Elon Musk, the Tesla Robotaxi event & the nature of increasing-returns tech monopolies
SpaceX might have cracked the code to a tech-enabled, real-world business model that has monopolistic increasing-returns cost curves - there is a chance that Tesla might follow.
On the 10th of October 2024, Tesla held its “We, Robot” demo day presentation in the Warner Bros Studio in Los Angeles. It’s the studio lot where science fiction classics like Blade Runner were shot in 1982. So Tesla took the occasion and presented its vision of a future that ‘takes the word fiction out of science fiction’. It demoed potential future autonomous taxis (Cybercabs) and autonomous shuttles (and humanoid robots) - without giving too many details about prices and timelines, but embedded into a vision of a future in which autonomous mobility will be the default, and around which whole cities and human life will adapt. Some people loved it (as always then it comes to Tesla). Some people were skeptical if this is ever going to happen anytime soon, and rightly so (as always when it comes to Elon). Some people called it vaporware and compared Tesla to Enron (as always when it comes to Tesla).
3 days later, on the 13th, SpaceX for the first time ever managed to catch the huge Starship booster out of the air while it was landing, with the so-called Chopsticks (that are part of the launch / landing tower). Understanding the second event matters a lot in order to make sense of the first one - mainly because SpaceX is further down its path to success, while playing a very similar kind of game.
Weak link engineering problems, strong-link technology problems
In order to frame the context and how to think about it in a productive way, the concept of weak-link vs. strong-link problems is useful.
In weak link problem spaces, the overall performance / outcome of the system is determined by the weakest link, so improvement comes from eliminating that bottleneck. It’s about lifting the floor, so-to-speak, about addressing the least-performing element in order to improve the whole. Think of a football (soccer) team: It does not matter much if a few superstars play for your team, as long as you have 2 or 3 total amateur players on board that for sure will decrease the overall chances of success. So you are probably better off to trade in your superstars against solid B+ level professionals and get rid of the amateurs in order to improve.
On the other hand, there are strong-link problems. Solving for the strong link in a problem space means raising the ceiling of the entire system. Optimizing the top-performing characteristic creates a disproportionate system-wide benefit because there is a nonlinear relationship between input and output of the system (and improvements at lower levels produce only minimal effects). Venture capital investing is a strong link example: it does not really matter if some of the investments of a venture fund go to 0 in value, as long as one or two investments return 10-100x of their invested money. Total fund success is determined by one or two most successful bets, not by many unsuccessful ones - mainly because venture-scale companies are extremely high growth opportunities by definition.
Weak-link problems often are engineering problems: while the overall maximum system remains the same, the weakest link needs to be fixed with cleverly engineered solutions. Strong-link problems on the other hand are not merely solved by solving singular engineering issues. They require more of an overall systems perspective that includes engineering and things such as
Manufacturing / operations & overall efficiency
Cost structures
Scaling path
Supply chain management
And so on (for other companies it might be things like GTM approach, superior customer support etc.)
which could be considered a holistic approach to technology (thanks to Abe Murry who shared some thoughts on X that inspired this train of thought).
The Starship system cost structure as a strong-link problem
The tower catch attempt of Starship in insolation was a mere weak-link-problem for SpaceX (even though a really hard one): After they were able to demonstrate the capability to launch the integrated stack of booster + ship in the first test flight (it exploded before reaching space), they then nailed stage separation on the 2nd attempt (both parts exploded), got the ship to orbit on flight 3 and, on flight 4, landed both the booster and the ship (which barely survived reentry) in the ocean. Flight 5 now was proof that they can land the booster again in a spectacular fashion, right where it had launched, fully intact and (basically) immediately reusable!
In a sense, the most spectacular thing of the integrated test flight #5, which was the booster catch, was the remaining weak link issue that was left to solve that enabled a much bigger framing: For SpaceX, the starship program as a whole is a strong-link problem! It aspires to “raise” the ceiling of spaceflight by orders of magnitude and towards what’s physically possible, with the defining system characteristic being an absurdly low cost structure!
