Intelligence as Infrastructure: How AI Is Rewiring the Economy

Guest

This conversation is going to be a really good one, a very timely one. I know the topic is intelligence as infrastructure—how AI is rewiring the economy—and there have been a lot of thought pieces lately on this very subject. So it'll be good to get your perspective on what you're seeing, talking to companies both inside the AI ecosystem as well as outside.

Leon

Leon, when we spoke earlier, you said that this may not be an issue now. This is something that you're looking at being an issue for unemployment in the future, in terms of AI working its way through the economy. How do you see this all playing out?

1. AI Turns Deflationary Over Time

I think approaching AI and its impact on the economy is such a broad topic that I could be here for 3 hours. But I think I'm going to be looking at it from a point of inflation and what the causation is.

I think long-term, this is a highly deflationary force. I just think it's as simple as we're going to get a lot more for a lot less. Take the health care industry as an example. I think it's like 18% of GDP and 10% of an individual's income. You're going to get a lot of it just through all these LLMs and things like that. It's really going to be disruptive.

Guest

What's the switch? Because we haven't really seen that yet. There are some companies, and there have been a lot of announcements lately from companies laying people off who say it's due to AI, but it hasn't really manifested in the broader economy yet.

Leon

There's a time lag. For example, I think in the short term, you have to be a little careful with the deflationary call, because in the short term, you can actually end up in an inflationary move.

If you look at it, the pricing of CPUs, memory, and just the infrastructure alone is skyrocketing. That's an input cost. On the labor market, the labor market is actually quite robust. If you look at software engineers, you would think this is the one area that would be highly disruptive. You'd just fire all these guys. That's not what's happening.

Last month, I think you had an 18% increase in software engineer hiring. I've been thinking about that more and more. Why is that? I don't think it's happening on the tech side, because they're actually laying off or being much more prudent about it. I think it's happening in the old economy.

As everyone is trying to put these LLMs in place and learn how to use them and how to implement them, they have to have someone help them. It's causing some of the hiring among these software engineers. Product developers are highly in demand. I think right now the labor market and a lot of your old economy are a little slower to fire and be disruptive about it.

So I think there's a time lag. First, you get inflation; the labor market is pretty strong. Then, for example, with each one of these models, every 3 months you get a tremendous advancement. As prompt engineering advances and you can prompt these models to tell them exactly the task you want to do, there may not be so much need for these guys.

I think first you get sticky inflation, and long-term this is a highly, highly deflationary move. You put robotics on top of it, and the labor market can look very different a few years from now, which will make the Fed's job challenging. I do not envy these guys over the next few years.

Guest

Yeah, because the tools at their disposal could be pretty limited. Alex, do you agree with this timeline?

2. AI Adoption Goes Straight Up

Alex

I have to say, I think Leon has really come up with the answer here. There are super-smart people on both sides who are saying it's going to destroy the job market. Others are saying it's going to be a huge boom. But I think he's dead-on: for the first time at Whale Rock, we want to be hiring. We're looking to hire Claude ninjas, and we know we need help to build these amazing things.

So you need to do a little bit of hiring before, and coding is the one area where it literally replaces labor, but then that's allowing people to build software where they never would have built it before. I think it's still a very hard question what it does to jobs. It is going to be incredibly, powerfully productive, but it takes time.

Really, we're in the first batter's box of AI. We've all been using AI, but it's just AI 1.0. It's a search engine on steroids. But now we see what business AI is going to be, and it's Claude Code or something like that plugged into all your data sources. Then you can build skills on it, and you can build agents that actually go out and do things.

There's just a tiny, tiny percentage of the white-collar population that's using AI in that very advanced way. Maybe like 10 basis points of the 1 billion white-collar workers. So, of course, we haven't seen any major productivity gains. But if you look carefully at what these people are doing, it's astonishing and astounding.

If you think of where we are in this whole AI story, those 10 basis points of people—Claude Code has 14 million DAUs, only 14 million. These are the people who are really using it for business every day, but they're not the 10 basis points. Those 10 basis points are burning 1,000 times as much compute and tokens as the rest of the people.

You're going to see those 14 million DAUs go to 500 million DAUs, and then you're going to see the portion of people that really use AI with 14 agents running things increase. That's all happening straight up. At Whale Rock, we talk about S-curve adoption. This is an L-curve, straight up, and it's just starting now.

All this CapEx that we've put in place—and the reason these chip stocks are going up—we have half of what we need from a compute standpoint right now before it's even started. So that's how it's sort of going to play out from that perspective.

