I was asked to come and explain everything that's happening in AI and technology, and not talk too fast, and not talk too long, take too long. I'll maybe manage two of those, maybe not all three, but just to begin with.
Sorry, I think it's kind of useful to sort of set a framing for how the tech industry tends to work, which is that every 10 or 15 years we have a new platform and all the innovation and the company creation happens around that. And so from the mid-60s to the late '70s we had mainframes, and at peak there were 70 or 80 mainframe, 70 or 80,000 mainframes in the world. Then from the late 70s through to the mid-90s the center of the tech industry was the PC. By the early 1990s there was as many as 70 or 80 million PCs in the world. That was a giant new industry. Then we had the web, and then smartphones. Today there's something in the order of five billion, maybe five and a half billion people in the world who have a smartphone. And now generative AI.
And the thing with each of these is that they follow a sort of S-curve. To begin with, they don't really work very well. Nobody's doing it. It seems kind of like a dumb idea. Generally, the important new thing comes as a stupid toy for rich people or children or both, and then it works, and I think you want to get a job in it or you want your children to get a job in it, and then you reach a point where it's boring and it's happened, and that's just the way the world has always been. The new iPhone is amazing and it's also boring because that's now become a mature industry. And then you start wondering, well, what's the next platform going to be? And it's pretty clear now that the next platform shift is generative AI. As you go through a platform shift,
hang on, can I? Yes. As you go through a platform shift, a couple of things happen. So firstly, all the innovation and investment and company creation moves over to the new platform. So anyone who's flown recently will be, or tried to pay a check into their bank, will notice that there are several companies that use mainframes. That's why you go and pay, go to a government website and it says, sorry, we're down for maintenance this week. That's still running on a mainframe. And so the old things are still there. PCs, big companies still run on mainframes. Big companies still run on PCs. The web isn't going anywhere. But all the innovation and all the new stuff happens on the new platform. Within the tech industry, everything changes and all of the barriers to entry and the basis of competition from the old platform stop mattering, and what matters is the new platform. And so you have new gatekeepers and you have new value capture and you have new businesses, and generally you find much, much larger markets. Mainframes sold to large corporations and governments. PCs sold to every corporation on earth and every middle-class family. The web expanded the market for both of those, and then smartphones took us from maybe a billion PCs in 2000 to, as I said, five billion smartphones, six billion smartphones today.
Outside of the tech industry, outside of Silicon Valley, people kind of scratch their head and look at this new thing and say, well, what does this mean to us? Is this just kind of a new tool that we'll all buy and use and then life will go on? Or is this going to create new revenue opportunities and new businesses and maybe new competitor threats? Or is this some kind of existential threat? This is actually going to change the whole nature of our industry and maybe put some of us out of business. There's an old joke that the newspaper industry looked at the internet and thought, this is going to be great, our printing bills will go down. And their printing bills did go down, but not quite for the reasons that they expected.
And as you go through these shifts, the people who are winning at the beginning are generally not the people who end up capturing all the value. So this is market share in PCs from the mid-70s when PCs first started appearing. There were dozens of companies making personal computers before IBM and before Apple, all of which are forgotten now. Apple talks about creating the personal computer revolution, but Apple basically lost PCs as well. The same thing applies to web browsers, applies to search, applies to social, applies to smartphones. In each of these platform shifts, the people who are leading at the beginning may not be the people who end up working out the right way to do it and the right way to capture the market and the right way to build a giant business. And the challenge here is that whenever this new thing happens, we don't really know how it's going to work. And we kind of forget now how unclear it was how the internet was going to work and how unclear it was how smartphones were going to work. There were all these kind of other ideas floating around. And if you remember back in the kind of the mid-'90s, which is I'm just old enough to remember what that was like, what computers were like then, we all kind of knew that everyone was going to have some kind of computer thing connected to some kind of network, but we didn't know it was going to be the internet. There were lots of other networks around. We didn't know it was going to be a PC. Everyone was talking about the information superhighway, which if you think back is saying the whole, the very term implies a sort of centralized corporate system that's run by the cable TV company instead of what happened, which was this radically decentralized system. It wasn't clear it was going to be the web. It wasn't clear how the web was going to work. Even when it was clear that it was going to be the web in say 1995, Google was five years away and Facebook was 10 years away. So you can kind of have 90% of how it's going to work and then still not know what's going to happen next.
