Hello. So, thanks for that introduction. I hope I don't need to say that this AI thing is kind of a big deal. People asked me to come and talk about it, and I've been calling this presentation 'AI Eats the World,' which is a shameless rip-off of Marc Andreessen saying that software was eating the world. But I kind of wanted to break it apart into three different questions. I'm going to talk about three things: I'm going to talk about capital, and deployment, and then change — like, how does this actually change things?
So the first thing is the stuff that everyone in this part of the world spends a lot of time worrying about, and also everyone in Wall Street spends a lot of time worrying about — like, how much money is being spent on this stuff. Most people in this room will know some version of this diagram. The tech industry sort of moves in platform shifts every 10 or 15 years. There's another one. And so we went from mainframes, which were maybe 80,000 mainframes in the world, to PCs and the web and smartphones. And you could probably put cloud and open source and SQL onto this as well. And now the new platform shift is generative AI.
And the thing about platform shifts is they reset everything in the industry. So all the new stuff gets built around this new thing. The old thing is still there. Like, PCs are still there. If you have a bank account, then you'll know that mainframes are still there. If you're filing with the IRS this week, you'll definitely know that mainframes are still there. But all the new — all the innovation happens on the new thing, and all the company creation and all the product creation inside technology — that changes everything. The old gatekeepers go away. There's new value capture. There are new companies. There are new markets.
Outside tech, people kind of scratch their heads and say, 'Okay, what does this mean for us? Is this a new tool? Is this new revenue? Is this some kind of existential threat?' I was at a meeting yesterday with a bank from Ecuador who were having their board meeting in New York City, and they were having exactly this kind of question: like, what does this mean for us? At the end of the day, we're a bank. Does this just kind of change how we do business? Does this create new competitive threats? Is this actually going to change what it means to be a bank, or is this just like more software that Microsoft will sell to us?
And so let's talk about Microsoft. This is Microsoft's share of computing sales. When the PC was the center of the tech industry, every PC being sold — or almost every PC being sold — ran Windows. Then when the center of gravity shifted to smartphones, that market share completely changed, and Microsoft spent 10 years being effectively irrelevant in the tech industry. And so that gets you to quotes like this from Sundar Pichai and from Zuck, both of them effectively saying that for their companies, the risk of missing this is much bigger than the risk of overinvesting for a couple of years.
And that overinvestment is pretty dramatic, if it's overinvestment. So it's also kind of worth looking at this chart, just remembering how small these companies used to be. But this year, those big four platform companies will spend something over 600, maybe $650 billion on capex, up from less than 100 during the pandemic. Oracle topping out there as well. We have OpenAI trying to join the club, sort of a flurry of press announcements and alliances of various kinds, because they don't have the cash flows to build this themselves. None of this seems to have resulted in spades actually going into the ground yet, but, you know, we'll see. Some of this might actually happen.
And about a third of that money drops through to Nvidia, which did close to 70, 80 billion in revenue just last quarter, and seems to keep actually accelerating upwards if anything. Now that flows through on the other side to construction. The US is now spending more money building data centers than building offices. I should point out this is only the construction, so this isn't including the compute. This is just the shell. So that number is going up and to the right. And in fact, you can see this in the trade statistics. This is US imports from Taiwan versus China.
And so I was at a macro conference in Singapore a couple of weeks ago where all the macro people are worrying about, like, Taiwanese GDP and exchange rates and how much this is kind of shifting global trade statistics. So I think most people in this room will know there's a bunch of bottlenecks here. Like, not only can you not get the compute fast enough, you can't get the power hookups and the power grid and the transformers fast enough. There's sort of an emerging moral panic around data centers, with people saying we don't want one of these near us, it's going to kill our children. And so there's a kind of a pig-in-the-python problem about building construction.
I started hearing people saying that data centers are transmitting infrasonic sound that damages your health, and also apparently they leak cyanide into the groundwater. So there's definitely a kind of a weird emerging concern around this, all of which basically means, as Kevin Scott said, it's really hard to build enough capacity, and it continues to be hard to build enough capacity.
