AI Eats the World by Benedict Evans

Live from Schenectady, New York, it's the Metrics Brothers, Growth and CAC.

And I'm Growth, better known as Ray Reich, founder and CEO of Benchmarkit.

And I'm CAC, better known as Dave Kellogg, EIR at Balderton Capital, independent consultants, and the author of CalBlog.

And together we are the Metrix Brothers.

And we go together like Níðhöggr and Yggdrasil.

Dave, I think you're breaking up. Did you cough or were those words?

Come on, Ray. Norse mythology. Níðhöggr is a fearsome dragon notorious for gnawing at the roots of the world tree, Yggdrasil.

You know, all I know about this Nordic stuff is what I learned from watching Ludofisk videos on Instagram. Why do people eat that stuff again?

Growth, growth, growth. We're not gonna talk about Ludofisk. We're using Norse mythology to talk about things that eat the world.

Ah, like the Andreessen Horowitz article that Marc Andreessen himself wrote and published called Software is Eating the World back in August of 2011?

There we go. Now you've got the spirit, Ray. Today we're gonna be talking about not that piece. We'll talk a little bit about that piece, but we're gonna talk about a new piece by Ben Evans called AI Eating the World. World, which is a new presentation published last month in Benedict Evans' ongoing biannual big presentation series that explores macro and strategic trends in the software industry, the tech industry.

You know, you introduced me to Benedict Evans, and I now, I just love his stuff. But before we dive into this, let's hear a word from our sponsor. The Metrics Brothers is proudly sponsored by Benchmarkit, the B2B AI and software industry's largest, broadest, and most contextualized benchmarking index. Dave Kellogg and Ray Groth Wright have put the Metrics Brothers podcast on hiatus for remainder of 2026. All episodes we have ever recorded are still available on your favorite podcasting app. You can reach Dave at Kellogg and Ray on LinkedIn at Ray Reich. Thank you so much for listening to the Metrics Brothers podcast.

Okay, so let's start with a quick reintroduction to Benedict Evans. He is an independent analyst and thought leader. He was a former partner at Andreessen Horowitz for 6 years. He's a Brit, I believe, currently based in New York City. And he's kind of a European Ben Thompson, right? Ben Thompson is the guy who writes Stratechery, kind of a big picture, top-down strategic thinker. I feel like Ben Evans is that, But he has this kind of European angle. He's better, in my opinion, he's better known in Europe than in the US. And our listeners may have heard of him before because we featured him in July, Ray, in our episode on AI metrics.

Well, let's talk a little bit more about this particular presentation. Where did it come from?

Sure. So the name is, in my opinion, obviously a play on the Marc Andreessen piece you talked about from August 20th, 2011. The date's going to matter in a minute. You'll see why. That was more than 14 years ago when Marc wrote that piece called Why Software Is Eating the World. That piece was written about 2.5 years before Evans joined a16z, by the way. He was not concurrently there with it, but he undoubtedly, undoubtedly knew about it. So I don't think the title of his thing, AI Is Eating the World, is any coincidence whatsoever. The Andreessen piece, let's go back because I read that piece again and you have to remember it was written, right? Do you remember what was going on in August 2011? Why do you think Andreessen wrote that piece?

Well, we were just coming off a major banking crisis, right? You know, when we had to give so much money to banks and automotive companies to bail them out.

Yep. So 3 things going on: there's a European sovereign debt crisis, the S&P downgrade of America's credit rating, uh, and the S&P 500 dropped 6.7% on August 8th, 2011. So Marc Andreessen published that piece 12 days after the S&P dropped 6.7%. So if you read that piece, it's a lot of macro trends, but mixed into it is, so if you see your 401 going up or down, right? So he's very much addressing kind of what was perceived as a bubble then. He argues that we were not in a bubble then. He looked at PE ratios as proof of that. Apple was trading at 15.2x earnings at that time. But he said, that's not really the point. That the reason I don't think it's a bubble, and I don't really want to talk about valuations and financial metrics, we need to understand what's happening here, people. And what's happening is we're in the middle of a dramatic and broad technological and economic shift in which software companies are poised to take over large swaths of the economy. And therefore, we're seeing all this craziness. The next sentence is, more and more major businesses and industries are being run on software. and delivered as online services from movies to agriculture to national defense. In short, in my opinion, hey people, don't pay attention to the noise. The stock market is noisy. The signal here is that we're in the midst of this massive transformation. Thoughts, Ray?

