Why I’m Still All-In on TSLA: Part 3 — Tesla AI/Robotics
Welcome to Part 3 of the series. Part 1 covers Tesla’s manufacturing advantage. Part 2 covers Tesla’s energy business. In this article, I will give my assessment of Tesla’s Artificial Intelligence efforts and my take on its bottom line.
Disclaimer: I’m not a financial advisor. These are just my opinions. Do your own homework, blah blah blah.
What is AI? ML?
I’m not a trained AI engineer, but I’ve worked on simple Machine Learning prototypes while at Google. Machine Learning is a subset of Artificial Intelligence where machines learn by itself from data. I’d like to think of ML like programming turned inside out.
Traditionally, a software engineer is given a set of inputs and outputs as “specifications”, writes the rules, and checks that the program generates the expected outputs given the expected inputs. We call this Test Driven Development. This is easy for things like well understood computations and if-then-else type logic, but much more difficult for problems like distinguishing a cat from a dog where the rules aren’t as obvious.
ML flips this around. Given lots of inputs, a software team creates a neural network, which is usually a complex set of matrix multiplications, so that each input generates the expected output, like identifying a cat vs. a dog. The set of multiplications have adjustable coefficients, possibly millions of them, and “training” tries random combinations to find the best results. The more pictures of cats and dogs, and faster training iterations, the better the “computer program”, aka neural network. It’s a bit like Darwinism of randomly mutating software.
In other words, the company with more data, better data and faster training hardware, wins.
Tesla’s Data Advantage
Most folks would consider Google’s Waymo as the leader in autonomous vehicles. Despite boasting orders of 62,000 Chrysler vans and 20,000 electric Jaguars, Waymo’s official fleet size in 2021 is about 600 cars. (source) Each car must be disassembled by hand and reconstructed with Waymo’s own sensors and computers, making them super complex to build, scale, and debug.
In reality, skilled disassembly is required. Engineers must take apart the cars and put them back together by hand. One misplaced wire can leave engineers puzzling for days over where the problem is. (source)
For argument sake, I’ll generously guess that each car costs $100K to purchase and construct, for a total vehicle hardware cost of around $60M. Cumulatively, Waymo cars have driven about 20 million miles. Just counting the vehicle hardware, it cost Waymo $60M to collect 20M miles of data. The real cost is probably significantly higher.
Tesla, on the other hand, has sold every car since 2017 with the self driving hardware and cameras. Cumulatively, that’s over one million cars by 2021, and they have logged 3B miles of data in 2020 alone. With an average selling price of $50K, multiplied by 20% profit margin, Tesla MADE at least $10B dollars while having the customers collect all those 3B data miles for FREE.
Did that sink in? Waymo has to pay to collect data, while Tesla is collecting 100x more data while making a huge profit!
Edge Case Training
If you watch the Autonomy Day video, you’ll learn that once Tesla identifies an edge case, such as boats on a trailer, bicycles on cars, Tesla can ask the fleet to send back examples to train the neural network. Waymo can only do this with 600 cars driving around Phoenix or San Francisco. Tesla can ask one million cars driving all over the world for this data!
When you consider Waymo’s linear growth (at best), and Tesla’s 50% CAGR, this gap will continue to grow exponentially. Tesla has such a huge data advantage that it’s not even fair. And the lead will only get bigger.
Tesla Dojo HW/SW Platform
Once you have all this data, you need to design a neural network and train it. This requires a super fast training computer. While the TeslaBot was all flash at Tesla’s AI Day, I think Dojo was the real star.
Let me start by going backwards. A computer has traditionally been a motherboard, CPU, memory, and specialized boards like a graphics card (GPU). The motherboard has an I/O bus that connects everything and allows data to flow as electric signals. For an electron, the classic computer is like the suburbs, where your house is driving distance from schools, work, and shopping. What you can get done quickly is pretty much limited by the traffic and distance on roads (the I/O bus).
In 2020, Apple introduced their M1 processor, which is an adaptation of its System-On-Chip design from their mobile devices to the desktop and laptop world. A SOC means that the entire computer as we know it, is now on a single chip. The CPU, GPU, memory, are all on a single microprocessor connected via the Fabric (like the I/O bus). The entire M1 SOC is manufactured using the TSMC’s latest 5nm process, meanwhile Intel and AMD are manufacturing just their CPUs at 10nm and 7nm. Can you see why Intel is in trouble?
This is a revolution in miniaturization, improving speed and efficiency while reducing heat and power. For an electron, it’s like moving from the suburbs to the Burj Khalifa, where you can live, work, and play in the same building.
What does this have to do with Tesla? Dojo is Tesla’s AI hardware revolution from the ground up, from the electron level at the bottom all the way to the machine learning software API at the top.