Guest

So how do you think about investing in an L-curve?

3. The AI Stack Captures Value

Alex

Nobody's ever seen anything like this, ever, in our entire careers. We were there for Internet 1.0. When I was at Fidelity, my first stock was Amazon. I remember at the time, in 1998, there were only 100 million internet users and only 2 million e-commerce users. I said, “It doesn't even need to grow for this stock to be a buy.”

This one's moving faster. The revenue growth that we're seeing at Anthropic, going from $100 million to $1 billion to $9 billion, and then it's already at $45 billion—it's going to be maybe $100 billion.

Scott Wapner

That's run rate?

Run rate. Last 12 months annualized. So it's not for the full year, but it's growing so fast, it's a good metric. 10x-ing at major, major scale. No one's ever seen anything like that.

So we think the foundation model layer—AI is a stack, with the chips at the bottom, the clouds in the middle, the foundation model companies above that, and then the applications on top. I think the 2 places that capture the most value in AI are the foundation models, where we own Google, OpenAI, and Anthropic, and then still at the chip layer.

Like I said, we're in a dramatic undersupply of chips, and there's dramatic growth ahead. But it's also the golden age of hardware, where there's now so much innovation.

Guest

I think the models heard you and wanted to also participate in the conversation.

Alex

Did they say? I couldn't hear them.

Guest

I don't know. I think they liked your thesis of the oligopoly of LLMs.

Alex

One of the Claude agents.

Scott Wapner

Yeah, exactly. What do you call them, a Claude ninja?

A Claude ninja.

I think the second thing that's so powerful here is that, between Anthropic and OpenAI toward year-end, you're going to be looking at $200 billion of revenue. You can break it down any way you want to; you can make a guess.

What's more interesting about it is the margin profile of these companies. Because they've been able to lock up compute—they were some of the first, and they already have it locked in for the next several years—this is going to be enormous incremental margin.

On a fixed-cost basis, the pricing per token is rising, and everything's rising. There was this huge debate a year ago, even 2 years ago, even 6 months ago: Where's all this CapEx going? What's the ROI? What is this all going to look like? Are they just wasting money?

You're going to look at profitability at Anthropic; it's staggering. You could be looking at something that's like 18 times earnings. So that argument would be put to bed, and the L-curve of adoption is so enormous that—I mean, I've traded several tech booms since 1998, when I started. There's never been anything like this.

Guest

So what do you make of some of the more legacy tech industries? A lot of people have been describing semiconductor moves as being parabolic—that's the term I keep hearing people use. The sector's down today, greater than the market, but is that something that you think is perhaps the best way to play this in the public markets right now? And how do you decipher within chips?

4. AI Drives a Hardware Boom

Alex

Look, it's definitely gone up a lot. So, like I said to you the other day, this was much easier a month ago.

Leon

But in some of the stocks, you probably have certain things that have gone 100 miles an hour in a 60-mile-an-hour zone, so there are going to be some accidents and someone’s going to get pulled over. But most of it is just on an incredible trajectory.

If you look at AI demand, it’s driving so much compute demand. First we started with GPUs, then we went to memory. Now it’s CPUs and networking chips. That, if you look at it, is just creating massive supply constraint.

If you think about the semiconductor industry, maybe just take the last decade: they’ve gone through so many booms and busts that most of these companies have gotten pretty disciplined about CapEx. I.e., they just haven’t spent. Since the last foundry and memory down cycle, no one spent, and it was really one big spender—and that’s Taiwan Semi.

Even if you break down their spend, you can make a very good argument that they’ve underspent significantly. You can use metrics like profitability over CapEx or revenue-growth-rate acceleration over CapEx. They’re all very anemic. You can actually see it now, with some of these announcements from Intel and Samsung and some of the lower-end stuff as far as the foundry competition, that maybe Taiwan Semi made a mistake.

They will have to rectify that. Now you look at the industry today, and it’s not just Taiwan Semi. You’ve got memory companies like Hynix, Micron, and SanDisk. What’s the last time we talked about a NAND cycle? It must be, I think, a decade ago. I don’t know.

Scott Wapner

Yeah, yeah.

Leon Cooperman

Right? And it’s a powerful cycle. The profitability of these businesses is absolutely enormous. If you look at forward CapEx indicators, historically, how profitable the customers are leads to forward CapEx. Intel is now—you know, this is a company that was dead for years—and it’s coming in, and the foundry business is starting to pick up customers.