Yeah, I would just take it a step further. As you know, I worked with Marc back in the '90s at Netscape, and I was always amazed at his knowledge of history of technology and software. And now here it is, 14 years later after writing this piece, And he highlighted things like how physical properties were going to be disintermediated. He talked about Borders versus Amazon, Blockbuster, and now we got Netflix. He also started to expand on things like taxis, which we had Uber, and hotels, we now have Airbnb. He even said, as I reread this also, that cars are becoming computers on wheels. And boy, with robotaxi and Waymo, we know that's true.

Yeah, I presume you saw the power failure yesterday in San Francisco.

I did. Waymos couldn't work. And then Elon Musk tweeted about, but we continued running.

Yeah. And like both of his cars, right? Or whatever, like compared to the scores of thousands of Waymo cars. But in any case, yeah, there was power failure in San Francisco yesterday and a lot of the Waymos hit stoplights that had no power and didn't know what to do. So they're actually clogging up the roads. All right. So let's talk now about this piece here, AI Eats the World. It was published last month, November 2025. It's 90 slides. So it's long, but it's one of those— there's not that many words per slide. You kind of click through. It flows very nicely. It surprisingly to me did not address the AI bubble. I thought, oh, you know, this must be kind of a sister piece to the Andreessen one that was talking all about the bubble in 2011. It's not. It's just talking about how AI is eating the world without any kind of bubble commentary. It's full of great insights. I think it's a must-read piece. It doesn't take long to click through it. And it's a broad industry piece. So it's actually hard to summarize. I'd say it's almost more flow than structure because there's no like topic slide where you're marching through. But in any case, the flow goes, we talk first about platform shifts. Then about the CapEx surge, including the famous circular revenue diagram that's floating around out there. Then the state of AI models and moats. Then we talk about value capture, then adoption, and then automation, where to me, I think of that section as kind of déjà vu all over again. He talks about the history of automation and where we've seen it before.

Yeah. The one thing that jumped out at me, and I know we're not going to cover it in detail, but slide number 9 was fascinating. It actually showed the percentage of heterosexual couples that met online over the last, you know, 15, 20 years. I didn't know it's over 60% of couples are meeting online today. I was pretty shocked by that.

I think 3 of my 4 kids, their relationships were online. So it's 75% in the Kellogg family. So yeah, and by the way, that slide makes a very important point. It's not just the data there, but the title of that slide is critical, which is When the dust settles, the world has changed. And I think that's arguably the most important point made in that entire section on platform shifts. So let me just review what he says in short. First, he goes, every 10 to 15 years, we get a platform shift in tech. In my life, from mainframes to PCs to web to smartphones to GenAI. And the same basic kind of metacycle repeats itself. We get the new thing invented. All the venture capital flows towards the new thing. All the hype flows towards the new thing. And then inside tech, there's a set of dynamics like who are the new leaders? Who are the new gatekeepers? What's the new way of capturing value? For example, during the internet, network effects became the dominant way of capturing value. Before that, we didn't talk much about network effects, right? Before 1998 or so. So you get new gatekeepers, new value capture and creation mechanisms. And then outside tech, right? It's nice that he does this because Sand Hill Road could be easily accused of navel-gazing. Like, how's everybody else going to adapt to this new technology? Is it a new tool? Is it a new revenue opportunity? Or is it an existential threat? So, those are the— he goes through that section. He talks about how leaders, early leaders often disappear. IBM and PCs, Yahoo and search. He talks about these platform shifts. often accompanied by financial bubbles, which they often are. And then he kind of concludes with the point Ray made, which is, but people remember with all this sea of hype, when the dust settles, the world changes. And it actually changes relatively slowly, right? The chart Ray's referring to is a 25-year period where it goes from 2% of people meeting online to 60%. Thoughts, Ray?

