Midnight Capital: The AI Compute Trade | Edited Transcript
A professionally copyedited transcript of Chris Barber's conversation: Midnight Capital: The AI Compute Trade.
This is a professionally copyedited transcript of Midnight Capital: The AI Compute Trade. It has been edited for readability and lightly formatted while preserving the substance of the discussion.
Chapter Timestamps
00:00 Memory earnings, SanDisk, and stocks that look parabolic only if earnings are ignored
01:10 The old memory-cycle model: surprise device demand and commodity shortages
02:14 Why the AI buildout may be less like a cycle and more like a structural demand ramp
03:05 Anthropic ARR as a clue for how large AI-lab demand could become
04:04 Hyperscaler capex and the natural expansion rate of cloud operating income
04:29 Why capex acceleration can slow even while absolute spend keeps rising
09:18 Debt markets versus equity holders as the real constraint on hyperscaler spending
12:19 Google Cloud backlog as the cleanest evidence of visible enterprise demand
14:40 Meta as the economic hurdle-rate test for internal AI compute ROI
18:09 Why Meta renting out compute would be a yellow flag, not an automatic bear case
20:20 Whether enterprise backlog customers are clearing ROI or still searching for it
25:42 Token spend, automation, and corporate labor substitution
29:27 How AI labs manufactured urgency across corporate America
31:21 Hyperscaler moats as cloud spend becomes a larger share of operating expense
35:44 Why Nvidia and CUDA matter when customers want cloud portability
37:25 The marginal value of intelligence and why frontier models may keep absorbing spend
40:01 Frontier models versus local AI at the high-value edge
41:12 Cloud-migration risk while everyone is still capacity constrained
43:15 Semis as S&P 500 weight and the dial-up analogy for the current AI stack
Made with: The Transcript Desk Chrome Extension
Full video:
Midnight Capital (Midnight_Captl) is a fund focused on the AI hardware stack, ex-Apple Global Sourcing.
Transcript
00:01-00:50
Speaker: On one hand, stocks have gone parabolic, but on the other hand, so have earnings. Take SanDisk, for example—we were talking about SanDisk. Their earnings year over year went up like 50 times. Literally 50x, not the stock price, I mean the company’s earnings. And it’s still trading at under 10 times forward P/E. I don’t know if you’re into finance, but that’s like half the market multiple on a forward basis. Meaning it’s a 50% discount on earnings compared to the S&P 500.
00:50-01:25
Speaker: So, yeah, the S&P average forward P/E is about 20, and SanDisk’s around 10, maybe even seven or eight, or possibly six. So roughly half the market, maybe even less. It’s a confusing situation because no one really knows when this cycle will end. Historically, memory companies are cyclical—they’re commodity businesses, after all.
01:25-01:50
Speaker: By the way, what used to drive those cycles? It sounds almost like ancient history now. Basically, Apple would release a new phone, and it would sell better than expected, but there wouldn’t be enough memory to meet demand. Full disclosure, I used to work at Apple—I was a supply chain sourcing manager there. A lot of my colleagues traveled internationally for that role, and I’m still connected with many of them.
01:50-02:25
Speaker: So, that’s what the cycles were back then. Maybe a laptop or a big upgrade cycle would unexpectedly spike demand, and suddenly there wasn’t enough memory. Companies had to scramble to buy more, and supply couldn’t keep up—that was what drove the cycles. It sounds so ridiculous now because we’re in a totally different world.
02:25-03:00
Speaker: If you listen to Jensen Huang—you know, I’m a big student of his—you hear him say we’re nowhere near the end of the cycle. Or maybe it’s not even a cycle. What’s more likely is that we’re still in the middle of a massive ramp-up in spending, and we’re nowhere close to the peak yet. Then, after the ramp, things will stabilize, and growth will be mostly hyperscaler-driven.
03:00-03:40
Speaker: Yeah, hyperscalers are going to be the biggest buyers moving forward. And maybe we can talk about this more, but looking at the revenue growth of some of the AI labs—take Anthropic this year, for example. I can’t remember exactly where they started the year, but I think they were around 10 billion dollars in annual recurring revenue at the beginning. Now, they’re at like 44 billion ARR—that’s more than 4x growth in just the first two quarters. If they double again from here… well, we’ve basically got the rest of the year left, so if they keep up this pace, that’s massive growth.
