So, thanks for that introduction. I was just listening to the previous presentation and it occurred to me that you've just had two northerners giving you very hard specific facts and now I'm the southerner here to kind of wave my hands and say how exciting this should all be. So no stereotypes here at all. But yes, as the introduction said, I'm a globally renowned thought leader and you're all very lucky that I'm here.
I was asked to explain what's going on in AI and as we just sort of heard, AI is a big subject and there's a lot going on. And so I wanted to sort of try and break my thoughts into three categories or sort of three conceptual things to think about. The first of them is capital because we can't get away from hearing about that. The second is to try and think about what deployment is going to look like. And then the third is to think about what all of this might actually change. Firstly, looking at capital, I actually started my career about 250 yards from here as a telecoms analyst in the dotcom bubble. So like share price go up, share price go down is not a new experience for me. But the underlying mechanic that we should be thinking about is that the technology industry tends to move in platform shifts. Every 10 or 15 years there's a new thing and that's what all the new stuff gets built around. So from the mid-60s through to the late 70s it was mainframes and then for another 10 or 15 years it was PCs. And then it was the web and then it was smartphones. And now it's generative AI and the old things don't go away. Like British Airways still runs on mainframes as far as I can tell.
But all the innovation, all the investment, all the company creation happens on the new thing. And so what happens in a platform shift is that all of the new stuff happens on the new thing. All the innovation, all the product creation, all the company building, all the value creation happens on the new thing. Within tech, everything changes. The old gatekeepers go away, new gatekeepers emerge. All the companies that were exciting and new become legacy and maybe fade away. And new companies get created, new markets get unlocked. Outside of technology, people kind of look at this and scratch their head and say, "Okay, well, what is this? What do we do with this? Is this just kind of a new tool that we'll use? Is this something new that our vendors are going to try and sell us? Is this a new revenue opportunity or is this actually some kind of existential threat that's going to change the nature of our market and change the nature of our business?"
And just to illustrate that point within the tech industry, this is Microsoft's market share. Going back to the beginning, when the PC was the computing industry, 90-95% of PCs ran Windows, Microsoft dominated the industry. Once we shifted to mobile and Microsoft missed mobile, their share of just new devices being sold dropped down to 10-15%. Nobody started a company to make Windows software since about 2001. The company's still there. And then they managed to jump onto cloud a little bit later, but that dominance went away. And that chart and many others like it explain these two quotes from the last year or so, both Sundar Pichai from Google and Mark Zuckerberg from Meta basically saying the risk of missing this is much bigger than the risk of overinvesting a little bit. There was a great quote from the TV show Silicon Valley, if anyone watched that, where the CEO who was supposed to be Larry Page said, "I don't want to live in a world where other people are making the world a better place more than we are." And that's some of what we're getting at here. It's a much bigger risk to miss this entirely and end up as IBM than to overinvest for a couple of years.
Now, what does that investment look like? Well, last year those big four companies spent about $400 billion building infrastructure. This year, based on the guidance we've had in the last couple of weeks, they'll spend at least 600 billion, some probably over 600 billion. The numbers for 2025 went up by about 50% in the course of the year from their guidance at the beginning of the year. Microsoft hasn't really given full year 2026 guidance yet. But these are pretty chunky numbers and just as the slide says for context, the global telecoms industry spends about $300 billion a year. Upstream oil and gas, which is digging holes in the ground, spends about $400 to $500 billion a year. So these are big numbers. About a third of that drops through to Nvidia. The numbers came out last night so I haven't updated my slide. Last quarter, Nvidia did, I think, $68 billion in that quarter. So the chart continues to go up and to the right. And Nvidia, I think it's funny, people still talk about Nvidia selling GPUs. Nvidia doesn't sell GPUs. Nvidia sells $50,000 computers. They're much more like kind of the new Sun Microsystems than anything else, although we kind of know how Sun Microsystems ended.