You’ve got to look at this landscape and say, you’ve gone from 1 spender—and, by the way, in that environment, they probably had all the power in negotiating with semiconductor equipment companies—to multiple spenders, all of which underspent. The forward metrics suggest they’re going to have to spend a lot.

I actually think people believe it’s like $120 billion or $130 billion of WFE. I think you’re going to reach a $300 billion mark over the next 3 to 4 years. Semiconductor equipment companies like the pricing, and all of the customers now have 70% to 80% margins. The memory guys are at 80%, Taiwan Semi is approaching 70%, and semiconductor equipment is at 50%.

So you have pricing power on top of it. These stocks may not screen as the cheapest things in the world right now, but I think the estimates are 50% to 70% too low. These businesses will also look less cyclical because a lot of the customers can now sign long-term agreements. Memory guys are signing LTAs right and left.

You now have much more visibility on the kind of CapEx you can put forth over the next 3 to 4 years. It’s just not as cyclical to them, so they’re going to do it. Maybe that could be reflected in the semiconductor equipment multiples also. They have pristine balance sheets; they can do M&A, and they can buy stock.

Look, I don’t know what the next 10% is, especially when you’ve had this kind of a move, but I’m guessing the next 50% to 100% is up.

Guest

Alex, you look like you’re not quite buying.

Alex

No, no, I fully agree. He articulated a lot of great points, but in addition to this, AI is the most compute-intensive thing we’ve ever seen, with shortages as far as the eye can see. We’re in a golden age of hardware.

For the last 40 years, hardware hasn’t changed. It’s been an x86 server that costs $2,000. Twenty or 30 companies can make it. Every little part in that server has been commoditized: the networking, the PCB, the power system, and the cooling system.

Compute basically grew 30%—compute demand, the bits—and that’s good, but Moore’s Law was going 30%, so there was really no growth. Everything was commoditized. All of a sudden, AI hits. Elon calls it a supersonic tsunami, and it really is, because it’s 10Xing every year with no end in sight.

The old compute can’t do these things, so you have to innovate at every single layer of these $300,000 massive server racks, which are now highly complex machinery. A printed circuit board, which used to be a total commodity, now has only 2 or 3 companies that can do it properly, and you’ve got to upgrade every year.

For example, networking speeds: it used to be 1 gig, and then 7 years later, you’d upgrade to 10 gig. Now you’re on 400 gig; next year it’s 800 gig, the year after that it’s 1.6 terabit, and the next year it’s 3.2 terabit. The people selling into that—there are only a few of them who can do it—are innovating hand in glove with Google and NVIDIA.

There’s less competition, higher margins, higher ASPs every year, and tremendous visibility. All these companies that nobody used to ever pay attention to are now golden, wonderful businesses. The earnings algorithm is: units growing 50%—that’s the end-user racks; ASPs growing 20% to 100%; gross margins rising 300, 400, 500 basis points; and visibility 3 or 4 years out.

You’re growing earnings 100% for the next 4 years, not to mention we’re in short supply of everything you’re making for the next 3 or 4 years. I’ve never seen anything like it, and the moves that we’re seeing are justified. It’s going to be bouncy, but AI is a compute problem first and foremost. It’s also a model problem, but I think it’s a phenomenal way to catch it, and the multiples haven’t caught up.

It’s all been earnings. In some cases, we’ve seen multiple expansion, but if you do a next-3- or 4-year kind of earnings analysis, it’s really powerful.

Scott Wapner

There’s no concern that AI gets more efficient and the computing problem goes away?

Alex

I think there are always going to be innovations. But in general, the tokens are growing. Tokens are the unit of compute in AI. They’re going 14X every year, and the chips basically get better 100%, maybe 200%.

You’re growing your CapEx and adding that to the base, so your computing estate can maybe grow 2 or 3X as efficiently with these innovations, but your token demand is 12Xing. Maybe you’ll get some efficiencies that can push up beyond that, but it’s still not going to be able to keep up.

Guest

What does this all mean for software? Software sold off in the first few months of the year and has rebounded about 20%. IGV, the ETF, over the last month feels like it’s reached an inflection point, where people are trying to figure out whether this is a zero-sum game as it pertains to AI versus software. I’m curious about your perspective.

5. Software Splits Into Winners

Leon Cooperman

I think it’s too broad to say software. There are different verticals within software.

Scott Wapner

Mhm.

Leon Cooperman

I think if you’re a horizontal application layer, there are some troubles there. There’s a lot of trouble there.

Scott Wapner

Mhm.