03:38-07:41
Chris Barber: If this year they double from here, they’re looking at a $100 billion ARR business. You don’t even have to stretch it much further to imagine Anthropic and OpenAI pulling in hundreds of billions annually. Those numbers are massive, but that seems like the direction we’re heading.
I think we’re about to see a huge buildout—bigger than most expect. This year, the four major hyperscalers will spend around $700 billion, and I expect next year they’ll hit a trillion, maybe even more. That growth will likely continue until it reaches two or three trillion a year.
You asked when hyperscaler capex growth will peak or have peaked—maybe around 2030 to 2032? What I’m saying is the absolute capex amount will keep increasing through then, but the growth rate, the second derivative, will start slowing down next year.
Here’s why: Right now, Google, Microsoft, Amazon—they’re all growing income from operations at roughly 25% annually. That 25% is what I consider the “natural” rate of capex expansion—that is, excluding debt, dilution, or other financial moves. They can grow capex by about 25% yearly without hurting their balance sheets.
But their latest earnings show accelerating growth: AWS is back in the 30%+ range; Google Cloud just posted 65% year-over-year growth; Azure is growing at 40%. So it’s possible that their natural capex growth rate could push up to somewhere around 30–35%.
Still, for context, last year these four collectively spent about half of what they will spend this year. That means capex growth year-over-year is roughly 100% this year—which they could afford because of ample free cash flow. But now, that free cash flow is down close to zero.
So if they want to grow capex faster than their natural rate—say, above that 25–40% range—they’ll start tapping debt markets and diluting shareholders. They’ll have to leverage their balance sheets and raise equity, which they can do but only up to a limit.
That’s why I believe the second derivative—the growth rate of growth—will slow down after next year. Still, in absolute dollars, we could easily see spending jump from $700 billion this year to a trillion next year.
07:41-08:10
Speaker: Yeah, I think this year looks like about a 50% growth rate. Maybe next year, too, it’ll be around 50%. Then the following year, we might hit about 1.3 trillion. That would mean the growth rate slows down to roughly 30%.
08:10-08:58
Speaker: So that’s kind of the big picture. I think the acceleration—the second derivative—slows down next year, but the absolute dollar amount spent on capex will keep growing through the decade, and maybe even beyond. That’s a huge piece of info. Linked to that is the size of these cloud businesses—AWS, Google Cloud, Azure. They’re going to grow into something unlike anything humankind has seen before. People talk about Big Tech being the size of nation-states—that’s true. But we’re about to go beyond that. These companies will be bigger than governments, which is a little scary to think about. That’s how I see capex evolving. The hyperscalers are going to become bigger than most people can comprehend right now.
08:58-09:59
Interviewer: There are two questions that come to mind. One is about the high-level driver of growth—the willingness of debt markets to lend to these companies. Who exactly is providing this debt? Is it BlackRock or similar asset managers? Or maybe sovereign wealth funds? And then my second question is about hyperscaler lock-in. These businesses benefit from deep lock-in once a customer picks a cloud provider. But is there a future where AI improves so much that it makes migrating between clouds much easier, which could then reduce margins?
10:00-11:34
Speaker: I don’t think debt markets will be the limiting factor. I think there’s nearly limitless demand to lend to these companies at their current leverage levels. Now, if the debt picture starts to look worse—say, if cloud margins shrink and debt service coverage ratios deteriorate—then debt markets could become a constraint. But I don’t see that happening. The real limiter will be equity markets. Equity investors will start to say, “Enough.” That depends on how revenues grow going forward.
Interviewer: You mean equity for the AI labs?
Speaker: No, I mean equity for the hyperscalers. We’re already seeing this with Meta, if you’ve been following them...
11:27-12:13
Chris Barber: When you look at Meta, their shareholder base is kind of restless. They’re basically saying they don’t want to put up with the spending. They don’t want to see it—they’re uncomfortable with all the money being thrown around. They especially don’t like that Meta is tapping the debt markets and piling debt onto their balance sheet. What the shareholders want is a clean ad business that doesn’t get bogged down in all this extra stuff.
