EP 188: The Wild West of Optical, Cisco's AI Hack, and Is Compute an Investable Asset

0:05Ben Bajarin: Hello, everyone. Welcome to another episode of The Circuit. I am Ben Bajarin.

0:11Jay Goldberg: Greetings, programs. I'm Jay Goldberg.

0:14Ben Bajarin: Well, we, as I think any long-term listener of this podcast has noted, we start talking a lot more about networking the last few years. We have joked and jested since day one. You've seen, all heard, who's been here from the beginning, Ben's evolution from networking's stupid, I don't want to talk about it, it's boring, to now it's the great interconnected networking and transceivers is the glue that makes AI interesting. But in that vein, a lot has been going on around this optical conversation, which we haven't spent a ton talking about optics. We have both written about it, and I know you and I have both independently researched this quite a bit.

1:03Ben Bajarin: But it's time to double-click, I think, a little bit more on networking. And we're going to lead into that through both Coherent and Lumentum's earnings this week that happened. Different tones, obviously different reactions. So I'll let you start with the both of those, and then we'll dig into some of the implications and observations around the networking and optical environment. Sure.

1:30Jay Goldberg: Lumentum and Coherent both reported this week. I don't have the exact numbers in front of me. Lumentum beat across the board, guided up. coherent was a revenue guide was good but uh struggled with margins and a few other things and there is a big divergence in their stock prices this week uh following those results and i think i think it's interesting because both of them are participating fully in in this uh ai build out, optical build out. It's big. We've talked about it a lot. So they have good revenue, but there is not all things are the equal. There's a lot of devil in the details of what they're actually selling.

2:15Jay Goldberg: And that is conveyed through their margin profiles. Yeah.

2:21Ben Bajarin: I mean, I think the, correct me if I'm wrong, I think most people were pushing on kind of their gross margins. You know, there was some interesting conversation on both of their calls about what was going on around both two inch wafers all you know all the way back to two inch wafers um four inch as well i think you know lumentum was pretty was pretty clear they were going to be able to use some of those uh ip from those two to also go into npo which i think made made people happy um but it's like the industry's trap you know moving into six inch wafers and six inch wafers is sort of the yield right and and and sure going to six inch is going to help but you've still got to you know squeeze everything you can out of your out of your four and even your

3:05Jay Goldberg: two inch right from what lumentum said right right so semiconductor economics 101 the the size of your wafer matters a lot the more wafers the more chips the bigger the wafer the more chips you can fit on it and so there's been a long history of moving towards 300 millimeter wafers for digital logic we're seeing this transition now to six inch wafers for optical components and um lumentum is says hey we've got four inch and we're doing really well with that and coherent uh sort of i guess perceiving that they were behind is pushing really hard into six inch wafers which theoretically should offer much much better gross margins because again for the same wafer you can get a lot more chips out of it but that's not actually playing out and there's lots of lots of debate about why that is exactly and i think the the thing to keep in mind is that six inch indium phosphate wafers indium phosphate wafers are are brand new it's it is not a mature ecosystem it's and these things take years to develop and so from you got to change the the tools and a lot of the fixtures within the tools and the the wafers themselves there's not actually a good supplier of there's only one supplier of indium phosphide six inch wafers and that's and they're very new to it too so seems to be some growing pains there i mean it's not you know this is normal stuff it's just like it would be better for coherent if this were a few years more advanced

4:40Ben Bajarin: 100 and and i think you know what there's there's two things i i keep coming back to um you know i I have been very vocal on the Twittersphere and others, and we will have guests in the upcoming months that can help us talk about the timing, adoption rate, attach of optical. But there's two things that I think are going on, which is leading to people to be a little bit too optimistic in the attach rate of all of this optical stuff. Again, outside of LPO, which is sort of the standard today, right? It's a giant plug. You go and stick it in and it's at the edge of the board. It's not into the board.

5:24Ben Bajarin: And that's what most people are just starting to ramp like today. Like your standard optical is LPO. So we're not even at NPO yet. And NPO is like the next version that takes this onto the board, but it is still not like into the die, which is the goal of CPO. and then past that is scale in, which is onto the substrate.

5:49Jay Goldberg: So, yeah, so there's three acronyms we've got to dive into. Yes, and that's what I wanted to do.

5:54Ben Bajarin: We need to clear this up.

5:55Jay Goldberg: Right. So there's LPO, which is what we've been using for a long time. Plugables. Plugable. You plug in an optical module into the server. The next step is you're going to plug NPO. You're going to plug the – instead of going to the server, it will go onto the board. onto the board near packaged optics near packaged optics on the board it's still some extra components in there the holy grail if you will will be co-packaged optics cpo which is where you you attach the fiber strand directly to the chip and there's all kinds of savings and interesting things important things about that but as you can imagine it is very very very complicated to manufacturer.