The same thing happened with mobile internet. People were excited about mobile internet for 10 years before the iPhone and no one knew how that was going to work or where the value was going to be. And if you talked about how mobile is going to work in 2000, you would have presumed that it would be a PC accessory. That was the whole conversation. Whereas what actually happened is today PCs are smartphone accessories, and you certainly wouldn't have proposed a failed PC company from the 80s run by some guy called Steve Jobs as the company that would take all of the value from it. The same thing now with generative AI. This is clearly a huge and important thing. We don't really know how any of it's going to work. There's lots of companies and ideas and acronyms floating around. Some of those will turn out to be the thing. Some of them won't. And we may look back in hindsight and realize we weren't even asking the right questions.
When you go through these cycles, another thing that happens is you tend to get a bubble. Not every time, but quite often. And this is a classic book about the history of bubbles called "This Time Is Different." There's a saying in finance that the most dangerous words in investing are "this time is different." And kind of this has a double meaning though, because it's always true. It always is different. Whatever is happening now will probably turn into a bubble at some point, but it won't be the same as the dot-com bubble, which after all was funded by IPOs and public market investing, whereas what's going on now is being funded by cash flows from giant corporations. Every bubble is different. Every bubble is driven by some fundamental change in the nature of society, but it can still be a bubble. That means we get an awful lot of noise. We get a lot of hype. We get a lot of anti-hype. We just heard about this study from MIT. If you actually go and get hold of a copy of the study, you find that it doesn't actually tell you how the 95% number is defined. It doesn't give you a sample size. Doesn't really have a clear definition of what success and failure are. And it isn't really from MIT either. But apart from that, it's a great study.
We also have a bunch of other kind of assertions floating around. This is Anthropic at the beginning of this year saying that within six months AI would be writing 90% of code. This is not like the most future-forward room, but I guarantee there isn't really any company now where AI is writing 90% of the code. And so there's a lot of kind of noise and hype that goes around. It's also worth kind of picking up that word disruption, which is a word that just kind of gets thrown onto the table and saying, well, that can mean different things. So Uber fundamentally changed the nature of taxis. It didn't sell software to taxi companies, and Uber demolished the existing taxi industry and created much larger demand for rides. But Airbnb completely changed what you meant when you said a hotel, but it didn't demolish the hotel industry. Very obviously, it didn't replace this hotel and the purpose that we're all here for this. And so Airbnb turned out to be mostly additive to the hotel industry, whereas Uber completely changed the nature of the taxi industry. So, you know, when you say disruption, you have to push a little bit harder and think, well, what is it exactly that you might mean by that?
However, if we step back from this, when the dust is settled, the world has changed. Online, this is kind of my favorite illustration of this. Online dating is now roughly 60% of all new relationships in the USA. This is something that's gone from being kind of a joke and kind of weird to being just a basic part of how the world works. If you want to meet somebody now, the best way to do it is to open an app because then you're not dependent on whether your colleague's best friend's sister might be looking for a relationship. You're not dependent on randomness. You actually solve a pooling problem. For gay relationships, obviously this number is closer to 90% because that's an even bigger pooling problem.
On the enterprise side, go back to those platform shifts. Each of these platform shifts have meant sort of an order of magnitude more tools and an order of magnitude more software. So when you rented a mainframe from IBM, you got one piece of software. In fact, there wasn't really such a thing as a software company. That's something that Microsoft invented. When we went to PCs, you had dozens of apps. Then when we went to cloud and software as a service, we go to probably 400 or 500 apps. That's how many apps a typical large corporation in the USA is using today. And using 400 or 500 apps sounds kind of crazy until you, and I kind of, the best way I can think of thinking about this is to remember or to think about when you make a list for when you're going to get married and you say we're going to have a small wedding, we'll have like a dozen people, and then you start making the list and you realize that you've got two brothers and she's got a sister and a brother and they're married and they've got children, and you discover that a wedding can have five people or 50 people and kind of nothing in between. And it's kind of the same when you're building software in a big company. You think, well, we've got Office and SAP and Salesforce. And then you start making the list and you realize, yes, we do have 400 or 500 applications. The interesting thing is that what all of those are doing in effect is unbundling Salesforce or Oracle or email or Excel. And everything that you do with Outreach, you could do in a spreadsheet and email, but that would be an incredibly bad idea. Every time that a task scales to a certain point, you discover that you want to build that out into a new piece of software, and SaaS made it possible to expand that massively.
