing service in Israel in 2015 and a pilot program in the Bay Area in 2016 with Waze Carpool. Tesla does have a plan to develop ride-sharing service, but it doesn’t have any advantage in this category. Apple doesn’t have any move other than investing $1 billion in Didi Chuxing, a Chinese ride-sharing company.

Market — Consumer adoption: Tesla has hundred thousands of loyal customers. It also takes an evolutionary approach in developing and releasing self-driving technology, which makes consumer adoption a little easier. Uber and Lyft are long-time ride-sharing service companies so their advantage is the trust they have established with the consumers on their TaaS model. When they eventually launch TaaS 2.0, at least their customers may give it a try. Google’s Waze Carpool and renowned expertise/effort in self-driving cars will increase the perceived reliability among consumers who are aware of Google’s long-time research and development of autonomous driving technology. Apple Car? I don’t know…

Market — Driver network: my previous analysis points out the importance of TaaS 1.5 as an intermediate stage, which will require a driver network to supplement fully autonomous cars. Uber and Lyft have one while others have to build one from scratch. Not an easy thing to do.


Now it’s clear that Google and Uber have solid lead in the race of revolutionizing transportation with self-driving-car-as-a-service. And it’s so interesting to see they are very different in terms of their advantages, almost complementary to each other. Google has advantages in self-driving software whereas Uber has advantages in TaaS technology and market. Yet both companies are making the right moves to nurture what they are weak at: Google testing Waze Carpool and Uber testing self-driving rides in Pittsburgh and SF. I bet they know very well what it takes to win the race and I’m writing this to make it crystal clear to everyone who’s passionate about self-driving cars like I am.

Me in a Google self-driving car at the Computer History Museum, Mountain View.