But if you step back from that, there's a kind of an algebra question, which is: okay, what is the steady-state capex? Because the number keeps going up. Are we going to stabilize at half a trillion dollars a year, a trillion dollars, $2 trillion every year? How much money are we actually talking about once everything settles down? And the problem is we don't really know. This is kind of like trying to predict internet bandwidth in, like, 1997. Like, you can make the spreadsheet, you know what all the rows would be, but you don't know what the values are.
So on the one hand, as we all know, and as especially we've seen in the last month or two, we have explosive demand, being both from more people but also accelerated by new technologies, particularly agentic. And the inference itself is apparently profitable, if you talk to anybody at these companies. And then the efficiency of the inference keeps improving — like, every efficiency sort of doubles every three months or so. Then there's all sorts of new stuff. So we have new chips, we have the question about how long the GPUs last. There's open source, there's edge. We of course might look at our phones this afternoon and discover a new paper that says you can get the same results for a tenth as much compute. We don't know.
And the thing, of course — the thing that sits beside all of this is that these models are only kind of relevant for, you know, pick a number: three months, six months, nine months, and then you need another model, because the frontier keeps moving, and so you're always chasing the frontier. So all of which is to say, as I said, it's kind of like trying to forecast bandwidth in 1997. You kind of know what the categories would be, but you don't really know where it's going to settle down.
Of course, one thing that does cap this is financial gravity. Meta will say that they will spend over half of their entire revenue this year on capex. Google and Microsoft not far behind. So they can't double it again because there just isn't any more money in order to build more of this infrastructure. So there's a certain point that you hit that you run out of cash. That results in stories like this, what Wall Street likes to call structure. So we have SPVs and debt and off-balance-sheet vehicles and all sorts of other kind of weird, complicated, exotic stuff, all of which tends to be the kind of thing that works really well when the bubble is inflating and then catches you painfully when the bubble deflates on the other side. So we'll kind of see where that settles out.
But meanwhile, as I think, and I don't need to tell anyone, particularly in this room at this event, the revenue is exploding and the usage is exploding. It's kind of important to point out, incidentally, that OpenAI and Anthropic don't account the revenue in the same way. So OpenAI is reporting net revenue and Anthropic is reporting gross revenue. So these aren't the same numbers, but obviously the growth for each of them is pretty spectacular, particularly for Anthropic.
And so that gets you to a capacity question. And I'm old enough to remember — I started my career as a telecoms analyst. I'm old enough to remember something else that had marginal cost and something else where you were consuming this weird unit where you didn't know what it meant. So back in the early days of mobile, mobile had marginal cost and there was very limited capacity, and they charged you per bit because that was the unit, and nobody really knew what a bit was. And then AT&T's network kind of collapsed because Apple made them give everybody flat-rate data.
And so first the network collapses and everyone gets really pissed off, and then AT&T says, 'Right, you guys, we're not giving you unlimited bandwidths anymore because we just can't, because if you're using this thing like a wire, the network will go down.' And so AT&T had to move away from giving everybody flat-rate data, and we moved back to a kind of a segmented model. We're obviously going through exactly the same question now. Instead of AT&T's network going down, it's Anthropic's network going down, and then Anthropic coming back and saying, 'Yeah, you guys who are spending $20 a month and using 10 grand of compute — like, we probably can't keep doing that,' because, like, we'll run out of money. And so, like, what's old is new.
And this gets you to, of course, the same sorts of questions that we had to wonder about then as well. How is it that you think about mapping — sorry, mapping that value. Meanwhile, of course, the models themselves continue to keep scaling, but keep commoditizing. And, you know, this is deliberately an extremely simple chart. I could probably make a presentation that was nothing but diagrams of benchmarks, but I've made this deliberately very simple because I think all you really need to know, if you're not at a frontier lab, is the models keep getting better, but they converge.
If you're deep in this, if you're using this stuff every single day, then you'll see big differences between the models, but most people aren't. Most people are using them a couple of days a week, and they all kind of look the same. And so we've got now three companies that are more or less exactly in line with each other. And it looks like Meta managed to just get back into the game after spending a couple of billion dollars to hire more engineers.
Now again, back to this question: where is this going to end? Historically, software — as the name, it's kind of the point of the word software was that there was no capital. It was soft. There was no hardware. So historically, software didn't have capex. Microsoft's business was putting a $1 CD in a $10 box and selling it to you for $500, which is kind of a good business. But you had network effects, and so that produced monopolies. So Microsoft, Google, Meta, and so on.