6:39Ben Bajarin: The one thing I want to point out about this, and I have had numerous meetings on this, and this is going to get hot at OCP summit here. Obviously, we just had OCP Taiwan, and I followed a whole host of notes that came out around that. There is the conversation to exactly like you said, which is the benefit to CPO is obviously you squeeze out more picajoules per bit yes and that that leads to economics i do think though right now it is highly debated how much of a savings that is and is it worth that additional headache over npo which again is is in its infancy like we're not even really at cpo despite people talking about that at any kind of scale it's very very small um and that dynamic i think is one of the things that It'll have to get worked over because it's so hard to do, to do CPO.

7:35Ben Bajarin: There's a whole lot of things, not just manufacturing, but reliability, like so much stuff's got to, got to get worked out that I guess my point is if the savings is not some significant percentage more versus the engineering effort, it's kind of why I think NPO kind of lasts for a long time. And I think this is debated right about how much savings are you actually going to get per the engineering effort that's going to need to go into cpo is still a debate and i also think it's just worth pointing out this is absolutely the wild wild west of connectivity ocp taiwan just had a standards paper come out it's like 300 long light matter helped lead that along with a host of other names uh lumentum is not on that list and i do not believe coherent is either from obviously they're not in a position to necessarily want a standard at this place given their their spec in um but it's the wild wild west everybody is doing their own implementations their own custom design for optical where they're using optical whether that's ocs with google or or you know lpos moving to npos uh today it's it's the wild wild west which makes this much much more difficult, much, much more unpredictable.

8:50Ben Bajarin: Um, we don't even know which direction this takes in, in fiber cables, in, um, a whole host of other protocols. So, so there's that right when any early cycle. And so people just keep getting ahead of their skis of the optical trade is coming. These companies are going to be gigantic. It's like, this is a very slow roll. And, um, And then the other part that I noticed is, you know, there's these conversations scale up, scale out, scale across. And what's interesting is like a lot of people are using scale up when sometimes they mean scale out. And that is also confusing people because scale up in some people's nomenclature can mean vertically.

9:38So within the rack and then also rack to rack.

9:42Ben Bajarin: so you're scaling up is north to south rack to rack but also east to west the neck the rack right next to you right because it needs to be say three meters right if you want to do um or shorter right to or longer sorry to go to optical and so then some people are like oh you know look at how many connections you're going to need inter rack that's going to just boom cpo and it's like no no no Inter-rack is so far, I mean, sorry, within the same rack is so far off from adoption. Rack to rack is starting to happen. But they're assuming when you say scale up, right? You know what I mean? Like I'm totally confusing everybody.

10:20Ben Bajarin: But my point is scaling with optical within the rack is much farther off. Rack to rack is what everybody is doing right now so that you can have the throughput you need at 800 Gi to 1.6 and beyond. at more than call it four meters, which is what AECs kind of tap out at. That I think is confusing people because other vendors use scale up to mean with inside one rack and scale out to mean rack to rack. And so we have this mixing of terminologies that's confusing people, leading them to assume that you're going to have far more optical connections when you're probably not because you don't need as many to go to the rack right next to you.

11:07Ben Bajarin: anyway i'm sorry i'm sorry for confusing everybody i'm just this the nomenclature is throwing people

11:13Jay Goldberg: off you are going to be fighting this fight you're gonna be tilting at this windmill for years

11:17Ben Bajarin: this is not the first time you brought this up i know or even the second no one's ever gonna agree on this i i i know but i think it's worth knowing because if you're gonna make an assumption of attach and optical content per rack you need to you need to know if it is within the rack within the vertical of the rack, which is not happening. It is rack to rack, or it is pod to pod. So at distance, 100 meters away. Very far less optical content in both of those two latter scenarios. So you need to know what we are connecting if you're going to make an assumption of optical content attached per rack.

11:57Jay Goldberg: That's a fair point. But I'm confused by all this because we're in this bizarro world where I always think of myself as the ultimate optical bear. Like I've talked about this before. I hate covering optical because I've covered it for years and it always disappoints. But like I'm more bullish on the optical build-out than you are. I actually, I think in 2028, we will start to see certainly NPO. Yeah, NPO for sure.

12:27Ben Bajarin: Absolutely.

12:28Jay Goldberg: And some CPO to some degree.

12:31Ben Bajarin: Right. Right. But again, horizontally, not vertically within the rack. You have to have NPO, perhaps. Yes. NPO can start to happen in the rack because it has to be serviceable. The density of connections out of 500 and something plus GPU rack is so much optical. You need so many more switches in that to make that happen. So the optical content, once that happens, is enormous. And it has to be serviceable, which is what NPO comes in. I can see NPO in that rack. My point is that you've got Lumentum out there going, we're doing CPO and we're starting to scale. And then people go, oh, by CPO, you mean – or they're saying scale up.

13:13Like Lumentum is saying scale up when they mean rack to rack, not within the rack. Because – and that's confusing people.

13:21Ben Bajarin: Like you'll see them make these diagrams. Like these bots or whoever they are on Twitter is like, look at how much fiber you're going to need in one vertical rack. and I was like, no. NPO, yes, not CPO. Within the same rack.

13:36Jay Goldberg: North to south. But it's coming. It's coming. It's coming. We can debate timing.

