EP 186: Mobile Bottlenecks, Skyrocketing AI CapEx, and the Compute Crunch

0:05Ben Bajarin: Hello, everyone. Welcome to another episode of The Circuit. I am Ben Bajarin. Hello, world. I'm Jay Goldberg. Well, we had another week, a week that felt like, I don't know, a year between news cycles, market cycles, earning cycles. I don't even know. But we're going to start and we're going to go over.

0:29Jay Goldberg: So many earnings. So many earnings. So many earnings this week.

0:33Ben Bajarin: As I was joking with Jay before the podcast, we have a backlog very similar to the hyperscaler backlog. We've got tons of demand for content and only a little bit amount of time slash gigawatts to, in my analogy, to cover. We're bandwidth limited. We're gigawatt and bandwidth constraint. So we're going to cover this in themes. We're going to group the earnings names we're going to talk to into themes. So our first theme is going to be mobile. And the three names we're looking at that for, even though two of them also want to talk about data center, but we're going to focus on mobile, is Arm, Qualcomm, and Apple, with a little bit of media tech thrown in, who also reported last night, but thematically.

1:20Ben Bajarin: Okay, so your top-line take, what did it tell you? I'll tell you my top lane take. And I think there's some very interesting supply chain nuggets in this theme.

1:31Jay Goldberg: So not surprising to anyone, the mobile market is weak. There's lots of debate about how the trajectory of it is. Is things getting a little bit better? Are things getting worse? Mobile market is soft. I don't think it's as bad as the worst case scenarios. We're having it down 20%. But it's definitely down. And now we're starting to ask questions about what demand looks like for mobile phones next year as well, because that looks like there's a possibility it will be down next year as well because memory costs are going to remain elevated. So a rough time for mobile, but some conflicting commentary about where that weakness is and what's going on among these companies.

2:17Ben Bajarin: Yeah, I think there's two central debates. One, are we at the bottom and starting to creep back upwards? Or two, is this a lasting cycle where two interesting things are compounding? One, prices are being increased. Okay, so we know that. And that's across everybody. That's not just Apple. So when people are like, oh, Apple's going to have to increase prices. It's going to hurt their sales. Like the conclusion is not that other people gain. The conclusion is that everybody is going to hit by this and thus turn off consumers across the board. So there is that dynamic. There's another dynamic where it's just tough in the supply chain, as I was saying.

3:00Ben Bajarin: I think it was very, very interesting that if you just look at this leading signal, which is Apple, who has the most scale of anyone who makes, let's say, premium-ish mobile phones. And if they are behind in the supply chain trying to catch up, good luck everybody else. Like, you are all dealing with a very serious supply chain problem if Apple is sitting there, right, trying to figure out how to make enough devices, which again, I think is interesting, right? The commentary is not necessarily nobody wants to buy phones. It's that prices are going up. We can't make as many phones as we want. And the dynamic there is a lot of supply chain driven, at least in terms of Apple.

3:55Ben Bajarin: I hear mixed things about Android. I think the low-end part of the market absolutely has nobody buying. I think the medium to high tier still has people interested in buying and we're up against supply chains to just – we just can't make the same number of phones we did two or three years ago. That's just the reality of it. So there's that as well. But I'm with you in terms of where does this go. But Apple's commentary I think was super interesting because it wasn't necessarily people aren't buying. It's that we're having trouble securing as well. One, they got demand wrong, which, okay, fine.

4:28Ben Bajarin: That's on you. Two, they're signaling the inability to get the demand that they need – sorry, the supply that they need for their demand, which I think is sort of a question of could they sell more phones if they got it? I don't know the answer to that, but it sounds like they're going to be constrained. But just not being able to make as many products because of the supply chain is going to have a drag on their numbers. so so just i'll back up a second here we had apple reported uh a good quarter but a weak

5:05Jay Goldberg: guidance essentially um i think revenue was off for both quarter and and guide profitability

5:11Ben Bajarin: looked a little bit better than expectations yeah so so mix mix yeah aft mixed results from apple

5:17Jay Goldberg: um qualcomm came in line on revenue for the quarter missed on eps and guided both below expectations for next quarter and then arm beat all around um both in terms of quarter and the guidance right now arm of course has a strong data center component to it that's driving numbers so they're doing well there but i think um as sometimes happens with companies the more they talked the more they kind of freak people out uh and and the sort of the center of concern around arm is just like is mobile because mobile is still a pretty significant portion of their revenue

6:02and and then apple apple i talked about and then yeah that's that's the third qualcomm apple ARM.

6:11Jay Goldberg: Yeah, we got all three. And I think what's... ARM in general just didn't have super positive things to say about mobile. I think there's a difference between what they were saying versus Qualcomm. Qualcomm seemed to be indicating that it was going to get worse from here. There was an ambiguous question in the Q &A where somebody asked about next year, and it kind of made it sound like they thought next year would be down further. Arm didn't get asked that question, didn't answer it, but they seemed to be of the take. A lot of their questions were, is mobile doing really badly? How long is it going to do badly for?

6:57And their response would be, data center.

7:00Jay Goldberg: Data center is doing really well. Really trying to sort of move the light away from them. Apple, I thought, was the most interesting in that, Like you said, I think they're supply constrained, especially I think what they're trying to do is warn a lot about next year, about the next cycle of phones. I would be somewhat surprised if the problem really is Apple can't get the supply it can. Maybe they're big enough that it's a problem for them. But by the same token, they're big enough that they should be able to get whatever supply they need. I think they're just trying to sort of dampen expectations for next year a little bit to make things a little bit easier for them down the road.

