Description

Tesla is the most controversial company I've seen in 30 years in the investment business.  Believers in Tesla are cult-like just as TSLAQ members adamantly believe Elon Musk belongs in prison.  It's reasonable to be on either side of this debate.

Tesla will become the most valuable company in the world if it successfully solves autonomous transportation.  If it does not, TSLA stock is going down 90%.

Autonomous transportation is really hard to solve.  Whoever succeeds will create trillions of dollars of value.  

I strongly encourage anyone who wants to debate this FIRST go to their nearest Tesla delivery center and test drive the Full Self Driving (FSD) software for themselves.   I've been using FSD Beta since its release in 2020, and frankly, it kinda sucked until a few months ago.  Then everything changed with a complete overhaul of the software from a heuristics/C++ code-based architecture to a full neural net design.

Tesla is the leading real-world robotics/AI company that is now on a path to solving autonomous transportation.  Many think that Telsa's Full Self Driving ("FSD") software is vaporware, mainly because it's been promised to drive better than a human for many years now, and has failed to deliver that promise thus far.

That perception is about to change.  Tesla began redesigning its autonomous driving software stack to an entirely "end to end" neural network architecture in the latter half of 2022.  Prior to this period, it was using primarily a heuristics/human code based architecture that was suboptimal in terms of reaching human level, and eventually, beyond human level driving capability.

FSD version 12 of the software of this full neural network architecture was released to the public a few months ago and it is an absolute game changer.  I have personally been using first Autopilot, then FSD Beta  since 2019, and was concerned with the rate of progress due to the limitations of the C++ code architecture.  That architecture is now fully neural network which is the reason I strongly believe FSD is now on the path to achieving super-human driving capability. 

"If you don't disrupt yourself, someone else will do it for you."

-Salim Ismail, from his book Exponential Organizations.

I began to learn in the Fall of 2022 that they were throwing away the old code-based system in favor of full neural nets.  This is when I became very bullish on Tesla.  Though it was very costly to do so, and would set the progress back a few years throwing away 8 years of software code, it was a bold decision; a "bet the company" decision.

Today Tesla is considered just a car company.  I strongly disagree.  It has an insurmountable lead in real-world driving data and AI architecture that is training their neural network driving models.  No other company has anything close to this.  There are today 5 million robot cars driving around the world capturing video data and sending it back to Tesla HQ to train their neural network AI driving models.

This new age of Artificial Intelligence that was launched a year ago with the release of Open AI's Large Language Model, Chat GPT-3, will change the economy in ways we can only imagine.

Beyond-human level intelligence will be brought to every machine and software application you could think of because of artificial intelligence.

My thinking is that machine intelligence is going to be a lot more difficult to build than software intelligence.  Since last year's Chat GPT moment,  there have been numerous copy cats that have been released such as Google's Gemini, Meta's Llama, Microsoft's Copilot, Perplexity's Claude 3, and X.ai's Grok.  They are all essentially equivalent in terms of accuracy, since, they're all trained on the same data set, namely every website that exists on the open internet.  The only differences among them is their speed and that comes down to size of their neural network models and differences in compute power.

Real-world, machine AI is going to be very different in my opinion.  Whereas software AI has very little barrier to entry as evidenced by the many LLM chatbots that have sprung up in just the last year;  real-world AI is extremely difficult both in terms of access to data, and in terms of manufacturing the machines at scale and cost-effectively.

Chatbots were the first application of "software" AI.  In my strong opinion, autonomous cars will be the first application of "real-world" AI and Tesla will be the clear winner.

No other automobile manufacturer in the world, except for China's BYD,  builds electric cars at a positive profit margin besides Tesla.   It's takes the scale of millions of electric vehicles produced per year to build EVs profitably.   It's the classic innovator's dilemma.  If you're Toyota, Ford or GM, and you're losing $50k on every electric vehicle produced due to your lack of scale, you stick to what satisfies your shareholders who demand profits; making profitable cars/trucks with combustion engines.

