Benedict Evans: AI Is Eating the World | hx Live NYC 2026 Keynote

Thanks a lot. So, thanks for the introduction. I guess that was a good introduction. Last insurance event I was at, I was the first speaker in the morning and all of the audience had spent the previous day out on a boat drinking heavily and getting very seasick, which I remarked was an odd choice for a room of insurance people. Hopefully you're all awake and not hung over. So I was asked to explain everything about AI and not talk too fast and only take half an hour and I'll probably manage two of those three, to the earlier point. But I think the place to start really is to say that everything at the moment is still questions. So most of you will be familiar with a framing like this, that the technology industry tends to move in platform shifts, that every 10 or 15 years there's some new technology model and everything gets rebuilt around that. And so from the mid-60s to the late '70s it was mainframes, and then PCs, and then the web, and then smartphones, and now it seems like it's AI. And as you go through each of those cycles, there are kind of multiple stages. There's the early stage where it doesn't really work very well. It kind of looks like a toy. The enterprise version is it won't scale. It's not useful. It's not clear what you would do with it. And then there's a period where it's very exciting and it's clear that there's a change. This is going to change everything. But it's really not clear how it's going to work. Like what are the building blocks? Who's going to win? What's the margin structure? What's the market structure going to look like? What are you actually going to buy? And then there's a period when it's kind of all done and it works, and it becomes kind of boring. So Apple is shipping everybody their new iPhone tomorrow and it will be amazing and also kind of boring. AI is kind of going through the early part of that where it shifted from being a dumb idea for rich teenagers to something that's very exciting but it's again not really clear how it's going to work and what the building blocks are going to look like. And so as you go through that process a couple of things happen. Firstly, inside technology, all of the innovation and investment and company creation happens around the new thing. The old thing is still there. Airlines run on mainframes. But all the new stuff happens on the new platform. Inside technology, that means that some of the old gatekeepers fall away. They become irrelevant. They become less important. New gatekeepers emerge. There's new value capture and there's potentially new markets. Outside technology, people kind of scratch their heads and say, okay, is this just kind of a new tool that Microsoft and Accenture are going to try and sell to us? Or is it new revenue or is this potentially some kind of existential threat to our business? And so that's kind of the setup I think for driving everything that you see on the front page of the Wall Street Journal at the moment, which is the build phase, the capital phase. So we have Sundar Pichai at Google and then Mark Zuckerberg at Meta both effectively saying that if you are a big technology incumbent, the risk of missing this is bigger than the risk of overinvesting, because if you overinvest you've got some empty data centers and if you're Zuck you've got preference shares so you won't get fired. But if you miss it, then you're the next IBM or you're the next Hewlett Packard or you're the next Microsoft. You're the company that becomes irrelevant for a decade or two or worse.