As we now go through to generative AI, we should expect as a default that we will just have an order of magnitude more software. There is, however, one way that this platform shift is different, which is that with all of the other things that I was talking about, we kind of knew what the physical limits were. Like we didn't in like 1992, in say, in like 2010, we were excited about the new iPhone and it was much better, but we didn't know what it was going to look like, but we knew it wouldn't fly. We knew it wouldn't have a one-year battery life or have a projector that could fill the wall. You kind of knew what the physical limits were for these technologies. In the late 90s you knew that telcos couldn't give everybody on Earth fiber next week. Whereas with generative AI, we don't have a good theoretical understanding of why these models work so well. We don't have a good theoretical understanding of what would happen next. And so the science kind of keeps moving without us really knowing why or what will happen next, which means that there's a very wide spread of possibilities for how much better they could get.
Hence this is Sergey Brin, co-founder of Google, saying basically we're on track to get AGI, which means HAL 9000. It means a computer that is actually as intelligent as a person in the same way that people are intelligent. On the other hand, Demis Hassabis, who created DeepMind, is one of the world's pioneering AI scientists, said no. This needs an unknown number of unknown breakthroughs, unknown number of years in the future, and we have no idea when we're going to get to AGI. Nobody really knows.
So what this means is, if I go back to my platform shift diagram, there's one possibility which is this is kind of going to play out like mobile and before that the web and before that PCs. It will change everything and in 10 or 15 years' time we'll have got used to it and there'll be something else. The other view is to say no, this is actually a much bigger shift. This is more like computing or electricity or the internal combustion engine or flight. It's a much bigger change in the nature of what we can do, what we can automate, how businesses and society can function. We don't really know. We have no theoretical model that will tell us which of those it is. But if this is only as big as mobile or only as big as the web or the PC, that kind of feels big enough to be excited about.
And so that gives us some questions. The first is, as I suggested earlier, how is this going to work? We don't really know how we're going to build this. We haven't quite got like a fully worked out sense of how these models are most useful, what kinds of tasks you can do with them. And then we have of course have to work out well where is the distribution, where is the new value capture, where is the new value destruction, where are the billion, trillion dollar companies going to get created and destroyed and changed around all of this stuff.
So let me go and talk about what's going on inside the tech industry. And I actually really only need one slide for this. This is Sundar Pichai from Google. The risk of underinvesting is bigger than the risk of overinvesting. There's a great quote from the TV show Silicon Valley about 10 years ago where the character who represented Google, the CEO of the company that was supposed to be called Google, 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 this kind of captures the issue here. If this is the center of the tech industry for the next 15 years and you miss it, that's a much bigger problem both from the lost opportunity and also the threat to your existing business. That's a bigger problem than overspending your capex for two or three years and pulling future capex forward. If you just spend the next five years' capex this year, that's a smaller problem than being out of position for the whole future of the tech industry for the next 10 or 15 years.
So what we've had since the last three years since ChatGPT launched is basically three years of FOMO, three years of fear of missing out. Step one is, oh my god, this works. This changes everything. Step two is this keeps getting better with more capex because the only thing we have found empirically is that bigger models, more compute produce better results. But then the question is, well, how much more capex and how much better? And that gets you to this chart. You may have seen versions of this. This is infrastructure spending by the big four companies plus Oracle and CoreWeave. They will spend something in the order of, this chart keeps changing. In fact, these numbers have gone up by 50% since the beginning of the year. They'll spend something in the order of $400 billion this year building infrastructure. And that number has now almost tripled in the last couple of years. A large portion of that money is dropping straight through to Nvidia, which is a kind of an astonishing chart and it's still going straight up and to the right. And Nvidia can't keep up.