7:50Jay Goldberg: But Apple certainly is struggling, I should say, around pricing, around demand. There's clearly some signs of softness there around iPhone shipments. I think what starts to get confusing is you start to dig into where is the weakness. On their call, Qualcomm said that what they're saying is consumers are moving towards lower-priced phones, low-tier Android, which is bad for Qualcomm because they don't compete as heavily as well, and that's not their strength in cheap Android phones.

8:30Jay Goldberg: but but even you see apple doing really well with pricing and margins i think there there's you know apple's not losing share um and so and then there's this question about like what is china demand actually look like uh like the china smartphone numbers i've seen coming out aren't terrible, right? There's decent demand in China for phones. And I think there's reason to be cautiously optimistic about like how some of the new phones, some of the new iPhones or Huawei has new phones coming out later this year. I think those high-end models will do fine. And I wonder going to Apple's commentary on supply constraints, are we going to have a situation in which they're going to launch a new phone and it's going to be sold out, lead times of three weeks, six weeks, long lead times for their flagship phones.

9:24Jay Goldberg: Because they always have a little bit of that launch until they get really ramped up. And I wonder how bad it's going to be this cycle. Also, I wonder personally, because I'm planning to buy a new iPhone, and so I'm going to have to buy it early.

9:37Ben Bajarin: Yeah. I mean, I think there's just so much embedded into this cycle that's fascinating. Because, you know, again, I come back to the supply chain, and Apple did a couple of things. they hinted at the, on their call, right? They basically came out and said, we wish there was more than three memory providers. We're exploring all options. They said that they are having trouble at the leading edge, getting foundry capacity, which then everybody was like, here's the door for Intel, which is a very fair point and true. But I think, again, when you look at Apple's scale and they are undoubtedly having problems getting some things, let alone getting them at the price that they want.

10:19Ben Bajarin: Like we talked about this before the irony that you, in fact, I said this, like I would, I would die on this hill as a matter of fact. And I also don't think it's that bold of a statement that Apple probably never ever anticipated a day where they would not have leverage on their supply chain. Like never. I think they ever saw this coming. And so now all of a sudden, oh, oh, hey, remember when you like really pressed us hard for widgets and for memory? We got a line of people willing to pay premiums. Are you willing to pay premiums? In fact, they dodged the LTA because the reality here is you have to be willing to sign an LTA at some reasonable price agreement, which again is duration, not necessarily price.

11:08Ben Bajarin: You're agreeing to capacity. Your price can go up or down within that's floor and ceiling. And, and I like Apple dodged that question. Like who's, who's going to pay the premium? Apple could probably, you're exactly right. Get what they want if they're willing to pay more. And I'm not sure they're willing to do that. So it's sort of like, okay, then somebody else absorbs that capacity. But, but, but this is, again, I just, I come back to the fundamental point that what's going on in AI compute is wreaking havoc on every other section of the industry. AI compute up to the right. A good contrast of this, like just to throw them in here, is MediaTek.

11:44Ben Bajarin: Nobody cared about MediaTek's mobile numbers. Everybody cared that management guided significant increase because of TPU wins. They didn't say TPU wins, but significant increase in data center demand because of their ASIC business. Like that's all anybody cared about because they're about to just make a ton of money in the custom ASIC side. And so they're getting a tailwind, mobile was crap versus others that are analyzed as highly mobile, which yes, for right now is ARM and Qualcomm and to a degree Apple. So anyway, I just think it's fascinating how much the supply chain is wreaking havoc on other industries, benefiting one that we'll talk about, right?

12:28Ben Bajarin: In AI with CapEx, But hurting a whole lot of other people because, again, this is the constraint of all constraints out there and ability to truly just get wafers and manufacture things.

12:44Jay Goldberg: I thought there was an interesting somebody asked

12:52hey you're so

12:53Jay Goldberg: everybody's so supply constrained at Foundry have you considered adding another Foundry partner and their answer was we already have two why would we need three and everyone's like wait wait wait

13:08Ben Bajarin: who's the second one right but that could be Samsung

13:12Jay Goldberg: I know they're working with Intel.

13:14Ben Bajarin: I know it is, but you're right. But I do think, I mean, yes. Yeah, anyway, Samsung is what it is. But, yes. But, I mean, same thing. This MediaTek came out, was very clear. Intel's a huge beneficiary of eMib for them. I think, again, everybody, the cat will be out of the bag. It's inevitable that Apple uses Intel. It's just a matter of when and for which trips. It might not be trailing edge. It might be lagging fine. But this just underscores you have to have more supply chain flexibility. You need more people in the game. And this tightness is not going to solve for the next few years. But it does open up the opportunity for a whole slew of other people in the supply chain because, like I said, AI is just wreaking havoc across the board.

14:07Yeah.

14:07Jay Goldberg: Yeah, there are also two interesting press stories that came out related to this around. One, we've known for a while, it's been reported in the press, that Apple is trying to lobby the U.S. government to get a waiver to buy Chinese memory. There was a report this week that now Micron is lobbying the U.S. government to not let Apple do that. Right. And so, OK, not surprising, but interesting. Interesting that it came out. Interesting that it got leaked. the other story was that now tsmc is ramping is exploring i should say a new packaging technology to compete with intel's emib which is like you know i don't know how much substance we put into that the difference between co-ops and emib is is pretty arcane um but the but just just the idea that for the first time in a long time, we have a headline saying Intel is the head of TSMC on some technical aspect of manufacturing.

15:09Yeah.

15:11Jay Goldberg: That was, I think we'll have more of those in the future.

15:15Ben Bajarin: No, I agree. But again, this is just, Intel has great stuff. I'm not going to say that. They have very good technology. The quality of their technology is what makes them a viable second source. but this is happening because we just don't have enough capacity anywhere for anyone and anyone who has it is going to benefit like any anything viable is going to benefit because we honestly like i saw that news too and so like copos so copos is fairly similar to emib t so it's not again like we're talking drastic differences here like tsmc has it it's just they can't make enough for their customers and other foundries will benefit, right?