And that all of this creates a kind of pig in the python problem in that the rest of the supply chain can't keep up with the desire to deploy capital. So we have bottlenecks in GPUs. We have bottlenecks now in memory. TSMC can't keep up with Nvidia's demand. Then power grids to support these data centers are backed up now for kind of 3 to 5 years. If you work at Siemens or GE and somebody from Google rings you up and says, "Hey, can you double your production of gas turbines next year?" The answer is no. And then of course there's increasingly a kind of political question around actually building these enormous data centers. So there's a challenge in just deploying that capital which gets you to quotes like this from Microsoft. Nobody's been able to build enough capacity in the last couple of years since ChatGPT came out. And then OpenAI wants to break in because at the moment of course they're renting capacity from Microsoft, Amazon, Google, anywhere they can get it. And so we've had a variety of very eye-catching announcements of how many gigawatts, how many billions, maybe trillions of dollars they spend, will spend, have committed to spend, would like to spend, none of which has actually resulted in spades in the ground yet. But we'll see what that looks like.
And there's some fairly basic kind of capex algebra that sits behind this. The first of them is that inference, that's to say running your existing model, is profitable and the cost of running the model goes down roughly every 3 months, halves roughly every three months. So the existing stuff is very profitable and works fine, but then you have explosive growth in actually running that stuff, particularly as you invent new things that you can do with it that you need more capacity. So the demand for capacity keeps going up. On the other hand the next model is much better and much more expensive to build. And so at the moment we're kind of chasing the horizon, chasing the frontier model. And in the meantime, I think somebody observed that at the moment a frontier model is the most expensive and shortest lived asset in history because it's incredibly expensive and after six months it's irrelevant because there's a new one. So we get this kind of question of well where is all of this going to settle? Where would equilibrium look like? How long do people chase the frontier? How long does the scaling work? One answer is that financial gravity at a certain point kind of kicks in. This is capex to sales for those big four companies.
We're all kind of used to the idea that these are asset-light companies. Some of us are old enough to remember when Microsoft's business was putting kind of a $1 CD in a $10 box and selling it to you for $500 and then doing it again two years later. That's not the business anymore. So these companies are now spending 30, 40, 50% of their revenue on capex this year, whereas telecoms companies again for comparison is typically 15 to 20%. And historically, you know, until kind of this year really that was coming out of cash flow because these are very profitable companies. Now, well, they don't actually have 50, 60% cash flow margins. Even Google doesn't have that much money. So we're starting to get a bunch of structure, a bunch of debt, a bunch of off-balance sheet vehicles. And of course that comes with this wonderful phrase circular revenue which is what we saw from OpenAI. But this is all kind of entertaining if you work on Wall Street or in the City and you probably get a bonus out of it. What it's actually delivering is two things.
Firstly, the chart keeps going up and to the right and you could have a one-hour conversation about exactly what the benchmarks mean. But the models keep getting better. However, there's an awful lot of them and there isn't really any fundamental differentiation. The models are becoming commodities and we've gone from OpenAI being way ahead of everybody else to now OpenAI being at best first amongst equals and certainly Google and Anthropic and arguably anything up to half a dozen companies are producing a frontier model every month or two. They leapfrog each other and if you're doing something very specific around say image generation or coding then you can see big differences between the models but for kind of general purposes they're all making roughly the same thing at roughly the same speed in roughly the same way.
That gets us to this question of where is all of this going to settle down. And you know as I sort of mentioned a moment ago historically the software industry was very capital light and it tended to have network effects particularly on the consumer side. So you had network effects around Windows and around Google search and around Facebook and around iOS. And so the network effects tended to produce monopolies. You know the phrase the great thing about monopolies is how many there are to choose from. And so that was the kind of the structure of the software industry. It looks today like LLMs might evolve maybe more like chips in that there isn't a network effect, but they're just really difficult and really expensive. And what happened in chips is we went from, you know, 30 years ago, 20 or 30 companies that were at the cutting edge to today there's really only one company that's at the cutting edge just because it's so expensive. And so it may be that we have this very capital intensive industry that doesn't have network effects and that has an oligopoly of people who are model providers.