This is also of course a TSMC chart because Nvidia uses TSMC and a bunch of Taiwanese contract manufacturers to actually produce the product and sell it to people, and those people can't keep up either. So you've got this enormous surge in investment in infrastructure flowing through to an investment in semiconductors. This is the CTO of Microsoft saying we basically haven't been able to build capacity fast enough since ChatGPT launched because these things are so computationally intensive and it's so difficult to build that much infrastructure that quickly. Meanwhile we had these three quotes from the last quarter's results in the last couple of weeks where depending on how you interpret these, all of these companies are expecting a further substantial increase next year. Some of them are actually expecting that the increase next year will be larger than the increase that we've had this year.
All of that gets you to a chart like this, like well what are we actually talking about here? These are various estimates for global data center capacity measured in gigawatts. 18 months ago nobody in the tech industry knew what a gigawatt was. My father used to work for GEC in the UK and I had to ring him up and say, so daddy, what exactly is a gigawatt? And so now suddenly we've got these estimates of gigawatts of data centers and some range of estimates for substantially more gigawatts of data centers, something like a tripling in global capacity, something like three, four trillion dollars maybe. These are very, very sort of impressionistic numbers. Somebody from KKR the other day coined the word bragawatt, meaning that these companies are kind of out bragging about how much infrastructure they're going to build, partly kind of as a competitive weapon to intimidate other people, like you can't possibly catch up with us because look at all these deals we've signed. The funny thing about a lot of these deals is they don't actually have time periods. So the other day OpenAI signed a deal for $38 billion of infrastructure with AWS, but they didn't actually say over what period that would be. Is that over 10 years? Is that over five years? Actually when the announcement came out I thought it was missing a zero. It was like only $38 billion, not $380.
Now it's worth kind of putting this in context though. $3 trillion, if you actually try and annualize that, it's probably five, six, $700 billion a year, which is a lot of money, but it's kind of what giant capital-intensive global industries spend. It's in the ballpark of what oil and gas or telecom spend. The US private sector, so UK, it's in the ballpark of what a large economy like the UK in total spend. So it's a lot of money but it's not like let's build a tower to the moon money. It's just like this is what oil and gas spends kind of money.
Now one important thing to understand about this is that this is mostly being funded out of cash flow. So the four big hyperscalers have become massively more profitable in the last five and 10 years. And effectively all of that increase in profitability is dropping through to capex. So this is free cash flow then with capex stacked on top of that. And you can see the variance in those companies. But you can also see effectively that the increase in profitability is dropping through to go to Nvidia. Now that's not necessarily going to carry on being true. Meta has raised I think $60 billion in the last couple of weeks, $30 billion with an SPV and $30 billion of debt. Oracle of course doesn't have this kind of money. Some of the estimates of Oracle's future cash flows are that it will be spending in excess of 100% of its cash flows building infrastructure, which obviously gets people in Wall Street very excited. So we are kind of getting to the end of this just coming out of cash flow.
And meanwhile OpenAI has decided that it wants to join this club without having a giant legacy cash-generative advertising or enterprise software business. So OpenAI said a couple of weeks ago they've now got estimates for about 30 gigawatts of capacity at something like $1.4 trillion. They said in a very kind of aspirational directional sense that they would like to be building a gigawatt of infrastructure a week at $20 billion per gigawatt. So that's about a trillion dollars a year. That really gets us into bragawatt territory. It's equivalent to about two-thirds of the total global data center base today. So these are very kind of sporting kind of numbers.