Electric drive-train cars are going to continue to gain market share until they're 100% of the fleet at some point in the next 10 years.  Why?  Not their smaller environmental footprint, though that's a positive externality.  It's because of cost.  Electric cars have only 10% of the parts of an combustion engine car.  No oil, no transmission, no engine.   Just electrons from the battery to the motor, and power out to turn the axles.  Much more efficient and cost effective than combustion engines.  The limiting factor so far has been the cost of the battery.  But, today, that cost is down to below $100 per kilowatt hour from $300 just 7 years ago.  The battery accounts for nearly 40% of the cost of an electric vehicle so as this cost continues to drop, in a few years, electric vehicles will be cheaper than their combustion engine equivalents.

Tesla is the leader in EV battery technology and scale. Scale is very important.  The more cars and batteries you manufacture, the more you benefit from economies of scale and no one has the scale that Tesla has in EVs and batteries.

"Be fearful when others are greedy and be greedy when others are fearful"

-Warren Buffett

Just a year and a half ago, Nvidia was just a company that made chips for gaming and cryptocurrency mining.  Today, investors have ultimately figured out that Nvidia is more of a software platform company that is providing the picks and shovels for the 21st century AI gold rush.  Nvidia stock is up 8 fold since it bottomed near $100 just 18 months ago.  Nvidia's business was suffering at that time even though, under the surface, they had been building an AI powerhouse on top of their CUDA software optimization layer for decades.  Unfortunately, I missed it, since I don't focus on chip companies.  I think Tesla has a good chance to do something similar over the next few years if I'm right about their real-world AI efforts.

Yes, Tesla's business is having a tough first half of 2024.  Today, Tesla appears to be just a car company similar to how a year and a half ago, Nvidia was just a gaming/crypto chip company.  Musk's political tweets turn away many potential customers and there's an air pocket right now before the next generation $25k robotaxi"for the masses" launches in 2025.

Tesla's stock is down from $300 last summer to $185 due to this temporary setback as the rest of the Magnificent 7 Mega-Caps have gone straight up.

I thought Q1 earnings  last week were going to be a disaster.  They actually were not that bad.   I was modeling a loss for Q1 given the 9% YoY decline in Q1 deliveries/revenue.  Tesla actually made $.45 in Non-GAAP EPS in Q1.  This is quite remarkable given the fixed cost deleveraging in a manufacturing business.

The TAM (Total Addressable Market) of real-world AI is easily in the tens of trillions of dollars.  I believe we're going to have a Chat GPT moment in real-world AI before the year is out; and it will be Tesla's autonomous driving platform that it's been building for almost a decade.  

Why this year?  Two words, transformers and compute power. 

Transformers, (the T in Chat GPT), ushered in the age of AI in 2017 when Google invented this  novel approach to deep learning. "Attention is All you Need" (PDF). Advances in Neural Information Processing Systems.

Sufficient compute power to fully utilize this revolutionary approach to AI took a few more years.  With Nvidia's release of the A100 Tensor Core GPU in 2020, it was now possible to train a multi-billion parameter AI model to memorize and understand all of Wikipedia and soon the entire internet with their next generation H100 GPU.

Tesla uses this same approach to deep learning utilizing transformers and immense compute power to train its autonomous driving platform.

Why can't someone else replicate this approach?  They certainly can and they are.  Google's Waymo, GM's Cruise, and many others are using this combination of transformers and massive compute power to train autonomous driving systems.

So why bullish on Tesla?

NO ONE HAS THE REAL WORLD  DATA THAT TESLA HAS. 

5 million Tesla cars driving around with 8 cameras each  ingesting video data and sending back to Tesla HQ to train their neural nets.