But then the question is well what does it get you to have a model? Is a model just commodity infrastructure or are you going to be able to build stuff on top of the model that captures value further up the stack? OpenAI would like to be Windows. They would like to provide that system that everybody has to use and everybody has to build on. At the moment, it looks more likely that these are going to look more like cloud in that there's three, four, five of them and you can pick and choose. And running a cloud doesn't give you value capture further up the stack. When you buy a SaaS application, you don't really care which cloud it runs on. Certainly, as a consumer, when you use an app, you don't care which cloud it runs on. Those aren't network effects.
What does that mean for the rest of us, though? Again, outside of the tech industry or outside of those hyperscalers, what does that mean? And it seems to me there's kind of two things to think about. One of them is we get more, better, faster, both cheaper and more expensive models depending on how you count it. But the value capture is very unclear. And the nature of the product is very unclear. What exactly are you supposed to do with these things? So that takes me to talking about deployment.
And this is an old quote from the movie industry. No one knows anything. This is William Goldman who wrote Butch Cassidy and the Sundance Kid and The Princess Bride and a bunch of other things. No one knows what's going to work. If we go back to thinking about the internet in the 90s, there were all these acronyms and technologies and concepts and companies that seemed like they were going to be a big deal and are now mostly forgotten. The same thing for mobile internet. i-mode, Java, WAP, J2ME. Anyone in this room know what J2ME was? This was a big deal. Nobody has a clue what J2ME was now. And Nokia, of course, Nokia and BlackBerry faded away. And the same thing now with generative AI. I don't want to single out one of these in particular. I just sort of point deterministically. We should just presume that a bunch of stuff that people are excited about in generative AI will not be the things that work and there will be other things that work that we're not thinking about now just because that's the way it always works.
Now you could look at this and say, "But Benedict, everybody's using this now and particularly everybody's using OpenAI. OpenAI has something over 900 million weekly active users, although only about 5% of those users find it valuable enough to pay for it. So they've got it. They've won the market." Well, yes and no. If you look at the data they gave at the end of last year, 80% of users sent less than a thousand messages last year, which works out at sort of three, four, five messages a day depending on how you count it, which is sort of implying people are using this once a day, maybe once every couple of days for 80% of the user base. There's a very small number of people for whom this has been completely transformative. For most people this is kind of useful. And if you look at survey data it all tends to say the same kind of thing. You've got 20, 30, 40, 50% of people who have used this are using it maybe once a month, maybe once a week. The number of people who are using it every single day looks completely different. It's sort of 5 to 10% of people. And you see the same thing in the enterprise. How many people are using this for work? And there's a big difference between have used it and are using it daily.
This again is very much like the internet if you remember that when people would talk about how many hits their web page had and then you discover that if your web page had 14 images on it that was 15 hits and if you added another item to your menu bar your hits went up 10% because now there was another image there. "Have used it", "this company has deployed AI" doesn't really mean anything. You kind of have to ask well how much, how deeply, what are you actually doing with it. And so the kind of the takeaway from these charts is that there are some people for whom this is already transformative, but there's a lot of people for whom it's kind of useful sometimes.
And a really interesting comparison to draw here, I think, is to go back and look at spreadsheets. So, this is a piece of software called VisiCalc, which was the first successful software spreadsheet. Before this, spreadsheets were paper. In fact, you can still go on Amazon and buy paper spreadsheets as a pad of grid paper. You should try this as a prank on your new hires. Like, no Excel for the first year. This is how you're doing your job. This setup, including a printer, incidentally, cost something over $10,000 adjusted for inflation. Well, you kind of forget how expensive computers used to be. And the point of this comparison is that if you saw this in 1978 and you're an accountant, it was life-changing. Like, you change the interest rate here and all the other numbers change. And you say that now and people say yes.