13:41Ben Bajarin: Not CPO.

13:42Jay Goldberg: CPO's coming too. Not in rack.

13:45Ben Bajarin: Not in rack. You can't service that fiber. It's going to die. It'll just kill boards and GPUs.

13:53Jay Goldberg: I still think it's coming. Now, here's the thing. It's optical. It's going to take longer than any of us expect. Yeah. Right. It's always going to take, it's always going to come later. Um, so be warned on that. But I think, I think it's possible.

14:09Ben Bajarin: Yeah. I mean, I'm not saying it doesn't show up. I'm just saying it doesn't scale. It's not going to be the vast number of attach of compute racks. It just doesn't scale. The manufacturing bottlenecks here are too deep. Like all you have to do is dig into how hard this is to manufacture. you can very quickly come up with some decent estimates of how many connections can be made. And GPU racks, accelerator racks, GPU racks will outstrip that by magnitudes of its ability to manufacture at rack scale.

14:40Jay Goldberg: Just so everyone knows, this is not just Ben and I having this debate. We have several guests lined up to take both our sides in this.

14:52Ben Bajarin: Okay, let's move on to Cisco. in the networking trend like that's optical let's move on to Cisco Cisco which I'd sort of stopped

15:00Jay Goldberg: paying attention to years ago because it got very very boring is not boring anymore um they they reported their quarter I I don't you know the numbers were I guess disappointing they looked okay to me at my eyeball but like the street it's down it's down a few percent this week on those numbers but I I read through their transcript and I thought there were some really interesting things about it, right? For one, they almost entirely use their own silicon now. Silicon One is their internally designed silicon. They were doing custom ASICs before we even called it custom ASICs. They've been doing their own silicon for 20 years, but it's now pretty much all of their portfolio runs on their own silicon, which I thought was pretty interesting.

15:47Jay Goldberg: It's much more advanced. They also went out of their way to say repeatedly that they're doing it themselves. They'd used MediaTek for many years, but they now have all the capability internally, which is something to keep in mind when you look at other companies' custom business. But I thought they're now getting a fairly healthy degree of AI attached. And I think the knock on them for the last, I don't know, pre-AI, 10 years leading up to that, was that they really didn't have a cloud, a presence in the cloud. Hyperscalers all wanted to build their own white box switches. And so Cisco was sort of shrinking, you know, shrinking relevance in the data center.

16:28Jay Goldberg: And AI seems to have changed that. They have a really good AI attach rate across all kinds of switches. Right. And all kinds of hyperscaler wins. They talked a lot about on-prem enterprise building up their own switches, using Cisco switches to build their on-prem AI clusters. And the thing I thought was most interesting, they talked a lot about their use of Mythos. They apparently have a license from Anthropic and they allow their customers to run basically security audits on enterprise IT systems to go through Mythos, like full, no guardrails kind of Mythos, like the full Anthropic. and that is exposing all kinds of security vulnerabilities.

17:16Jay Goldberg: And I thought this was a pretty smart sales tactic for Cisco's security business is because they allow you to do this audit probably for cheap or for free. And of course, you're going to find all kinds of vulnerabilities and then Cisco would be happy to patch those for you for a fee. And I think where this is particularly interesting, you think about like Cisco stuff, there's Cisco switches everywhere. Some have been around for a very long time and they're all end of life and Cisco stopped patching them years ago. And now people are finding vulnerabilities in sort of core switches. And I think it's a very smart sales tactic on Cisco Spark because now they're able to charge for hardware that they haven't had on license in 10 years.

18:00Jay Goldberg: But it's an interesting AI use case where Anthropic and Cisco are working together. It's pretty clever.

18:10Ben Bajarin: um so i want to talk about this a second because i have heard similar use cases now popping up all over the place in some cso cyber security conversations that they're using mythos for um uh for security vulnerabilities product testing um trying to find vulnerabilities etc And I actually somewhat predicted this in my cybersecurity report where I said that the frontier labs are going to be very well positioned to work with these companies on their security and cybersecurity efforts. The question I have is, and we have no evidence of this, but I'm just throwing this out as a theory. But I have to imagine Anthropic and to the degree that OpenAEI know this, like, yes, we know they charge these large enterprise licenses.

19:01Ben Bajarin: So all of these businesses are buying a bulk license and paying them a bunch of money. But I have to imagine like that they could charge or they are charging or they should charge more for the cybersecurity efforts or a model. A model. Let's just say that they develop a new one. So Mythos is great, but they develop another one just for cybersecurity. I feel like you could charge more for that. Like that would be an economically higher license given the problem it's solving for the company. but we don't know what the economics of that look like yet.

19:33Jay Goldberg: Yeah, we don't have a good picture of Anthropics economics and their business model. You know, they filed to go public, and so hopefully soon we will get to look at all that juicy good data and learn about it more. My strong suspicion is that they are under-monetizing Mythos at this point just because it's new and they probably don't know how to charge for it. but you could absolutely see the value like in the cisco case is a good example of how how they can do this because you know anthropic story has long been that they are very focused on selling ai to the enterprise right as opposed to open ai which is doing a lot of different things including a big focus on consumer and i think the sort of um the heuristic that people have in their minds is that anthropic is doing it better than open ai i don't think that's true but I think that's the bias a lot of people have.

