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

[techmaven]

One thing I will note Craig Irwin the Wedbush analyst said, “I have the data, right. $240 per KWh is what they’re paying for sells out of Japan. The costs at the Gigafactory is higher.”

Ok, so is Craig Irwin lying? Is he misinformed? Or is Tesla lying? How can he just present this as a fact if it’s not true? Should Tesla IR be contacted and perhaps they can correct this false information?

Several of us have looked at METI's report used by Irwin for his analysis. His problem is that he didn't properly pull out the power cells that are also exported from Japan. Tesla uses high energy cells from Panasonic that have a very high specific energy density and low cost per kWh. Hybrids and plug-in hybrids typically use high power cells, where the cells can generate a lot of power per unit of weight or unit of volume. These type of cells are much higher cost per kWh. Once one pulls out the high power cells, then the cost of the high energy cells makes sense. The METI information for 2018 corresponds to about 12.4 GWh of cell production. They have only one category for battery cells for automobiles, so this covers pouch, prismatic, and cylindrical cells. Panasonic has multiple battery plants in Japan and you can see that in their investor reports. It helps to know that Panasonic supplies Toyota with automotive battery cells for their hybrid and plug-in hybrid vehicles that use lithium ion cells. Those are high power density cells.

The value of the cells is listed in yen and at today's conversion rate is $2.81 billion. That's $224/kWh and the source of Irwin's misunderstood figure. We can estimate that Tesla bought about 8.5 GWh of Panasonic's cell output. That leaves 3.9 GWh of cells that didn't go to Tesla. If Tesla was buying the cells at $135/kWh, then the value of the rest of the 3.9 GWh would be $410/kWh. That's about right for top quality power cells.

Here's also some more info from @vgrinshpun :
Vladimir Grinshpun on Twitter

Furthermore, the cost of the cells at GF1 are certainly lower. They have much higher production rates and they have utilized many additional strategies for reducing cost including the amount of energy required to make the cells. The cathode oven, for example, has a 80% reduction in energy usage. That energy is one of the large components of the cost of the manufacturing part of the cost of a cell.

 

[Pras]

I have another wild theory. Tesla will release the numbers exactly at the same hour:min of 500k tweet. To show the difference between a material news and a non material.

(Again, don’t take this theory any seriously)

 

[Singer3000]

I have another wild theory. Tesla will release the numbers exactly at the same hour of 500k tweet. To show the difference between a material news and a non material.

(Again, don’t take this theory any seriously)

What time was that then? Do I have time to make myself a cup of tea first?

 

[SteveG3]

Several of us have looked at the METI's report used by Irwin for his analysis. His problem is that he didn't properly pull out the power cells that are also exported from Japan. Tesla uses high energy cells from Panasonic that have a very high specific energy density and low cost per kWh. Hybrids and plug-in hybrids typically use high power cells, where the cells can generate a lot of power per unit of weight or unit of volume. These type of cells are much higher cost per kWh. Once one pulls out the high power cells, then the cost of the high energy cells makes sense. The METI information for 2018 corresponds to about 12.4 GWh of cell production. They have only one category for battery cells for automobiles, so this covers pouch, prismatic, and cylindrical cells. Panasonic has multiple battery plants in Japan and you can see that in their investor reports. It helps to know that Panasonic supplies Toyota with automotive battery cells for their hybrid and plug-in hybrid vehicles that use lithium ion cells. Those are high power density cells.

The value of the cells is listed in yen and at today's conversion rate is $2.81 billion. That's $224/kWh and the source of Irwin's misunderstood figure. We can estimate that Tesla bought about 8.5 GWh of Panasonic's cell output. That leaves 3.9 GWh of cells that didn't go to Tesla. If Tesla was buying the cells at $135/kWh, then the value of the rest of the 3.9 GWh would be $410/kWh. That's about right for top quality power cells.