Yeah, the only thing I would say is, This is a personal opinion. I think it's so much more than a platform shift. And that's because what we have the opportunity to displace now is labor costs over time. And labor costs represents like 55 to 60% of the US GDP. So from my perspective, I believe it goes beyond platform shift and it goes into where's the work being done and by who and what.

Yeah, that's actually the point he makes on slide 12, right? Which is, is this another platform shift? Or a whole lot more. And I think Gray is clearly from that last comment in the whole lot more camp. And I think I'm with you. At minimum, it's another major platform shift and it could be a whole lot more. First, 'cause we don't really know what the technology can do yet, right? We're just kind of learning about it. Like, I think we understood relatively quickly what client-server could do, right? Or what a smartphone could do, right? It wasn't that hard to imagine what you could do with a powerful computer in your pocket. I think with AI, there's just a ton of questions about how far the technology can go, which also gives it the potential to be a whole lot more.

Yeah. One more thing I'll add. You did an amazing session at SaaSmetricspalooza a few months ago about selling work versus selling software. And I think you highlighted that the target addressable market is, you know, maybe 5x bigger when you're selling work versus selling software. And that kind of started generating a lot of my ideas.

Absolutely. You're not only going after the pure labor market, but you're going after the entire SI market, right? The TAM just explodes when you sell work, not software. And thanks for referencing that presentation. The second part of Evans's deck talks about the CapEx surge largely driven by FOMO, fear of missing out, probably best exemplified by a Sundar quote from Google, the risk of underinvesting is higher than the risk of overinvesting. Right, which has really been the dominant philosophy. The tech leader CapEx spend has gone in 2021 from basically unreadable. If you look at their chart, you could just see the thinnest little pink line. Oh, sorry, that's 2010. In 2010 to, what was it, $100 billion in 2021 to $400 billion plus in 2025. So that's just fascinating to look at just the scale of this CapEx. You know, another tidbit he puts up there is that data center construction has recently passed office construction. And I think there's 2 factors going on there. First, we're building a ton of data centers, and second, we need fewer offices. But in any case, those 2 lines have crossed for the first time ever. He also talks about how Nvidia cannot keep up with demand and how TSMC can't keep up with Nvidia, right? Which is super interesting. He had an interesting quote in there where he said Nvidia was trying to build the new Sun Microsystems, which I thought was a really strange quote. I think I get it spiritually, but Nvidia is already so, so, so, so much more than Sun Microsystems ever was, right? In my mind, Sun was kind of a chirpy company, a little bit on the sidelines of the internet, trying to make themselves relevant by saying we're the dot in dot-com, right? And they didn't really succeed. People used a lot of their servers to power the internet, but they never got positioned as a strategic player. Whereas I think NVIDIA is absolutely, absolutely, absolutely kind of the strategic player right now in AI. So he also talks about power constraints, hearkening back to our episode, Ray, on gigawatts Manhattans. And he dropped a great term in there, bragawatts, which I thought was pretty hilarious. Do you have thoughts on this section, Ray?

No, I just, you know, when I was at Netscape, we actually worked with Sun Microsystems when they introduced the Java Station. And they embedded our e-commerce software. And then in late '98, when AOL bought Netscape, Sun actually took our e-commerce portfolio and integrated that into their software stack. So they did try, they just never made it.

