Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

[Fact Checking]

What would stop people with a lot of capital, buying all the vehicles and immediately reselling them for $$$?

Not reselling them, but renting them out with annual re-pricing - so that they can take advantage of the appreciation.

Will the development of FSD and perhaps a monopoly on the market make tesla increase the price of their cars significantly?

Tesla is unlikely to have a monopoly in the long run (Intel's MobilEye will certainly be selling turn-key FSD solutions, and I suspect Apple isn't sitting still either, and I'd be rather disappointed in Bezos as well if he wasn't taking the next trillion dollar high-tech market seriously either), and it's an interesting question of economics of how it is going to play out if there's multiple FSD offerings competing against each other. Will they race to the commodity bottom? Or will they just have pricing similar to human taxi driver pricing and pocket the non-existent 'labor' cost?

If history is any indicator then in new high-tech markets there's always one premium product category capturing much of the profits.

Anyway, very little of this is going to be realized until FSD capabilities are more real. The moment they are, it might be a gold rush like AirBnb was (and is) a gold rush - with the difference that the 'real estate' that can be bought is priced constantly and is still up for grabs.

I'd expect Tesla to start raising the price, but not of the cars, but of commercial FSD licenses. There might be a consumer FSD license that doesn't include robotaxi revenue generation. We already saw this earlier this year, when Tesla first removed commercial licensing from their FSD conditions and then added them back.

 

[CorneliusXX]

I agree about the Model Y and the Semi, but I was a little taken aback by Elon making the following statement in the Q1 earnings call:

Back in 2016 or so, I clearly remember Elon making the claim that the Solar Roof would last the lifetime of the house. That's a significantly better value proposition than 30 years! A regular, cheap composite roof should be able to last for 30+ years, and it can be walked on and modified over the years as needed.

Further, solar cell technology continues to improve. In 20 years, if we can substantially increase our solar generation by replacing existing panels with newer ones, it'd be a lot easier to do so with standard solar panels.

I think the Solar Roof is going to be a nice niche product for high end homes. Personally, I love the look of slate tile, and installing a slate-style Solar Roof could be the ticket for some. But I can't see us considering a Solar Roof or recommending it to most people unless and until the costs come way, way down.

On the other hand, I'm very happy to see Tesla's new, lower pricing and simpler options on residential solar arrays! Whether this generates significant profit or not, it's the direction that the solar industry needs to move in. I also think that maintaining visibility in the solar and storage business, in addition to EVs, helps Tesla to have a great marketing story (generate, store, and use sustainable energy) that's central to the company's mission. With Tesla's lower prices on panels, though, it's even harder to justify the expense of a Solar Roof.

They're building the Roadster of solar tiles at the moment. As the product develops they will produce a MS and M3 of solar tiles.

The raw material cost is very low for a solar roof as it's basically just made out of sand, lower prices will come with design and manufacturing progress.

 

[dqd88]

FSD just increased to $6k. A few hours ago it was $5k. Shucks. Hope it doesn't keep going up as I want to buy in a few months.

 

[T@oo]

FSD just increased to $6k. A few hours ago it was $5k. Shucks. Hope it doesn't keep going up as I want to buy in a few months.

Oh dammit. That was supposed to be postponed to the 10th. Grrr.

 

[Alphacrux]

I'd expect Tesla to start raising the price, but not of the cars, but of commercial FSD licenses. There might be a consumer FSD license that doesn't include robotaxi revenue generation. We already saw this earlier this year, when Tesla first removed commercial licensing from their FSD conditions and then added them back.

So in this sense, robotaxi revenue generation is offered to all Tesla buyers initially, like free supercharging was, then withdrawn after a certain date (with existing owners grandfathered in, of course). Then a current owner’s car would have “extra” value, just like free supercharging does now in used Tesla’s. If the owner does not want to rent out his Tesla in the robotaxi service, he could sell his vehicle for a profit.

This is an interesting scenario. I’m in the process of selling my 2015 AP1 S with free supercharging (which transfers) and getting a new S with FSD. So if Elon’s vision comes through (a big “if” to many), I could actually be increasing my potential profit?

 

[Fact Checking]

This is an interesting scenario. I’m in the process of selling my 2015 AP1 S with free supercharging (which transfers) and getting a new S with FSD. So if Elon’s vision comes through (a big “if” to many), I could actually be increasing my potential profit?

Several layers of 'if', but yes, that's a possibility. Also, the new S with the new drive train has the 1-million-miles design, which might independently lift the depreciation curve as well - beyond the higher performance and the increased 200 kW charging rate. But not advice, as usual ...

 

[Fact Checking]

FSD just increased to $6k. A few hours ago it was $5k. Shucks. Hope it doesn't keep going up as I want to buy in a few months.

The price of the FSD option will likely go up as Tesla keeps introducing new FSD technologies. I pointed out potential FSD price hikes last October.