With the first successful landing of its Falcon 9 rocket in 2016, SpaceX initiated a totally disruptive cost curve trajectory in the launch industry that is now topped off by Starship. The Starship system architecture in fact specifically aims for a system-level, globally (almost) optimal cost structure. Here is how SpaceX is trying to get there, among other things:
Cheap materials (cheap to build means cheap to iterate & test)
Modularity that facilitates economies of scale (e.g. 33 engines for the booster, the booster being made of 33 identical stainless steel rings, (almost) same engines for booster and ship etc → a lot of scalable output)
Designed for manufacturing at scale → Starship factory with a high degree of vertical integration
Fully reusable - both stages will land again (which they both already managed to do)
Rapidly reusable - scalable ground operations (think: the scale of the fuel tanks, fueling operations and logistics, stacking the next ship on top of the booster with the chopsticks etc.)
Cheap & effective fuel (It burns liquid oxygen and methane)
Available R&D infrastructure that allows for fast and cheap testing (thanks to F9 & FH) → virtualization, HITL testing, simulation infrastructure etc.
Booster catch – no landing legs for the ship, which makes the ship easier to build and cheaper, because overall operations get faster (Tower = “Stage 0”)
It’s a highly integrated approach. Such systematic and system-level cost structure considerations can’t be resolved in hindsight after having a functional technology in place. It has to be part of the product specs and execution from the get go. It’s deliberate, it’s a strategy (Founder Mode™ Elon Musk Edition).
Technology cost structures, scale & simplicity
The outsized value of technology that we see in the world today comes from scale. Scale will bring costs down (if the business model is designed the right way), which will bring demand up, which will bring costs further down, which will increase demand further and so on. Quantity is a quality in itself!
It’s how Moore’s law has worked over the last 50 years, and why computation is now as cheap as it gets at scale. It was never a law of nature, but always driven by the human aspiration (of initially Intel) to push towards the limits of the possible, which drove demand, which increased factory utilization, which in turn made chips cheaper and kicked in a virtuous cycle. Yes, chip fabs are insanely expensive (TSCM chip factories cost on the order 20 billion USD), but each chip produced only costs on the order of a few dollars (it’s basically sand). Software is even more extreme: It can cost millions or tens of millions to build the first version of a software product, but then making the first copy (on a CD disk in the past, now as a new instance in the cloud) is almost free. Certain technology domains follow the logic of increasing returns (the bigger you get, the more you make, as W. Brian Arthur wrote in his seminal piece in 1989), with software being the ultimate instance of this principle, because it literally has a (basically) zero marginal costs structure.
The fundamentally limiting factor to low cost hardware at scale is production speed, mainly because the capex for the factories is so high and hence depreciation is one of the biggest cost blocks per unit. So making more units with the same equipment pays off and makes each marginal unit cheaper (Software: Instant copy for the cost of electricity, thanks to things like virtualization and the cloud; Cars: A modern car factory costs billions, so you better make sure that you make a lot of units consistently in order to depreciate the investment over more units). It’s why Musk keeps pushing for speed all the time, in all of his businesses:
If the design takes a long time to build, it's the wrong design. [...] And therefore it must be modified, so that progress gets accelerated.
SpaceX, just like any Musk company, is therefore good at throughput maximizing: The faster you can deliver your product/service at scale, the better your costs will look like - at scale! But in order to build something fast, it needs to be easy to build. Complains Musk:
Over and over, the tendency is to complicate things.
Therefore striving for simplicity and with that for better scalability is now the #1 operating approach at all Musk companies, which he calls “The Algorithm”: Design a technology for scalability from day 1 - make it easy to produce, which requires making it simple, which in turn will benefit its scalability, due to many different reasons:
Materials that are easier to handle
less parts / less processes
Easier production methods
Designs that are easier to automate / designed for automation
less labor / lower-skilled labor required for the remaining manual work
More modularity
Etc
That is why “deleting parts or processes” comes first at SpaceX, and only after that do they optimize the product or component and eventually automate production. First make it simple, then make a lot of it. What is crucial to understand here is that technologies are recursive: a rocket is a technology, which consists of engines, tanks, flaps etc which are technologies as well. So are fuel tank farms and launch operations procedures. So scalability is recursive in this case as well: it’s not only that SpaceX works in a simpler overall system, but that each element of the system is pushed towards simplicity as well.