Now this is what that fear of missing out translates into. This year those big four companies combined will spend something over 700 billion on capex, up from less than 100 a couple of years ago. And that's just those four. Add in Oracle and China and you probably get to closer to a trillion dollars this year. Just for comparison, this is not out of scale for what global infrastructure-heavy businesses spend. Telecoms globally is about $300 billion. Oil and gas is close to a trillion depending on how you define it. But it's clearly massively more than the tech industry has ever spent before. If we go back and look at a longer-term sequence here, this is US investment in compute back since the 1960s as a share of GDP. And so you can see sort of in the middle the deployment phase of PCs. You can see the dot-com bubble and then you can see this enormous surge in the last couple of quarters. You can also see down at the bottom the increase in the physical construction of data centers, which is just the physical construction. It's not the compute itself. So you kind of need to stack the two of those together. So we're getting to something like one and a half percent of GDP going onto this stuff, and still going upwards. Something the vast majority of that drops through to the semiconductor industry. Semiconductors have kind of always been a cyclical industry, hence the chart on the right here. But as you can see, if this is a cycle, it's a much bigger cycle than the industry's gone through before. And if you pull those two together, so here we have trailing 12-month free cash flow for all of the hyperscalers versus the semiconductor industry versus Nvidia. In fact, I need to update this chart for the last quarter because I think Nvidia has now passed all — Nvidia free cash flow has just passed all of the hyperscalers combined, as all of the money kind of flows through here. Now, it turns out it's actually quite difficult to spend that much money that quickly. And so there's a massive supply crunch in semiconductors, both GPUs and now memory, which is making all of our computers more expensive and our phones more expensive. There is a multi-year backlog in power. Turns out you actually can't ring up Siemens and ask them to make 50 gas turbines instead of 25 this year. And so all of the power infrastructure is backed up till 2030 or longer. There's a growing kind of community political backlash. I'll probably stand down so people on this side can see the slides. There's a kind of a growing political backlash around data center construction, particularly in the USA. Reminds me quite a bit of the backlash around 5G, if you remember that, in that a large portion of it is based on sort of assertions of fact that aren't really true, like this stuff uses lots of water, which it kind of doesn't, and is in some senses a proxy for kind of other political questions rather than the data centers themselves. But all of that means that it's been effectively difficult or impossible to build anything like as much infrastructure as these companies would like to do. It's also becoming increasingly difficult to pay for it. Until relatively recently these were very cash-generative companies that didn't have much capex and so they funded it from capex. This year three of the four big platform companies will have capex to sales over 50 percent. And they can't keep increasing the capex. So we have this wonderful word structure from our friends in Wall Street and so they move to leases and debt and SPVs and various kinds of off-balance-sheet vehicles and future purchase commitments and leases that haven't started. So you are building an enormous kind of financial infrastructure, a capital structure around that in order to fund that spending. If you're Oracle, of course, it's much simpler. You just gear the out of your legacy business to buy the new thing.

I'm not sure if that's a technical term in finance, but like it seems like the most appropriate way to describe this. This is a very cash-generative, very infrastructure-like business, and they suddenly decided they wanted to be a data center company. So they're looking at something over 100 percent capex to sales in the next couple of years. Nvidia I think is a little bit more interesting. As many of you will have seen, Nvidia has been using its balance sheet as a weapon, had $130 billion of free cash flow in the last 12 months and it's sort of deploying that all over the industry. It was credit guarantees, financing, paying for data centers, paying for power, all sorts of things. And kind of there's a sort of an obvious explanation which is they're accelerating deployment and financing their customers. Indeed, we used to call this vendor financing. There's a slightly more cynical view, which is to say this is circular revenue. The money is kind of going in big circles to pump up Nvidia's revenues. I think the more interesting analysis is to say what they're actually doing is fragmenting the customer base and weakening their competitors because all the money is going to people who wouldn't otherwise be able to buy as much as quickly, which means they're creating more customers and a broader customer base with therefore less purchasing power and also making it harder for people to buy competing products. So they're using capital as a weapon to strengthen their own competitive position at the cost of their competitors and their customers. Now I've mentioned growth a couple of times. This is the growth we're talking about. These are very, very non-GAAP numbers. So these are round numbers reported at scheduled events and they are annualized previous four-week revenue multiplied by 13, sometimes net, sometimes gross. Not entirely clear, which is why we were looking forward to the Anthropic S-1. But that's kind of irrelevant because the point here is the growth. So Anthropic has gone from basically nothing to probably hitting $75 billion of annualized revenue in the course of this year. The two of them combined probably a hundred. There was an enormous product-market fit almost entirely for software development and a bunch of other stuff kind of somewhere behind that. And that demand is driving this enormous compute demand which is driving all of this capex in order to support that.