15:57Ben Bajarin: Just pure and simple. The Apple stuff on Chinese memory, I agree. And I really don't think the government's going to want that to happen. I don't think Micron's leverage or lobbying at all makes any difference. I don't think Micron, I mean, I don't think the US government is going to endorse that in any way, shape, or form. And to be honest with you, if things keep going the way they're going with these model app things, it kind of feels like they're going to crack down even harder on China across the board. And that's going to hurt US companies trying to use Chinese memory, not Chinese companies using Chinese memory.

16:30Ben Bajarin: But I think that's the argument in favor of it. I think that's a pretty compelling story there.

16:41Jay Goldberg: Ultimately, we're asking a political question, which we are not equipped to answer because it's not going to be decided on a rational, commercial basis. It's going to be sort of politics at the moment. 100 i so i i who knows who knows right but but the need is is immense um and ymcc has capacity and cxmt has capacity well cxmc doesn't have capacity but they're adding a ton very quickly

17:09Ben Bajarin: so yeah yeah yeah no and everybody i mean we're tracking that i do think it will be interesting to see who does get those wins because if they again predominantly skew toward Chinese OEMs who I think are going to get really hit hard in this mobile crunch um that demand for Chinese memory has to go beyond mobile and that part with the talk we're having I'm not sure how that shakes out So we'll see. We'll see. Okay. Yeah. Mobile. Interesting. Stay tuned for that. I don't think the worst is over, but I don't know if we have quite enough to say that upside is coming other than Apple will probably sell a whole lot of phones in the fall because they usually do at any price that they come, including a very expensive foldable that will sell like hotcakes.

18:05Jay Goldberg: um okay two two things that will tie us into the next uh on data center one is apple signal they're increasing capex not huge but like they were they were they're clearly going to spend more on capex heavy implication that that's going to go towards ai servers for some purpose their own chip yeah right yeah we we've known for a long time that they're working on their own server chip for presumably inference. That hasn't come to light yet, which means I think they're trying to figure, they probably have, trying like everybody else to figure out what the right architecture is. But if sort of the new Siri AI stuff they're working on with Google really starts to take off, we can assume they're going to be inference bound just like everybody else.

18:59Jay Goldberg: The other thing too is, I thought it was really interesting to compare Qualcomm and ARM's data center commentary.

19:08Jay Goldberg: Qualcomm has a custom ASIC they're doing this year for one of the hyperscalers, maybe more. Lots of questions about what custom really means there. It feels like the word custom is doing a lot of work. It's a lot of IP from the customer. Their own merchant accelerator and their own emergent CPU are coming next year and the year after, respectively. And they got asked, like, you know, what's the status of this? You launched this AI stuff a month and change ago. And they said, now we're trying to get, we're talking to customers about potentially doing evaluations, right? Implying that they don't have silicon ready to actually do those because it hasn't been taped out yet.

19:55Jay Goldberg: But also, it was a very loose answer. It didn't sound like people were knocking on the doors. In contrast to Arm, who had a launch event a few months ago, and they went out of their way to repeatedly say, we're seeing new customers come on board. We have new customers signed up. We're raising our prices. So very, very tangible progress for them on their AI chip, which again is coming out later this year, but they have a whole family coming out. And so very, very tangible, solid traction on their part. Lots of interest for their AGI chip in the data center because it's a CPU and those are, you know, agentec CPUs are a big deal right now.

20:44Ben Bajarin: Yeah. Yeah, agree. I mean, again, right, with everybody in that space, even ARM, I thought it was interesting that some questions sort of circled around, like, any chance you could get any more capacity to up that number? Like, just the recognition that there is demand for so much additional stuff. Yeah.

21:04Jay Goldberg: I mean, that was a lot of the sort of questions that ARM got was, like, your data center's business is doing really well. Do you have capacity for that? Exactly. And they have a lot now that the CFO was repeatedly saying, oh, that's my job. That's my number one task right now is finding more capacity. Yeah. I mean, it's not there, Jay.

21:28Ben Bajarin: It's not there. There is no capacity. Like, you guys. I just don't think anybody truly – I mean, I'm not sure. I'm not saying. I'm overgeneralizing anybody. I don't think the masses just truly understand how big of a problem this is. in the supply chain. Like it is, it is a nightmare. Um, I feel bad for everybody and there's no way, there's no way around it. Like we need a good analogy. I don't know. I don't know what it is. Um, but so, okay. So tie that CapEx again, not enough capacity, not enough gigawatts, lots of money, increased backlog of the hyperscalers. Yes. They're going to keep spending because they've got a massive line of customers waiting to give them money and no place to, and there's no room at the end.

22:15So I don't know.

22:16Ben Bajarin: What was your, what was your, I mean, let's throw, we talked about Google, but let's just lump everybody into the same thing. Hyperscaler commentary. What did we take away? Are there still concerns in the market about monetization? I think there are, but I thought Andy Jassy was like pretty, pretty like strong on this. And to be honest with you, like if anybody knows how to invest CapEx and turn around and turn that into money, it's Amazon. It's not like this is their first rodeo.

22:45Jay Goldberg: Yeah. So the headline is both Amazon and Microsoft reported good numbers. I don't have all of them in front of me, but good numbers. And for us, what matters most is they said all the right things about CapEx. They said there's – I think all four of the reporters now, they've all either said that they were going to the high end of their range or they were increasing it a little bit. And all of them repeated again and again, they're very, very, it was very constrained on compute. And they haven't given explicit guidance for CapEx for next year, but they're telegraphing that the numbers will be, CapEx numbers will be going up, will be very strong next year, which is what the semi-complex kind of needed, wanted to hear.