It looks like LLMs are very capital intensive, as we know, but don't appear to have network effects. And so it looks like this might play out more like the semiconductor industry, or more like, say, airlines, where with each generation it gets harder and more expensive, but you don't have a winner-takes-all effect. And so you kind of narrow into having something between three and six companies that are making these things.
But then you get a question on top of that, which is: do you have value capture? Can the LLM actually capture value further up the stack the way Microsoft did, or is it commodity infrastructure that sits at the bottom of the stack like Intel or AWS? And there's no network effect. There's no leverage further up the stack. Everybody else builds further up.
Now, Sam Altman a couple of weeks ago said he thinks that intelligence is going to be like a utility, and there'll be electricity or like electricity or water. And you think, well, hang on a second. Nobody in this room is paying a percentage of their revenue or their net income to their electricity company or their water company. That's not how it works. And so a utility is typically not a company that captures value or has high margins.
In fact, going back to talking about data, this is a chart of mobile data consumption back since 2010. It's increased something over one and a half thousand times, maybe 2,000 times, depending on where you are. On the other hand, telecom stocks have been flat. And what happened in telecoms was that the usage just grew and grew and grew, and the telecoms companies spent about 15% of their revenue, 15 to 20% of their revenue every year on capex to support that. The revenue was flat, the profits went down, and the stocks were flat.
And so providing commodity infrastructure at the bottom of the stack tends not to be a particularly great business, and it tends not to allow you to capture value further up the stack. So zooming out one more level, where has all of this got us? Well, the models keep getting better, and the models seem to be commodities. And for both of those, it's important to say so far, because we don't know what's going to happen next week or next year. But the value capture is unclear, and it's very unclear also what the products are going to be that use this.
So that takes me to talking about deployment. So first of all, well, 'Hey Benedict, everyone's already using this.' Well, sorry, I'm English, so I'm allowed to swear. Sort of yes and no. OpenAI say they've got about 900 million weekly active users, although that's weekly active, not daily active, and only about 5% of those people are paying.
So they did a big marketing push at the end of last year where they gave everyone a kind of a cute graphic of how many tokens they were using. And if you looked at all those graphics, you discovered that you're in the top 20% of OpenAI users if you sent a thousand messages last year. So you kind of scratch your head and think, okay, so 80% of ChatGPT users sent less than a thousand prompts in the whole year. That's not particularly high usage for this stuff.
If you go and look at survey data, and there's all sorts of surveys around how much people use this, they all kind of point to the same thing, which is that something like 10, 15% of people, maybe 20% depending on the demographic, are using this every day. But many more people are using it kind of every week. This is a very interesting thing to think about here. If this is a complete transformation in what compute is, why is it that you can't think of anything to do with it this week, and maybe next week you might use it for something?
Why is it that even for teenagers, two-thirds of people are using it every day, and a third of people can't think of anything to do with it today? You see the same thing in the enterprise, at work. Far more people are using this kind of every week or two — 'maybe next week I'll think of something to do with it' — a much smaller number of people are finding something to do with this every day. You see the same thing even in the tech industry. So even within technology, there are far more people who are weekly active users and monthly active users than are daily active users.
So this is kind of an interesting puzzle. I mean, you can clearly say a kind of glass-half-empty, glass-half-full comparison here and say, like, come on, like, there's 10 or 15%, 20% of people are using this thing already, which is fine, but that's not what we're aiming for. Or if we think this is the future of computing, why do people use it and say, 'Okay'?
And I think a really kind of useful way to understand this, which certainly one person in the room will have been old enough to remember, is to remember spreadsheets. And so this is an Apple II from the late '70s running VisiCalc, which was the first successful software spreadsheet. Before this, spreadsheets were paper. In fact, you can go on Amazon now and buy a pad of spreadsheets. You should try this as a prank on your new hires: you give them a pad of this and say, 'No computers for you for the first month, you're doing all your spreadsheets on a piece of paper.'