12:13-12:59
Chris Barber: What they really want are two things: either buybacks or increased dividends. And if Meta is going to spend that much money, they want a clear line of sight on where it’s going and why, so they can be confident in the growth it’s supposed to deliver. That’s why Google’s stock has done so well—it’s because the growth story is really clear. Google’s backlog looks like a hockey stick shooting straight up. There’s a very clear reason why Google is spending the money.
12:59-14:12
Chris Barber: When I say backlog, I’m talking about their enterprise customers in Google Cloud who are lined up for accelerated computing. For example, say Boeing hosts all their infrastructure on Google Cloud and they have projects where they want to use AI to speed up something in their business. They want enterprise-grade SLAs and support, the kind of thing you only get from hyperscalers. So Boeing goes to Google and says, “Hey, we need X amount of compute capacity,” and Google tells them, “Sorry, we’re sold out, but here’s a contract that gets you in our backlog.”
This backlog is then reported to Wall Street every quarter, and all the major cloud providers do this. If you look at Google Cloud’s backlog, it’s a literal hockey stick—just exploding upward. Investors see this and think, “Okay, now we know why Google is spending this money, because they’re working through this huge customer demand.”
14:12-15:04
Chris Barber: With Meta, though, it’s different. They don’t have a cloud business. There’s no backlog there, no comparable demand line like Google. It sounds like they’ve talked about maybe selling compute if needed, but I’ve been in the camp that thinks Meta won’t go down that route. In fact, it would be fundamentally bearish for the whole AI sector if Meta started renting out compute. Meta should theoretically be in the best position—because they’re essentially a software company—to leverage all this accelerated computing themselves.
So if Meta can’t make use of it internally, it strikes me like some kind of economic hurdle rate test...
15:05-16:01
Chris Barber: The basic question is whether Meta can actually earn on the compute—the money they’re spending on compute. Can they make more than what it would cost to rent that compute from one of the hyperscalers? For example, let’s say AWS has an operating margin around 30%, maybe creeping up to 35% across the big four cloud providers. That margin creeping up is pretty interesting and ties back to the natural rate of capex expansion—if cloud margins go up, it gives them more room to spend.
16:01-17:00
Chris Barber: So, if the margin is about 30% for AWS and a fully loaded cost to rent a GPU is something like $100 an hour, including Amazon’s 30% margin, then the question is: can Meta earn more than that $100 per hour by using the same GPU internally? Does it make sense economically? Meta should be one of the best positioned companies in the world to pull that off—they have all the talent, tons of Silicon Valley engineers, so they should be able to optimize for that.
You’re mostly thinking in terms of ad algorithm optimization, right? Yeah, because that’s 95% of the business, but also things like recommendation algorithms or ad matching. And beyond that, there’s value in other bets too, like increasing software development velocity—stuff that has real enterprise value. That’s why Meta spends so much building that stuff out.
17:00-18:06
Chris Barber: Meta should be one of the best companies to use accelerated computing and clear that economic hurdle rate. Historically, I thought, if Meta decided to become a cloud provider, that would be scary for me because I’d have to ask: if Meta can’t clear that economic hurdle rate, why would anyone else? Why would Boeing be able to do it? It’s a key thing to watch.
18:06-18:48
Chris Barber: Wall Street would love to see Meta do this—it’s a popular opinion. People would say, “I’d love it if Meta became a cloud provider, it would be amazing,” and it would probably boost the stock because spending would get more visible. Meta has already mentioned on earnings calls that many customers have approached them asking if Meta has excess compute capacity—and if they can provide compute.
So yes, that idea is out there, Wall Street loves it, and people like the concept. But for me, if...
18:50-19:10
Speaker 1: People would love it. But if I see that happen, I start having real questions about what’s actually going on here.
19:01-19:11
Speaker 2: Yeah. It reminds me of XAI Gro, for example, where they had excess compute they clearly couldn’t fully use.
19:15-19:40
Speaker 1: Right. And to me, there’s a valid argument that just because that happens doesn’t automatically mean there’s something wrong with the whole story. At a minimum, it’s a yellow flag—something I’d want to dig into deeply. Because, like I said, if Meta can’t clear that economic hurdle rate, then who can?
19:54-20:10
Speaker 1: Now, it might be Anthropic or OpenAI who can, which could be totally fine. I’m not saying this is a red flag so severe that you should sell everything if it happens—it’s just a pause for thought for me.