But you could take this thing and do a project that would take you a week and it would be finished in 10 minutes. In fact, there are stories from this period about accountants who would be given a one-month project and they'd do it in a week and then go and play golf for three weeks and they couldn't go back to the client because the client just wouldn't believe they'd done the work properly. So, they had to go and play golf just to establish client trust. So this was life-changing if you're an accountant. Now imagine you're a lawyer and you see this. Well, you'd say, "Well, that's very cool, and my accountant should see this. My finance team should see this. And maybe next week I'll use it to do my timesheets. But this isn't what I do every day. I don't do numbers. I don't do spreadsheets. Give me a word processor. That will change my life. But if you're a lawyer, a spreadsheet isn't actually very interesting." And I think that's the kind of spread we see at the moment with generative AI. There are some people who really, really have this use case right now and it's completely transformative. And there's a lot of other people who pick it up and say, "Well, it's kind of useful. Maybe I'll do something with it next week."
And if we kind of generalize this point, with every new technology, there's always a couple of phases that you go through. To begin with, you use it for the obvious stuff that you can see. You use it for the problems you already know you have. Then over time you start working out new things or other people work out new things and show them to you. You create new use cases. You create new possibilities. That also of course means that you go from cost-saving in the first step to new revenue opportunities in the second step. Then the third step is every now and then someone will come along and actually redefine the market. Someone will come along and disrupt you. That's the thing that's impossible to predict and of course it doesn't happen in every industry. It doesn't happen on the same timeline. But at the moment we're very much in that kind of step one.
If we look at the things that are really, really working for generative AI, the first of them as I think everyone here will know is software development and then marketing and customer support and sort of automation of back office processes. Those are the places where this is really, really working. This is data from OpenAI from January. Where are their weekly active users by industry? They tried analyzing just the email domains that people had used to register. Unsurprisingly, it's technology followed by professional services and everything else kind of drifting off in a long tail. This adoption of software development just by itself, I think is kind of hugely consequential. And it's again, it's kind of interesting to compare this with AWS because what AWS did was it was an order of magnitude, maybe two orders of magnitude change in how quickly and easily you could deploy software. The famous example here is that Instagram had something like 50 million users and seven employees when Meta bought it because they didn't need to buy all the servers and build all the databases themselves. They could get that as marginal cost on AWS.
And vibe coding and now coding agents in the last couple of months are very clearly a transformative change in how quickly and easily you can build software and what levels of abstraction you think about in building software. However, what does that do to software development? Well, obviously in the last couple of weeks, we've had people saying, "Oh my god, everyone's just going to vibe code SAP." I always describe this as saying kind of imagine the people in the office, like the TV show The Office. Imagine like Gareth and Fat Keith coming into the office one morning and saying, "Hey, we're not going to use Salesforce anymore. We're just going to write our own." It's probably not going to work very well. What is more useful, I think, is to break this apart.
So clearly this is just a fundamental change in how quickly you can build software that you might have wanted to build 5, 10, 15 years ago. Now you can just build it much faster and that changes a bunch of things. Secondly, again very obviously there's a bunch of actually new things that you couldn't have done with software at any cost 5, 10, 15 years ago and that will manifest both as features in existing products and as new products and as people trying to unbundle existing companies. Thirdly, there will be some places where it turns out that AI is a way better way of solving that problem than traditional software where people have very large, very complicated expert networks and expert systems. It may be that LLM is just a much easier, much better way to automate that. And so there will be some places where legacy software gets disrupted by some new AI-based thing.
What's not going to happen is everyone will write their own software. The point of this slide is like writing software, writing the code isn't the hard part of building software. We've already got lots of software. We already have way too much software. In fact, for anybody who's actually trying to buy this stuff, writing the software isn't the hard part. The hard part is working out what it's supposed to do and working out who would want it and who would use it and working out what the actual problem is. Two famous quotes here from Steve Jobs, like people don't know what they want until you show it to them. People who are working inside a company don't really know what SAP does, don't really know what their software does and certainly wouldn't know what it is that they would want to build to replace it. You can't start with the technology and work back to the experience. You have to start with the experience. You have to start with the problem that you're trying to solve. And working out that that problem exists is the hard part of software. You know, the whole reason that you have go-to-market and sales is that most of the users don't actually know that this is a problem and don't really understand why they would buy it. They're not going to go off and build something to solve it for themselves.
