Tesla Stock (TSLA): Wall Street Is Asleep To Tesla’s Technical Advances In Autonomy

I’ve been banging the table for five and a half years about Tesla’s (NASDAQ:TSLA) clear advantage in autonomous driving: lots and lots of vehicles from which to collect training data for its neural networks. My fear was that a competitor like GM/Cruise (GM) or a partnership between Waymo (GOOG, GOOGL) and a large automaker would copy Tesla’s strategy of having lots and lots of vehicles while Tesla’s production numbers were still small and before the company could reap the benefits of its (still, to this day) unique strategy of having a large-scale vehicle fleet.

Instead, competitors have twiddled their thumbs for five years while Tesla has never stopped running. In terms of useful data collection, the gap between Tesla and its rivals has never been wider and all signs point to it continuing to widen at least over the next 1-2 years and probably for longer.

Still, this is an investment thesis grounded in general, widely-held theoretical principles of deep learning and that’s esoteric enough to leave most automotive analysts and retail investors agnostic at best. In recent months, however, evidence of a more concrete nature has emerged that should convince some of the doubters.

First and simplest of all, we can see the actual real-world performance of autonomous urban and suburban driving in customers’ cars. Crucially, we can also see the rate of improvement between software updates, which come frequently. Here’s a recent example.

The most hardcore Tesla aficionados await every major software update with bated breath, but the rest of the world is out of touch. Unless you follow how the software evolves over time, your sense of its abilities will be stuck in the past and you will have no sense of the rate of improvement.

Second, although the talks can at times be highly technical and opaque to laypeople, Tesla discloses a laudable amount of information about their R&D progress and how their autonomous driving system works through talks published on YouTube. This includes not just Tesla’s highly publicized official company events, but also talks at academic conferences like the Conference on Computer Vision and Pattern Recognition (CVPR).

Tesla’s most recent CVPR presentation bears close attention. For years, Tesla established itself as a fast copier. That is, it’s been known to grab recent, cutting-edge research papers from the deep learning literature and quickly implement them into its own software. However, Tesla has not earned a reputation as a research innovator in its own right, i.e., a company that generates original research ideas not thought of first elsewhere.

With Tesla’s new CVPR presentation, that changes. Tesla’s new invention is a kind of neural network that it calls occupancy networks. Occupancy networks divide the world around the car in voxels, which can be thought of as 3D pixels. The occupancy networks assign each voxel a probability that it is occupied by a physical object or obstacle, such as a vehicle, a fence, a lamppost, or a traffic cone. This new approach to detecting obstacles is a wholesale upgrade to the previously existing approaches that are still the industry standard. The full CVPR talk is on YouTube.

Since at least Tesla’s first AI Day a year ago, the company has been making a push to in-house more world-class AI talent, with the singular focus of pushing forward the state of the art in autonomous driving (well, and, I suppose also prototyping a humanoid robot). I see occupancy networks as a decisive victory in this pursuit. Tesla has shown it can not only copy the best research, it can improve on it. A short clip with a visualization of Tesla’s occupancy networks is viewable here:

Anyone who believed that Waymo had an unquestioned lead in AI R&D as it pertains to autonomous driving should re-evaluate that belief now. Alphabet does not have a monopoly on great AI researchers and only a fraction of its research output directly pertains to autonomous cars.

Frustratingly to me, with a few exceptions such as Alex Potter, Pierre Ferragu, and Adam Jonas, Wall Street analysts who write and speak publicly about Tesla have failed to grasp the magnitude of the technical advances that Tesla’s AI team is making and the implications these advances will have for Tesla’s consumer business. CEO Elon Musk’s stated aim is to make the full suite of city driving features available for purchase to any North American customer (providing they have a Tesla no older than about five years) by the end of 2022. Given Musk’s general tendency toward over-optimism, the wise thing to do, in my opinion, is to take this timeline with a grain of salt and to evaluate its plausibility by closely watching the Tesla software updates as they roll out.

The point at which the Full Self-Driving Beta software is available for purchase by a broad base of customers is the point the esoterica of deep learning research becomes a company-defining product for the world’s most valuable automaker. In every practical sense, Tesla will be a company that sells integrated hardware and software product that is defined at least as much by the software as by the hardware (and probably more so). The tired analogy is Apple (AAPL), but the shoe fits.

Valuation models need to adapt to account for software revenue with software margins. Few analysts (with Alex Potter being the most noteworthy exception) have been daring enough to do this. My comparative advantage is translating knowledge about deep learning into investment theses, not granular valuation modeling, so I’ll leave that task to others.

A final word on robotaxis. All that I’ve written above is agnostic on the topic of robotaxis. However, I would consider it prudent for analysts and investors to start taking this possibility more seriously. Perhaps it would be wise to append an expected value calculation to your valuation models. It may be useful to think of robotaxis as a low-probability (but not so low that it’s prudent to ignore), high-impact possibility. With the upside for robotaxis, under an optimistic scenario, being astronomical, it seems imprudent to me to simply ignore them as a possibility altogether.