23:37Jay Goldberg: But I think there's a lot of variation now among those four for how the results did. Certainly, you can compare Microsoft and Meta. Meta, there's lots of questions now about how much capacity they have. Did they overbuild? What's this whole thing about them selling excess capacity or leasing excess capacity to somebody else? I don't think they answered that to anyone's satisfaction on the call. There's lots of questions about their monetization of AI. Microsoft, I think, did the opposite. They made it sound like they were selling what they could. Their results looked pretty good. And I pair those two because we've had this consistent pattern for the last three quarters where one of them does well.

24:26Jay Goldberg: Both of them raise CapEx. One of them does well. and people say, oh, yay, CapEx is good. And the other one does poorly and CapEx is punished, right? And we've just been alternating for the last quarter. It was Meta did well. Everyone liked their numbers. And so it was happy for their spend increase. Microsoft was the opposite. It's going back and forth. It's kind of, which I think is kind of silly. Like it's, we just don't know. We cannot, it's very hard to measure ROI on AI based on quarterly results. And I think it's, you can read into it what you want. If you want to be bearish, you can look at meta.

25:08Jay Goldberg: If you want to be bullish, you can look at Amazon. I think of all of the hyperscalers, Amazon is the most rational right now. Amazon has the clearest case for spending because they are spending to hold on to AWS. They make a ton of money on AWS. And if they're not careful, they can lose it in AI and they're not going to let that happen. And so for them, it makes lots of sense to be spending at these levels. For Meta, we talked about this a lot. Meta is an advertising company, and they don't seem to recognize that. And so they're going to have to sort themselves out. And, you know, Microsoft is what you make of it.

25:49Jay Goldberg: I don't think their story is totally clear, but they seem to be doing okay.

25:53Ben Bajarin: Yeah. Yeah, I think, I mean, you're right, right? My whole thesis for Amazon was like, if this company can just get its AI compute strategy under control, they're well positioned because Jassy pointed this out on the call. And this was my thesis all along, which was that the cloud, which is AWS, that was the market share leader by a landslide for general purpose CSP cloud workflows, meaning all vast majority of enterprises who brought their workloads to the cloud, brought them to AWS. us. Why would you want to separate those workflows in an era of agentic AI to another cloud? Like keep them on the same platform with the same software stitching.

26:32Ben Bajarin: If your email is there and your databases is there and whatnot, like bring your agentic right stuff to that. It makes a ton of sense, but, but they were behind in compute. Now I do think this is an interesting point that, that I haven't seen enough people pick up on, but, but I want to make it. So, you know, in a training world. And somebody might argue with me on this, but, but I would contend that Microsoft benefited a lot in the training cycle because they were oriented well for training. Um, Amazon, I think is oriented well for inference. Now that doesn't mean that Microsoft is all of a sudden not going to be good for inference.

27:10Ben Bajarin: I'm not saying that, but that if you look at the way they architected themselves. The inference cloud, I think Amazon is very, very, very well positioned for this. But in an inference world, you, the customer, right, are not deciding where that infrastructure runs, right? Kind of like we talked about the Anthropic stuff with AMD last week. So in training, yes, you go, I need to train. It's going to be on NVIDIA. And I think NVIDIA has won that for the foreseeable future. That's not going to change. But inference is more like, hey, I got my data, I got my databases, I just need to run Agentic AI on this, I'm bringing the workload, here's the token price I want to pay, or whatever, and then the hyperscaler, or the cloud provider, decides where that goes.

27:57Ben Bajarin: The takeaway for me then is compute is not the decision of where to run commute is much more heterogeneous, and so therefore more transparent to the end customer, what the hyperscaler is going to then do is decide which one for which workload gets me the best margins. Does that mean I run it on NVIDIA or my own custom silicon, etc.? But that's going to be their decision. So their infrastructure choices will be highly, highly tied to what they know their inference margins are for workloads and the abundance of workloads they believe they will get in average from those customers. And this is how I think you now start thinking about how do they monetize this CapEx?

28:52How do they start to return an investment on compute, which Andy Jassy said is under three years.

28:58Ben Bajarin: And our plug for our gigawattonomics model confirmed that. I can also tell you his assumptions and how much it will take to pay back a PowerShell now that we know he's confirmed it's under three years. But the bottom line is that they orchestrate that workload for margins. And so I say all that to say, like, does that change the way we think about infrastructure? So, like, yes, NVIDIA might still win on absolute TCO.

29:32Ben Bajarin: But whose volume mix shows up in a hyperscaler when margins is really the central scale platform as to where they bring and push those workloads to make the most money, the most margin per token?

29:49Jay Goldberg: I think that's a very good point is as we move towards inference, we've known for a long time that inference is going to be much more heterogeneous than training was. that opens the door for NVIDIA share to be different, right? They're basically 100 % share of training, so it has to go down because there's going to be other people in the mix. And I think we've been saying this for years now, is the question will become how much does it go down, right? If it's 70%, that's one outcome. If it's 30%, 40%, that's a very, very different outlook for NVIDIA.

30:32Jay Goldberg: and we just we don't know the answer to that because there's lots of there's lots of reasons to say that nvidia still has certain a lot of advantages here um but it sort of explains why they're spending so much on rack level system level designs because they're trying to eke out every every little uh ounce of gain possible right and remember all the hyperscalers have these massive Excel models for total cost of ownership for TCO that goes out to three decimal places, and they look at every last little thing inside the data center that they can tweak, arguably, NVIDIA has an advantage there in that they can tweak everything to get the most gain.

