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

[Fact Checking]

I think the worries about production are mostly gone now. Deliveries, too. I think the real remaining issue is profitability. Everyone expects that there is a big demand for the Model 3 SR. Musk has already indicated that Tesla can't make that car at a decent profit, so high deliveries just mean more red ink. Having many cars in the pipeline to Europe won't help, even though people will understand it.

Note that the SR+ is +$2,500 over the SR, a big chunk of which is profit margin - and SR+ take rate appears to be very high. (It's a really good deal as well.) 75% of Model 3 buyers also select a color other than black, which adds another $1,500. That's a ~$3,000 cash margin improvement over the $35k entry model margin straight away, just in the SR/SR+ space.

But there's another, very important point that I think many have missed: my guess is that a big reason why Elon was noncommittal about SR margins is the significant fixed-size depreciation and amortization overhead.

The background is that Tesla just built the new Standard Range battery pack assembly line at the Gigafactory, a Grohmann Machine the size of a football field. That was probably significant capex cash expense, and once they start selling products they'll have to start amortizing the line, and they'll probably do it on the usual straight line basis.

This necessarily means that the initial GAAP margins on SR and even SR+ will be awful: if the new line has cost them $200m and they are amortizing it over 10 quarters then the straight-line A&D cost is around $20m per quarter. If they deliver 2,000 units this quarter then that's a per unit GAAP overhead of $10,000 (!), which results in significantly negative margins. They'll have to get up to the ~2-3k units/week, 25k-40k units/quarter volume to reduce this GAAP expense to a tolerable $500-$800 range. That is not going to happen until the summer.

(Maybe @brian45011, @ReflexFunds and @schonelucht can confirm.)

So what I'd concentrate on for the SR and SR+ in Q1 financial metrics is the cash margin, with warranty reserves and other high probability future costs imputed as a cash expense, and I think Elon already confirmed that even the base $35k sales are generating cash for Tesla. That's what matters as long as Tesla is cash flow positive - scaling up SR volume this year will solve the GAAP space margins and profit metrics.

 

[MABO]

Yeah, but note that even the CoGs argument isn't really accurate: Tesla's ASIC is 20 times faster than their Nvidia GP106 based board. The proper CoGs comparison would be either a lot more GP106's (not feasible for power envelope reasons), or one of Nvidia's next generation GPUs - which adds a zero to the price tag...

So this is significant CoG savings if we compare it to the price of a platform able to provide th computing power required by the FSD features Tesla has planned.

Respectfully, I don't think that is a very practical option for Tesla and therefore not a good comparison. When we're training/testing large models at work, we go to great lengths to avoid splitting the models between GPUs. It can absolutely be done, but it's inefficient, even in a desktop/datacenter environment. In an embedded environment you're much better off having everything in one chip and the memory for the weights as close to the computation as possible (compute-in-memory being the best).

IMHO, if they didn't design their own NN engine, they would likely have moved to a (single) larger and newer Nvidia chip. Remember, Nvidia isn't sitting still either.

Regardless, I suspect that we can agree that short term costs (both ATL with IC/package/test and BTL with added R&D teams) are far outweighed by the benefit.

 

[jerry33]

What is the benefit to shorting at all ? That we don't have stock market bubbles … ? Obviously we do ...

The fact is - Wall St needs much stricter regulation. Which means things like shorting should be illegal.

But, but, wouldn't that cut into Wall Street's commissions? Isn't the whole point of the stock market, as it's practiced today, to enrich Wall Street?

 

[Carl Raymond]

Note that the SR+ is +$2,500 over the SR, a big chunk of which is profit margin - and SR+ take rate appears to be very high. (It's a really good deal as well.) 75% of Model 3 buyers also select a color other than black, which adds another $1,500. That's a ~$3,000 cash margin improvement over the $35k entry model margin straight away, just in the SR/SR+ space.

But there's another, very important point that I think many have missed: my guess is that a big reason why Elon was noncommittal about SR margins is the significant fixed-size depreciation and amortization overhead.

The background is that Tesla just built the new Standard Range battery pack assembly line at the Gigafactory, a Grohmann Machine the size of a football field. That was probably significant capex cash expense, and once they start selling products they'll have to start amortizing the line, and they'll probably do it on the usual straight line basis.

This necessarily means that the initial GAAP margins on SR and even SR+ will be awful: if the new line has cost them $200m and they are amortizing it over 10 quarters then the straight-line A&D cost is around $20m per quarter. If they deliver 2,000 units this quarter then that's a per unit GAAP overhead of $10,000 (!), which results in significantly negative margins. They'll have to get up to the ~2-3k units/week, 25k-40k units/quarter volume to reduce this GAAP expense to a tolerable $500-$800 range. That is not going to happen until the summer.

