Even though I strive to intellectually understand the technology that makes vehicle autonomy (whether full or partial) work, seeing the technology actually work is still surprising to me. The latest surprise was Navigate on Autopilot. Here's how Tesla (TSLA) describes the feature:
Navigate on Autopilot is an active guidance feature for Enhanced Autopilot that, with driver supervision, guides a car from a highway’s on-ramp to off-ramp, including suggesting and making lane changes, navigating highway interchanges, and taking exits.
Take a look at Navigate on Autopilot in action:
We have crossed a subtle threshold. Until now, Autopilot was just a system that assisted the driver. Navigate on Autopilot feels like a precursor to fully autonomous highway driving. It’s still far from that point, but you can see the bones. The driver is now just the babysitter for the car: watching that it doesn’t do anything stupid or dangerous, giving it permission to change lanes, and correcting its mistakes.
Here's how Navigate on Autopilot fares in heavy traffic:
Yes, the system is still buggy and has a few predictable failure modes, like when a highway has a dedicated exit lane. This video shows some of its smart behaviours, like understanding HOV lane lines, and some of its imperfections, like inexplicably cancelling lane changes:
The important thing to remember about Navigate on Autopilot is that this is the worst it’s ever going to be. It’s only going to get better from here.
It’s especially impressive when you consider that Tesla’s Hardware 2 suite launched just two years ago, and Tesla pretty much had to start over from scratch. In two years, we’ve gone from having no working features to Navigate on Autopilot.
Looking forward, progress is going to happen faster in the next two years than it did in the last two years. The progress rate is not steady, but accelerating. Hardware 2 Teslas have gone from driving 1 million miles per day in early 2017 to close to 10 million per day today.
Human programmers don’t do most of the coding for Autopilot. Neural networks do. Feed them data, and they code better. Tesla is presumably building its dataset at an accelerating rate, which should contribute to better neural network performance.
A bigger neural network can also do a better job. That’s why Tesla is going to launch the Hardware 3 computer with 10x more compute power, and will offer to retrofit all Hardware 2 cars with it. Here’s what Tesla’s Director of AI, Andrej Karpathy, said about it:
This upgrade allows us to not just run the current neural networks faster, but more importantly, it will allow us to deploy much larger, computationally more expensive networks to the fleet. The reason this is important is that, it is a common finding in the industry and that we see this as well, is that as you make the networks bigger by adding more neurons, the accuracy of all their predictions increases with the added capacity.
So in other words, we are currently at a place where we trained large neural networks that work very well, but we are not able to deploy them to the fleet due to computational constraints. So, all of this will change with the next iteration of the hardware. And it's a massive step improvement in the compute capability. And the team is incredibly excited to get these networks out there.
It’s not just a bigger dataset. It’s also bigger neural networks.
An order of magnitude increase in daily miles. An order of magnitude increase in computing power, which will enable a much larger neural network. The fundamental inputs are getting bigger.
Elon Musk recently did an interview with Kara Swisher, and he reiterated his estimated timeline for full autonomy:
I think we’ll get to full self-driving next year. As a generalized solution, I think. ... Like we’re on track to do that next year.
Elon’s timelines are famously overoptimistic. For example, he talked about a fully autonomous coast-to-coast demo drive by the end of 2017. That’s coming up on a year late. It would certainly not shock me if Tesla did not reach a generalized solution for full self-driving next year.
We don’t need to worry about how late Elon might be. Even if we don’t see full autonomy until 2025, that’s still incredible, and in the meantime, partial autonomy will continue to evolve and become more useful. Look at where Navigate at Autopilot is today, in version one. Even if in two years full autonomy is still far off, partial autonomy is going to have a profound effect on the experience of driving.
If it turns out that a generalized, scalable solution to full autonomy is still several years away for all companies, that is advantageous to Tesla relative to other companies. Competitors like Waymo (GOOG, GOOGL), Cruise (GM), and Zoox are pursuing an all-or-nothing approach. That would mean several more years of nothing for them. Tesla, on the other hand, will leverage the value of neural networks to maintain or increase its popularity and pricing power, and to increase its software revenue, which is high-margin.
The only big competitor I can think of in partial autonomy is Mobileye (INTC). Mobileye’s philosophy and business has more overlap with Tesla’s than other companies. Both companies believe that cameras alone are sufficient to do full autonomy better than humans drive cars. Both companies are working on the evolution of driver assistance systems into fully autonomous systems. Yet Mobileye has structural disadvantages against Tesla.