31:15Jay Goldberg: The hyperscalers can probably construct their own level systems that are competitive there, but it's much tougher. And I think if you sort of reflect back on what AMD was talking about last week, where they have very, very compelling performance for Instinct,

31:36Jay Goldberg: I think that certainly makes NVIDIA's job harder to move that way. Yeah.

31:43Ben Bajarin: Yeah, and I think -

31:44Jay Goldberg: To say nothing of custom.

31:45Ben Bajarin: No, agreed. And that's why I think like, once you start to change this, and obviously everybody's been pushing this language for some time, right? Who has the highest TCO? NVIDIA is like, you know, betting the farm on this, right? That they are the best TCO, the best tokens per watt within a power budget. And so my point is that that can be true. And if it is true, then you will continue to be adopted at scale as long as what you are giving the hyperscalers is the best margin on those tokens. And so as you look for this over the next two years, it's really like the proof's in the pudding, The proof's in the volume.

32:21Ben Bajarin: If you are continually a NVIDIA sold and scaled at volume in those who I think are going to be the most margin sensitive, which is the hyperscalers, then you have the best TCO. It's really going to come down to proofs in the pudding and they will deploy and scale what they think is best for their margins. And I just think that's a much more interesting way to look at this going forward because the economic proof is there. And that's what NVIDIA is saying. And so if they're right, then they'll continue to be installed, deployed at scale.

33:00Ben Bajarin: Okay, are we missing anything on the hyperscalers?

33:05Jay Goldberg: On the hyperscalers, data center, we get Extron report, very strong numbers. They're a semi-cap equipment maker that's very important for certain parts of data center, AI data center, had incredible numbers. I only mentioned that because that's like sort of one of the first obscure stocks that people start looking at when they're looking at supply chain constraints. This is one of the first ones that always comes up. You can get a lot more obscure than this, but it's already pretty obscure. They had very strong numbers. Bloom Energy, very controversial stock that makes fuel cells for behind-the-meter data centers.

33:51Jay Goldberg: That stock was getting just beat up last month. We'll come to why. But it's now reversed itself on some very, very positive, very bullish commentary on their call. Yeah. Yeah. Full disclosure, I own Bloom Energy in my personal account, but I don't know why. I'm not an expert on the energy system, so don't take that as investment advice. Okay.

34:20Ben Bajarin: So let me brain dump on this because I think this is actually super interesting. So backtrack. Around six to eight months ago, we were looking very heavily into what would it take to do behind the meter. Like what are the problems with behind the meter? And one of them that kind of stood out to me through a host of conversations with our friendlies in power and utilities and in electrical was training workloads are very bursty. So often a training workload can send upwards of 30 % voltage to your system and behind the meter itself alone cannot absorb that, can't normalize that, and needs redundancies, whether that be batteries or a grid connection in order to help balance that out.

35:02Ben Bajarin: um so so largely like my sort of working going into this which was why i thought that grace blackwell was so interesting when when uh nvidia talked about the thermal modeling that they essentially could say we could keep it under 25 and i had heard 25 was the kind of bar you were looking for that if it could stay spiky under that then but then behind the meter was sustainable i'll go forward i could do 100 behind the meter um okay now inference though is not bursty Inference is very smooth. It's very normal. And so inference data centers are going to be much easier to go behind the meter than a whole host of other solutions out there, right?

35:42Ben Bajarin: But where I think this is interesting is two things. So one, their call, I thought, was just absolutely brilliant because the CEO basically was like, any question that came up, he's like, we've solved it. It's answered. Moved on. Next question. And he's like, the tone of confidence. we've got all the hyperscalers now. Nine months ago, we only had one. I don't know. There's no problem with scandium. There's enough scandium for everything we need on the gigawatt on the planet. Next question. Like it was just bullishness after bullishness. And they will, again, continue to make a whole host of power slash solutions, right?

36:20Ben Bajarin: And they just can't need enough. But I say all of that to say what makes Behind the Meter interesting. And again, and it's not just bloom it's it's anybody that's doing this natural gas is behind the meter there's lots of natural gas available on tons of sites you can go get modular data centers as long as you're not trying to build a gigawatt you want to do anything under a couple hundred megawatts you can build these you can get gas turbines you can do behind the meter actually accelerates like for behind the meter to be viable accelerates the ability to go get watts and deploy compute because building the shell and getting the land is hard, but it's not the hardest part of this equation, right?

37:00Ben Bajarin: Connection to the grid is the hardest part to this equation. But that, again, doesn't mean that the grid goes away. What I think that is interesting is it creates that bridge. So go deploy a couple hundred megawatts wherever you want, start behind the meter, and then you'll eventually add the grid and now you have more capacity, right, into that. And so it gives them a time to power, which is sort of the point, right? And in fact, it gives a whole lot of other people the ability to go time to power, not just your NeoClouds or your hyperscalers. The names that I'm interested with are just those that want to build shells and manage power.

37:33Ben Bajarin: Like they can go do these things and get behind the meter. And I think more hyperscalers will continue to partner with those names that are just PowerShell landlords not trying to sell GPUs. Behind the meter just brings out a lot of opportunity to bring power on faster, is my point. Now, again, you're up against how much you can do, but the equation has changed here once BTM becomes viable. And even everybody we've talked to in utilities is like it's much more viable now than it was six months ago and will become more so over the next 12 to 18 months.

38:06Jay Goldberg: Yeah, I mean I don't know enough to comment on Bloom specifically, but more broadly, I think you're right. The power options are growing. I think there's a good likelihood that just timing-wise, things get a little tight this year. But I don't think it's an unsolvable problem. It just needs – I think that supply chain is adapting very rapidly.