Musk is bringing the increasing returns / zero-marginal cost logic (huge first copy costs, high risk - low to zero marginal costs, huge competitive advantage at scale if it works) from chips and software (and basically any knowledge-heavy business domain) to more traditional industries like automotive and aerospace. It’s not only that software is eating the world, but as well software-like economics start eating hardware domains with historically more traditional economics. Or, to state it more precisely: Silicon Valley is eating the rest of the world.
The Tesla Cybercab system as a strong-link-problem framing
Back to Tesla’s “We, Robot” event. Yes, it was a demo event. Yes, some of the humanoid robots were tele-operated. No, we can’t be sure whether the RoboVan will ever become reality. Yes, it remains to be seen if they can achieve level 5 autonomy anytime soon (or even before 2027, as advertised. Given that Elon time” is a meme now even in the Tesla community, it probably won’t be available much sooner than 2027). But here is what actually matters to understand:
What was presented that day was not so much a weak-link problem of level 5 autonomy. What was shown was the solution to a strong-link problem: the lowest-cost, end-2-end autonomous mobility service possible that can scale with a baked-in increasing returns logic.
Tesla is not only building the (relatively low-cost) sub-30.000 $ car that will be manufactured via a manufacturing innovation called the unboxed process (thanks to a high degree of vertical integration, design for scalability, own batteries etc). It also owns and operates a global supercharger network (and claims to charge the robotaxis via inductive charging, so no person would be needed to plug them in for recharging → delete parts or processes). Those are all elements of the problem space, and Tesla has solved them more or less already.
Level 5 autonomy and the bitter lesson of scaling AI systems
Enter the self-driving problem: The ability to go fully level 5 autonomous, without pedals and a steering wheel, is the puzzle piece that would raise the system ceiling to new heights and cause outsized improvements. It would introduce a similar, massively disruptive cost trajectory to autonomous transportation as Falcon 9 and Starship did for space launch services in the last decade.
But it’s not solved yet. It is the current - and probably super hard - weakest-link problem that Tesla is working on. There is a possibility that it might never get solved, even though that’s unlikely. If it gets solved however, it will get solved through a combination of data, algorithms and computation. And this factor combination can be approached in two main ways:
Competitors like Waymo already offer level 4 autonomy solutions in the market that make use of many sensor types like radar and LIDAR, with a much more hard-coded, rule-based system. They work under certain boundary conditions (like geofencing, in certain cities). It’s a feasible and operational approach, but very expensive in the margins (Waymo cars apparently cost around 200.000 USD each at the moment).
Tesla’s approach is - of course - the simpler one: No expensive radar or LIDAR for inference in the fleet. Just camera vision and AI at inference time, and a neural network architecture that makes the car learn how to drive from scratch, without any human teaching, just by looking at millions of miles of human driver video data. Delete parts or processes and transform a hardware problem into a software problem. Simpler, more fundamental solutions lead to more scalability and therefore better cost competitiveness. That has been very true for SpaceX so far.
It’s unclear if this approach will work for FSD as well. But here is what we know has worked in the past for AI problems though - it was a bitter lessons for humans:
We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning.
This short section of Richard Sutton’s essay called “The Bitter Lesson” hints at this: When it comes to progress in AI systems, more compute, more data and pure learning algorithm approaches always have outperformed more human intelligence built into the solution design via expert rules - period. The “simpler” approach always was the more scalable one in the long run, and more scalable means cheaper at scale (even though it might be expensive in the short term: NVIDIA’s Jensen Huang said in a recent interview that the Tesla AI GPU cluster is “easily the fastest supercomputer on the planet as one cluster”, consisting of 100.000 H100 & H200 GPUs. This has cost Tesla easily a billion dollars, and it might get bigger in the near future).