But as the price goes up I'd expect there to be more levels to it: for example "base", "consumer", "professional" and "commercial". It would be a marketing/pricing mistake to sell any of the individual FSD components for much more than ~$5,000, as it wouldn't allow budget constrained customers to stretch as long as they can.

So I'd expect Tesla to accommodate regular car owners: if for example in two years if commercial level FSD costs $19,900 then Tesla should make $5k, $10k and $15k variants available as well - why lose a lower price sale if someone doesn't have budget (or interest) for the full $20k option?

Tesla might also offer revenue sharing tiers - for example with a $20k FSD option Tesla keeps 40% of the robotaxi revenue, but pay $50k upfront if Tesla only keeps 25%.

I.e. since Tesla FSD is going to be a software licensing plan, not a physical option, Tesla has literally an infinite amount of pricing options available to front-load revenue or earn it later on, if they so desire.

I also expect 'FSD jail-breaking' to be a very lucrative business - but since Tesla's new AI chip has built-in crypto it's probably a losing battle in the long run - so to run Tesla's FSD chip you'll need valid a license.

 

[ReflexFunds]

Actually, every nine requires more edge cases to be solved (not sure whether exponentially higher or linearly). This is because each of those edge cases happen more infrequently than the earlier ones (if they have prioritized correctly !). Essentially this is the "long tail" problem.

This is why i said it wasn't intuitive. My instinct was also to assume the above, but I don't think it is correct.

The probability of no disengagements per mile is the sum over all possible driving scenarios of: Probability per mile of this driving scenario occurring * probability of handling this scenario correctly. For simplicity, lets assume each scenario has a 10% failure rate.

If the current top 1000 problem cases identified by Tesla's problem identification system on average happen once every 10000 miles, then the probability of an error per mile across all of these 1000 cases is 1/10000 * 1000 * 10% = 1%. So improving failure rate on these first 1000 cases from 10% to well below 1% reduces your overall probability of failure per mile by c.1%. This could be your second 9 of accuracy (rounded to 99%), but only if the probability of error per mile summed over all remaining problem cases on your list add to less than 0.5%. Otherwise you will have to move further down your list to get to 99% accuracy, maybe even to problem number 5000-10000. But it is here where you meet the long tail - every 9 has some more common problems which give you large reward for solving followed by a long tail of diminishing returns. But solving the next 9 does not have a longer long tail than the previous 9. If the first 9 took 5k scenarios to solve, then the next 9 should also take c.5k scenarios. This is because the decreased frequency of events now you have moved further down your problem list scales in-line with the 10x smaller problem you are now solving (now only need to solve problems summing to 0.1-0.4%, rather than 1%-4% previously).

Right - they are basically solving for accidents (or mistakes) per million miles. This is where it becomes increasingly difficult to figure out what are the next top 100 cases to solve. They get a lot of disengagements / shadow errors from the fleet. Millions every day. How do they analyze and then synthesize the data without watching each video and manually categorizing ?

Karpathy was talking in the youtube video about how the problem has shifted from software algos to data. These are the kinds of data issues that need to be solved - and automated - in order to make FSD possible.

I think we can guess roughly how Tesla's automatic data filter works and/or how they plan to scale it up in future.

At different stages of the development cycle, Tesla is collecting high level summary data on 1) when shadow mode executes a manoeuvre which varies significantly from the human, 2) when a human disengages Autopilot and 3) full sensor data when a car has an accident. From the shadow mode and disengagement summary data, Tesla can try to group the events into categories of the most common problems (this could just be type of location and speed and type of disengagement data, or it could be more detailed info about what the NN sees). Within these broad categories, Tesla can ask the car to start collecting more information when similar situations are next encountered by the fleet. From this they can sort the new data to split the broad category into smaller subsets and better define the problems. Tesla can use this to build a priority problem list of these subcategories. Tesla can then write heuristics for cars to collect data in these situations, or it can train a basic NN to recognise these situations in sensor data as a trigger to start collecting data. When Tesla start getting data back, they annotate this and feed it back to the recognition neural net. This makes the cars better at detecting the situation they are targeting, so the quality of data sent back by the fleet improves. At this stage they should be getting very high quality, relevant information from the fleet for the problems at the top of its priority list. Now humans have to review and annotate the data, and update the heuristics and retrain the neural nets.

As Tesla starts progressing through the march of 9s, shadow mode/real world disengagements will become less and less frequent. This means Tesla can start collecting more and more information from each of these events as their baseline (if these events are now 10x less common, Tesla can collect 10x more data from each event without increasing its overall data usage). This will make the priority problem identification easier as they move forward.

 

[lklundin]

OT


Sadly, printed bond coupons were eliminated in the US in the early 2000s IIRC. I remember clipping coupons too, though we did it at the bank since we kept our bonds in the safety deposit box.

In my native language it's all but a (mostly forgotten) saying, a "kuponklipper" (Coupon Clipper) is a person able to live off of their wealth.

 

[vigleik]

Oh dammit. That was supposed to be postponed to the 10th. Grrr.

...for those unable to order by May 1.

Like me.