This Apple II, incidentally, with enough memory and a disk drive and a monitor and a printer, cost something over $10,000 adjusted for inflation. So computers were really expensive. Now, but the reason I show this is: imagine you're an accountant seeing this. You change the interest rate here and all the other numbers change, and that would have been a week of work, and it does a week of work in 10 minutes.
And in fact, Dan Bricklin, the co-founder, has all these stories about CPAs who would say, you know, 'I had a one-month project and I did it in a week, and then I had to go and play golf for three weeks because otherwise the client wouldn't believe I'd done the work.' This is like an extreme case of usage-based pricing. [laughter]
Now imagine you're a lawyer and you see this, and you'd think, well, that's very clever, and my accountant should see this, but that's not what I do. I mean, I might do a timesheet next week. Maybe I could use it for that, but that's not what I do every day. And I think that's what explains the charts — the last half dozen slides of charts — is there are some people who look at this and this is life-changing, based on what they do, what their job is, what industry they're in, what they do every day. There are other people who will look at it and think, well, that's kind of useful, and they need something else. They need it to be delivered in a different way.
Now this is a kind of a classic framework for how every new technology works. To begin with, you use it for the stuff you can already see and the problems you already know you have. And it takes longer to invent new things that you could do with it, and to create new use cases and new revenue, and then longer still to disrupt, you know, to come in and completely change the question. Right now, we're all kind of in this category, using it for the problems that we already know we have.
This is US Fed data surveying US CFOs on what people have been doing with it in 2025, what they think might happen in 2026. And I think it's kind of interesting to see it's kind of concentrated in stuff that's obvious that you would do with this, like analysis, maybe making more slides, maybe better customer service. It's much harder to use this to save money. It's much harder to use this to create new revenue — that just kind of takes longer.
There's also things, incidentally, that are much easier to measure. The stuff that's being deployed now is easy to deploy but kind of hard to measure. It's kind of hard to measure the economic benefit of getting better analysis, at least hard to measure it straight away. And I think that's kind of mapped to thinking about, well, where are the places where people are feeling like the accountant looking at a spreadsheet? And very obviously it's in software development. It's also in marketing and customer support and back office.
But my old shop, a16z, looked at their portfolio companies' data and said, yeah, all the revenue — vast majority of the revenue for this right now is in software development. That's where this is really working. I probably don't need to show anyone in this room a version of this quote from Andrej Karpathy. Anyone in this room who doesn't know who Andrej Karpathy is, raise your hand. You'll be fired afterwards. You're not paying attention to all of this stuff.
Clearly this is a sort of generational shift in abstraction, shift in cost of writing software. But what does that mean? Well, how do you unpack that? Clearly, it's going to be much quicker and much faster to write software that you would have thought of 10 years ago. You can write quicker. Clearly, there are new capabilities for new stuff that you can do with LLMs that you couldn't have done before. Clearly also there will be some legacy apps, like big complex rules-based systems, where generative AI or AI in general is a better, cheaper, more efficient way of solving that.
Is this going to replace all of your apps? [laughter] This is kind of such a weird conversation because it's like there isn't anyone who actually believes that, and yet you still kind of have to say there are kind of somehow people who kind of believe that. I always feel like this is kind of a good reply to that question. This is some of the marketing map of the marketing technology industry. Every one of these logos is a marketing software company. We don't have a shortage of software, but writing the software isn't the hard part. It's working out what the software should be doing and how you would get people to use it. That's the hard part.
In the meanwhile, even when something is old and well understood, it still takes time. So enterprise deployment of cloud, which is like an old idea from 25 years ago, is still only sort of 30, 35% of enterprise workflows. Meanwhile, the entertaining part of this slide is that CIOs always had this kind of pathetic hope that they'll persuade the CFO to give them the budget to replatform, and it's like, 'Yeah, in three years' time I'll finally talk them into giving them money to move to the cloud,' and somehow it never happens. So they kind of get stuck in the middle.
But then I think there's a more important quote here from Steve Jobs, which is to say you can't just give people the raw technology. You can't throw people a chatbot and say, 'There you are. Work out what to do with it.' That works for some people, but for most people you've got to actually work out what product or problem you're trying to solve, and work out what experience you want to give to people.