And of course that gets us to this phrase circular revenue. So without these companies, this company actually having its own money, we've got lots of kind of money going around in circles and lots of kind of mutual deals. So as the sort of slide title said, it's also of course worth pointing out of course where some of this money is ultimately coming from. So Nvidia has said that they will invest $100 billion in OpenAI to build infrastructure. Nvidia's cash flow, of course Nvidia does have that cash flow. It had $70 billion of free cash flow in the last 12 months. But that money is coming from well Microsoft and Amazon and Meta. So Nvidia is taking money from Microsoft and giving it to OpenAI to compete with Microsoft, and OpenAI is then taking that money and giving it to AMD to buy chips to compete with Nvidia and also Broadcom to design its own chips to compete with Nvidia. And meanwhile taking equity in AMD, which is always an entertaining part of the cycle. I buy something from you and take stock from you, because the fact that I bought from you has made you more valuable.
So yes, this does kind of, some of this points to a certain point in the cycle. But let's kind of step back from this and think, well, if you're not an Nvidia shareholder or a Google shareholder, like so what? Why do I care about all of this? What have we actually got from all of this money? Well, the models are still getting better. We have many more models. We have many more acronyms. We have open source models. China has roughly caught up with those models. But we don't have any apparent differentiation. There are no apparent barriers to entry other than just spending the money. There isn't really any sense of where there would be a moat and there's not really any sense of where there would be kind of value capture in building these things or where indeed what the products are going to look like.
This is a chart of how many models have been produced in the last three years. The vertical axis here is a kind of an aggregate of a bunch of different benchmarks. And the kind of the takeaway from, and there's many other ways that you can do these. Part of what this chart is hiding is that you have many models at different prices. So the reason that you, for example, you see a string of OpenAI models with different scores is that the ones with lower scores are cheaper because they're smaller to run. But the real kind of takeaway from this is that to begin with OpenAI were the first people to get this working and then there were lots of other people trying and now OpenAI is sort of at best first amongst equals. We have Chinese models that are sometimes at the top of the scores. We have open source models that are sometimes at the top of the benchmark scores. There are lots and lots and lots of models. And the benchmark scores for all of those tend to converge.
You'll notice here I've truncated the axis, which I don't normally like to do. But the point of this is that every week there's another model. Every week the people who are at the top of the benchmark charts are changing and they're all kind of within 5% of each other or even a couple of percentage points of each other depending on which benchmark scores you want to use. Now if you dig into very specific use cases, so if you look at coding or making video or making images, there's much more divergence and there are clear leaders. But from kind of general purpose models for doing kind of general purpose tasks, there's anything from half a dozen to a dozen models from different companies that are all roughly the same.
Now what's interesting is that there's a lot of divergence in consumer adoption. So even though OpenAI's models is roughly the same as Google and as Meta, and as Claude, well Meta's fall away a little bit, and also as Claude, you can see Claude has got completely different kind of consumer adoption to OpenAI. And this kind of prompts a question: is this about the technology or is it about brand or is it about distribution and product? Like what is the basis of competition here? You're spending hundreds of billions of dollars to get basically the same product and then you compete with stuff that you have to build on top of it, which means some people have started talking about this as a new browser wars. Younger people in the room, like ask your parents what Netscape was.
Where again you had fundamentally a commodity product that was competing on brand and distribution, and indeed there's a lot of kind of corporate strategy going on. So the classic function, one of the obvious things that happens in a platform shift is the incumbents try and make it a feature. You hear about the sexy new thing and the legacy company says, yes, we have that, and if we don't, we'll have it next year. That was what Microsoft did in the 90s with the internet, it's what Google's done, it's what Meta. And so all of these companies are now basically spraying LLMs all over their products. You can see the same thing from Salesforce and from Oracle and from SAP. Every incumbent is saying, yes, we'll make this a feature, we can do this. Meanwhile the cloud service providers say, well, you want to sell cloud, we can sell cloud, that's what we do. Amazon is very happy to sell you more cloud to run models on. And of course, OpenAI wants to break in, as I sort of suggested earlier, and it's kind of worth just kind of
Pulling all of those OpenAI threads together, really what OpenAI is doing is a sort of full-stack attack on the market. So they're building infrastructure, they're building an advertising model and e-commerce models, they're building new distribution platforms, their own app platform, social video, building a web browser, and then they bought Jony Ive and they're building robots and they're trying to cure cancer. So kind of everything — they're trying to be Apple and Google and Microsoft and Meta all at the same time, and issuing more stock every day to do that.