That gets you to this hilarious story from Accenture, sorry if there's anyone from Accenture in the room, that they're penalizing senior employees if they don't use AI enough, which says to me that they haven't built enough tools that are useful to senior employees. Maybe if you allowed them to use it to write performance reviews then they'd have massive adoption.
So how do you actually build AI software? What would you do with this? And it seems to me there's a bunch of problems that you have to reconcile. On the one hand, you have these new capabilities that you simply couldn't do with software before. On the other hand, this stuff doesn't work like software. It doesn't work the way we expect software to work for all of our lives. It doesn't produce deterministic answers. Traditional software is deterministic. LLMs are probabilistic. So the best framing that I heard of this a couple of years ago is somebody said that generative AI tends to be good at things that computers are bad at and bad at things that computers are good at. And if you try and use it for stuff that computers are good at you will say well this doesn't work, which is kind of like picking up a PC in 1980 and saying does this have the same reliability of a mainframe or looking at the web in 1995 and saying can I use this to build AutoCAD? The answer was well no but it does something else.
And so you kind of have to reconcile the capabilities with firstly the error rate and secondly what people kind of call the blank screen problem which is you just go to ChatGPT and think well what should I ask? And then there's a bunch of stuff that you build around that. So you think about the right use cases, about where the human in the loop goes, about what the tooling and the process and the data and the GUI and the guardrails and the go-to-market should actually look like, which of course is how software companies work. That's what software companies always do.
How would you work out what you would do with this? Well one answer is that you can go to your systems integrator, so you can go back to Accenture and you can joke to them that they're not allowed to use AI, but Accenture said last quarter that they had done something over $2 billion of new generative AI bookings. They also said that they're going to stop reporting this number because effectively everything is now generative AI so it's no longer a meaningful number.
Another answer might be everyone does pilots. This is survey data over the last three years from Bain. Again you can see the pilots are concentrated in technology, in customer service, in marketing. You can also see well you do development, you do pilots and then you move stuff into production and that kind of takes a while. Stuff doesn't go into production immediately and sometimes the pilots don't work. But over time that gets deployed out into the company.
The other side of this again is that every startup right now is out there trying to build stuff with AI. This is Y Combinator which is a big Silicon Valley incubator, effectively every startup today is trying to build something with AI. And what all of those companies do in effect is they unbundle something. In the past that meant they tried to unbundle Excel or email into a company. They try to unbundle Oracle into a company. They try to unbundle Google or Microsoft or SAP into some dedicated company, which is the reason the typical big company today uses 400 to 500 SaaS applications. It's the reason why you saw all of those different marketing tech companies, they're all unbundling something out of Google or something out of each other. And today they're all trying to unbundle ChatGPT, they're all trying to find some use case or some specific point of leverage to build a new thing.
Now when you do that, when you go out and deploy things, then sometimes that can change everything. This is a chart about barcodes. So this is the number of average SKUs stocked in a US supermarket back since the 1950s. And in 1974 they deployed barcodes and UPCs and laser scanners. And that meant that you didn't need to do a stock take every Friday night with people walking around the store with a paper clipboard. But once you could do that, then you could completely change how you thought about your inventory because you didn't risk running out of stock in the same way and you didn't need to carry the same amount of inventory and you could use EDI to automatically reorder products from the manufacturers and you shift the whole nature of the capital structure, the inventory, the deployment model for the grocery business and you get at a peak to something like five...