All those links chained together - if successful - would create the (almost) lowest possible cost structure imaginable for a fully autonomous, generalized (software) and scalable (hardware) robotaxi service. It is the remaining weak link, the decisive one. There is a certain probability that the approach won’t work anytime soon. Let’s assume it will work though - what would that mean?
Moats vs. tech trees and the nature of technology
In 2022 Elon Musk tweeted:
Musk’s tweet was sent in the context of a debate about moats (barrier of entries that existing businesses set up) and about their defensibility over time. Musk’s position is clear: Tech trees are the thing to build. Such trees map the dependencies between different technologies and therefore basically point out which new technologies get unlocked once certain other technologies have been built and added to the “tech tool box”. It’s like a directed technology dependency graph (see image). It's a view on technology that is adopted from strategy computer games
But it’s not just a computer game analogy. It fundamentally reflects a view of technology as a collective pool of building blocks which together form a complex adaptive system:
Different tech domains evolve over time, collectively, driven by collective human invention (e.g. the open source movement), and the entirety of the collection of building blocks works like a complementary toy box that over time enables you to do more fancy things (and gets filled with new stuff almost daily). Technological progress in some sections of the tree beget progress in other sections (and vice versa). It’s a view of the world as a complex-adaptive system that is made up from technologies - they are core to economic development and progress, not the outcome. So why does this matter for the debate about viable autonomous mobility offerings?
Consider this: SpaceX is currently on track of delivering 90% of all mass to orbit in the year 2024. It’s delivering more payload than any other company or country, and obviously more than all of them combined. And here is the thing: This de facto monopoly position does not even factor Starship into account, but has been achieved merely with the partially reusable Falcon 9 / Falcon Heavy architecture, which is already absurdly cheap. Until today it is the only booster architecture that has been reflown multiple times - more than 7 years after its first landing (RocketLab recovered one of its boosters successfully though). People would consider this low-cost advantage as a moat, in a more traditional perspective on technology.
But SpaceX did not stop there: With Starship it will basically have a launch monopoly that can't be outcompeted anymore for the foreseeable future (if we ignore antitrust or foreign state actors for the time being). As Isaiah Taylor remarked on Twitter:
So goes his reasoning: In order to compete with SpaceX for launch services, you have to basically build your own version of the Starship system (SpaceX spent on the order of 10 billion USD on development). So far, so good. Here is where it gets tricky: In order to build a Starship equivalent, you have to have a profitable Falcon 9 system in place first. Not only was Falcon the money printer that funded Starship until today. It also enabled SpaceX to build the capabilities in design, manufacturing, launch operations, testing infrastructure, software stack etc. that you need in order to pull off such a hugely ambitious and complicated feat such as the Starship in the first place. It’s not just about the billions of dollars R&D expenses, but the human skills that need to be developed and honed while building it. Nobody else in the industry has that capability. SpaceX is already the cheapest provider, and will be the cheapest provider by far once Starship is fully online (in ~1-3 years).
And that is where the monopoly lock-in happens. SpaceX’s tech tree is now simply too powerful and advanced for others to realistically compete with it anytime soon. And it turns out that your existing tech tree capabilities can serve as a boot loader that unlocks further (probably unplanned) value creation options: Thanks to its low marginal cost launch platforms, SpaceX was able to launch a satellite internet network that now is the de facto global monopolist, simply because there is no viable alternative available that can deliver high-speed, low-latency connectivity at scale around the globe. It now has around 4 million subscribers globally, which probably translates into hundreds of millions USD of revenue with absurdly high operating margins at scale. Starlink now is the second money printer that gives SpaceX enough runway to push the Starship system to where it needs to be - at a minimal marginal cost.