20:10-20:27
Speaker 2: Right. So what about those big GCP backlog customers, like Boeing? What exactly are they doing, and does it clear their economic hurdle rates today?
20:27-20:40
Speaker 1: That’s the big unknown. Honestly, I don’t think even they know. That’s one of the last real open questions here.
20:40-21:03
Speaker 1: It’s a good question: how much of this spending is just tinkering or testing ideas, and how much is genuinely generating positive ROI versus spending in search of ROI?
21:03-21:17
Speaker 2: Yeah, I was just having this conversation with someone who leaned bearish on several things because of exactly that—the view that a lot of this is spending looking for ROI, not actually positive ROI.
21:17-21:44
Speaker 1: I was more sympathetic to that argument when the numbers were smaller. But looking at S&P 500 data—excluding the big tech stocks—the earnings of the rest are accelerating. It’s not bulletproof, but it’s a strong indicator. If this spending wasn’t ROI positive, expenses would be outpacing earnings.
21:44-22:12
Speaker 1: So the fact that, on average, earnings are still growing is the biggest data point to me that the story holds. Yes, some aren’t hitting the economic hurdle rate for sure, but at the aggregate level, my observations suggest that this isn’t the general case.
22:12-22:47
Speaker 1: It also reminds me of the web and mobile era. A lot of early spending was speculative, testing new ground. But then there comes a point when it becomes fiscally irresponsible not to have a solid website or mobile app—like if you’re Expedia—because your competition has one and customers expect it.
22:49-23:11
Speaker: Competition is fierce, and customers want to interact through these channels or similar ones. So I can imagine something like this happening with a lot of AI spend—starting off speculative, then competitors add features, and if they can’t keep up, they’re at a disadvantage. It’s kind of a Red Queen race.
23:11-23:33
Speaker: Yeah, I think we’re still a bit away from the customer side of things changing much. You see some early signs—like Amazon’s “Roffus,” which is supposed to help with stuff; Walmart has something similar. There are these customer-facing AI tools popping up.
23:33-23:48
Speaker: Amazon claims it’s driving 70% revenue growth, but honestly, I don’t buy that. I’ve used it; it’s fine, but it’s nothing groundbreaking. This isn’t the main story with companies like those in the S&P 493. What’s really happening is efficiency gains.
23:48-24:06
Speaker: Teams are doing more with less—maybe building products that look like regular software we’d use but built with AI, and they’re monetizing it. It’s not about flashy AI products like “Roffus,” it’s more that teams are moving faster through their agendas.
24:06-24:30
Speaker: So, that creates the same effect: if you’re more efficient and need to spend less on headcount, you’re forcing your competitors to match that efficiency. That definitely happens.
24:30-24:55
Speaker: Right now, corporate America is in chaos over this—it’s intense pressure at every level. Teams are feeling the weight of this push to do more with less. That’s basically what’s driving all the spend on companies like Anthropic—they have to do more with the same or fewer resources because their competitors are doing the same.
24:55-25:17
Speaker: Actually, I think the key phrase is “do more.” The focus isn’t just doing everything with less, but genuinely doing more overall. It’s pretty painful inside these organizations. The pressure flows down from the CEO all the way to every team member.
25:17-25:35
Speaker: Everyone’s pushing their direct reports, saying, “You’re not getting more budgets to hire; you just have to get more done with the tools you’ve got.” I even tweeted about this. For example, Apple’s been super conservative with AI—like 6 to 12 months ago, their sourcing team didn’t have access to AI tools because of data privacy concerns related to handling sensitive information.
25:35-26:12
Speaker: Now, if you lose a headcount and the leadership looks at your team’s AI token spend—and it’s basically zero—and you ask to backfill that role, the message is clear: You lost someone, they left or rotated out, and before you just backfill, leadership wants to know what you’ve done with AI. The focus is shifting from just replacing people to using AI effectively first.
26:35-27:02
Speaker 1: The point isn’t about replacing people with AI. Sure, AI can replace someone, but that’s not the main issue. The real point is that AI actually ends up creating more work—it piles more tasks onto people’s plates. They can use AI to help manage the workload, but combined with that, there’s an expectation that everyone produces more output.