20:30Jay Goldberg: And because of that focus on the enterprise. And in fairness, OpenAI has a pretty good enterprise business now too. But both of them, how are they charging for these things? Where is the money actually coming from? Is it millions of us signing up for$200 a month or are they actually generating far more money from their enterprise-grade API? I'm super curious about that. We just don't know. There's just not a lot of data out there about it for obvious reasons.

20:59Ben Bajarin: But I think it makes sense, right? Logically, the point is if you have a model, and let's say that it's mythos class or beyond, and you're positioning it to enterprises as, hey, as a part of our license, this one is for your cybersecurity efforts to defend yourself against randos who can get an open model and start attacking and try to hold your business ransom. and for your own software or hardware, in this case, chip vulnerability work, that's a more valuable model than what you'd give for summarizing a Teams meeting.

21:36Jay Goldberg: Make me another PowerPoint. I mean, we haven't really touched on this, but there's obviously been this big debate going on right now about Chinese models, open source models, and I don't think we need to get into it all because it gets very complicated very quickly. But again, the sort of the understanding in lots of circles right now is that Chinese open source models are almost as good as Frontier Labs up to and including Mythos. And those there are free, I say with air quotes, free. And so, I mean, it's a headline in the Financial Times today is there's a price war going on between Anthropic and OpenAI.

22:16Jay Goldberg: I don't I don't think we know any how it is going to play out. the Chinese models may not stay open for free. The frontier labs are still moving forward pretty aggressively. But it's, it's, it's all wrapped up in that debate. And I think this is, this is one of these big areas where I just say like, we don't know how this is all going to play out yet. There's, it can go a lot of different directions. Right.

22:41Ben Bajarin: Yeah. Um, okay. So interesting stuff, networking security all those things let's talk to let's pivot to applied materials um aim at so i mean it was extremely bullish they're they're benefiting from something i've been i've been thinking about and i and i tweeted this called uh lta water falling which i think this is very interesting as it relates to kind of all we've talked about, right? That the dynamics of this industry has changed around these kind of LTAs or as SanDisk likes to call them, new business models. Micron likes to call them SCAs, whatever. These are just customers coming to them going, I'd like to secure capacity for a long period of time.

23:35Ben Bajarin: I'm either going to tell you exactly what my visibility is, or in some cases with Micron and memory and even SanDisk, prepayments are happening to secure that so that the company, Micron, maybe your favorite CPU company or whoever, can go and secure downstream their supply chain LTAs they need as well for substrates, for wafers, for foundry, etc. So you're getting this kind of like company A who's very big, right? They're the key point supplier, gets an LTA. They're like, cool, I'm going to go sign LTAs with my suppliers because I need to make sure I have the scale I need to meet the commitment I just gave to one customer.

24:13Ben Bajarin: And so you're getting to go down. So I thought it was interesting that even at the wafer equipment level, this is sort of the first time we've heard this type of language around wafer equipment from AMAT, that they have eight quarters of visibility. And what's interesting is that in prior cycles, they've only had two quarters of visibility. It was the best their customers gave them. Now they have eight. To me, this is very telling. their business is growing. Revenues were up. I think the one knock, which I have a theory on that I haven't told you yet, so you're going to hear it all fresh for the first time, is they were very hesitant to guide.

24:52Ben Bajarin: Again, last time they said, we think we're going to grow more than 30%. They're just saying now meaningfully more than our last number, but they didn't want to give away a full-year guide for next year yet. Now they have an October investor day coming, But I have a theory as to why they did not want to do that right now, and they probably will at their investor day, which I'll save after you give your take on Applied.

25:17Jay Goldberg: All right. So let me back up. Applied materials reported yesterday. This is Friday, August 14th. And I think they delivered very strong numbers. They, they, uh, revenues way above revenue is a little bit above expectations. Earnings for both the current and the guided quarter were significantly above expectations. So they're doing well and they're raising prices effectively or selling better, more expensive systems. Really good numbers, right? Objectively in absolute terms, very strong numbers is the fastest sequential growth they've ever seen in revenue growth, right? Historic, right? Right. There's a cycle.

25:57Jay Goldberg: We all know about it, and they're benefiting from it. But it wasn't enough.

26:03Ben Bajarin: It's never enough, Jay. It's never enough.

26:06Jay Goldberg: I was pulling my hair out listening to the questions on the call. They were all some variation of, yeah, good quarter, but we want more. I'm just like, dude. It was very telling. There was one question somebody asked, I won't name the analyst. He said, your customers are all talking about five-year long-term agreements. but you're only talking about three-year long-term agreements. I'm just like, 12 months ago, it was unheard of to have a one-year long-term agreement. Now we're debating three years versus five years. And so, yeah, the street wanted more. Stock is down 5 % today. Expectations were apparently higher than street-printed consensus.