times more products being sold because you had automated the way you thought about managing your products. This is not really what anybody was, what was on anyone's roadmap when they deployed barcodes. They just thought it was going to save them money. But once you've done that cost saving, then that transformed how everything works. And that takes me to just thinking about what that might mean now. How is it that you change the questions? Just sort of kicking around thinking about how stuff has moved in the last couple of months, couple of years. Today you can be on a Zoom call and you can say, 'Well, show me the data that we're talking about on this call.' And that wasn't something you could have done a year ago. Now you can ask that question. You could build software that could answer that. But then you could also push this way further and say, 'Well, hang on a second. This system has every call in our company and it's had every call for the last two years.' And that's data. That's data we've never had before. So now we can say, well, what are people talking about on all of their calls? What are the patterns within that? That's data you never had. And then you can push a little bit further and say, 'No, I want you to answer questions that I would never have thought about before.' The questions maybe that wouldn't have been questions. Pull the whole thing inside out and answer stuff in a completely different way.
And that gets me to talking about change. I think one of the kind of feelings that I encounter talking to big companies now around AI is to go back to this framing and say, well, we've deployed a bunch of stuff. We've got generative search and we've got a SKU tagging project and the marketing team are all over it. We've built a bunch of things. Are we done now? Everybody's got a chatbot. Next quarter everybody will quietly remove the chatbot and pretend that never existed. But meanwhile we've deployed some stuff, but what's next? What are the actual innovations and potentially the disruption that will come from this? Because it feels like we're not done yet. And as sort of Yogi Berra — I always thought Yogi Berra was a cartoon character. I then discovered that he was actually an American basketball coach. No, baseball coach, sorry. And he's famous for saying this: predictions are hard, especially about the future. Imagine being in 1997 and saying, 'What does the internet mean for our company?' What would you have got? You'd have got some of it maybe. Think about how many things you wouldn't have got then, and that's kind of where we are again with where we were with mobile in 2005 or 2010. It's where we are with AI today. But I think there's a couple of places that one can at least go and look and think about what might change. The first of them is to think about what the internet changed and how the internet changed information and changed aggregation and changed e-commerce, and to think about how LLMs might change that. The second is everything that the internet did to physical assets, the way that it changed what it meant to be a physical asset. LLMs are probably going to have a similar kind of change. And then the third is to talk about price elasticity.
So first of all, talking about aggregators, it's always sort of intrigued me that Amazon and Google and Meta don't really know what you buy and they don't really know why. Amazon doesn't know what's in that box. It just knows it's a scoop. Amazon has 700, 800, 900 million products on sale. They sit in warehouses that look like this. Rather than mess about with copyright, I just got ChatGPT to make me an image. Amazon knows that you bought this there and other people who bought this bought that. So you might buy that, but it doesn't know what this or that are. This is why you get these jokes like, you know, hey Amazon, I bought a toilet seat, I'm not collecting toilet seats. You don't need to give me more of them. My favorite one here is Amazon emailed me to tell me that as somebody who's bought books by Tacitus, you might like to know that he has a new book out. Which is either a time machine or an amazing archaeological discovery or just a system that doesn't actually know what things are. It just does correlation. The same thing for Google. The same thing for Meta. They don't know what any of those things are. They're just doing correlation.
Now if we think about what you can do with a system that knows what those things are, or that has a sort of fundamentally different level of correlation, a much higher level of abstraction. Today you can go to ChatGPT or Gemini or Claude and say, give me some Christmas ideas for a teenager who likes UFC, which is the fighting sport, and football, and it will suggest toy presents. It will suggest different things. It will suggest courses and classes and TV subscriptions. That's not a Google search. It's not an Amazon search. It's going to get you very different kinds of results. It's a question you couldn't have answered before. Secondly though, you can kind of push further and think — who's looking at these slides — there's an old line that a computer should never ask you a question if it should be able to work out the answer. And when people said that, what they were thinking about was like the computer shouldn't ask you what sound card you have. The computer shouldn't ask you whether you have a color screen. The computer shouldn't ask you what printer you have. It should work it out. But now, why do I have to go to my fridge and then look at the recipe and then go to Instacart or Deliveroo or Ocado and work out what I should be buying? Why don't I just take a picture of the recipe and say, 'Hey, just buy this for me.' So, you get these kind of successive levels of automation. Then you can push it a little bit further and think, well, look at my Instacart, look at my Instagram, look at my home, look at some data about me, and then you work it out. You tell me what coats I should buy. You can look at my Instagram. You tell me what cushion or lamp I should buy. You look at my home. You shift those questions of what you can buy, and of course that flows through to advertising, to commerce, to retail, to marketing, to everything else.