(Maybe @brian45011, @ReflexFunds and @schonelucht can confirm.)

So what I'd concentrate on for the SR and SR+ in Q1 financial metrics is the cash margin, and I think Elon already confirmed that even the base $35k sales are generating cash for Tesla. That's what matters as long as Tesla is cash flow positive - scaling up SR volume this year will solve the GAAP space margins and profit metrics.

But they own Grohmann, right. That provides a lot of flexibility in how the accounting is done. They only have to attribute the hardware and the new subfloor as capex. Grohmann’s costs, all the labour and software from employees, is already baked in. $200m sounds like a lot of hardware. Grohmann seem to achieve stuff with simple actuators and customised off the shelf components rather than throwing top of the line robots at a problem.

 

[Fact Checking]

Respectfully, I don't think that is a very practical option for Tesla and therefore not a good comparison. When we're training/testing large models at work, we go to great lengths to avoid splitting the models between GPUs. It can absolutely be done, but it's inefficient, even in a desktop/datacenter environment. In an embedded environment you're much better off having everything in one chip and the memory for the weights as close to the computation as possible (compute-in-memory being the best).

Training != inference computing.

Tesla's AI chip is basically a neural network inference computing co-processor. (It probably cannot even be used for training.)

Inference computing is a massively parallel problem that can be distributed among GPUs just fine: the easiest to implement model would be to have two identical boards for example on the Tesla Semi Truck which has 20 cameras, each board processing 10 camera video feeds at full frame rate.

In principle they could also do multi-GPU boards, but I'd advise against that: they should just use 2-3 Model 3 computing boards as-is and improve economies of scale.

 

[mongo]

This necessarily means that the initial GAAP margins on SR and even SR+ will be awful: if the new line has cost them $200m and they are expensing it over 10 quarters then the straight-line A&D cost is around $20m per quarter.

10 quarters? that is only 2.5 years, the minimum length I see in my quick Google is 3 years/ 12 quarters. With some things extending to 15 years. Learn About Straight-Line Depreciation Method on Income Statements

MACRS lists manufacturing equipment for automobiles (37.11) at 12 years. Per MACRS Asset Life table
At 10 years/ 40 quarters, that drops the amount to $5 million per quarter, or $1,000 per car at a 5k/ quarter 500/wk rate SR/SR+ rate.

 

[Nakata]

Last push of quarter has started in the Netherlands.

Yesterday 163 deliveries.
https://opendata.rdw.nl/Voertuigen/Tesla-Model-3-delivery-NL/76t3-9ba8

On the dutch forum is a lot of activity for picking up cars and home deliveries.
Belgium is same story.

 

[Fact Checking]

At 10 years/ 40 quarters, that drops the amount to $5 million per quarter, or $1,000 per car at a 5k/ quarter 500/wk rate SR/SR+ rate.

We can attempt to calculate the average depreciation schedule, in Q4 Tesla had depreciation of $500m, while property, pant and equipment was 11.3b on the balance sheet. This suggests average D&A count of ~22 quarters, but buildings would probably have higher amortization schedules: 15 to 30 years, which draws the average up.

Anyway, 10, 15 or even 20 quarter schedules doesn't really matter to the outcome of the basic math: only hundreds of SR units were announced so far, so even if we generously assume a very fast ramp-up to say 1000-2000 units in Q1, even 20-quarter $10m of amortization overhead splits up only between these units, with $5k-$10k per unit D&A expenses. Even with 2k-3k units (which I don't see happening) the GAAP margin would still be negative.

But this doesn't really matter, this is normal ramp-up math when equipment built for 3k/week rate is only running at 10% its planned capacity. It just explains why Elon was reluctant to give SR margin figures.

 

[MABO]

Training != inference computing.

Tesla's AI chip is basically a neural network inference computing co-processor. (It probably cannot even be used for training.)

Agree. But that makes the splitting of the model even harder to take. Rapid (low-latency) inference requirements are precisely what make splitting the model across GPUs problematic. It can be done, but I feel it is an engineering choice that Tesla would have been unlikely to make, even if they didn't do their own accelerator.

 

[mongo]

We can attempt to calculate the average depreciation schedule, in Q4 Tesla had depreciation of $500m, while property, pant and equipment was 11.3b on the balance sheet. This suggests average D&A count of ~22 quarters, but buildings would probably have higher amortization schedules: 15 to 30 years, which draws the average up.