So let's talk for a minute then before we leave the CapEx surge about who could afford to play at this table. I always think of like a casino where the minimum bet is really large, and can you afford to play at that table? And the people at the table, the hyperscalers, have plus or minus $100 billion in cash flow every year. And even them, it's pressing their abilities, right, to be able to play at this table, which to me always begs the question, how the hell could a VC-backed company like OpenAI even sit down at a table where every other player has, you know, $100 billion plus, plus or minus, in free cash flow? And the answer is they have to do a lot of funky stuff, right? You know, like you've seen the circled revenue chart. We won't describe it here. But that's just a depiction of funk in AI CapEx spend with OpenAI in the middle of it. The other one I think is super interesting is that Larry Ellison, in my opinion, got headfaked. And Larry Ellison is no dummy, right? I think he's charging so hard into AI because he kind of almost missed the cloud, right? He was the guy in the early 2010s going, the cloud is gibberish. The cloud makes no sense to me, right? The only thing he got right on cloud was to invest money in NetSuite. But everything else he got wrong in the early 2010s. And I think like he's now saying, I don't want that to happen again in AI. I want to be relevant. In some sense, I want to buy myself a seat at the AI table, right? So he did this deal with OpenAI where they committed— OpenAI committed $300 billion in spend to him. It looked very real, of course, until OpenAI subsequently made $1.4 trillion in other commitments. Which immediately raised the question, Are any of these real? So Oracle now has $523 billion in RPO on the balance sheet, of which $300 billion is from one customer, OpenAI. And Oracle's CDSs, their credit default swaps, hit an all-time price high because people are worried that their debt isn't good. Basically, OpenAI in some sense headfaked them into borrowing money to go build data centers for them. And now everyone's going, can they actually write the check they promised? Thoughts here, Ray?

Yeah, in fact, that $300 billion deal with OpenAI, they went out and committed to a huge data center in Detroit to serve that. They raised $10 billion of debt from Blue Owl Capital, and now Blue Owl has actually pulled that because they're not sure that OpenAI deal will ever become reality.

Yeah, so it is interesting. I don't know, I don't have a great feeling personally about OpenAI or Sam Altman. I never had. He's a little bit of a 3-card monte dealer in my mind. And we'll see how this all ends up. But Oracle's, I mean, their stock has been cut pretty dramatically as a result of all this. Like it peaked on the day of their OpenAI deal and it's been coming down ever since.

In fact, Dave, I do wanna remind our listening audience, we do not provide any investment advice and we encourage you to listen to our disclosure at the end of this session.

Thank you, Ray. Every time we mention something on finance, we try to include that. But yeah. So in any case, the next section of all this is what have we gotten for all this money? So we're spending hundreds of billions as an industry in CapEx, right? Closing in on trillions over time. And do we have clarity on moats? What's going to provide sustainable competitive advantage for players? Or do we have clarity on product and value capture? And the answer he gives is basically no. I mean, I think in some ways it's even stronger than no. I feel like he Either uses or alludes to the dreaded C-word, commoditization of models. He's like, we have more and more models, but they seem to be more and more similar. So our model's commoditizing. And he emphasizes that there's more kind of experimental over critical usage. One, for the metrics geeks out there, one of the interesting things he catches is that ChatGPT— OpenAI is releasing ChatGPT user stats. As WAOs, not DAOs, right? As weekly active users, not daily active users, which would imply that it looks worse if you're looking at daily active users, which would imply that it's not mission critical if you're not using it every day, right? If you're using it once a week, it's kind of a curiosity, right? So, um, I thought that was a good snag on the metrics funds. Thoughts on, on this section, Ray?

I don't know.

Is it—

would this be a good time to talk about my concern with the conversion rates from all these experimental Proof of concepts to real production AI rollouts.

You're a little ahead of us, but why not go, go for it. You're in adoption. But look, before you hit adoption, I'll just say the last thing he offers in this section is he quotes Jim Barksdale, a famous quote. There are 2 ways to make money, bundling and unbundling. And OpenAI is actually doing experiments on both sides. He feels like the coming wave of startups is going to be focused mostly on unbundling, but there's not a lot more there to that section. And then, Ray, we go to adoption. Take it.

Yeah, so I'm just going to kind of go into what I've been looking at. But he said most successful use cases today are they're absorbing AI, but they're not really getting ROI because they're asking, you know, what do I need to automate? In fact, I don't want to take your line here, Dave. I think you talked about something about Dr. Evil equals Palantir. What do you mean by that?