26:53Jay Goldberg: So it goes. What's your theory on?

26:56Ben Bajarin: Okay. Well, two things. I'll give you my theory, but let's say, let's do this. The two parts from their call that I also think were interesting was, and this is going to lead to my theory. So one was, they said customers are finding creative ways to free up clean room space for more equipments. So that's just reorganizing, I don't know, moving a machine somewhere, decommissioning something somewhat. They're like figuring out how to say I've got this cleaner in space. How can I get a couple more, you know, a couple more tools in here in order to do that? So there was that. The other one they said was customers are – or they alluded to pulling in some construction timelines to get their fabs out sooner.

27:50Ben Bajarin: Cool. All right. So you got those two things. Also, wafer per square footage, right? So they have machines. They have a whole line of machines that gives you better wafer output per square footage. That's starting to come into play for some of their things. Cool. Okay. I think that they did not want to give this number yet because of Intel. I think Intel's capital raise just happened, right? And they have not told them how many machines they want yet over the next two years because they first needed to get the money. Then they need to go very quickly, place those orders, which is going to happen before their October day.

28:30Ben Bajarin: and I think Intel is also doing what I just said, which is going to use some of this CapEx to accelerate data center or cleanroom construction. And so I think they were waiting. That's the variable. Intel is the variable you just raised. You're going to spend this CapEx capacity. You needed this money that just happened in order to figure out how much faster your fabs can come online and two, how much machines you're going to need, and Intel has not told them yet. So that's my theory, is that Intel is the variable. Intel will 100 % do all of this before the end of October, because they have to, and then AMAT will know how much Intel is going to need.

29:16Jay Goldberg: I think that's true. I definitely, my spidey sense during their call was There was like the unspoken, the unspoken guest on the call was Intel. Like the one thing we're going to talk around and not actually name. I think you're, I think you're right. I mean, in there, I mean, another simpler alternative is that they just didn't want to guide for the year because companies don't like to guide too, too far in advance. But yeah, I think Intel is, is, is a prime, prime cause of uncertainty at this point. so okay i buy that um okay good well i'm glad you don't know anyway i i think it just makes sense

29:59Ben Bajarin: i think it's logical but but but i say all that to say it felt like even the notes i read afterwards um like that seemed to be the only real negative like yes you were like hey they expected more i agree but they wanted they felt like they were being conservative in their guide and my point and they were and my point is that's because of intel and at their october day this will all be

30:21Jay Goldberg: settled yeah and just to be clear when we were talking about conservative on their guidance they're still guiding to record record everything it's just not record that's not record enough

30:31Ben Bajarin: just you know you know to be honest with you maybe you could just briefly explain this to me so

30:40Ben Bajarin: maybe we've talked about this but like so i tweeted out i was like look i think broad consensus right now is moving wafer equipment toward 300 billion or north of 300 billion by 2030. Yes, there are some things that could move that forward. But then somebody tweeted me back and they were like, the buy side's at 300 billion for 2028. And I was like, why is the buy side kind of consistently more aggressive than the sell side? I don't understand that.

31:14so so a couple things one is i think this is a symptom of just wall street in general where

31:22Jay Goldberg: and this this applies as much to optical we were talking about before as it does to semi-cap is there is a tendency a bias on the street to expect things to happen sooner than they actually do because we're all we're all on 90 day cycles um we're all thinking 90 days and And we're like, oh, that's six quarters away. That's a long time, right? Remember, most people don't have models that go beyond two years. So Wall Street's in a hurry, right? That's the bias. As to why there's a divergence between the buy side and the sell side, I think that's because the sell side is in public and will be more likely held to account for that number, right?

32:06If I put out a number now for a prediction for a year from now, somebody's going to come and ask me a year. They said, hey, you said this a year ago.

32:15Jay Goldberg: And when I say somebody, it could easily be like I'm on CNBC and somebody asked me a question about that. A reporter asked me like, hey, didn't you predict this? I get those questions a lot. You predicted something last year. Did it come true or not? And for the most part, the buy side doesn't make public predictions. right and just the idea that there is a by-side consensus is a little a little hazy too because it's not published anywhere it's just like 10 guys in a room five guys at a bar coming up with a number so i think that's that's the disconnect there right but but okay fine but

32:58Ben Bajarin: But like I guess I just don't understand why that – I don't understand why that makes such a big deal and can actually impact a market or take it off. Like that their expectations weren't met and so it's down.

33:13Jay Goldberg: Well, because the buy side are the people with money and so what they are expecting matters to their stock purchase decisions or stock sale decisions, right?

33:27and you know that's they have expectations and they matter in the degree that if those

33:34Jay Goldberg: expectations aren't met then they're gonna change their investment decisions

33:40Ben Bajarin: got it okay all right well anyway that's that um good stuff with wafer equipment they've got like i said october investor day they've got epic coming up as a launch um i just you know the point the point i want to end on on this is like even even even with sandisk's day yesterday everybody just keeps talking about how in in this in the deep of the weeds of this supply chain all of their customer engagements have changed and i just this isn't like this is not lip service to the degree that people think it's like, this is a structurally changing dynamic of how you need to shore up your supply chain.