And so there's a sort of a set of questions around how you build AI software, what would that mean, because you have like these transformative capabilities, you have this jagged frontier problem, which I think is a much better way of describing it than error rate or hallucination. There are some things that just work, and there are some things that kind of don't work, and there are some things that look like they worked except they didn't. And no one in the science or anywhere else really knows what all of those are, or what the pattern is, or how that's going to change next week.
And so there are some people who are sitting in the bits of the jagged frontier that work really, really well. And there are other people who kind of pick it up and say, 'Yeah, I asked it the question and it kind of didn't give me what I needed.' You also, of course, have the blank screen problem. You give somebody a blank prompt and say, 'There you are. Be more productive.' And they're like, 'Okay, what do I type? What do I do with this?'
And of course, this is why we have software companies. You have use cases, tooling, graphical user interface, process, go-to-market, workflows. You actually work out what you do with this, and you work out how to wrap the jagged frontier and the blank screen in product. How would you work out what that is? One answer is you go to your systems integrator. Accenture say they did almost $250 billion of new generative AI contracts. They actually stopped reporting this number because they say basically all their business is generative AI now.
Another answer is you do a bunch of pilots. So every big company is doing pilots. Some of those pilots go into production, some of them don't. People who say there was an MIT study that said — are like an interesting tale for people who really haven't tried to think about this stuff at all. Some pilots work, some pilots don't. Again, those pilots tend to be concentrated in those early obvious use cases.
And of course this is also why software companies exist, why startups exist. A startup exists to find those use cases, unbundle them out of incumbents, work out new use cases, and take them to market. And, you know, if you're going through Y Combinator now and you're not an AI startup, then obviously all of your friends are going to kind of point at you and laugh, behind your back obviously.
So then we get to this kind of question though of like, 'Wait, what was our budget?' Which is, you know, going back to build and going back to capacity. I just kind of saw this story yesterday. Uber said they blew through their whole year's AI budget in the first three months of this year. All sorts of similar stories coming from every part of the industry now, like this really, really works, but it costs way more money than we expected, and there's an ROI there, but that doesn't mean we've necessarily got that budget.
And so there's a sort of a set of questions here which, as I mentioned earlier, kind of apply to that question of data pricing 10, 15 years ago in much the same way that they apply to token pricing today. In that you have marginal cost, and the SaaS industry is not really used to having marginal cost. We're not used to it costing money every time the user presses OK. That hasn't really happened in the computing industry since mainframes.
And then you have limited capacity, right? It doesn't matter how much money you're willing to spend. You actually just can't get enough compute. And then the token and the bit is kind of unintuitive. No one really knows what a token or a bit is. It's unpredictable. It doesn't map very well to the business value. If you think about the way mobile worked, like this website might use 500k and that website might use 5 meg. You wouldn't be able to tell when you were doing it, and you wouldn't really be able to see why this use case should cost more than that use case.
So there's a lot of kind of basic problems in trying to use tokens as a pricing system. And so you get this question: well, how is it that we should charge for this? How is it that people should use this? So I showed this slide earlier, but it's probably better to use it here. This was exactly the problem that we had with mobile. You can't just give everybody flat-rate unlimited data because you've got this very, very steep pricing curve.
And if you give everybody the price that would — you can't charge everybody 20 grand, but you can't give the people who are using 20 grand a $50 monthly bill. So how do you work out how to segment this? How do you split up the pricing curve? How do you break that apart to try and map your marginal cost against the usage and against the value and against the use cases and the user understanding?
So we've got a bunch of fairly obvious conversations around this. Now you can find content marketing on this from all the big venture capitalists. So you have consumption-based pricing, and we've just kind of talked about the problems with consumption. You can also do task completion. So you can go in and map workflows, and you say, well, we'll charge you this much each time you successfully finish this task or go through this workflow. You can also do outcome-based pricing, which is to say we'll take a revenue share or profit share, or you'll pay us a percentage.
And then you've got all sorts of kind of new questions about how this is going to work. Do you run it on the edge? Do you sell used? Dell are now saying that they think they're going to make lots of money selling companies computers so that they can run enterprise versions of OpenClaw, which sounds completely insane to me, but like, we'll see. You clearly are already running models on your phone. You're running models on your phone without even knowing about it. And then there'll be all sorts of new stuff that comes along afterwards.