Yeah, so sorry. So on the adoption point, Ray, I thought you were going to go off to the other thing, but on adoption, He's saying basically the most successful use cases today are absorption, i.e., easy and obvious use cases, which tend to drive productivity, right? They're not driving massive breakthroughs. They're just increasing productivity, but they're easy and obvious. Let's automate accounts payable matching invoices to purchase orders. What my friends at Vic.ai do, right? That's an easy and obvious use case. And he's saying those are the kind of use cases people are using. And then beyond that, he says, well, what should we automate? He gives 2 answers. One, Ask your SI, and he's got a slide with data on SI projects and how much AI is taking over the SI agenda, the system integrator. And the other one he had, which was the funny one, which was ask Dr. Evil, i.e., Palantir. So he's the guy who called Palantir Dr. Evil, but he was using them as an example of where basically their sales as an example of people needing not just help automating, but needing knowing what to automate. Back to you.

Got you. Yeah, what I wanted to really talk about was, and I don't want to give too much time to the MIT report because it's been debunked by a lot of people, where they said only 5% of AI projects, and these are really proof of concepts or pilots, go into production. But IDC actually said that only about 12% make it across the production threshold. And I think Gartner said it's closer to like 30%. So the huge issue, and I think 2026 is going to be more about how do I actually convert these experimental dollars going into AI projects to real production that deliver ROI. And we haven't seen it yet.

Yeah. So I think if Ben were here, this is also my viewpoint. I think he'd say something back roughly, we're not going to measure a new technology by percent of pilots succeeding. We're going to measure it by the business impact of those that do. Right. And I think in some ways people are focused on the wrong metric, and this is the nature of innovation and experimentation, which is You know, we get a new tool, we use it for 10 different things. It works for 3 of them. It doesn't work for 7. Basically, don't count strikeouts, right? Count runs or count grand slams. I think that would be the answer to that. I think there's also, as you mentioned, I'm writing a blog post. Who knows when I'm going to get it done on the MIT report because I got tired of hearing about it. But I do think it's suspect for many reasons, including the fact that the authors have got a skin in the game. But in any case, you know, I think Ben remains quite optimistic that this will be highly transformative and generate a lot of business value.

Yeah, I know we've only got a few minutes left here, but I truly believe that we need to go beyond that personal productivity. And you and I covered this also in our ROI with AI episode, that we need to start looking at where are we getting the biggest returns as measured by reducing costs, Or increasing revenue. And I hope 2026 is when we start to see that, Dave.

Yeah, I mean, I, I, I'm going to be critical here, Ray. That, that statement bugs me because like productivity does both, right? People are a little too poo-poo on productivity. What it was, employee costs are 40 to 60% of a company's costs. And if I can make your employees more productive, if I can make your salespeople more productive, then you're going to sell more. If I make everybody else more productive, I'm going to reduce your cost. So, I do not poo-poo productivity as some people do. But I think I also understand where you're coming from, which is— and my answer will just be, and actually, we'll go back to the deck here. This stuff takes time. And I can't remember the example he had. Oh, I know what it is. It was cloud. He's like, the cloud is now officially old and boring, right? People are tired of talking about the cloud. You can't get a media story on the cloud. You certainly couldn't raise money for a new cloud company.

Right.

But it's still only 30% of enterprise workflows, right? Like 20-plus years in, 25 years in. And that's the part where if you're young or you haven't been around for a while, like, Ray, I think I've told you this one before, but I thought the relational database party was dead in 1990. 1990. It was a billion-dollar market at the time. And I just like, oh, this party's over. It was a great run. And the older people around me were like, you're insane. This is just like, it's 2 minutes after sunrise on the relational database And like, just because you were there at the Big Bang, you think this is old, but it's not. So, some of it's a matter of perspective. And I do, I'm a huge believer in the business value of AI. And look, the old quote back in the '80s and '90s was, you could see computers everywhere but in the productivity statistics. And computers helped us. It might have been hard to measure, but computers clearly, clearly helped. So, some of this stuff, Just because it's hard to measure doesn't mean there's no value. And by the way, just because it increases table stakes doesn't mean there's not value either.