Dojo From the Bottom Up
Tesla’s engineers determined the longest distance an electron can travel per clock cycle in a modern 2GHz microprocessor, and used that to determine the maximum dimensions of their training node that contains memory, CPU, and I/O that’s optimized for machine learning. Even though the training node is the smallest unit of compute, a node is capable of 1024 GigaFLOPS, or 1 TeraFlops. By comparison, a new PS5 or XBox is up to 12 TFLOPs.
A training node is like one of the tiny CPUs or GPUs inside the Apple M1. A full Dojo Chip, aka the D1 Chip, has 354 training nodes, and are fully bi-directionally connected for high throughput. The result is a whopping 362 TFLOPs with a dedicated I/O ring at 4TBs per edge. Tesla claims this is more than twice the throughput of the fastest network switch chips. This chip, built on 7nm, is about 1"x1", and is completely purposely built for machine learning. (Dojo will get faster when it moves to 5nm or 3nm architecture)
Not impressed yet? Just wait. Instead of connecting D1 Chips together on separate servers, what if they were just printed on the same wafer? That’s correct. What’s cooler than System On Chip? System on Wafer. Tesla is able to etch 25 D1 Chips on a single silicon wafer with shared memory and fast I/O. Also known as a Training Tile, each tile is capable of 9PFLOPs, or 9000TFLOPs, as powerful as 900 PS5s.
The training tile has amazing engineering, including a 3D design that powers the tile from the bottom and heat dissipation from the top. This revolutionary design is how TSMC was able to build the Cerebras wafer scale super computer. BTW, TSMC also builds all of Apple’s latest A, S, and M chips, plus AMD’s latest Ryzen chips. Talk about an important company for Taiwan, and the world…
But wait, there’s more. Training Tiles are infinitely scalable. What happens when you build a data center out of them? You get the ExaPOD. 120 Training Tiles networked together for 1.1 ExaFLOPs (1 million TFLOPs). Mic dropped.
The hardware part of Dojo was amazing. Tesla really showed how they thought from the electron up to build the ExaPOD. But what is equally impressive is the software stack. Dojo is scalable at the D1 Chip level, so you can create virtual training “servers” called Dojo Processing Units that is “right sized” for your training model.
The way they do this is by building an entire software stack, including the neural net compiler, which can figure out your computing needs, secure the hardware resources, run the processing in parallel, and return the results.
This reminds me of Borg, Google’s cluster management system. At Google, we don’t physically touch the hardware when we launch a service. We simply write a Borg Configuration Language script that we need to run X instances each in US East, US West, Europe, and Asia, and that each instance needs Y cores and Z GBs of memory. Borg will automatically find the resources in those data centers, reserve the hardware, copy over the binary, deploy and manage the live instances.
In other words, Dojo can provide machine learning as a service. Google is already doing this with their Tensor Processing Units. I believe Amazon is offering the same. In my opinion, no one will be buying GPUs to build their own ML hardware in the future. All of it will be outsourced to whom ever can offer the highest performance per watt.
Note that Tesla’s numbers are theoretical at the moment. Tesla has not built an actual ExaPOD, vs. Google, who has built out a TPU v4 supercomputer that will be soon available to Google Cloud customers.
Dojo will supercharge Tesla’s ability to train ML models faster based on the vast amounts of data collected by the growing fleet of a million+ cars.
Once a new version of the ML model is ready, it needs to be downloaded to the car so the car can drive by itself in real time.
The Car’s Full Self Driving Computer
Each autopilot enabled Tesla car has 8 cameras: 3 front, 4 side, 1 rear. The car uses input from the 8 cameras to figure out how to drive. Believe it or not, this is all done by a small computer board that fits behind the glove box, rather than a bulky computer that takes up half the trunk.
Announced in 2019 at Autonomy Day, Tesla set out to create its own neural net processing chip because commercially, there wasn’t one that had great performance and was power efficient. You don’t want to lose half your driving range due to the self driving computer! Tesla went to the drawing board to do what it always does. Design a solution that is specific to its needs while saving money build vs. buy.
Tesla’s FSD Computer accepts the camera and sensor inputs from the right, then a pair of redundant Tesla FSD chips in the middle to process the input, and on the left is the power supply and vehicle controls. The pair of FSD chips checks each other’s output before sending vehicle controls, and in the event one fails, the other one keeps working.
There are a lot of details in the video, but the highlights are: Tesla’s vertically integrated FSD computer, compared to Nvidia’s solution, can provide 144 TOPs (trillion operations per second) vs. 22, while using 72W vs. 500W. More compute, less energy. For an autonomous robotaxi running all day, this translates to more drivable range on the same battery. In the end, Tesla’s own computer costs 20% less than buying from Nvidia. Multiplied by millions of cars, the cost savings alone paid for the R&D effort.