But I think there's a bunch of kind of sort of puzzles in all of this, because it's all very kind of clever to kind of break apart the pricing and look at that pricing curve and try and segment it. But there's a couple of challenges here. The first of them is that as soon as you are saying, 'I have marginal cost, but I'm not going to give you a price that directly reflects that marginal cost,' then you're taking a risk because now you've got uncapped cost but capped revenue.
And so you need to think about: well, am I going to throttle my users? Am I going to tell my users that they can — well, I'm going to charge them this per workflow, but I might cap the number of workflows they can use if they burnt too many tokens to do it? If I'm going to do a revenue share, then now I'm taking two kinds of risk because I'm taking the risk on the token and I'm also taking the risk on what value my customer is creating from this. Suddenly, I've kind of turned my pricing team into a hedge fund. And I'm kind of mapping my usage risk against my users' business risk, my users' costs, my users' revenue, and do I really want to be that deeply embedded into my users' P&L?
Then there's a sort of a second set of questions around how many use cases exactly are you going to be able to separate out and turn into a unique billing code? And then are you going to go to the user and say, 'Your user type, your user did this thing 745,000 times, so this is the price'? Some of you may remember seeing phone bills from the late 2000s, and they would give you a separate charge for each data session.
So you got your smartphone, and there would be 15 pages of data bills because it would say you used 17k at 3:45 in the afternoon. Do you really want to start trying to segment all of those individual use cases and all of those individual workflows? Are you actually going to be able to separate them out?
For value-based pricing, for outcome-based pricing, there are clearly some industries where you can tie a specific interaction to specific revenue, most obviously in the advertising industry. Although the fact that I gave you that entire screen of millions of marketing technology companies suggested actually it's not very easy to tie revenue back to value. Attribution is quite hard.
But there's a huge number of kind of functions where a company buys a piece of SaaS and they think it's valuable, but they wouldn't be able to tie like a dollar-and-cents value to every single time the user clicked OK inside of SAP or inside of Workday or inside any of your products. So it's like it's theoretically great, and yes, there are some companies that are going to be able to do this. We should also, of course, presume everything's going to change three or four times more this year.
I made a bunch of these slides last week because this was kind of become a new and growing question, and all of this will change again in a month and six months and then next year as well. But I also think, like, having said, hey, it's kind of complicated and it's kind of hard to work out how this is going to work, we should also go to one of the great American business gurus of the 20th century, which is Michael Kolion, [laughter] said, 'If anything in life is certain, it's that compute gets cheap and everything moves to flat-rate bundles.' And in general, a seat-based pricing model kind of is a usage-based model and a value-based model.
So this year, stuff's going to get really hard and complicated, and Anthropic is going to be running around with their hair on fire trying to get more servers. On like a three- or four- or five-year view, we should probably presume that nobody will know what tokens cost anymore. Just as today, nobody in a SaaS company really knows what AWS is charging them for each unit of pricing, unless they're kind of very close to that metal.
So that takes me to talking about change. What else will happen with this stuff? I've had a lot of conversations with company C-level boards where they said, 'Okay, we've done a bunch of stuff. We've deployed some things. We've got review summarization. We've got natural language search. We've got review tagging. We've got a SKU system going in. What else is there? It feels like that's not enough.'
And this is a classic, because I was British and I grew up in Britain, I thought Yogi Bear was only a cartoon bear. I had no idea that he was also a business guru. But this is kind of a useful way to talk about this stuff. Because imagine trying to ask any of these questions about the internet in 1997. Like, you would have got like some of it, but would you have been able — how much would you have missed? And so we should just kind of presume we don't know, but we can look for questions.
And there's three questions I think are kind of interesting to think about here. The first is to talk about unbundling. The second is to talk about price elasticity. And then to think again about what the new questions might be. So Jim Barksdale gave us this classic quote: there's two ways you can make money. You can bundle or you can unbundle.
One of the ways you can think about the internet is that it kind of unbundled, or decoupled, a bunch of businesses where you had a physical asset, but the physical asset wasn't really the thing. If you were a record company, you had to make little pieces of plastic, but that wasn't actually the product. If you were a newspaper, you were a light manufacturing and trucking company, but that wasn't actually the product. And the internet kind of split those apart and completely changed how those business models worked.