Now, the other side of this is of course the models. You may be surprised I've got 17 slides into this without showing you a chart of AI models. This is deliberately a very simple one, because what this is showing you is there are lots of models and they're all pretty close together. If you dig right into the details for specific use cases, you can find differences. But what's going on is that OpenAI and Anthropic leapfrog each other every couple of weeks. Google is a couple of weeks behind. The Chinese a couple of weeks behind. Most interestingly, I think Meta and SpaceX both dropped off the scale entirely last year, went out and spent a couple of billion dollars to build a new team and then got straight back up to the top of the leaderboards, which tells us at the moment we don't really seem to have any barriers to entry other than willingness to spend money and ability to hire the right people. I think a more interesting way to split this is to look at the price versus the performance. And so, as you'd expect, there's a very steep price-performance curve. And so the real question is like how many people need to be at the head of that curve? How many people need absolutely the latest model at the highest price from the smallest number of suppliers? And how many people are just fine with the model from six months ago that's a fraction of the price and comes from many suppliers, maybe open source, maybe you can run it on your own compute? How far can you unbundle, disaggregate, fragment that usage and split it into things that you can run for much cheaper than buying the frontier model? All of that means we have this kind of question about how this is going to split. Right now you have massively more demand than supply. If you're a model lab, you can sell more or less anything you can make. You can price at ROI. You have very high gross margins. Sort of apparent rumor is that Anthropic is going to report 80 percent gross margins, which is great for now. But then how is all of that going to evolve? The one thing we know about crunches is that they tend to be followed by gluts. So that may pass out on the left into the capex catches up and the models become increasingly commoditized and you have a pricing collapse. The other alternative is it may be that you get positive price elasticity as more supply rises, more demand follows, and the models themselves manage to avoid becoming commodities and manage to move up the stack and become more and more value-added to capture more value up the stack. We don't quite know how that's going to work but we kind of have to ask the question. It's not clear. This is kind of the worst-case scenario. This is what happened to mobile networks. So on the left, mobile data traffic has exploded up one and a half or two thousand times since the emergence of smartphones. On the other hand, the stocks have basically gone nowhere. The mobile networks globally have revenue of about a trillion dollars. They have capex of about $200 billion. Everybody in the world is paying for this. There's about five billion, six billion people with a smartphone. And they don't make any money from it because all the value is further up the stack from Google and Meta and Uber and Amazon and Instacart and everybody else. They became commodity infrastructure. And so we can kind of plot this out a little bit. It's maybe that all of this evolves like the software industry where, you know, Microsoft put a $1 CD in a $10 box and sold it to you for $500, which had pretty good gross margins because they had network effects and the network effects gave them monopolies. It may work like chips where there isn't a network effect, but you had escalating cost and complexity and that meant you went from dozens of chip companies to basically just TSMC at the frontier. It may work like mobile, which I just mentioned. We don't know where this is going to go. There may be a path to pricing power or value capture. It's not quite clear what that would look like. And so a sort of a provisional thesis for how this is going to work out, like what's going to happen with all of this money, is all of this money is going to buy us an enormous amount of infrastructure with lots of models competing at every price point and capability. It seems like chat itself, the model itself, is a bad user interface and you need an enormous number of use cases and products built on top that the labs won't be able to do themselves. And so it looks like this is going to end up as commodity infrastructure with the value moving further up the stack. And there will be many interesting conversations about the return on capital of all of that data center construction. But for everybody else, this will be infra that you buy at cost and use to make something else.

And so that takes me to sort of a more present conversation. What's happening to deployment? How are people actually using this stuff? And first I think we should probably just sit and take stock. The models are amazing and the models keep getting better. They remain jagged. Jagged, I think, is a better concept than hallucinations. There's things that they can do really, really well and things that they can't. And it's not very intuitive which or predictable which, and it changes all the time, and you can get the model to do something transformative and then you try something that a five-year-old would manage and they fail, and those frontiers keep changing. Secondly, the adoption is very, very broad. OpenAI talks about having a billion weekly active users, but shallow. They're only weekly active users. And the use cases and the deployment models are still pretty unclear. This is all very 1997. This is all very amazing. But we don't quite know how it's going to work. And then there's a sort of a broader question which is, okay, what does actual real transformation for consumers or for enterprises or for industries actually look like beyond just giving everybody ChatGPT? Digging into a little bit, this is what I mean by broad but shallow. On the one hand, 50, 60 percent of people say that they're using AI. On the other hand, depending on how you ask the question, something between five and 15 percent of people are using it every day. So if you go back and think about previous technologies, if you say, are you using the internet? Yes. Have you used it this week? No. Okay. Well, those are both important answers. It hasn't moved to the point that it's an essential part of daily life for very many people. The usage is concentrated. The same thing if you dig into individual industries, again, this is very sort of glass half empty, glass half full. On the one hand, amazing growth. On the other hand, yes, but most people are only using this every couple of days or every couple of weeks rather than every single day, even inside the tech industry.