And it's worth kind of saying we've got some kind of very kind of exciting numbers, and you know it's kind of easy to look at charts like this where OpenAI is reporting 800 million weekly active users and say, wow, like everyone's using this already right now. But I think we should also kind of step back from this and think, well, actually the big questions remain wide open. Like, how actually do you compete when the products are commodities, when they're all roughly the same? How actually do you compete with the model, and how much product do you have to build on top of this?
Can you just kind of give everybody a chatbot and say, there you are, get on with it, or do you have to build product and tooling and interface and workflow and integration around that? That's exactly what we just heard. You can't give everybody in a regulated industry a chatbot and say, off you go. It has to be linked into tooling and processing and compliance and domain-specific knowledge. And how is it that you make this useful for people, even if that chatbot does have domain-specific knowledge? How do people work out what they're supposed to do with it? This gets us to charts like this.
And we can see kind of many variants of the same thing. This is a very kind of glass-half-full, glass-half-empty kind of chart. On the one hand, something like 5 to 10% of people in developed markets are using a chatbot, an AI chatbot, every day. On the other hand, 20 or 30% of people are using them, know what it is, they've got the account, they know how to use it, and they can only think of something to do with it once a week or once a month. And if this is kind of like the transformative future of computing that changes the world, why is it that you can only think of something to do with it once a week or once a month?
Now you could say, and as I said, there's many other sets of data that show roughly the same kind of thing. This is a spread of different surveys from different people that all kind of point in roughly the same direction. Sort of 5 or 10% of people using this every day, another 20 or 30% of people using it every week. And there's a question here: is this just early? Well, maybe, but that seems like a cop-out. It's more interesting to think about use cases. How many use cases are an easy, obvious fit for this technology?
How many people have a job where it's easy and obvious to think of what you would do with this? Who has a flexible job where they're free to decide how they do their work and they spend a lot of time thinking about how they could optimize their work and do it differently? Some people have that kind of job, but many people don't have that kind of job. And for everyone else, how much do you need to help them work out what you should do with this? So there's two observations from Steve Jobs: people don't know what they want until you show it to them, and you've got to start with the experience and work back to the technology.
You can't start with the technology and give that to people and say, "There you go, you work out what the experience is." That means, of course, that you have to build use cases and you have to build products. That's what startups are doing now. This is Y Combinator, a big Silicon Valley incubator, and basically every startup in Silicon Valley today is an AI startup. And what all of those companies are doing — what startups have always done — is try and find problems and find use cases and find things that you didn't realize were a problem or that you didn't realize you could automate.
And they unbundle. They find something that you're doing in email or Excel or Salesforce or Oracle and they unbundle it into some dedicated vertical tool. Now, of course, they find things that you were doing in ChatGPT and they unbundle that into a dedicated vertical tool. This is a classic quote from Jim Barksdale: there's two ways you can make money — you can bundle or you can unbundle. What startups tend to do is unbundle incumbents and turn them into new use cases. What OpenAI is doing, of course, is both of those at the same time.
So they make a social video app on the left here. On the right, they make an app platform — you can use other people's services inside ChatGPT. They make a web browser. There will be more stuff too. They're kind of trying all of these things all at once, trying to work out the right experience, the right distribution, the right combination of product. And so that gets me to go back to that question I posed at the beginning: where's the value capture? Is it in having the best models? Probably not. Is it in having the most capital? Maybe.
One of the things that we saw in semiconductors and we also saw in aerospace is that if it just becomes progressively more expensive to do the thing with each generation, that of itself can mean that you end up with two or three companies instead of 50. Is it then about having proprietary vertical data? Is it about having domain-specific knowledge, domain-specific data that's built around a specific problem? Is it conversely around the distribution, the go-to-market, the sales, the support, the user experience that you build around that?
In other words, it feels like most of the way that you capture value is that you go out and build a traditional software company, because the underlying model itself is just going to be commodity infrastructure that's sold at marginal cost. And what matters is the product and the use case and the problem that you're solving on top of all of that. And that takes me to what happens outside of technology. And so everybody in this room and everybody in every big company in the last couple of years has had this question: what's our AI strategy?