So what is it that generative AI does? What is it that it decouples or unbundles? I always used to describe machine learning as saying, well, this gives you infinite interns, right? You can listen to every call coming into the call center and tell me if the customer is angry. You can look at every credit card transaction and say, is there something weird about that? You don't need to be an expert. You could just kind of get an intern to do that.
So what is it that infinite interns unbundle now? Like, seems like a kind of a basic way to think about this is: what is it in your company that's really boring, but it isn't actually the product? How can you break those apart? What industries need a whole bunch of boring work, but that's not actually what they do? And so how can you split those things apart?
This, of course, is also a price elasticity question, which is why people started talking about the Jevons paradox, which is just price elasticity. If you make it much cheaper to do something, do you do the same work for less money, or do you do more work for the same amount of money, or do you do more work for more money because now you have an ROI? If that got really cheap, was that cost base your moat? The moat for newspapers was having to print all that paper. Can somebody unbundle you?
But then there's a kind of a more profound, kind of a further-out kind of a question, which is: there was this stuff that's expensive, and now it will get free. So what does that make possible that was just impossible before? And this is what happened to the newspaper industry. This is US consumption of newsprint. Newspapers were a manufacturing industry. And so when you didn't have to do that, when that moat got automated away, as we all know, the newspaper industry kind of collapsed.
This is a more extreme example. This is comparing people to steam engines in Britain in the 19th century. Steam engines gave Britain the equivalent of something like 250 million extra people. Interestingly, those people eat coal rather than eating grains, which is another kind of interesting shift. You didn't need to grow more food for them. You could just dig the food out of the ground.
But this kind of makes some new things possible. Imagine if you'd wanted to make a train from London to Scotland in 1800. It doesn't matter how much money you have or how many horses you buy. You can't do that. You couldn't buy 20,000 horses to make a train. There were things that were impossible because they were impossibly expensive, and then steam engines made them free or cheap. And so the same thing now happens with AI. The stuff that's impossibly expensive that will now get free or cheap.
Sometimes just automating things changes everything. So this is a chart about barcodes. The US grocery industry deployed barcodes and UPCs in the mid-'70s. And once you don't need to do a paper stock take every Friday night, you can completely change how you run the business, and you can have five times more SKUs, and in the end that gets you to Walmart. It kind of changes how the grocery business works.
But I kind of want to come back to my intern analogy for a moment here because there's kind of two parts to this. One side of this is: what if you had a million interns and they each do something that's kind of dumb, like listen to every call and tell me if the customer's angry? But there's another kind of intern, which is: what if you had one intern that listened to all the calls? Because then they're going to see patterns that no person would have seen before.
This is what DeepMind had a couple of years ago. They discovered that human retinas are different between men and women, which doctors didn't actually know before. And you could only do that if you had one system that looked at a million pictures of retinas. So there's an old saying in the tech industry that a computer should never be asked a question if it should be able to work out the answer. And when people said that, they were talking about like, 'What printer do you have?' Like, 'The computer should know what printer I've got. It's plugged in. God damn it.' We've all — somehow that hasn't been solved now, incidentally.
But now there's a different kind of question that people can ask. So, you know, you're on a Zoom call, you're not paying attention. 'Show me the data that Bob's talking about.' Well, that will work now, but push that further. How about: what did every — because you're recording all of your Zoom calls. So what are all the things that people have talked about on calls in the last six months? What are all the things that people complained about? Why are people canceling their credit cards? If you've listened to every single call coming into our call center, what kind of insights do you get when you can hear everything and see everything?
One thing I think about a lot is e-commerce and advertising and marketing. Because it seems to me that Amazon and Google and Meta don't really know what anything is. They know it's a SKU. They know the SKU. They know the metadata the manufacturer typed in. They know it's a book, but they don't know what a book is. Everybody here has probably got some version of the crazy Amazon recommendation, like, 'Hey, I bought a toilet seat. I'm not collecting toilet seats.'