Secondly, I talked about the unbundling of physical assets, and I think everyone kind of understands this point, that if you were in the newspaper business or the TV business before the internet, you were doing two things. You were making TV shows or journalism, or you were running a bank or you were running a taxi, but then you owned a physical asset that you needed to have to deliver that. So if you were a retail bank, you were a bank, but you were also a specialty retailer. You had half a fleet of shops, and today you don't. The need to have those physical assets became separated from the underlying actual product that you were delivering with that, and that rolled across a very broad category of different products. Now, how do we think about what LLMs might split apart? And the way I would kind of think about this is: who has a building full of people doing something really, really boring in order to deliver their product? In the US that would be health insurance. It's not a concept that anyone here is familiar with, fortunately, but I live in New York, so I have to know about this. But there's a whole category of industry where you need to have a bunch of people moving pieces of paper around, looking at screens, doing something that's kind of boring and pointless in order to deliver the product. What is it that those will change? How many of those are there that will get changed? And of course, that wasn't obvious for the internet. There's a joke that newspapers looked at the internet and thought this is going to be great because their printing bills will go down. And that was obviously true, but also kind of not quite the point.
And if we kind of extend this a little bit, the way I used to talk about machine learning was that AI gives you infinite interns. So, you would like to listen to every call coming into the call center and tell me if the customer is angry or if the service agent is rude. Look at every credit card transaction and tell me if there's something weird about this transaction. Maybe you could extend this now and say it gives you infinite associates. Part of the point of this analogy is like you don't need to be an expert. You just need to be a mammal in order to do this. But on the other hand, you're still going to need to check the output. You wouldn't take the work that your intern had done and just give it to the client. But it would still be really useful if you had 10,000 interns in your office who you could just send off to go and do stuff for you. And so this gets us to kind of a price elasticity question. Like if you can do the same amount of work with fewer people, if it becomes way cheaper and way quicker and way easier to do any given piece of work, do you do the same amount of work and just have way fewer people, or do you do more work with the same number of people, or with more people? One of the ways I express this, kind of relevant to where we're having this discussion today, is the joke I used to make was that young people won't believe this, but before Excel, junior investment bankers worked really long hours. Now, thanks to Excel, Goldman's associates get their work finished at lunchtime on Fridays and can go home. Same thing with lawyers. Before word processors, junior lawyers worked really long hours. Well, why is that not what happened? Why is it that Excel resulted in more bankers and more accountants working longer hours? Secondly, was employing lots of people, was having to have a whole bunch of people doing that, your barrier to entry? Was that your barrier against competition? And then thirdly, what becomes possible if you don't need millions of people? What were things that you would never have done, you'd never have thought of doing, because the only way you could do that was to have like a million people working on it? And now you can just get the machine to do it.
This is an illustration of my Excel joke. This is the number of people employed as accountants and auditors in the USA as a percentage of the population. And so this goes through adding machines and then it goes through mainframes and then it goes through PCs. And the process continues after 2000 as we go to the cloud. Computers and automation didn't result in fewer accountants and auditors. They resulted in more, maybe just because it was easier to try and trick them as well. Price elasticity does not necessarily result in you doing less. Sometimes it results in you doing more. This is the same thing since 2000. The average cost of a US audit has been basically flat over the last 20 years, even as again we've had successive waves of automation and technology rolling through the industry that didn't actually change the cost of what it got done because you just did different stuff.
