But those FSD subscriptions would be almost pure profit. In 2024, Tesla made about $7 billion in profit. Fill in the blank how much profit FSD subscriptions would be, but the number is huge. Game changer.
Despite what you might have read on this board, delivering on those numbers makes a huge difference for the investment case if they deliver this year or five years from now.
Well, my understanding is that Tesla has about 8.5 million cars on the road today and about 12% of them subscribe to FSD. So that’s about 1 million FSD subscribers, so $1.2 billion per year in revenue.
MSFT is mostly a software company and has 35% net margins. FSD requires specialized hardware to run it, so I don’t think the margin will be as good as MSFT’s is.
But even if you give TSLA 50% margins on FSD, that’s only an extra $0.6 billion in profit or a little less than 10% increase in TSLA’s profit from 2024. Still, pretty much a nothingburger given TSLA’s 300 PE.
I don’t believe that TSLA decreasing the FSD price to $100 per month is going to result in any meaningful increase in the FSD take rate so I expect FSD to have a minimal impact on TSLA’s profits.
Sorry, I was getting ahead of myself. I was assuming that Tesla obtains true L4. At that point (one would assume) FSD subscriptions are flying off the shelves. Why wouldn’t you buy one?
The incremental cost between supervised FSD and unsupervised FSD is presumably tiny. So I think there is a big unlocked profit potential here.
I wouldn’t buy one because I don’t mind driving. I’ve watched a lot of FSD videos and it just does too much stuff that would annoy me like going too fast or too slow for my preferences, stopping too long at a stop sign, stuff like that. Even if TSLA does achieve FSD unsupervised, it’s still going to do things that I personally wouldn’t prefer.
I disagree that FSD subscriptions are going to fly off the shelves if TSLA achieves FSD unsupervised. From what I’ve seen, FSD can mostly drive unsupervised now. So, maybe in next week’s earnings report Musk will suprise us and say that the take rate on FSD has gone to something like 50% given the price decrease and the improvements in FSD, but I doubt it.
Edit: Let’s assume that TSLA sells 2 million cars this year (an optimistic projection) and that 50% (wildly, wildly optimistic) subscribe to FSD. That’s 1 million new FSD subscribers and $1.2 billion in added revenue. At 50% net margin that would be $600 million in additional profit or less than a 10% increase. Barely a blip.
I was too harsh in calling FSD a flop, but it’s certainly not the wild success that bulls predicted and doesn’t justify a 300 PE.
Assuming they make their goal of running FSD on standard equipment, what is the COGS on FSD?
Well, you may never be satisfied, but given a control for more or less aggressive, lots of people are likely to think “good enough”.
The reality is that if there’s a new, more efficient processor (en suite) then they’re going to upgrade, and some portion of the currently installed base is going to be left behind.
Since even Tesla now admits that it’s older models won’t be able to use the newest software, it seems unlikely that they will be able to just “flip the switch” as has been promised for so long.
Once they do achieve it, I believe it won’t be with the current hardware, and they will be essentially beginning the hardware race from that point. (There is an ancillary factor here, which might actually be the primary factor, and that’s getting the regulation portfolio in order. That could take a couple years, by which time there is sure to be a new hardware suite.)
In addition to Goofy’s comment above, this question may not even have a relevant answer. Any more than you can ask what the COGS is for airbags, or the air conditioning system, or the rear seats, or automatic windows.
One thing you might have noticed about the major AI models - they’ve all kind of converged. You have a bunch of cutting-edge AI companies all doing relatively separate AI ventures trying to invent the AI God Thingy. But they’ve all ended up with models with roughly similar architectures and capabilities and features. There are modest differences, but they’re all kind of the same.
This is kind of a “no special sauce” situation, right now. Everyone’s AI product is roughly as good as the general overall level of humanity’s technical knowledge in the field allows. Everyone’s pushing the frontier of what the tech is capable of forward. But generally, it’s the frontier of our level of computing technology/level of AI theory/level of programming knowledge that’s setting the boundary for everyone.
So - what if this same dynamic applies to real world AI, like self-driving? In that scenario, Tesla won’t be able to make their goal of “running FSD on standard equipment” until the general state of AI knowledge has reached the point where unsupervised AI driving is feasible (“unsupervised” meaning, presumably, being able to operate in a fashion that neither Tesla nor Waymo have today - the car drives itself without having to be part of a TaaS network to provide support in the event the car encounters an edge case).
But once the general state of AI knowledge has finally reached that point, then lots of people will be doing it. Just like with LLM models. Once the tech and the theory reached the point where one could build an LLM, you have more than a half dozen that all do the same thing the same way. The brain behind these things is shaped by whatever our global skill is, and the brain gets commoditized almost instantly.
And if that’s the path, then Tesla won’t be in a position to run FSD on standard equipment until the general state of global technology in the field has advanced sufficiently. And at that point, most car companies will also be able to offer an AI driver on standard equipment. In which case, they will. And AI drivers won’t be a product that’s “sold separately,” but something that’s just part of the car - without its own COGS.
Actually, Tesla just patented some AI software that many expect to enable FSD to work on HW3 hardware vehicles:
TL:DW; HW3 (from 2018) uses 8-bit pipelines, while the newer HW4 uses 16-bits. This patent is for a process to split 16-bit numbers (directly supported by HW4) into two 8-bit numbers through a convolution process, run each of those through the two HW3 GPUs on the board, then recombine them.