Mobileye is bound by a glacial three-year update cycle with auto OEMs. The manufacturers want to sample Mobileye’s chips three years before the cars launch, and the software only gets updated when it ships with a new chip. Tesla’s hardware update cadence is faster, and it puts out software updates every few weeks. Don’t take my word for it, hear it from Mobileye’s CTO Amnon Shashua:
Companies like Tesla or NIO, they skip the use of a Tier 1. They become their own Tier 1. I think this is something that is possible to do when you have small scale... That allows them to move faster.
With Tesla, the first Autopilot was introduced in November 2014. It took about a year of development. With regular OEMs, it took closer to three years. That there are less players in the loop, in the chain, accelerates things.
So, keep an eye on Navigate on Autopilot. Look for a significant update by the end of the year, and see how much it improves things. Most importantly of all, wait for the launch of Hardware 3 and the deployment of the new neural networks it will enable. That will be a sudden step change in capability.
A high-level theoretical question about Tesla is: why won’t its products become commoditized? My answer is that its structural advantages in developing and deploying autonomy (partial or full) will differentiate its software. Navigate on Autopilot is the strongest evidence for that thesis yet.
An element of almost every short thesis I've read or heard is that new electric vehicle models from incumbent automakers will significantly hurt Tesla's demand or margins. The idea is that 1) Tesla enjoys atypical demand and margins by being differentiated from other automakers, 2) that differentiation results from selling high-performance, long-range electric vehicles, and 3) that differentiation will disappear as high-performance, long-range electric vehicles become more commonplace. I have three responses to that:
1. The rest of the global automotive supply chain seems to lag Tesla on battery costs and battery efficiency. This means a worse product for the same cost to produce, or the same product for a higher cost to produce.
2. Once auto manufacturers and suppliers cross the critical threshold where an electric vehicle has a lower total cost of ownership than a gasoline vehicle, demand will rapidly shift away from gasoline vehicles toward electric vehicles. Some estimates say the Model 3 has already crossed that threshold, but to confirm these estimates actual data from Model 3 owners is needed.
A recent survey found that 20% of American drivers want an electric vehicle, up from 15% in the previous year. If manufacturers can reach a lower cost of ownership than gasoline cars, and if ubiquitous fast charging is made available, the tipping point for consumer demand could happen quickly. How long before more drivers want an electric car than a gasoline one?
3. Navigate on Autopilot and future advances will continue to differentiate Teslas even if electric vehicle technology won't. As explored above, Tesla has structural advantages in partial autonomy that it will be difficult for competitors like Mobileye to overcome.
So, for three independent reasons, I don't think new electric vehicle models from incumbent automakers are a good reason to short Tesla. Tesla has short to medium-term advantages in batteries and long-term advantages in partial autonomy, and total electric vehicle demand will increase rapidly when electric vehicles beat gasoline vehicles on price-performance, which will probably happen soon, at least for Tesla. I predict that demand for Tesla's vehicles will only increase over time.
If Tesla can traverse the distance from partial autonomy to full autonomy, the long-term opportunity is immense. Suppose that full autonomy arrives in 2025, rather than 2019. Using ARK Invest's modelling, that puts the net present value of the global autonomous ride-hailing market at $1.26 trillion on a 15-year time horizon. Suppose you assign a 50% chance to Tesla staking out a 10% global market share. That's an expected net present value of $63 billion, about $4.5 billion higher than Tesla's current market cap (at time of writing).
This bet requires 1) a very long time horizon, 2) a very high risk tolerance, and 3) a certain level of credence that full autonomy is a realizable goal in the not-too-distant future.
If these assumptions turn out to be conservative, then the opportunity is larger. For instance, if Elon is right that full autonomy will happen next year, the expected net present value is $100 billion. If Tesla has a 1 in 3 chance of capturing a 0% market share, a 1 in 3 chance of capturing a 10% market share, and a 1 in 3 chance of capturing a 20% market share, the expected net present value is $200 billion. And so on.
Yes, this idea is highly speculative and uncertain. Technological progress is a poorly understood and largely unpredictable phenomenon. However, were it proven and certain, there would be no investment opportunity. That's the market.
Photo by Vlad Tchompalov
Disclaimer: This article is not investment advice. Please educate yourself about the risks and consider consulting a professional advisor before making investment decisions.
Disclosure: I am/we are long TSLA.
I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.