Leon Cooperman

But I think if you’re a data-driven business or infrastructure software, you can succeed, and you can be very successful. We can debate some of the multiples being paid in the market for winners versus losers, but now I look at a company like Datadog. That’s a unique asset.

Again, is it 30 times or 40 times? The market will kind of get that, but they are at the heart of actually benefiting from everything that’s happening.

Guest

Mhm.

Leon

So I think just saying all of software—that happened in January and February. You just had anything that had the word “software” attached to it. Now we’ve seen a lot of separation over the last month and a half.

Guest

Yeah. People kind of start to—you know, everyone goes through the rubble and figures out which ones are which.

Alex

I think the decline is largely justified. Basically, the old way of doing code is pen and paper, or horse and buggy. The new way of code—it’s not a car, it’s not a jet engine; it’s the transporter from Star Trek. It’s such a massive change in how software is getting sold.

The good news for software owners is that software tends to be very sticky. It will probably take time. Nobody wants to rip out their existing system, but in the back of your mind, you’re thinking, in 1, 2, 3, or 4 years, could that really change? Maybe it could.

In the near term, they have a problem in that software used to be at the top of the CIO’s list. Now AI is at the top. Everyone’s spending all this money on tokens, and that’s taking budget away from software.

The software companies themselves—we thought they would be able to build great AI applications, sell them, and get money for that—but that’s been kind of a fail so far. Maybe it’s just a matter of time, but maybe it’s a culture thing. They don’t have the right people.

It’s very hard. It’s a different sales process because you’re selling a service, not software, and it’s a different business model. I don’t think software’s going to be bouncing anytime soon, but we’re watching it really carefully because we might see a few software companies actually develop and benefit from AI.

Leon mentioned Datadog. A lot of the big model companies, like Anthropic, are using Datadog’s tools. So that’s also a pretty good tell.

Scott Wapner

That’s my cue. Yeah. In the remaining time, let’s talk stocks. What do you think are the best ways to play this? What did you call it? Supersonic—

Alex

Tsunami.

Guest

Tsunami.

6. Investors Choose Their AI Winners

See? Leon’s term. I’ll just start with 2. I’ll start with a small one that you haven’t heard of and a big one that’s easy to buy or think about, but the first one is TTMI.

And they make printed circuit boards, which used to be the biggest commodity of all time. But as these AI chips and servers are growing, demanding more power, needing more signal integrity, and running much, much faster and hotter, they need more and more printed circuit boards. So there's a tremendous unit-growth story, and then the printed circuit boards themselves are getting much more complicated. They used to just have 10 layers, and now they're going to 20, 30, 40, even 120.

That's causing ASPs to rise, and there are very few companies that can do these highly complex printed circuit boards. TTM Technologies is one of them, and they make them for Google and NVIDIA. They just won NVIDIA, and they also do them for other AI companies. Then they have 40% of their business, which is defense. And there's a huge upcycle in defense. They've won business with the Iron Dome contract, and you know how defense is getting so electronicized.

The second one is just Google. It's simple. They've won AI. They're the only public company with a foundation model. Their Google TPU chips are phenomenal. They're now powering Anthropic, and other people are using them besides Google.

Search is actually getting accelerated, and they've got so many other assets like YouTube, Gmail, and Google Sheet. They're going to infuse AI. The stock is very cheap, and we're going to see revenues accelerate at Google. So it could easily be up 50%. I don't see very much downside.

Leon

I actually like analog. I like semiconductor equipment. I brought it up earlier. Lam Research happens to be my favorite because I just think they have such a high exposure to memory, and that's where the market's still skeptical. There's been a ton of lack of spending, and I think this is more of a mid-2027-to-2028 story where I think there's just going to be a boom in spending. These guys, I think the Street may be 50% to 70% too low. I think they're going to do $55 billion of revenue, and margins will go significantly higher.

That's kind of one of my favorites in that, and I actually think the analog semiconductor sector is quite interesting. There's a decent chance that this could look like memory from a pricing standpoint, in terms of how tight things are. Someone pretty smart pitched what was likely Infineon earlier at the conference. I like that one. I think Texas Instruments is very good.

I think likely Renesas in Asia is quite interesting. I think it's just going to stay really tight for a while, and I think if you find ones with the AI power angle attached to them, you're going to have significant upside.

So my mother's in the audience. I think Leon's mother's in the audience, and I bought the SMH semiconductor ETF for my mom a while back, and she's going to keep holding it after what Leon said.

Leon

See, that's what good sons do. They buy their moms ETFs. Happy Mother's Day, by the way.

Guest

All right, thank you guys so much. Really appreciate it. Thank you.