27:02-27:25
Speaker 1: And that expectation is extremely stressful. This is a really uncomfortable moment in time. People working at big companies right now are feeling deeply uneasy about all this. It’s because these new demands? They’re just not reasonable.
27:25-27:43
Speaker 1: But it’s not about whether these expectations are fair or not. That’s missing the point. The expectations have simply changed—and that’s what’s driving earnings. From a shareholder perspective, we’re definitely seeing this show up in the numbers.
27:43-28:03
Speaker 1: It’s complicated though, because this isn’t an entirely positive story. People are being affected by it—not always in a good way. Everyone’s getting squeezed, expected to do more work with the same or even fewer resources.
28:03-28:19
Speaker 2: Yeah, regardless of whether you use AI or not, it doesn’t really matter. You have to figure out how to make it work.
28:19-28:33
Speaker 1: Right, and that plays perfectly into all these AI product pitches. They slot neatly into the narrative.
28:33-28:38
Speaker 2: By the way, are you okay to keep going for about 10 more minutes? We’re at halfway.
28:38-28:55
Speaker 1: Yeah, absolutely. This is a good conversation. I also think there’s a lot of pressure at the C-suite level about how executives are actually transforming their businesses. That’s why you’re seeing companies like Palantir do so well—they’re being asked what they’re really doing to change things.
28:55-29:11
Speaker 1: It’s just pressure everywhere right now. So much pressure at every level.
29:11-29:19
Speaker 2: It’s fascinating. It’s almost like the labs did a great job manufacturing a sense of urgency about all this—and now that pressure has completely cascaded through corporate America.
29:19-29:33
Speaker 1: Exactly. Especially for people outside the U.S., the American corporate system can seem like a monstrous machine—usually slow and inert. But these labs managed to create enough fear and chaos to actually get that monster moving.
29:33-29:49
Speaker 2: That’s a great way to put it.
Speaker 1: Yeah, that’s exactly it. And another example is the Mythos thing with the U.S. government—they’ve stirred up a similar level of urgency that’s causing...
30:01-30:30
Speaker: So basically, what’s happening is you’re forcing change, right? And there’s this paranoia, this pressure, this fear. It’s a moment of extreme, extreme change.
30:30-31:00
Speaker: Whether that fear is justified or reasonable doesn’t even matter. That’s completely beside the point. The fear started back around 2012 or so, like with Less Wrong, and now it’s finally spread to enterprises all over the country. It’s in full flight. Completely accepted. If you’re not bought in or are skeptical, well, there’s the door.
31:00-31:30
Speaker: Yeah, it’s a crazy moment in time. Another thing about the hyperscalers—here’s how I think about them—they have strong operating modes. Once you invest in a particular cloud, everything runs through that platform. You hire people with expertise on it. You use compatible tools and services. So there’s a massive long tail of lock-in.
31:30-32:03
Speaker: That’s probably the strongest argument against AI disrupting that lock-in. But then, I also think about how the way people use computers is going to change a lot over the next few years. Most of that lock-in exists because humans are directly using the systems; but maybe if AI agents or automated systems start driving how computers get used, that lock-in might not hold as tightly.
32:06-32:34
Speaker 1: So, how do you think about this? I’m curious if viewers see it the same way or differently. Historically, if you look at what percentage of a typical company’s operating expenses is cloud spend, it’s been pretty small—like, maybe two to four percent.
32:34-32:53
Speaker 1: At that level, it wasn’t really worth paying much attention to or worrying about lock-in. But that’s changing now because this spend is no longer just part of IT expenses; it’s starting to creep into labor costs and other people expenses too.
32:53-33:02
Speaker 2: That actually reminds me of something. Maybe I could ask that as our last question? Or we can keep talking longer if you want—I’m good either way.
Speaker 1: Yeah, I’m happy to go till about 45 minutes if that works.
Speaker 2: Sounds good.
33:02-33:31
Speaker 1: Okay, great. So, the question I want to bring up—sort of the ‘milkshake’ question for us—is: what do you see on the horizon that might change this dynamic? Is there anything that could break this trend? Are we facing a long-term under-supply of compute for the next seven years or more?