Here's also some more info from @vgrinshpun :
Vladimir Grinshpun on Twitter

Furthermore, the cost of the cells at GF1 are certainly lower. They have much higher production rates and they have utilized many additional strategies for reducing cost including the amount of energy required to make the cells. The cathode oven, for example, has a 80% reduction in energy usage. That energy is one of the large components of the cost of making the cell.

Very helpful. Tasha Keeney of Ark didn't seemed caught off guard re rebutting Irwin's claim. Maybe we could try to get this analysis to her and/or Kathy Woods of Ark via social media.

 

[madodel]

I’m 10000% certain he’s not intentionally mocking her. I also think she’s not very bright, and there’s a high probability of user error.

I believe they accidentally pushed the production/delivery numbers to her car and blanked the screen to prevent her from telling the SEC. Otherwise the SEC will require every OTA to be monitored by a lawyer.

 

[1kEE]

What time was that then? Do I have time to make myself a cup of tea first?

4:15 pst if I remember correct.... not quite 4:20 lol

 

[KarenRei]

FactChecking's "after the hearing" hypothesis looks more plausible by the hour

 

[MP3Mike]

How does one get invited to the "Autonomy Investor Day?" I'm an investor, will I get an invitation or do I have to register somewhere?

Probably the same way you get invited to ask questions on the quarterly investors call.

 

[Unpilot]

Since I am checking this thread for news and I am on my phone to do it ..I am wearing a groove in the center of my screen from all the refreshing

 

[Pras]

What time was that then? Do I have time to make myself a cup of tea first?

7.15 est

 

[ZeApelido]

You need to read Tyler Hamilton's book, "The Secret Race." There is literally not a single top cyclist at the world level who is not doping.

Actually I did read it. And unfortunately others weren't put on a pedestal as an American hero like that dude.

Also way OT, but with expertise in exercise physiology, it's known that dopers don't all respond similarly to performance enhancers. So if all of them were not doping, Lance likely wouldn't have won.

 

[Doggydogworld]

Not exact cost, but he has said they are narrowing in on $100/kWh and that doesn’t jive at all with Craig Irwin saying unquestionably that he has the data and they are at $240/kWh from Japan and higher at GF1.

I think that's a pack level cost, but it's hard to say. CATL's numbers are consistent with $200+, but that includes a lot of bus batteries that are more like $500/kWh Powerpacks than car batteries.

I don't see $200+/kWh at the cell level in Panasonic's numbers. $150, maybe, give or take. Of course Tesla pays some of the true cell cost out of other pockets, e.g. building costs, so their $150 or whatever isn't directly comparable to a true arm's length cell cost.​

That would put M3 LR pack at > $18,000. Ludicrous

​

Well, something was responsible for Model 3 COGS being $46k in Q4 instead of ~36k. It wasn't fixed costs or quasi-fixed costs like vehicle production labor, those are only a few thousand per car now. Module/pack assembly COGS were inflated by their 'semi-automated' process, but again that can't be more than a grand or so. They were getting full volume breaks on parts and materials by Q4.Warranty reserve was down below $2k.

What else could have pushed COGS so high?

 

[BornToFly]

I think the webcast on April 19th on the Autonomous driving advancements could be a good SP catalyst. If it is really impressive, it should wake up Wallstreet to the fact that TSLA should be given a significant value to its self driving similar to Waymo, and that means a SP more than double what it is now.

 

[Antares Nebula]

I have another wild theory. Tesla will release the numbers exactly at the same hour:min of 500k tweet. To show the difference between a material news and a non material.

(Again, don’t take this theory any seriously)

Ha, ha... interesting. That tweet was on a Tuesday, so ideally they would have done it yesterday. I think the time was 7:15 eastern

At least we're not going to see any CNBC TV coverage today on the report. Bloomberg maybe.

Curiouser and curiouser.

 

[bdy0627]

I think the webcast on April 19th on the Autonomous driving advancements could be a good SP catalyst. If it is really impressive, it should wake up Wallstreet to the fact that TSLA should be given a significant value to its self driving similar to Waymo, and that means a SP more than double what it is now.