Yeah, Dave, you know what? You're absolutely right. Even like e-commerce, which I was kind of involved in back in the '90s, it's only 16.3% of total retail sales, and it's 20 to 25 years later.

Yeah, it's just incredible how long it takes. Like, for example, I mean, look, last example of this thing, folks, but mainframes. I've been hearing about the death of mainframes since I started in IT, which is 1983, probably 1985, depending on when you count.

All right.

Mainframe's dying. Well, sure, taking its time, right, about dying. And I think people in tech get confused about not growing anymore versus dying or not being the next thing. So, in any case, let's wrap this up by talking about automation. He's got— one of the things I like about Ben's decks is he goes and finds good artifacts. So, he's got a slide rule ad that basically says, hey, get our slide rule. It's the equivalent of 150 extra engineers. He's got a 1955 report on automation and technical change from the government.

Yeah.

And he does a little case study on the elevator operator employment. I use telephone operators when I do this myself. But I think it's a great little case study where, you know, remember people back in the past, once in a blue moon, you'll still find one, by the way, an elevator with an operator, like at a department store or something, right? But back in the day, every elevator had an operator. And yeah, that was automated by the push-button panel.

Mm-hmm.

And I guess elevator operator, elevator attendant employment peaked around 90,000 people in 1950.

Right.

And now it's literally probably down to like 10 people, not 10,000, but like 10 people.

Yeah. The one thing I want to add to this, cause I really was thinking about this for today's episode. And I thought about what we talked about earlier with robotaxis and Waymo. I didn't know this, but we have 9.4 million Uber drivers right now. My question is, when I look at this chart in 20 years for Uber drivers, how many are we going to have left? And I think it's going to be a lot.

Yeah, are they going to be the next elevator attendant? Yeah, it's a good point. And look, I think Ben would say pretty clearly, yes, right? If you want to conceptualize the whole last section on automation, it's when automation works, it disappears. Just like elevator operators disappeared, just like switchboard operators disappeared. probably just like car operators will disappear as we go forward. So, he ends with one of my favorite quotes from a guy named Larry Tesler, who worked among other places at Xerox PARC, was an early— somebody who worked early on Smalltalk. He also worked at Apple and some other great companies. And his quote was, I think people call it Tesler's theorem, AI is whatever machines can't do yet. And he said that in 1970. And I think, I still think it applies today. Any thoughts on wrapping up here, Ray?

I don't think I can end any better than that Larry Tesler quote.

Okay. Well, thanks for joining us, folks. I think Ben Evans has put out another great presentation. I think it's when you read his stuff, it gives you a chance to kind of zoom out and get the top-down strategic view, help you think about the right questions to be worrying about right now. And given that we're going into the holiday break, I hope you have some time to read it during this period.

Okay, Dave. And by the way, Merry Christmas. Happy holidays to you and yours.

To you and yours as well, Ray. Take care. The Metrics Brothers AI, SaaS, and More is a production of the Metrics Brothers Growth and CAC and a member of the Bench Market Podcast Network. By accessing this podcast, you acknowledge that the Metrics Brothers make no warranty, guarantee, or representation as to the accuracy or sufficiency of the information provided or the humor content of the jokes told. Ray, information, opinions, and recommendations presented are, according to our spouses, probably wrong. and provided for information purposes only. This podcast should not be considered professional or for that matter, unprofessional advice. We disclaim any and all liability for any direct, indirect, indirect, misdirect, incidental, special, ordinary, consequential, inconsequential, or other damages arising out of any use of, or God help you, reliance upon the information presented here. Ray Grothreich is based in New York City and available on X @rayreich. Dave Kellogg is based in Silicon Valley and available on X @davekellogg. Kellogg. Schedectity, which is French for unspellable, is not our actual production location. Thanks for listening.