33:31-33:55
Speaker 2: Okay, let me answer that by finishing my thoughts on hyperscalers. The thing is, cloud spend as a percentage of operating expenses is no longer stuck at two to four percent—it’s growing. It could creep up to maybe 10 or even 20 percent, depending on how...
33:52-34:21
Speaker 1: It depends. If you’re OpenAI, it’s a lot. Speaker 2: Yeah, definitely. But even if you’re, you know... say you’re Boeing, for example— Speaker 1: Right, Boeing. Speaker 2: Boeing has a research and development budget, so— Speaker 1: Yeah. Speaker 2: The question is, what percent of that budget will be spent on tokens in the future? That percentage is going to increase. The amount spent on compute, specifically on tokens, is going to grow. All of this will happen through hyperscalers, through your cloud providers.
34:21-34:51
Speaker 1: Do you think there’s a saturation point for models? Like, could they hit a limit where they’re basically exceeding some threshold, almost like you could categorize the different token tasks by how much intelligence each token actually needs? Then, is a lot of Boeing’s work going to saturate at a level where something like a Kimmy K3—or whatever it ends up being—would be sufficient? Or will Boeing mostly need frontier-level tokens?
34:51-35:36
Speaker 2: Okay, well, let me finish that point about the hyperscalers because I think it’s important. Speaker 1: Sure, go ahead. Speaker 2: As you can tell, I’m pretty ADHD; my brain just jumps around sometimes. Speaker 1: No, all good, happy to get this on video. It’s been on my mind for a while, too. It’s great to talk about it now. Speaker 2: Right, so I think as the percentage of your cloud spend in your overall operating expenses goes up—from two or four percent to more like ten or twenty percent—then the concept of lock-in becomes much more significant. Because now, instead of just a small sliver of your operating expenses, cloud costs are a big chunk of your total overhead. That changes the game.
35:38-36:43
Speaker: One thing that often gets overlooked is how much operating expenses in your business are being driven by this. To me, that’s a hugely underrated reason why Nvidia, especially Nvidia, is so important. The hyperscalers always say, “Oh, we’re locked into Nvidia.” Sure, they admit that, but their end customers actually benefit from it. If those customers build on top of Nvidia’s CUDA platform, their software can move between Google, Amazon, and Microsoft clouds with relative ease.
But if they build on Google’s TPU accelerator, which is their own hardware, it’s a whole different story. They can’t simply move to Microsoft or Amazon because everything’s tied to TPU’s architecture.
The hyperscalers have a really loud voice and a big megaphone. Everyone listens when they talk about customer choice and lock-in, but what I think their end customers care about—and will care about even more—is which platform their solution is built on, especially as the percentage of overall operating expenses spent on AI computing keeps growing. That’s really the core point I wanted to make about the hyperscalers and lock-in.
36:43-37:28
Speaker: You also asked if companies like Boeing will reach a point where their AI models become smart enough that they stop spending more. And honestly, if we look at it on a case-by-case basis, sure, there is a saturation point. For example, I used to work at General Mills, and there’s definitely a limit where they’re not going to pour in infinite spending—on tokens or compute or whatever it is.
But looking at the bigger picture, what we’re really talking about here is problem-solving. And to me, unless we solve all the key problems out there, there’s probably a lot more intelligence and compute spend ahead of us.
37:29-39:15
Chris Barber: I think there’s still a lot more intelligence that can be invested overall. And when it comes to value, this is something I feel really strongly about: the marginal value of intelligence is way higher than most people realize. A model that’s only 5% smarter isn’t just 5% more valuable — in some cases, it can be worth ten times as much.
Take quantitative trading, for example. You might know about high-frequency trading, where firms set up data centers extremely close to exchanges so they can execute orders a nanosecond faster than competitors. They’re willing to spend tens of billions of dollars on this because even a tiny edge is worth an almost infinite amount of money.
Now, I don’t want to equate every situation with that extreme case, but it illustrates the point: the marginal value of intelligence is huge. Inside organizations, this plays out a lot too. At Apple or other big companies, there are teams full of extremely talented, hardworking people who produce great work. But you still have to differentiate their contributions.
Not everyone gets the same performance review. The gap in pay based on performance can be significant—for instance, one person might receive a $100,000 RSU package, and another might get $185,000. If you look at their actual work, it might seem pretty similar on the surface...