I was planning to pay for part of my Model 3 with Tesla stock, but I don’t want to sell any of my stock at this price. I’m really not sure what to do if orders open up in Australia in the next few days, especially if I have a very short window for ordering FSD at the old price.

 

[lklundin]

OT

(In the old old days, your personal vehicle doubled as a work vehicle, and was called a "horse")

In a not-so-well developed Asian part of Turkey I have heard a horse being referred to as an "Anatolian Mercedes".

 

[KarenRei]

In a not-so-well developed Asian part of Turkey I have heard a horse being referred to as an "Anatolian Mercedes".

My Model 3 is going to be named Syndargæðingurinn, which is about the closest I can come to "The Thoroughbred of Sin". If in the future I ever own multiple Teslas, they'll all be given horse related names (generally from Norse mythology).

The car is the horse of our era

 

[Artful Dodger]

Sooo GF 3 is almost done..lol.

RayTesla on Twitter

Ray4Tesla's GF3 timeline for completion:

• May - main buildings
• June - flooring & roads
• July - utility lines
• August - production lines
• September - tooling

First Model 3 roll out towards the end of September.

 

[JohnDinger]

Now that the capital raise is done

Source on it closing? I'm curious what coupon rates they got for their converts.

 

[Dan Detweiler]

FSD just increased to $6k. A few hours ago it was $5k. Shucks. Hope it doesn't keep going up as I want to buy in a few months.

The closer we get to feature complete FSD the higher it will go. This kind of ability won't go cheap.

Dan

 

[j6Lpi429@3j]

I just had a question on a probable future scenario, curious on peoples thoughts. With the success of FSD, we are all talking about Tesla vehicles being able to create passive income of up to $30000 a year. This would undoubtedly increase the market value of the vehicles as soon as they are bought from the company, right?

So a model 3 costing $42K with FSD, would immediately have a resale value of way higher than $42K as soon as it has been bought. What would stop people with a lot of capital, buying all the vehicles and immediately reselling them for $$$? Will the development of FSD and perhaps a monopoly on the market make tesla increase the price of their cars significantly?

Yes. They would be nuts not to cash in on that premium.

 

[Swampgator]

Of course, so am I. Everyone thinks that.

I have 35 years of driving with zero accidents to back it up.
And I will not decapitate myself by driving under a semi.......ever.

 

[Cobbler]

Ray4Tesla's GF3 timeline for completion:

• May - main buildings
• June - flooring & roads
• July - utility lines
• August - production lines
• September - tooling

First Model 3 roll out towards the end of September.

The machine that builds the machine:
GF3 tooling would be version 1.0?
Fremont m3 production line is 0.5 and has some flaws (hence ga4 tent). Major revamp on a clean slate. Will be interesting to see how much more efficient the line will be.

 

[Artful Dodger]

Climate national emergency decision overshadowed in UK by minister sacking and now now local elections.
UK Parliament declares climate emergency
Combined with London emissions zone, UK getting fairly serious now.
Brexit not in the news - EU will go ballistic soon for lack of progress!

Really serious. /s

"This proposal, which demonstrates the will of the Commons on the issue but does not legally compel the government to act, was approved without a vote."
​

SS/DD.

 

[aubreymcfato]

So 'exponential' is a bit of a hyperbole as obviously there's limiting factors, such as the size of the universe. But it's definitely faster than linear, and I see three major areas where this really matters:

1)

There's one fundamental metric that is driving the improvements, it's the growing Tesla fleet - which will be growing exponentially for the next 3 years and longer (we hope). Both the NN training data that can be received and the quality of the training data scales exponentially.

The important 'edge cases' still have to be created manually, but once created I believe a 'campaign' of training data collection can probably be launched again with much lower costs and overhead - and the results will thus scale up with fleet size.

I.e. the edge cases that they are specifically training the networks for will semi-automatically scale up with fleet size.

2)

Another part where the exponential fleet helps is the disengagement summaries. While it doesn't result in direct training data such as images, it sure comes with GPS coordinates and other metrics that allow Tesla to identify unknown edge cases or weird spots on the road network that Tesla vehicles encounter.

I.e. if the fleet's size increases exponentially, so does its statistical sampling ability and its 'sensitivity' for edge cases increase exponentially as well.

3)

Finally, an exponentially growing fleet will drive new FSD revenue as well. Large parts of Autopilot are in reality prototype FSD functionality, packaged in driver assistance and convenience features. These do sell the cars, they sell upgrades - and all of that income stream scales with fleet size.

So even if some of the NN training capacity doesn't scale up with fleet size (we still have only one Andrej Karpathy today, as we had two years ago ), there's a fleet-proportional revenue stream plus increasing take-rates of the FSD option that will finance that part.

To add to this, the more a driver uses AP, the more it will use it.
Not sure if there are differences in quality, for Tesla, gethering data during AP or in shadow mode.
If there is, then you have a positive feedback and more data.

Furthermore, more global sales means much more edge cases, which is very good for a necessary varied dataset.