May not achieve 100% of FSD v14 on this older hardware, but might come close enough for SFSD use.
Let’s assume 100% margin on FSD. If we further assume 2 million cars sold this year (an optimistic projection like I mentioned above) and a more realistic take rate on FSD like 12% (matches the historical take rate) then we get 240,000 new FSD subscriptions this year, which only gives us $288 million in additional FSD revenue/profit this year.
That’s just a paltry 4% increase in TSLA’s net income year-over-year. So FSD is basically a nothingburger. That’s what I see as the bear case for FSD.
Watching YouTube videos is not the same as experiencing it first hand. Go to a Tesla store and get a demo drive. I’ve moved on from my Model X, but a friend with a Model S says he uses all the time when he’s not in a rush to get somewhere.
Depending on just what is approved (L3 vs L4, limits of the ODD, etc.), UFSD could drive sales of not just the FSD subscriptions, but for new Tesla vehicles. Being able to text while driving, even watch a movie, etc. will change a lot of commutes for people.
In the meantime, the focus is on Robotaxis. SFSD or UFSD are just kickers at this point. BTW, your math assumes no increase in subscriptions for existing vehicles, which is unlikely.
Interesting that you chose to group two of my bullets together and not address the second one, on costs, after saying “Waymo is losing huge amounts of money.”
My first point is relevant to gauging Tesla’s progress. Proving that Tesla’s technology - software and hardware - is viable and is approved in one market paves the way for additional markets. You’ve made the argument that Tesla’s hardware won’t pass regulatory scrutiny, and approval in its first market will clearly address that argument.
We all know that Tesla is behind in terms of robotaxi deployments. Waymo’s roll-out has been slow: just 4 new cities in 5 years. Here’s the number of vehicle currently deployed:
That’s 2500 vehicles total. Waymo acquired 3500 iPaces from Jaguar, with no more being made. So, another 1000 vehicles to go. Further expansion will require the new Zeekr and/or Hyundai vehicles. What’s Waymo’s retraining time for new vehicles with new sensor suites? What’s the re-certification process for these new vehicles It’ll be interesting to monitor Waymo’s progress and see if they really have hit an deployment inflection point.
But, that doesn’t mean Tesla won’t be making progress, or that Tesla, with its much lower CAPEX per vehicles and ability to mass produce them, won’t catch up and surpass Waymo at some point in the not too distant future.
Again, to my OP, Tesla getting approval in one city takes yet another claim of the doubters away and will boost the stock price.
Back to costs, the analysis I linked to in another thread shows that Tesla’s lower capital costs will mean it is potentially more profitable than either Waymo or Uber. We’ll have to see as things play out to see if that analysis is correct or not.
As for Waymo losing money, clearly not everyone thinks that’s going to continue.
Clearly, investors with money think there’s a profitable business in robotaxis. But, interesting that Google isn’t just backing the whole thing.
I posted the question for cars they will sell, not for existing cars. That’s yet another potential market, even if is has some hardware retrofit requirements.
You keep trying to spin out of this. I made a very simple clear statement:
There is no debate on this. It’s a fact. Stop trying to spin your way out of it, first obscuring the very nature of my point by not talking about either robotaxis or autonomy for POVs, but instead glossing over it with some generalized talk about “AVs,” and now talking “order progress.”
You may disagree that my point is important, but you can’t deny that it’s real and factual. Actually, I guess you can, in which case our discussions are over.
No thanks, not interested. Why would I want to pay $100/month to sit in the driver seat of a car and not drive? Makes no sense. Musk would have to pay me if he wants me to be an FSD beta tester.
In the meantime, the focus is on Robotaxis. SFSD or UFSD are just kickers at this point. BTW, your math assumes no increase in subscriptions for existing vehicles, which is unlikely.
Ah yes, now that FSD has proven to be a revenue/income nothingburger Musk will want to switch the focus to the pre-revenue robotaxi. Keeping the focus on pre-revenue boondoggles is the perfect way to keep a story stock afloat.
I’ve had trouble finding the number of Tesla cars on the road that are FSD capable, but let’s assume 7 million. If 5% of those who aren’t already subscribed to FSD subscribe, that is 350,000 new subscribers or an extra $420 million in revenue/net income.
Adding that to the $288 million in ‘new-car’ subscription revenue I calculated above gets us to $708 million in extra revenue/income or a 10% increase in net income. Again, barely a blip.
Each of which has a list of parts, a cost for those parts, and assembly cost that can be prorated, together making a COGS. And NOT including the R&D necessary to come up with the current design.
Sure - but is that something that can be done for something like FSD? FSD needs the entire car to operate. It’s not a physical “part” of the car, since we’re talking about a function that runs on standard equipment - not an optional installation. Does it have its own COGS if it’s not a separate option that gets added or not added to the vehicle?
Can be? Sure. Currently the incremental cost to add FSD is $0. All the hardware is in the base car.
I guess I’m asking more of an accounting question. I thought COGS applied to…well, “GS.” Goods sold. If something isn’t an actual “good” - if it’s not something that’s being manufactured or made for the customer from other materials - is COGS still a relevant concept from an accounting perspective?
Oh, sorry - I was talking about solving the march of 9s to L4 driving and solving autonomy with human-level safety.
Are you talking about human-driven taxis?