I would definitely bet on a nice climb for a few days leading up to. Shorts aren't going to want the exposure and longs won't want to miss the potential climb. Watch out for selling on the news though. Maybe not if it's really extraordinary. Good reason to buy any dip from the deliveries report.

 

[uselesslogin]

FactChecking's "after the hearing" hypothesis looks more plausible by the hour

The only thing I don't get with that theory is could we really expect any kind of ruling tomorrow? I would think she would have to take some time with any ruling even if her mind is already made up one way or another.

 

[oneday]

I think that's a pack level cost, but it's hard to say. CATL's numbers are consistent with $200+, but that includes a lot of bus batteries that are more like $500/kWh Powerpacks than car batteries.

I don't see $200+/kWh at the cell level in Panasonic's numbers. $150, maybe, give or take. Of course Tesla pays some of the true cell cost out of other pockets, e.g. building costs, so their $150 or whatever isn't directly comparable to a true arm's length cell cost.​

​

Well, something was responsible for Model 3 COGS being $46k in Q4 instead of ~36k. It wasn't fixed costs or quasi-fixed costs like vehicle production labor, those are only a few thousand per car now. Module/pack assembly COGS were inflated by their 'semi-automated' process, but again that can't be more than a grand or so. They were getting full volume breaks on parts and materials by Q4.Warranty reserve was down below $2k.

What else could have pushed COGS so high?

I should have wrote M3 LR pack > $18k for just the cells. As to COGS, I’m no expert on any of that information so I really have nothing for you. Others, hopefully can chime in or there might be a model somewhere. I was only pointing out that $18k for just the cells with nothing else is absolutely ludicrous.

 

[MarcusMaximus]

They've mapped the world. But perhaps "the world is not enough"?

Accurate labeling and variety matters a lot more than sheer quantity.

​

Near worthless? You're overfitting your argument to fit your thesis


So if Tesla changes camera vendors, or makes something like a pickup truck with different camera locations, they have to throw it all out and start from scratch? C'mon. Image NNs aren't nearly that sensor-dependent.


One trivially simple approach.:
- Have 4 boxes they know have road signs, 4 that don't and 1 that's unknown
- Pass you if you get the 8 known boxes correct
- Use your answer on the 9th plus the answers from 20 other people to train their NN

A few things here:

1. No, the world is not enough. Because they’ve mapped it *once*(ok, probably a couple times, but relatively few per location). The data needed is how to handle driving in different conditions/events, not just where objects were at some particular points in time.

2. Yes, near worthless. I’ll go into why next.

3. Kind of. SGD(stochastic gradient descent, the algorithm underpinning most deep learning techniques) follows the shortest path it can find to reducing error rate on the data it has seen. If all the data it’s seen is from one particular sensor in one particular orientation, it will happily rely on any idiosyncrasies of that to produce a model that works very well for that particular configuration and falls apart completely for anything else. Since all google’s streetview data is gathered from a very particular sensor, any network you train where that is the sole/large majority source of the data is going to learn the world through that lens(pun intended). You can get around that problem through a couple strategies:

a. Use a variety of data from different configurations
b. Have the implementing hardware exactly match what was used for training
c. A mix of a and b; having a small set of configurations and training with equal weight given to each.

Tesla likely uses c. This would explain why both the Model 3 and AP2.5 S/X models shipped with autopilot deactivated initially. Google, at best, could try to use a, if they wanted to use streetview data(though I doubt they would; the Waymo data they have is far more useful), making it a small portion of their dataset mix.

 

[EVNow]

The only thing I don't get with that theory is could we really expect any kind of ruling tomorrow? I would think she would have to take some time with any ruling even if her mind is already made up one way or another.

Not the ruling - oral argument and any questions the judge asks can be used to interpret what the outcome might be.

 

[KarenRei]

Not the ruling - oral argument and any questions the judge asks can be used to interpret what the outcome might be.

Exactly. It denies the SEC a chance to try to pull a "Well, they produced X vehicles this quarter, and if you multiply that by four, it's not 'around 500k'..." argument during the hearing.