And I think there's a really kind of general conceptual point to think about here, which is: what was the actual job to be done? What were you really doing? What were you really buying? Were you buying the endpoint to a logistics chain? Were you buying the answer? Were you just wanting somebody to get something done? In which case, you went to Amazon or now you'll go to ChatGPT. You don't really care where you buy it or what it is. You just want the thing. You just want the answer. You just want the spreadsheet. You just want the box. Or were you more interested in experience and taste and judgment and preference and expertise and authenticity? Did you want something more than just automating that thing? In which case, guess what? Most people still buy stuff not on the internet. Amazon didn't conquer all of retail. You'll still want something else.
And then just some final comments, and I think we have time for some questions. As I've been talking about AI for the last couple of years, it seems to me that all questions about AI have one of two answers. The answer is either: how did it work the last five times? So, should we go to Microsoft or Oracle or to a new vendor or buy it ourselves? Well, how did you do cloud? How did you do SaaS? How did you do the internet? How did you do mobile? How should we work out what the right places to deploy this are? What should we spend? What are the best divisions to look at first? Well, how did you do cloud? How did you do mobile? How did you do machine learning? The other category of question is: no one knows. So, how long will the scaling continue? When will the bubble pop? How good will these models get? Will it kill us all or turn us all into pets? Like, check back in five years' time. We'll find out.
But meanwhile, we should also remember how many times we've been here before. This is an IBM ad from the early 1950s that said, 'An IBM electronic calculator is like having 150 extra engineers.' This is basically every AI pitch today. Except that most of the AI pitches today hopefully aren't just a sea of white men, all with receding hairlines, all wearing exactly the same clothes. There is actually a chart you can find of tie sales in America by color and shirt sales in America. Basically, in the 1960s, everybody in America wore a white shirt and a black tie, and then it all kind of changes. Sorry, that's completely irrelevant, but just kind of interesting to this picture. Conformism, I suppose, is probably the point.
This is a US government report from the mid-50s on what at that time was called automation, and they then were just starting to call it technology instead. And one of the things that they called out from automation, amongst all sorts of other kind of more obvious industries, was electronically controlled elevators. Now, what is an electronically controlled elevator? Well, the answer is it's an elevator. And this is what they're getting at. This is the elevator in my apartment building in Manhattan. We have what's called an attended elevator, which means that you walk in and the doorman says, 'Hello, Mr. Evans.' And they get into the elevator and they pull the lever and they drive you to your floor. This is basically a tram. There's a forward button. You push it for forward, you push it back. It's an accelerator and a brake. You have to be sure to give him a good tip at Christmas. Otherwise, he might not come to your floor quickly when you press the buzzer. This is literally true. They might go to the other floor first and you're listening to the elevator going past and like, obviously I need to give him a bigger tip next year. This is what people meant by an elevator.
And this is the number of people employed as elevator attendants in America. And this is the perfect chart of automation because the first half is America deploys lots of elevators and the second half is that they automate all of them. It costs about half a million dollars to automate an elevator, I've discovered, because it's in my assessments, which is why we haven't done it yet. But the point of this is like, to begin with this was exciting and it was technology and automation, and then it just disappeared. When was the last time you used a lift? You didn't get into it and say, 'Ah, I'm going to use an automatic elevator now with artificial intelligence, an electronic elevator.' No, it's just a lift.
And this is a famous quote from an AI researcher called Larry Tesler from 1970 who said, 'AI is whatever the machines can't do yet. Once it works, people say, well, that's not AI anymore. That's just software.' I really think this is what's happened with machine learning. Like 15 years ago, image recognition really didn't work. 10 years ago, machine learning means it works, and that's AI. Today I hear people saying, 'Oh, that's not AI. That's just image recognition.' Well, wait a minute. 15 years ago that was witchcraft. Like now it's just software. It's not AI, that's just image recognition. That's just voice recognition. And it's kind of like the word technology. Technology is whatever's new. AI is whatever doesn't quite work yet. Once it's been around for 10, 15 years, it's just the way the world's always been. And we move on to something else. And with that, I will say thank you. And I think we've got five or 10 minutes for questions.