38:33Jay Goldberg: And we probably have some points next year where we just can't get enough. But a quarter later, six months later, everything arrives and we turn things on. But it's really – it's fun to watch just how quickly that sector, which hasn't really changed a lot in our lifetimes, is now advancing so rapidly and all kinds of new stuff coming out.

39:00Ben Bajarin: Yep, yep. With the big constraint, I don't want to leave anybody with not – that we don't understand this, is labor. everybody we talk to in the electrical trade is pretty much ripping their hair out because they don't have enough labor they have so much demand to do power and electrical and they just they don't have the people um and that that to me you don't just print new humans like this is a huge problem that i think you have to really really understand that that labor if anything is the bottleneck of bottlenecks in power and utility that I don't know how we get around. And it's getting worse.

39:44Jay Goldberg: Hey kids, thinking about your career, have you considered being an electrician?

39:50Ben Bajarin: Yep. Generational wealth awaits you once you get out of apprenticeship school where you will do labor for practically free and then make a lot of money. But again, I think within the vein of this sort of overall CapEx trend,

40:14Ben Bajarin: everybody keeps emphasizing, right? We don't have enough compute. We don't have enough power. We don't have enough supply chain to make the compute that's going to run on the power that's going to serve the AI. we just don't have enough of anything and you know economics aside that that dynamic right remains that there's such demand for this workload and uh and we got to just keep racing um and we'll be up in constraints right in in the time frame uh and that's just that's just the dynamic so So CapEx again to the moon for like the seventh time that we've talked about this. Did you read anything into the depreciation, particularly of Microsoft, to data centers too longer?

41:05Ben Bajarin: Some people were like, this is a terrible sign. Fine. It's a shell. I get it. But I saw some controversy over that.

41:14Jay Goldberg: So I think that was the other sort of – that's where we started this week actually before earnings. We had two very sizable debt deals come to market for data centers. One for OpenAI, backed by SoftBank. The other, that one's in Texas, Sopapia. And the other one is in Ohio, which is, what am I getting them backwards? There's another one out there too, which is backed by NVIDIA. Massive, massive amounts of debt going into this. And I think that's starting to spook people, right? Because we now have reports from the journal saying that NVIDIA is talking about supplying close to$600 billion in various forms of debt commitments or financing commitments to support this data center.

42:13Jay Goldberg: That is in addition to their NeoCloud partnership financing vehicle that they announced a couple weeks ago. These are just big numbers, right? These are big numbers that NVIDIA is throwing around. And I think that is at least getting the sort of fixed income investors to ask tougher questions because they don't really care about AI. They care about their coupon getting paid. Then you had the Sopopia deal in Texas, which is very similar to the Beignet deal for Meta over in Louisiana that was signed in October, I think. So just about six months ago. Very, very similar deal, very similar structure as far as we know, but they're paying almost a point more in interest.

43:07Jay Goldberg: So if you have similar deals and one is a higher interest rate, that sort of says, seems to be signaling that the market is getting a little tougher. Now, again, there's a lot of devil in the details. We don't know all the terms of the deal. It's not that big a difference. So maybe it's just explained by things we don't see. But it is a serious question. They have two very similar properties. Why is one peg a higher interest rate? That implies a worse credit risk. So that's a little bit of a concern. And then you have Microsoft, who is reclassifying their operating leases into financing leases.

43:44Jay Goldberg: This is a subject I hated in MBA accounting class. But basically, this allows them to have less of their balance sheet, less damage, less impact on their balance sheet to the capex that they're bringing on. and um i i think what like i mean at one level it's an accounting difference like who cares but at another i think it's it's very telling that we now have both meta and microsoft and to some degree google all sort of going out of their way to keep debt off their balance sheet um right they still have it but it's not on the balance sheet and like in meta's case we know they're paying a point a full point higher because of the structures they're putting in place to keep that off their balance sheet and it just it they come across as very self-conscious right like they're they're a little embarrassed and so they don't want to show it the full impact on their balance sheet we all know it's out there but it's i don't know it's not a great signal it's not a great signal i don't know i don't want to read too much into it but like clearly the financing teams at these companies are thinking long and hard about what this looks like a few years down the road as opposed to nvidia who's just like who seems content on spending i don't know we'll see they're not keeping it on they don't have it on their balance sheet either but they they're if they really deploy 600 billion dollars that's gonna have to show up somewhere Yeah.

45:27Yeah.

45:28Ben Bajarin: Agreed. All right. So lastly, let's just talk about the wild market swings that happened because apparently it's explainable now. I mean, you and I talked about this. We didn't talk about it on the pod publicly, but it seemed like it was a little bit abnormal, like something more than just a revaluation correction, right? Because fundamentals didn't change. So it seems like we have some insights now as to at least some answers as to what exactly happened, particularly with Leopold's fold and situation awareness. So what was the take? What's everybody landing on what happened?

46:11Jay Goldberg: So, all right, this is a very contentious subject. I'm going to be a little careful how to explain it. There was a very large hedge fund by a new fund manager. Leo Offenbender, Leopold, whose fund is called Situational Awareness. And I think he was one of the early OpenAI employees. I forget where he was. But he was very, very knowledgeable, deep AI roots. And he started a hedge fund to benefit from that. And he had an incredible run. He started with his own money. And it went up. I think overall his fund was up at its peak like 2 ,200 % in just a few years. Just an incredible run. He called a lot of names that no one had heard of before, like Bloom Energy and a bunch of others.

47:00Jay Goldberg: Just had an incredible run. But he's also taken on a lot of outside funds. And this happens with hedge funds a lot as they get so big. He apparently used some very advanced trading strategies, which is essentially a form of debt, complicated option stuff, but he levered up pretty heavily.