Anyway, 10, 15 or even 20 quarter schedules doesn't really matter to the outcome of the basic math: only hundreds of SR units were announced so far, so even if we generously assume a very fast ramp-up to say 1000-2000 units in Q1, even 20-quarter $10m of amortization overhead splits up only between these units, with $5k-$10k per unit D&A expenses. Even with 2k-3k units (which I don't see happening) the GAAP margin would still be negative.

But this doesn't really matter, this is normal ramp-up math when equipment built for 3k/week rate is only running at 10% its planned capacity.

Or they can calculate depreciation based on unit production life. You can't do this for taxes (IRS requires MACRS or such), but you can use it for your accounting.
How the Unit of Production Method Works
Units of Production Depreciation: How to Calculate & Formula
How do I compute the units of production method of depreciation? | AccountingCoach

 

[Tslynk67]

best headline I've seen so far today...

 

[KarenRei]

best headline I've seen so far today...

Money? Meh, who wants that. Annoying stuff.

 

[mongo]

A 40% rise puts us back at $360 or so, quite apt to go full circle...

Off topic:

Agree. But that makes the splitting of the model even harder to take. Rapid (low-latency) inference requirements are precisely what make splitting the model across GPUs problematic. It can be done, but I feel it is an engineering choice that Tesla would have been unlikely to make, even if they didn't do their own accelerator.

Easy split: two chips running the same NN in lock step for safety.
Harder split: two (or more) chips running different layers of the NN. Data transfer between them would be bottle neck, but if the first level of chips handles the image processes/ categorization from the cameras (x cameras per chip), the amount of data passed from them to the driving logic chip would be less. (assuming # of neurons per layer decreases as you traverse the NN)

 

[KarenRei]

A 40% rise puts us back at $360 or so, quite apt to go full circle...

Off topic:


Easy split: two chips running the same NN in lock step for safety.
Harder split: two (or more) chips running different layers of the NN. Data transfer between them would be bottle neck, but if the first level of chips handles the image processes/ categorization from the cameras (x cameras per chip), the amount of data passed from them to the driving logic chip would be less. (assuming # of neurons per layer decreases as you traverse the NN)

If they do split the NN up into two separate chips, may I suggest naming them "Wintermute" and "Neuromancer", and their onboard controller "3Jane"?

 

[dmvevguy]

I would love to see a 40% rise in stock price today! Tesla is seriously executing and I can't keep up with the latest software updates, V3 updates, FSD, China GF 24x7, it's just moving forward greatly! Triple bottom bullish pattern, just hope we have great delivery numbers and that "tiny" profit for this Q1!

 

[Tslynk67]

Tesla should put out a “German Edition” that reduces the front fascia by 5 cm.

To be fair, whole EU would prefer that mod...

 

[KarenRei]

To be fair, whole EU would prefer that mod...

Would hinder aero (=range, charge time on trips, etc) and safety (you'd have to significantly shorten the four-tube-profile alumium segments between the bumper and the frame), including not just passenger safety, but also making low-speed accidents more likely to damage the car.

But Tesla could make it easy to re-upgrade. Maybe even just have the normal front-end be a "free optional upgrade" that can be selected at purchase time, if the wording of the law would allow that.

 

[Tslynk67]

OT, but especially for @Curt Renz - European MP's just voted to approve abolishing daylight-saving. No more clock changes as of 2021!

Now we just need to ensure they stick to winter time and not summer - humans need morning light and dark at night for better circadian rhythms.

Euro MPs vote to end summer time clock changes

 

[Tslynk67]

Nice - now let's get to 5k for the end of the month!

 

[lklundin]

Or they could just issue this guidance:

"The future is uncertain, there's only one thing certain about it: us trying to predict anything will only result in lawsuits.

But we are confident in guiding the following about Tesla's performance this year:

1. Tesla is in the business of making more cars, and this year we'll probably make between zero and 1 billion cars.
2. We have several factories under construction and have more planned, which might or might not be built by the end of this millennium.
3. Not all of them will be on Earth.
   
4. There are a lot of idiots in politics who make our life harder, and their actions might or might not affect demand for our products and our expenses of making them.
5. As things change so will we - we are adaptable.
6. Don't buy our stock. Really.
7. If you did buy our stock, see point 6 and shut up.
   
8. So long and thanks for all the fish. Tesla out."

... and then they could just concentrate on execution, cash generation and reporting the results, quarter after quarter.

Well, if your TMC posts and Elon Musk's reply to Judge Nathan mean anything, then Tesla's next Earnings Call will be quite a sensation.