The Starship moment & monopolies in the world of increasing returns
You could argue that the same reasoning applies to Tesla: when it was just a car company with a successful but niche-like Model S, it was the fancy new kid on the block but largely irrelevant in terms of volume. But it learned how to design and manufacture EVs. Over time Tesla added an ever increasing set of tech branches & twigs to its tech tree: the supercharger network, a proprietary battery production, after 2018 the affordable Models 3 and Y, which were produced at scale on several continents (and more broadly available, with many EV-specific improvements over time) and many manufacturing innovations (like die casting and the unboxed process). In 2016, ground work on FSD was laid, and it was deliberately decided to not go with sensors at inference time in the cars, but with vision only (because it is simpler at scale).
Each of those technologies was considered not to be defensible enough over time: Others will get good at batteries too, others will get good at software too, others will work on autonomy too. There is no moat, Tesla will become commoditized. For each of the individual technologies, this might be true: other legacy car companies might have solid EVs in place by now, there are many other leading battery companies, and Tesla’s market share (for car sales) in many markets has stagnated.
But just as SpaceX has shown us: Moats for isolated products don’t matter much if you work towards the lowest cost structure possible at scale, with the help of your tech tree! Outside of China, only Tesla, due to its vertically integrated approach, has now access to all technologies necessary in order to pull the Robotaxi fleet off. In order to compete with them, you would have to have not only a profitable EV manufacturing business in place (which most legacy OEMs don’t), but as well billions of miles of actually driving data, a global supercharging network plus all the skills and capabilities that allow you to build and operate all of those technologies, at scale - including a giant 100.000 GPU supercomputer cluster (and a guaranteed 200 MW power supply by the end of 2024, which roughly translates into the combined output of about 66 large offshore wind turbines and could power ~150.000 households). Good luck!
So as already said before: It remains to be seen when and if Tesla will reach level 5 autonomy required for a scalable robotaxi service. If it doesn’t, there will not be any special long-term value baked into Tesla as a tech business, which is why Musk already said years ago that
Solving Full Self-Driving...is really the difference between Tesla being worth a lot of money and being worth basically zero.
It will just be another successful EV company among many others.
But if it does solve level 5 autonomy, it will all of a sudden be in a position that is similar to where SpaceX is now: basically hard to compete with, because the end-2-end system operates close to a globally optimal cost structure level - at scale. It will be a business that has been designed with a built-in increasing returns cost structure: It gets better at scale, technologically and economically, and therefore has characteristics of a winner-takes-most market (ever wondered why there are only two smartphone operating systems, or basically only one or two SaaS providers per vertical, or only very few CPU or GPU makers?).
People have repeatedly compared Tesla cars to the iPhone, and this comparison was used again after the Robotaxi event (“the iPhone moment”). I’m not going into the details why this comparison is probably only valid to some degree (the iPhone did reach 50% market share, but not because it was cheap, but because it was a great smartphone and because the OS platform was governed by an increasing returns logic (the typical network effect: the more phones, the more users, which makes it more attractive for app developers, which leads to more apps, which attracts more users etc.). So Apple’s success was still governed by the network effects of its software platform, just like with Microsoft & Windows in the past).
But in a sense, the robotaxi event might in hindsight be considered as something closer to a “Starship moment”: Starship as a system totally will disrupt human spaceflight, on many dimensions, due to its (absolutely) superior increasing returns cost structure. SpaceX’s missing puzzle piece was the tower catch, which made nailing this technology on the first attempt so special for the entire company in October 2024. Tesla’s Cybercab fleet on the other hand has the potential to completely disrupt individual ground transportation as we know it, by making it fully autonomous, with a potentially (absolutely) superior increasing returns cost structure. Tesla’s missing puzzle piece will be level 5 autonomy with camera vision and AI only.
So it will be interesting to come back to the 10/10 event in LA one day in the future and reconsider whether by then we still need a fantasy movie study to shoot science fiction movies about full autonomy or whether we already live in such a world by then.
Some more reading material
My book “Teslafication” - see www.teslaficationbook.com
A great piece of Ben Thompson from Stratechery: https://stratechery.com/2024/elon-dreams-and-bitter-lessons/
Brian Arthur - The nature of technology: what it is and how it evolves