47:27Jay Goldberg: And through a series of trades, a bunch of trades appear to have gone against him. And in particular, the fear that interest rates were going to go up, that sort of catalyzed a liquidity crunch for the fund, and then they had to sell out. And so it looks a lot like the run of the markets in sort of April, May, June, where things were just going up every day after day after day across everything

48:03may have in part been explained by situational awareness taking on those levered trades. And then the recent downturn we've seen for the last month of everything selling off for no apparent reason may similarly be tied to unwinding of that fund.

48:24Jay Goldberg: That's oversimplifying it. And again, I want to be careful here. We don't know all the details.

48:31Jay Goldberg: but it's it is it is very possible in a market that is you know for a small number of players to have an outsized impact um and that's some of that going on and and we saw like after he he sold his fund to citadel uh the next day a lot of the stocks that he had sort of most favored were up very, very strongly. And there's press reports to tie those two together. I don't know. I think everything's much more complicated than this. But to me, it sort of highlights the thing that I took away from this. It highlights that debt is dangerous. Debt is dangerous because it's on a clock. If you borrow money, you have to pay interest.

49:18Jay Goldberg: You have to pay it back at a specific specific time. You buy a stock and it doesn't go up right now, you can hold on to it longer. The company doesn't have to do anything. But when there's debt, you have to pay that. And I think that's really, really important on a week when we have a few hundred billion dollars of debt coming to market. When times are good, debt amps things up. It's pro-cyclical. But when times are bad, it's pro-cyclical in a bad way. And it makes things much more painful and much worse. And I think that's another reason that sort of people are taking – it's another sign that people are a little bit like scratching their heads, asking tougher questions.

50:02Jay Goldberg: I don't know which way the market's going to go. I'm just saying that these are real – these are things we have to consider. Yeah.

50:11Ben Bajarin: Yeah, and I think you have to grasp the point which you say, which is that a small number of players can have an outside impact. And I think that dynamic doesn't go away. I think it was also compounded by, you know, there's a lot of hefty retailers out there who can see trading volume early and chase that volume. And while they're not going to cause that swing by themselves, they can pack onto it when it happens. And I think a lot of that went into it as well. but what I'm interested in is and one, two, I've had a lot of conversations with friendlies in finance across the board and many people land at the same place that you did this was an unwinding, people saw it it explains a lot of what happened but now it's interesting, it feels like and I've had a good same group of finance I've had conversations with that as they're starting to you know again pile back in and buy they are um doing so with a different set of conviction on who the players are going to be going forward like making not that not that the thesis has changed a bit but that some of the bets have changed than before and so as they get back into the market this is uh certain themes they're kind of like we've got deeper conviction now than we did a year ago and so i'm curious to see where this lands is that re-rotation comes back comes back in

51:42yeah

51:43Jay Goldberg: I don't know the answer to that but I

51:51yeah

51:52Jay Goldberg: I don't know what's going to happen next I'm confused because I see both sides of this I'm very conflicted I think

52:07Jay Goldberg: I see the same things data you do I read your stuff all the time of course and I hear the same things everything is sold out everywhere across the semiconductor supply chain everything is sold out and it's sold out into next year possibly the year beyond and for some things probably even further out that's very bullish fundamentals um what what that means for the stocks we can we can that's a separate debate is all the upside priced in again that's separate debate but i see all those fundamentals at the same time like there's so many like warning signs out there about like there are certain areas which are clearly in bubble bubble territory and i i you know there are lots of things out there that make me cautious.

53:01Jay Goldberg: Personally, I think things are going to get even more confusing, because if we really have hundreds of billions of dollars of debt, that's just going to turbocharge all of this. If there is a bubble, it can get even bubblier, even crazier. But at the same time, there's a very reasonable argument you can make that, no, we're still in rational territory because demand is so strong. And the hyperscale doesn't keep spending. And I think at some point this year, we'll get much better insight into Anthropic and OpenAI's user numbers. Certainly, all the reports I've seen, those are still moving incredibly.

53:49Jay Goldberg: Their numbers are moving up into the right very sharply. So there's a lot of demand and demand out there. um it's a it's a pretty it's pretty precarious times right we're sort of dependent on hyper sailors continuing to spend and the frontier labs continuing to see just surging usage growth um but you know things i mean i think i i think the at the end of the day like it's it's hard

54:20Ben Bajarin: it's hard to not still see that there's a structural dynamic here. I get capital can get spooked. I get capital can not believe it. None of that changes the demand profile. So I understand that piece. And again, if it's just a continued, you know, another three to four months of just really helping people understand why it's structural, how it's structural, we just have to go through that cycle. Um, but I mean, no, there's no sign that we're overbuilding anything at this point, right? Every sign is the opposite. And so you just take that, take that with what you will. And, uh, you know, we, we keep riding.

55:06So the, the world is

55:10Ben Bajarin: short compute remains the same. I did, I mean, again, I did think it was interesting. Like I get that people are like, oh, management can sometimes mislead, but like Jassy to say, you know, that they think that AWS cloud, like cloud business could be a trillion dollars. Like they have to see how much enterprises are going to spend on this, how much is also fueling the flywheel of value in the, in, in that, in, in a company who has either a public product or just internally. Like that's a pretty bold comment. And if you just do the numbers of what does this need to be in annual software spend to pay for itself, it's not that much more than a trillion dollars.