34:24And the best supply chains in the world are making these longer term

34:28Ben Bajarin: agreements. And so that they have the demand, like that's, that's so much more structural than, do you know what I mean? Then, then cyclical, not to say that doesn't, I just, I just think people have to grasp how important the customer dynamic changes. And that's, that's significant. And I just think that that gets missed in a lot of this analysis.

34:49Jay Goldberg: Again, I think there is a disconnect in timing. Long term means different things to different people. And I don't want to be totally skeptical about long term agreements. I think there is a qualitative difference between this and past cycles.

35:10Jay Goldberg: But we haven't tested any of those. right we'll see what this looks like in three years four years in the meantime it's great and it's going to be different than the last past cycles i agree with that a lot but it's not going to we're not ending cyclicality right we're we're changing the the curve maybe

35:30Ben Bajarin: yeah structural floors new structural floors okay all right let's uh let's move to nvidia and neoclouds um i i haven't talked to you about this so i'll be very curious on your take um i watched the cnbc roundtable with jensen and all these ceos um

35:56this this is very con this was controversial a lot was done by nvidia and others to kind of

36:04Ben Bajarin: make it less controversial. But I think there's something interesting here. I sort of disagree with one of Jensen's points, so I want to talk about that. But I want your first reaction.

36:16Jay Goldberg: Right. So the big news at the start of the week, really still dominating news cycles, is NVIDIA and six large financial institutions are putting together a platform, air quotes again, a platform to provide up to$500 billion of data center financing. And that's a big number. NVIDIA seems to be saying that they're going to effectively back 25 % of that, which is significant. But beyond that, we don't know a lot. They have not revealed many details, not least because they haven't actually finalized the agreements.

37:01Jay Goldberg: and lots can still change, right? It is interesting to me. I don't quite understand why NVIDIA did it right now. Like, why not wait a couple of weeks until everything's signed? Because like, let's face it, NVIDIA will press release before things are totally ready. And sometimes they never get there, right? That's what happened with OpenAI last year. So the deal could end up being very different. But let's just take it at face value. Nvidia is launching$500 billion of financing to build Nvidia-based data centers. And it's a big number and has people freaked out. Okay.

37:42Ben Bajarin: But what did you take of Jensen's points about, I'm trying to address the circularity by pointing out that we aren't just all selling to each other. we're selling largely to, you know, sovereigns, to nations, to enterprises, which is the point I keep making. Like when people are like, oh gosh, they're just all circling to each other. It's like, yes, but this crew of infrastructure providers, labs and model companies is gearing up to sell to 300 million plus enterprises worldwide, ex-China, I'm not including China in that number, and all of their employees as the tip of the spear of adopters. Like it's B to B to B to B to B.

38:23Ben Bajarin: like long long sets of B2Bs and so that's a very big customer base who's going to pay the open AI's of the world, the anthropologists of the world to run compute on the infrastructure that passes to NVIDIA etc etc etc so I often feel that gets discussed, Jensen just you know in passing does that but he didn't deeply dive on that but he sort of made that point so anyway I feel like he tried to do that But I also wonder if like, you know, again, part of this is just he does want that diversity of NVIDIA GPUs out there in the world from as many sources as possible so that there's just a giant pool of NVIDIA compute that can be there to absorb the coming demand of all of the enterprise.

39:13Jay Goldberg: So I'm of two minds about this or there's two sides to this. One is to some degree, this is just vendor financing. You go buy a car and you are borrowing money from Toyota or GM, whoever's selling you that car. All right. That's pretty straightforward. And NVIDIA can also argue that they're going to probably have revenue share in this and they're investing in this great hot new category. Okay. Okay, the flip side of this is, it's like I keep saying, there is a very fine line between enabling your customer to buy something, enabling demand, and then actually creating that demand where it doesn't exist.

39:54Jay Goldberg: And I'm talking about demand for compute systems. Whole other debate about what the end of all those B2s are, where the end demand is there. That's a whole different topic. I'm just saying if NVIDIA did not step in to provide financing for these deals, the deals presumably would not get done. And so there you either have to believe that one possibility is that, and this is what NVIDIA is arguing, is that the market, the capital markets, do not understand this opportunity and they're mispricing it. And so NVIDIA is going to step in and bridge that gap to fix the pricing until everyone else figures out how genius it is.

40:34the other alternative is that there is like a lot of bubbling going on,

40:40Jay Goldberg: right? And Jensen actually posted on Twitter a little bit of more about this than was in the press release. And one of the, one of the things he actually asked that question is this circular financing? And then he has a whole bunch of words, none of which are actually answering that question. And so I think there is a degree of circularity here. You borrow money from NVIDIA to buy NVIDIA products. How bad is that? There's two sides to that. But I think just the sheer size of it, I mean, half a trillion dollars is a lot of money. It has people a little bit on edge.