Every car built since April 2019 has the new FSD computer. Those built between late 2016 thru 2019 can be retrofitted, as was our 2018 Model 3. We are talking about 2 million cars by end of 2021. Each of them capable of running the latest FSD software in shadow mode, sending training data back to Tesla’s data centers. All they are waiting for, to wake up, is the final software release.
But, But, Lidar!
Tesla has taken an unconventional vision based approach to solving full self driving, and that has ruffled a lot of feathers in the AI/robotics establishment.
For background, I recommend reading “Driven” by Alex Davies or watching this documentary.
In the 2005 DARPA Grand Challenge, Stanford’s Stanley used an innovative vision+lidar based approach to beat out the CMU’s predominantly lidar based vehicles (Sandstorm, H1ghlander). Eventually, Sebastian Thrun, Chris Urmson, and Anthony Levandowski all joined Google’s self driving program called Chauffeur, now named Waymo. Waymo leaned heavily on Velodyne’s lidar system and pre-mapped high density maps. They were able to quickly get a prototype going, but found it difficult to solve all the edge cases for city driving for true autonomy.
Eventually they all left Waymo, and Urmson and Levandowski both started new autonomous trucking companies, Aurora and Otto, because highway driving was easier. Even Uber, with its deep pockets, cancelled its own autonomous driving program after infamously killing a pedestrian, to partner with Aurora.
Warning: speculation ahead. As a technology, lidar still hasn’t solved autonomous driving after a dozen years. Lidar is like a bat’s sonar, where you “see” by signal reflection. Lidar can’t distinguish a red, yellow, or green light. It can’t read the road markings or speed limit signs. You still need a vision based system for that. While some of this can be solved with high definition maps, can you imagine the effort required to map every single mile in the world, and keep it updated? Is this really scalable, cost efficient, and “realtime” to deal with construction zones and traffic accidents?
Tesla takes a different approach. The entire driving experience and infrastructure is design for humans, who drive with nothing but a pair of eyes and a brain (and ears too!). Once a person learn the basics of driving, they can drive in new places that they haven’t seen before. They learn how to anticipate other drivers and plan the optimal drivable path.
Tesla’s solution? A car with cameras, computer vision, and deep learning AI computer brain. In fact, all US made 3s and Ys since May 2021 no longer has radar. This pure vision based system is called Tesla Vision.
To complete the recent history in autonomous vehicles, the disgraced Levandowski eventually plead guilty for selling Google’s self driving intellectual property to Uber. In retrospect, he thinks lidar was the wrong approach. (source)
In December 2018, (Levandowski) dismissed lidar, a technology he had long evangelized, as a “crutch.” The real key to the self-driving future, he said, was computer vision and deep learning, a view shared by Tesla CEO Elon Musk.
Tesla Full Self Driving Beta
As you can see, Tesla has the entire self driving eco system, from auto manufacturing to data gathering to ML training. For a few years now, Tesla has had a traffic aware cruise control system with automatic lane change, and with GPS data overlaid on top to do fully automated highway driving from on ramp to off ramp. We use these features all the time on our Model 3.
About a year ago Tesla release the Beta of the city driving portion of Full Self Driving to 2000 early adopters (videos). They have helped the engineering team tune the system with lots of data and disengagements. In August 2021, Tesla gave us a preview of how the AI is learning the same driving skills as us humans.
To me, the highlights were:
Memory. Tesla’s system can “remember” what it’s seen even if it’s in the past, or hidden behind other objects. For example, that there was a lane marking 50 ft in the past that this is a left turn only lane, or there is a bicyclist to the side of the road even though it is currently blocked by a truck. There’s even velocity tracking to anticipate where these objects will be in the future.
Planning. Tesla will run scenarios for thousands of different maneuvers to drive the car, such as changing multiple lanes to make a left turn. The car will pick the best move based on safety, comfort, and efficiency. Hey, that’s what I do as a Lyft driver as well!
Prediction. Not only will Tesla cars pick the best maneuver for itself, it is also looking at other cars on the road, including their front tire angle and velocity, for example, to know how to share a one lane road with on coming traffic.
These improvements give me confidence that Tesla is on the right track. Will they solve full autonomy? While I can’t guarantee it, I think they are closer than anyone else. It’s certainly a very difficult problem, but AI is capable of learning at exponential rates.
By end of Q3 2021, FSD 10.1, which merges the highway and city driving into a single software stack, will be launched. Tesla would also open the beta to anyone who either paid for full self driving at $10K, or $199 monthly subscription. This is super exciting as it will unlock years of deferred revenue. 2021 Q4 earnings could be a good surprise. (Not Financial Advice!)