My favorite is that as somebody who's bought books by Cicero, you might like to know that he has a new book out, which is either a miracle of archaeology or not very good data. LLMs kind of know what those things are. They don't have to just do, well, you bought X and other people who bought X bought Y. They might know what X and Y are. Well, at least they know them at a much higher level of abstraction or much higher level of correlation.
So go back to this question again. Obviously now you can use AI to say, well, what's that code and where will I buy it? But now as everyone thinks about agentic commerce, you can say, find me 20 more options and tell me the pros and cons. And that will kind of work now. But then you should probably think about pushing this further to a different kind of question, like: look at my Instagram. Pick some clothes that you think would freshen my look, that will move me forward a little bit. That's not a query you can type into Amazon or Google or Meta today.
Now again, we should be conscious that like this is all super, super early. Like even people under 30 are not replacing Google with ChatGPT. There's this kind of weird meme that like no one uses Google anymore. Like no, everyone uses Google. Like a tiny number of people have switched their behavior away so far. This is still early.
But you kind of get back again, I think, to thinking about unbundling. How do you separate out the hard part from the easy part? What's the bit that was boring and time-consuming and can now get automated from the place where you're actually adding the value? What was the real job that the customer was coming to you for?
And just some kind of concluding comments, and then I'll kind of stop talking. I just want to kind of go back to this slide for a minute because one of the interesting recurring kind of things I see in the technology conversation is like, this is amazingly completely different and nothing like this has ever happened before. And I always think, yes, just like the last five times. Like the internet was quite a big deal, actually. Mobile was kind of cool, it kind of changed things. And so you can believe that this is going to be absolutely different while also kind of remembering that other stuff was different too.
In fact, this is an IBM ad from 1950 where they say that an IBM electronic calculator — because they didn't call it a computer yet — is like having 150 extra engineers. Raise your hand if that's your company's pitch. 'We'll give you 150 extra engineers.' Hopefully not all white men wearing white shirts. [laughter] Although I do actually have my father's slide rule. I need to remind myself to ask him how I should use it.
And so I speak at a lot of events like this, and it seems to me that all AI questions have kind of one of two answers. The answer is either 'no one knows' — and like there's a broad class of question about where the science is going to go where no one knows. Or the other answer is, well, what happened the last time? How did you do SaaS? How did you do cloud? How did you do mobile?
Should we hire Accenture to build something for us, or buy it from Microsoft, or hire a startup, or should we build it ourselves? Well, how did you do cloud or SaaS or mobile or open source or the web? That's kind of the same sorts of questions we asked before. Now part of that, of course, is no one knows.
I started my career in 1999 when there were lots of words like this floating around. Lots of people were excited about all these things, and none of them worked. The same thing with mobile. There were lots of ideas floating around before the iPhone, indeed after the iPhone, that didn't work. And so, just deterministically, we should presume that a bunch of stuff that people are excited about now in AI won't work, just because half of the stuff that people try doesn't work. We should presume some of this isn't going to happen.
Then we should presume we really don't know what this is going to look like in three or four years' time. There's also a question of, well, how big will the effect be? Which I think is kind of fascinating here because I grew up, you know, spent years hearing people say, well, Uber doesn't sell software to taxi companies and Airbnb doesn't sell software to hotels. Those were different things, which is kind of true, but the impact of Uber on taxis was radically different from the impact of Airbnb on hotels. And the reason for that was, well, they're different things, and those markets are different.
Sometimes the whole thing gets automated away. So elevators used to be a manual system where there was an accelerator and a brake, and then that all got automated away into a button. Other times the job turned out to be something else. So we spent a hundred years automating accounting, and the result was more accountants. For you, that may or may not be a good thing.
And sometimes the questions change, and I think this is kind of a — there's a fascinating point to make here. Just this is the revenue of the music industry. The first half of this chart is what happens if you don't have to pay $15 to get a track. And the second half is what happens if $15 a month gets you all the music that there is. And that's something that was impossible before the internet. It would have been impossibly expensive. And then it became free or nearly free. And so how did that change how everything else worked?
So then some concluding thoughts. Well, we don't know how any of this is going to work. We can think about what you and your competitors are going to do with this. We can think about what will unlock or break or change in the business model, but we should also just presume everything's going to change over and over again for the next couple of years until things settle down. And with that, I'll say thank you.