41:20Ben Bajarin: Yeah. So I'll say two things. I completely am on board that this infrastructure makes money, that there is an attractive return on capital. And I'm saying this broadly, not just for NVIDIA. I'm saying this for CustomASIC. So the hyperscalers and others, I get the sensitivity of the neoclowns because their business is just AI. But again, I think over the lifespan of those assets, they will make money. I guess where I'm not totally solved, and I would say I'll explain why, is the idea that compute is an investable asset, as Jensen says, because I think that necessitates that they're never good enough, that compute is never good enough.

42:15Ben Bajarin: because I feel like it means that if you're going to keep investing in compute and compute is an investable asset, it has to maintain next generation's attractiveness to put into your data center. Not that I can build it all and then it just sits there and depreciates and it's good enough and I don't need next generation, next year's products anymore. Maybe I'm wrong, but that's how I'm thinking about it, that it kind of has to, I have to keep investing in it. It has to be good enough to push the TCO that I keep investing in it. I don't invest in it and stop investing in it, is my point.

42:52Jay Goldberg: I think this is a fundamental dilemma at the heart of NVIDIA's business model now, which is they need to keep coming out with new systems. They need to keep raising prices on that, selling them as much as they can to keep numbers moving. But at the same time, they are taxing the balance sheet of their customers to a high degree. And that sounds very bearish, but I think there is a plausible strategic rationale here, which is NVIDIA wants the NeoClouds to exist. Because they would rather sell to 200 customers than to five customers who have all the power. It's a very, very different dynamic. And so I get, I mean, part of my issue here is that there's a lot of like bubble talk in all of this discussion.

43:48Jay Goldberg: Rightfully so. Like a lot of Jensen's tweet was kind of bubble speak. And so I think that rightfully makes people a little nervous. But there is a strategic rationale here, which is, I mean, it's become clear over the last month is the race for AI is as much about capital as is about technology. We have everybody going into debt, burning cash flow, free cash flow into negative territory to fund this build out. And I don't think the payout is clear. I'm not saying it can't exist. I just don't think it's clear. You have to put some pretty heroic assumptions into your models to really get these to pay off.

44:27Jay Goldberg: So that being said, we did have CoreWeave and Nebius, the two leading neoclouds, report this week. Both had good numbers. Both saw very strong positive stock reaction. And I think there is a short-term case here and a long-term case. The short term is if you can buy GPUs today and get them deployed, you can pay those systems back in under two years. Nebius said this on their call. It used to take them about three years to get payback. Now they're at two-year paybacks. Both companies talked a lot about how they're renewing Hopper and Ampere systems for much higher rates than previously. Prices are going up.

45:07Jay Goldberg: And so I don't disagree. In the near term, if you have Compute, you can sell it. Right? But less clear to me is the extent to which those two in particular are very dependent on a small number of customers, the hyperscalers and the frontier labs, who are all now building their own data centers. And so it gets a little murky because we have so many layers in this. Jensen likes to talk about layers. There's a lot of layers in this. And I think there's clearly lots of enthusiasm right now. But if you're talking about multi-year horizons and trillions of dollars of CapEx for the foreseeable future, I understand why people are getting a little wobbly on this.

45:59Ben Bajarin: yeah yeah i mean i mean i think again i think it comes back to the amount of the return which again has a number of variables which is obviously utilization of the gpu price that you can charge and all of those being normalized above your hard costs in construction and power over the length of time, right? Your fixed one-time cost versus what's amortized over the length of your GPU value for those particular assets. And again, we have a gigawattonomics model that we've talked about before. I can run all these assumptions down to it and, again, come back with all the assumptions you need to validate CoreWeave, right?

46:45Ben Bajarin: How much do they need to charge per hour? at what utilization, like we could back into that, right? We could even show how much it makes. In fact, I tweeted this, how much that on a 10-year, assume a 10-year life, which again, Microsoft and others are saying they're going to extend this to 20 or 25 years. I can't remember, but assume a 10-year's life and average economics. So$6.50 an hour, 75 % utilization. and we came back with well over, just over 200 % return on that capital, right? In that one scenario. But you're right. The issue is, do those investments stop or do those investments have to keep continuing?

47:28Ben Bajarin: And I guess, unless again, as you tell me I'm interpreting this wrong, but Jensen saying that compute is an investable asset, I think has to assume that it keeps going. It doesn't invest and stop and depreciate.

47:46Jay Goldberg: Apple Farmer says apples cure all illness, right?

47:54Jay Goldberg: Look, I'm not saying that this is all bubbly and this is never going to happen. What I'm saying is nobody knows. Nobody knows where this is going to be in five years from now. And the problem we're having now is we're starting to sign long-term debt agreements, which extend out five and ten years. And I think it's very, very hard to forecast. Absolutely. Today, I can build a model that shows payback on these systems is very strong in two years. Right. But the scale we're going to, especially with prices increasing, like remember Jensen saying prices per gigawatt are doubling. you know at you're signing these long long-term agreements i think you know nobody should have anybody who expresses a high degree of confidence that this is you know going to be saleable we're still going to be selling ampere systems five years from now maybe but nobody knows that and so i think there's you know and again we're dealing with bond investors bond investors don't care about ai they care about when they're gonna get paid back and so far they're willing to take that bet because the near-term numbers are so good you get to pay everything back in two years but i think with the numbers getting as big as they are that gets more challenging yeah i mean generally

49:06Ben Bajarin: like i've talked with people who work with some of these gpu as a service companies on rates of return and they have been seeing their uh their as aided financing be paid back in three years pretty consistently. So again, that's happening. I think the point that we're dancing around those, again, like when you say something's an investable asset and I get your point, right? About farming and, and, and apples, but next year's our next year's apples. I don't, I don't take my kind of completely bork this analogy. I don't, I don't, I don't replace portions of my apple farm with new trees that double performance of my apples.