It's just a tiny little bit like that moment when everyone said, "What's our metaverse strategy?" Obviously nobody in this room had to do a metaverse strategy, but there were other people who'd had to spend a lot of time talking about metaverse. And the classic pattern for a big company is that you kind of go through a couple of steps. So to begin with, you absorb it. You use this to automate obvious, easy use cases, problems that you can already see, things where this is a really clear, easy, obvious fit.
A little bit later on, you start that, and that's kind of generally cost-saving, and you could call that bottom-line innovation. Then over time you start working out new products, new things that you could build, new revenue opportunities, new possibilities, new ways that you can bundle and unbundle. You actually start innovating as opposed to, say, using this to save money. And every now and then somebody will come along and disrupt, someone will come along and turn your whole business model inside out and say, we're not going to do it like that, we're going to redefine what this problem even is.
What's happened so far in the last two and three years is mostly that first step. It's mostly absorb. So we have people using this for software development, people using this for marketing, for customer support, and then for automation of point solutions inside every big company, inside every big enterprise workflow. It's worth just kind of pausing on coding because it does feel like, if nothing else happens — and lots will happen — this feels a little bit like the new AWS.
And what I mean by that is that AWS was an order-of-magnitude change in how hard it was to get software out of the door, how hard it was to build and deploy software. And it feels like writing software using AI as an assistant may be another of those changes in the abstraction layer that makes it an order of magnitude easier and quicker to build and deploy software. So the great joke here is, you know, AI means that no one will need any more software. Look, it makes it easier and quicker to write more software.
After that, how do you know what you should automate? Well, one answer is you hire some consultants. You talk about digital transformation. In fact, if you say digital transformation three times, then a partner from Accenture will appear in a puff of smoke and offer you a contract. And Accenture is now booking getting on for $2 billion of new deals around quote-unquote generative AI every quarter. And everybody has a pilot.
This is survey data from Bain. They're just about to put out the latest data point. I was hoping I'd have it yesterday. I think it's going to come out tomorrow. Everybody has done pilots. Some of those pilots have gone into production. Some of those pilots have gone out of production as they said, actually, that didn't work quite as well as we thought it was going to. Those pilots have been concentrated in the fields that I talked about earlier: so software development, customer service, and marketing.
And not every pilot works. That's kind of the nature of technology. The point of doing a pilot is that it may not work. And the reasons that they tend not to work are partly things that are specific to generative AI, so things like error rates, partly issues to do with integration with existing data. One of the interesting phenomena that I've heard about a lot is that companies finally sort out their data story and get everything in one place because they want to do AI.
And having done that, they can then do a bunch of stuff with SQL that they kind of wanted to do 20 years ago, but they never managed to get all of the data story lined up. So once you've sorted out your legacy systems, you do a lot more legacy stuff even without having to do AI. But again, the story across all of this is: how is it that you work out how to deploy new technology? Why do technology projects fail? Why do they succeed? This probably isn't any different from any of those.
I've had a question a couple of times in the last couple of months from CEOs at big companies that are sort of along the lines of, "How do I make my people use AI?" And I find this a very puzzling question because to me this is like if you'd said to them, "How do you make people use the web?" You imagine if you'd gone to your company in 1996 and said, "We're going to put a web browser and an internet connection on every desktop in the company. What's that going to do to our ROI?" Well, you probably wouldn't be able to measure it.
And if you did, your ROI would be negative because everyone would just be messing about looking at stuff on the internet instead of working. And you actually had to kind of tie that into use cases and workflows in ways that this was going to make this work better. You have to work out how you use the technology rather than just kind of spraying it at people. All of that means that the future can take a while. This is a CIO survey for Fortune 100 companies from Morgan Stanley.
Something like a quarter to a third of people have got at least one project in deployment, and a bit over half expect to have something in deployment by the end of this year. But another quarter still don't plan to deploy anything at the moment. This reminds me a little bit of a conversation I had a long time ago with the CEO of a utility company, a big water company. I gave a big presentation about machine learning and he said, that was very interesting, I think maybe innovation will be one of our top five priorities next year.