You might wonder, how will all these self driving cars be insured in the future? Tesla has that covered as well. Currently only available in CA, Tesla is entering the insurance business. In fact, Tesla will use its insurance calculator to determine who gets to access the FSD open beta.
Why does Tesla offer insurance? One, Tesla always vertically integrate, thus it makes sense to offer insurance at the point of sale of a vehicle. Plus, it adds an entire new line of business to generate profit.
Two, Tesla can calculate your insurance premiums more precisely than other companies based on vehicle telemetry. Basically Tesla knows if you are a safe driver or not. Insurance is based on data, and Tesla can get more detailed data than anyone else.
Three, Tesla vehicles with autopilot are way safer than regular cars in terms of accidents per miles driven, and this safety ratio will only get better with FSD. The savings on insurance payouts should be passed onto owners via lower premiums.
Tesla Insurance is being expanded to other US states, and I personally feel this is a big deal. If Tesla has 70% market share of EVs, and eventually EVs become 100% of all new car sales, what will happen to all the other insurance companies? As the safer Tesla drivers exit their legacy insurance company, insurers have no choice but to increase premiums to cover the remaining drivers, which encourages more people to move to the Tesla platform. Vicious cycle.
“Obviously, insurance is substantial. So, insurance could very well be, I don’t know, 30%, 40% of the value of the car business, frankly,” — Elon Musk (source)
Subscription Revenue, Profits, & Market Cap
Let’s take a moment to consider the impact of FSD subscription and insurance to Tesla’s market cap. Let’s say Tesla grows at 50% from 2020 to 2030, reaching 20M of yearly production near the end. That would be about 75M cars made in total. For argument sake, due to car accidents or what not, there are 60M active cars.
Let’s say 75% of the cars subscribe to FSD and Tesla insurance, so $200 + $50 = $250 a month. 60M cars x .75 x $250 a month x 12 months = $135B annually of recurring software profits. Multiply by P/E of 10, that’s worth $1.35T of market cap!
Suppose one day Tesla solves autonomous driving, it will create the Tesla Network, a ridesharing network where owners can add their own vehicles and earn passive income. Tesla is full of moon shots, and this is the big one.
You might think this will never happen for a number of reasons. Maybe Tesla won’t solve autonomy, but let’s say they do, eventually. As explained in this article, Tesla has invested in the full stack of autonomous driving, and they have a better chance than most at succeeding.
Maybe you won’t put your vehicle in the network, but many others will do the math and realize that they can buy a second or third Tesla with the earnings, effectively running their own taxi fleet and retiring from their day job. What does your car do 90% of the time? Nothing, it’s parked.
Will users adopt the Tesla Network? If Tesla can beat Uber and Lyft by just a few dollars per ride, everyone will switch. Consumers in this space are very price sensitive. I should know. I drive for Lyft in my retirement.
In America, a personal car costs $.62 per mile to operate. A ride share cost $2–$3 per mile. Elon believes that a Robotaxi can cost less than $.18 to operate per mile (with overhead like insurance and tires). Based on 90K miles driven a year, a single car can gross ~$30K a year for you, before mileage tax deductions.
Ark Invest has also done a study. They find that it’s reasonable to charge $.60 per mile for an autonomous taxi. Since the mileage tax deduction is currently at $.56/mile, I would agree with Ark that $.60 is more likely than $.18. At $.60/mile, Ark argues that there is a $3T+ addressable market. At $.25/mile, that goes up to $11T+. Subtracting $.18/mile cost of operation, that’s $2.1T to $7.7T of profits. If Tesla charges 30% platform fee like Uber and Lyft, factor in a P/E of 10, Robotaxi is potentially worth $2.1T x .3 x 10 = $6T+ in market cap!
The market cap numbers here are nutty, but what type of investor are you? Are you a play it safe guy who is happy with 10% yearly returns on an index fund? There’s nothing wrong with that, but you will never become wealthy (unless you already are, haha). Or are you the type that wants to invest in companies that can grow 1000x, 10000x? If you can find the right company, you only need to be right once to achieve generational wealth, like those who bought and held Microsoft, Apple, and Amazon for 20+ years.
Tesla is on track for an open beta of its FSD program by end of Q3 2021. As a software engineer, when we go to an open beta, it means the app is almost ready. For Google Meet, we spent about 12 to 18 months in open beta before the full launch. Obviously, FSD is a much more difficult problem to solve, but it means Tesla is pretty confident in its abilities, and optimistically, could be 18 to 24 months away from regulatory approval. ML could learn a lot in 18-24 months.
Remember that before 1903, no expert was able to build a flying machine. By December that year, a pair of outsiders who were bicycle mechanics, achieved first flight. 66 years later, humans landed on the moon (Hello TeslaBot). That was over half a century ago. Think about how fast technology advances now. Never bet against human ingenuity, especially Elon Musk.