49:54Ben Bajarin: I just, I just, but do you see what I'm saying? Like you have to keep your, your, anyway.

50:01Jay Goldberg: Yeah, that's right. I, and I just know when you, when you build out these models, that becomes a major factor, right? At some point you need to get to a steady state where they're generating cashflow. And look, they're like Amazon did that. AWS did that. They went through years and years of losing money and now it's hugely profitable, but but it was on a much slower upgrade cycle and, and the, the, the expense was much lower relative to the economy.

50:30Ben Bajarin: Yeah. I mean, but, but at the same time, like I don't think Amazon, I'm going to mention this specifically, like wouldn't disagree that as they were building AWS and they continue to sustain and run AWS, that their infrastructure and continuing to invest in that infrastructure was not a valuable use of their money. They monetized that to run Amazon.com and host people's email and SaaS servers and whatnot. And that business was sustainable and they made money on it. So AI is similar. the other point I'll make is I think I think he like I think NVIDIA likes this deal this idea because

51:17Ben Bajarin: it widens the pool beyond the NeoClouds and I do think there is some sensitivity around the NeoClouds wanting to compete with the hyperscalers that is hurting some of their potential relationships long-term with those hyperscalers and the hyperscalers and other labs are sitting there saying we need more compute so if it was there which is why i think you know the likes of of crusoe and lambda and some of these others are attractive as partners for said hyperscalers or frontier labs for compute um that i think that they they can have pretty attractive economics for their margins for people who want to do kind of spill over from their own infrastructure into some of these, these new data centers that are going to get built that are really just pure AI factories, right?

52:13Ben Bajarin: Not trying to be a, like a neocloud is like not trying to be a hyperscaler. So I think there's a dynamic here at play as well.

52:21Jay Goldberg: So, yeah, and I think, I mean, this is where I start to really push back on Jensen's language, because he's not saying that compute is an investable asset. He's saying compute is a whole new class of investments. You'll have stocks, bonds, currency, real estate, and compute. And I don't think that's true, because of all the reasons you just mentioned, right? Land doesn't get obsolete, right?

52:54Jay Goldberg: you don't need to keep, you can't make new land. Like it's just, it's not, it's not the same category of things. This is debt. It's a form of debt, right? Maybe a little equity mixed in, who knows? But I think there is a real tension here where NVIDIA wants everyone to keep upgrading and buying new systems every year. And it's getting hard to do that. It's getting very expensive.

53:19Ben Bajarin: yeah i know i mean i i think this happens for the foreseeable future because the tco of each new generation and then revenue per watt with the latest generation is going to be good my point though is over the course of 10 years is that true that in 2035 next year's compute is 2030 40 percent better than last year's i again again maybe i guess that maybe i'm thinking about this wrong and that's not the right way to think about it and you're just going to do rip and replace and replace 20 of your infrastructure every six years fine but but but my hunch is that to make the claim what you're saying to some degree necessitates very little obsolescence

54:09Jay Goldberg: i think in i i think it's it's too hard to say too far out because i i could see it either way I'm not playing bear here. I'm just saying if AI keeps improving at the rate it's improving and becomes more and more useful to people and it actually starts coming up with drug cancer cures, then we could very well be in a place where we still don't have enough compute 10 years from now. Yeah. But my intuition is that it will not be an even move. It won't keep going up into the right forever. And we could very well be in a position where at some point in the future where we have too much compute, at least for the moment.

54:55Jay Goldberg: And I just don't think anybody knows. That's the honest answer. Nobody knows. Right.

55:01Ben Bajarin: And I think it's that bit that weirds everybody out, to be honest with you. And I totally get it. I get the tension. Like I said, I can see it both ways. Like I pointed out before, if this workload is like graphics and visual arts, um, it just keeps going every year. You need more because you're chasing fidelity. And if you're chasing AGI and we don't really get there, or it's always the next bit of infrastructure away, I get it. Um, so you're right. But I think that's the sensitivity to it is it's, it's hard to grasp the vision when you just don't know, like logically you could see both sides.

55:37Ben Bajarin: So time will tell. Right. Time will tell. But, you know, I mean, for the foreseeable future, we're compute constrained. So that is at least our common denominator. So, all right, everybody. Thanks for listening. And we will talk to you next week.

55:56Jay Goldberg: Thank you, everybody. Click like, subscribe, tell your friends.

56:05you