And even if competitors had the data, which they don't, they don't manufacture the robot cars at scale or profitably.  Waymo, arguably closest to Tesla in the autonomous driving race, only has approximately 1,000 cars driving around.  And their cars are not everywhere in the world like Tesla's.  Instead they're geo-fenced in easy driving environments such as Phoenix AZ, San Francisco CA, and Austin TX.  And Waymo cars cost over $100,000 to produce and Waymo doesn't manufacture them, partners do it for me.

Tesla sells the cars at a profit and every car they sell contains all the equipment necessary to train the FSD system regardless of whether a buyer subscribes to the software or not.  They also get the data regardless of whether FSD is turned on or not.  Today, only about 8% of the Tesla cars on the road have FSD turned on.  But, they get data from 100% of the cars.  

Not only does Tesla have the data advantage from 5 million cars driving around, they have another huge advantage,  the human driver supervisors.  Every driver of those 5 million Tesla cars, is a free software debugger for Tesla.  The billions of miles of real-world driving data that is sent back to Tesla HQ is extremely important to training the AI models, but what is even more important is the data generated from when the human drivers dis-engage FSD.

The 900,000 human drivers who use FSD is the most valuable asset to Tesla.  When they dis-engage FSD, they notify Tesla HQ that they disagreed with the autonomous action the car took.  This video data is flagged to Tesla HQ and the neural network polls the rest of the fleet to send back similar looking scenes to the ones that are flagged as a disengagement.

It really is amazing that Tesla charges FSD subscribers $8.000 or $99/mo. for the software when they actually gain a huge asset for every buyer of the software since the most important data that improves the system is the disengagement data provided by the human drivers.

This is creating an important network effect and costs less than zero for Tesla to create it.   As more drivers subscribe to FSD,  more disengagement data is sent to Tesla HQ to improve FSD.  The better FSD gets, the more people will want to subscribe to FSD, leading to more subscribers, more disengagement data, and further improvements to the FSD system, etc., etc.

FSD is a game changer.  40,000 Americans and 2,000,000 people around the world die from car crashes each and every year.  

I have a Google Pixel phone instead of an iPhone because my Android experience is not much different from an iOS experience.

It's going to be very different when it comes to choosing your autonomous driving system.  No one is going to want the second best autonomous driving system.  Loved ones' lives are at risk.

It's going to be a 'winner takes all' market I believe.  That market is in the trillions of dollars when you think about the size of the world's road-based transportation market.  Trillions.

When FSD gets to a level that is superior to human level driving, anyone who wants to get it will have to buy a Tesla car. At that moment that FSD reaches super-human capability, which I believe happens before the end of 2024, investors will wake up to what Tesla's been building.

Think the iPhone.  It's not better than a Samsung or a Lenovo, etc in terms of nuts and bolts of the actual phone.  It's what you can do with it.  The App Store makes everyone buy an iPhone.

I think the same thing happens with Tesla.  FSD is to Tesla as the App Store is to iPhone.

Everyone will want a Tesla to get access to FSD.

And that's only the first application of real-world AI.  The next application will be humanoid robots.  Then we're talking about tens of trillions of dollars of market value.  Though it's certainly early in that race, I think Tesla has advantages there as well.  They've already began manufacturing humanoid robots and are optimizing for low cost and power efficiency.   The same system architecture Tesla is using for FSD, they're using for humanoid robots.  That's very important when it comes to scale.  Other companies have to build a brand new system for humanoid robots whereas Tesla does not; its system for robots is already in place and at scale with regards to compute and neural network architecture optimization. 

Tesla's transition of it's autonomous driving software stack to full 'end to end' neural nets, combined with 900,000 human supervisors that label edge cases by disengaging FSD, will lead to a flywheel allowing FSD to rapidly improve in the months ahead.  A year from now, FSD will be 5X safer than an average human driver.  

Consumers will want FSD.  And the only way to get it is to buy a Tesla car....

The iPhone moment for the transportation industry.

Catalyst

FSD gets beyond human level safety within a year and millions of consumers order a Tesla car to get access to it.