39:15-41:05
Speaker 1: These people might be really similar—like almost identical. Maybe the difference comes down to just one meeting during the year where one person showed up better, and the other person had a meeting that was just... okay. Not terrible, but not great either.
And that small difference? It’s worth 85% more money to the company. That’s huge.
So we already have a pretty good analogy in our own human enterprises that shows how much the marginal value of intelligence really is. It’s way bigger than most people realize.
I think there will definitely be strong demand for frontier AI. By the way, you can see this in places like the Forbes rich list—power really is distributed unevenly.
That’s why sometimes I chuckle when I hear people talk about local AI or open-source models like that’s the future. Don’t get me wrong, like you said, if you can break down a task and handle part of it on local compute, sure, that can be useful and important. But to think you’re going to ditch frontier AI entirely and move everything local? You’re fundamentally at a disadvantage. Why would you do that just to save a little on costs?
The value of frontier AI is way bigger, and that’s exactly why all the demand we’re seeing is still focused on the frontier. For example, if you look at OpenAI or Claude, that’s where the vast majority of tokens get used—not so much with DeepSeek or Quen. Those local or alternative models are being used, sure, but they’re not really competing with OpenAI or Claude.
41:05-42:54
Interviewer and Guest: Interviewer: Competing with OpenAI and Anthropic—I think that’s just going to keep happening.
Guest: Yeah, definitely. And then, does AI introduce new ways of using computers? How does that intersect with how easy it is to migrate between clouds? Does it make switching easier or not so much?
Interviewer: Good question. I think the jury’s still out on that. What makes this technology so disruptive is that, at the end of the day, everything is software. Even hardware boils down to software—you start with a CAD design, then it gets developed into chips using tools like Cadence and Synopsys, and finally manufactured by companies like TSMC. All the machines themselves are digitally created first before being physically built. So really, everything eventually comes back to software.
This software-centric nature is what makes it such a disruptive concept. Could easier migration between clouds happen? Sure, it’s possible. That’s partly why we’ve seen such a derating in valuations recently. But there’s another factor with the clouds—they’re all capacity constrained. So even if you could theoretically switch from Amazon to Google easily, if Google doesn’t have the capacity, there’s no point in making the move.
Until this capacity challenge is solved, I don’t think cloud migration presents much of a risk. For now, capacity constraints keep customers locked in.
Interviewer: Yeah, makes sense.
Guest: So until we figure out this capacity issue, cloud switching isn’t a serious concern at all.
Interviewer: Cool. I think that’s a good spot to end things.
42:56-43:00
Speaker 1: I think this is a good place to wrap it up. Speaker 2: Yeah, definitely. Speaker 1: Yeah. Speaker 2: This has been great. This stuff is just wild.
43:00-43:07
Speaker 2: The world is changing so fast, it’s crazy. One thing I wanted to mention— Speaker 1: Yeah? Speaker 2: I was looking it up before joining, the percentage of the S&P 500 that was semiconductors before ChatGPT…
43:07-43:29
Speaker 2: It was like 6%. Now it’s up to 22%. So it’s gone from 6% to 22% of the S&P 500 over just four years, since 2022 when ChatGPT came out.
43:29-43:41
Speaker 2: That’s literally showing how much the world is shifting. November 30th, 2022 — that’s going to be a historic moment. Speaker 1: Totally. Speaker 2: Our kids are going to learn about it in school from their AI teachers.
43:41-43:57
Speaker 1: That’s wild. Do you have kids yourself? Speaker 2: Yeah, I have three. Speaker 1: I have an eight-month-old baby girl. It just feels crazy thinking about...
43:57-44:19
Speaker 1: You know, the stuff we think of now — like Mythos, Ver Rubin, Nvidia systems, all this frontier cutting-edge tech — it’s going to feel like ancient history soon, like dial-up internet. Speaker 2: Exactly. It’s going to be dial-up by the time our kids grow up. Speaker 1: Mythos is going to be dial-up for them.
44:19-44:40
Speaker 1: And just thinking about having to wait for tokens, how slow it once was... Speaker 2: Right, I remember when you used to get 50 or 100 tokens per second — it wasn’t instant. Speaker 1: Now I’m getting a million tokens per second. What are you talking about?
Made with: The Transcript Desk Chrome Extension