55:47Ben Bajarin: Like, cause again, it just, it just compounds over time. So I don't know. I mean, anyway, like I said, I just buttoned this up. Like Amazon of anybody gets like has so much history in the cloud business i just feel like if they're confident like they have to know something like they see something yes the other clouds do too but like amazon amazon knows this and their pursuit of this should tell you everything about what you know the structural cycle of this of this space so anyway that's my closing rant i mean i i take your point i just think like i said at the earlier on amazon is the most rational player in the hyperscaler market

56:37Jay Goldberg: right now because i'll you you have amazon i agree with you what they know what they're doing and they seem to be spending appropriately set against meta who doesn't check all those boxes um agree and is now and right and like is is you know talking about leasing out capacity that that one they gotta they gotta they gotta explain that one better because it's spooking people

57:02Ben Bajarin: no i'm with you but but but again i think and and maybe this is just not going to help the thesis for them like we are not in a situation right now where anything really relative to consumer is going to benefit AI. Like even your recommendation engine, like fine. The entire reason that we focus so much on studying the enterprise customer and the enterprise as the customer who adopts this and diffuses it and is a capital source is because that's exactly how I think this plays out. They are the first person at scale who will spend real dollars on AI annually because of the business value. And so therefore, they're the customer who comes to open AI and Anthropic and NWS and GCP and Azure to run their workloads.

57:54Ben Bajarin: This is enterprise is the funding source. So in this thesis, anybody who's well positioned to capitalize on this enterprise build out cycle is the early winners. And that's not meta unless meta sells that infrastructure to somebody else. I just don't see the consumer side here really anything like I see the enterprise cycle where the amount of enterprise demand that we hit, like how much they're about to run and how short they are currently on compute, it's just staggering. And they will be the ones that roll this out. And so Amazon is well positioned for that. Again, so are the others. I don't think Meta has a very clear strategy on how to benefit on an early cycle, which is the enterprise is your adopter fueling a cycle, not the consumer.

58:48Jay Goldberg: I think that would be a little alarming to me if I couldn't explain it away to some pretty idiosyncratic factors about meta in particular. Like, I actually think there is a pretty compelling consumer-centered opportunity for meta in particular, for companies in general, but for meta in particular around AI usage for consumer products. And the consumer doesn't necessarily know that it's AI, but there is, like, they, meta could do this, right? Medica do this. They just, for their own reasons, haven't quite figured it out yet. And just to debate this, I get what you're saying. Enterprises are doing a lot with AI, but we've yet to see a company really come out and say, hey, we had this massive, massive change in our business because of AI.

59:51Jay Goldberg: I'm talking about outside of semis.

59:57Lots of companies are doing lots of things that are chipping away around the edges,

1:00:04Jay Goldberg: but we haven't seen somebody come out and say, hey, our earnings are up 20 % because we spent on AI. I'm not saying it can't happen, but we just haven't seen that yet. I think we need more of those fruit points to keep it going, to keep it solid.

1:00:19Ben Bajarin: Yeah. I mean, you do hear pockets. I would agree with you. They wouldn't take it to earnings, but you're seeing very widely publicized productivity gains. So take that for what it is. Your people are more efficient, are doing better. Anything that, again, what we see is we see more than enough evidence that they will continue to spend more per employee in a dollar amount in their enterprise. And then to your meta point, though. It's like, okay, let's just, let's just play this out. Let's say an enterprise for the average of their organization to spending, you know, hundreds to thousands of dollars per employee for AI for the productivity gains.

1:00:53Ben Bajarin: A human consumer is not adding a thousand dollars per one person to meta in, in AI value dollars. Like the economic, it might, is it adding$5 for meta? Okay. Even then, sure. You have, what do they have? Billions of people. Great. Five times billions equals X number. Does that cover your CapEx? My point is the dollar numbers per person in enterprise to fund this is different than the small dollars that you get, but you get larger scale in consumer. So I'm not disagreeing with you, but I'm just saying if the tip of the iceberg is a whole lot of money per employee, that's easier economics than.

1:01:34Jay Goldberg: Yeah.

1:01:35Ben Bajarin: But at what return?

1:01:37Jay Goldberg: right I think that's we probably just get real along but like the debate I think the debate there is enterprises are spending all this money they're going to need a return and we haven't really seen that because remember we're dealing at like the amount of money we're talking about is economy wide it's like GDP scale productivity improvements we need to see to justify the AI spend and I don't think we're seeing that in the enterprise

1:02:04Ben Bajarin: yet um but how would you but how would you go about then saying like okay what's their product what's their rationale on returns to spending you know fifteen hundred dollars per person on a computer

1:02:18Jay Goldberg: right or services costs or sap right well so here's the thing i'm not saying it can't happen right i'm just saying we haven't seen the proof points yet i'm perfectly willing to believe there are lots of enterprises who are sitting on big big gains because of ai or can make it happen We just haven't seen that in their numbers yet. And so either it's sort of in development and we'll see that at some point, hopefully soon, next year, or they're just cannibalizing non-AI IT spend. And so we're spending$1 ,000 per employee on tokens. We're going to cut our SAP budget. I mean, that's exactly what happened with IBM.

1:03:00Jay Goldberg: That's likely, yes. I agree with you.

1:03:02Ben Bajarin: It comes from other sources. It comes from other sources. Versus revenue or profit, yeah.

1:03:07Jay Goldberg: And if that's the case, that's not a net gain for the economy.

1:03:12Jay Goldberg: And I think the spend on AI is so big that you sort of have to see a broader acceleration in corporate earnings as a result of AI. And we just haven't seen that yet. And I'm not going to get into it, but I do think there is a case to be made for a consumer to be much more compelling. but that's up to Apple and Meta and a couple others to sort of lead the way. Maybe Google some too. Thank Google. Yeah.

1:03:40Ben Bajarin: All right. Thanks for listening, everybody. We will talk to you next week. I am going to be at Flash Memory Summit to have a handful of meetings there. We can talk about that and I'm sure more interesting things will happen because it does every single week. So thanks for listening. We'll talk to you next week. Thank you for listening, everybody. See you next time.