Which sounds insane until you think, well, this is actually a water company. Maybe they've got other priorities besides innovation. And so the new thing can take a while. And we kind of forget that none of this stuff existed three years ago, but then the new thing can always take a while. My favorite illustration of this is that if you're in the tech industry, cloud is kind of old and boring and finished. And it's a story from the 2000s, but it's still only about a third of enterprise workloads.
The other part of this story is that the CIO always hopes that they'll be able to get budget to move another 10 or 20%, and the CEO always says no. Now finally the CEOs are starting to say yes and move things to the cloud. Now when you go through this process, sometimes just automation can be a big deal. This is a chart of the SKUs in stock at US supermarkets back since the 1950s. And this is really a chart of barcodes and mainframes and EDI.
This is a chart that says once you can automate your stock take and you don't need to send people around the supermarket every Friday night with pen and paper to count what you've got, then it becomes possible to have 10 or 20 times more product in the store. And then you can change all sorts of other things about how your store works as well, and about how your suppliers work. And in the end, that gets you Walmart. It gets you a complete change in what it means to say supermarket.
And so you go from, well, we're just going to automate our stock takes, to changing the nature of the grocery industry. However, you can also, I think, step back from this and say, well, we've all kind of deployed a few things. Is that it? This is a conversation I had with a big company CMO earlier this year. He said, we've all had five or 10 AI presentations. We've had the one from Microsoft and Google. We've had the one from Bain or McKinsey. We've had the one from OpenAI and from our ad agency, and we've deployed some things. And is that it? It feels like there should be more than this.
And so I think again we kind of have to step back and think we don't know how this is going to work. This is kind of the comment I made at the beginning: imagine trying to work out what the internet meant for your business in 1997, what mobile meant for your business in 2010. And you'd get some of it. This is an attempt I made a couple of weeks ago to think, well, what are some of the really profound building blocks that will move? What happens if content creation becomes free, or certain kinds of content creation become free? If video creation becomes free?
What industries are protected by having lots of really, really boring desk work? What companies are protected, where their entire business is processing lots of really, really boring desk work? What industries have a business model that's based around having a building full of associates that they sell to you at a markup with a partner on top checking what they're doing? How do those industries change? How do we change the world of advertising and e-commerce if Google and Meta and Amazon have a fundamentally different understanding of what they have and what they know about you?
One of the ways I used to talk about the last wave of machine learning was that it gave 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. You don't need an expert for that. You just kind of need a 10 or 15-year-old for that. You would like — if you go back to the genesis of image recognition in the 1960s — it was, can we please try and spot tanks in satellite photographs? And that turned out to be a 50-year problem rather than a six-month problem.
But if you had enough interns, you could just look at all the satellite photographs and spot the tanks. What we have now is kind of pulling that inside out. What if you had infinite interns in your company? You could ask them to make you stuff instead of just recognize stuff. And part of the point of that analogy is you probably have to check it and you probably have to explain it to them before you just ship it to your clients. So that gets to another sort of set of questions around what does it mean to automate?
Do you do massively less work with the same number of people? Do you do the same work, or much more work, or different kinds of work with the same number of people? Or do you solve completely new problems that you could never have done before because you didn't have a million interns working for your company? One place to just kind of drill down on a little bit is to think about what that means for advertising and for marketing and for consumer. And I apologize for a very consumer use case, but this is kind of the stuff that's working right now.
So the global ad industry is about a trillion dollars. Outside of China, it's 800 or 900 billion dollars. About half of that is Google and Meta and Amazon. Those companies are going through the first of these stages. So they're using this to optimize their existing business. Google and Meta talking about significant uplifts in their ad business just from applying AI to it. Their clients are using this already to transform how they create assets.
Producing kind of five or 10 times or 100 times more imagery, more imagery and more videos for their advertising using these tools. There was a joke a couple of years ago that AI meant that half of AI would be turning three bullet points into an email and the other half would be turning emails into three bullet points. And this is kind of what's happening with advertising now. Half of AI will turn three bullet points into 300 ads. Then the other side of that is that the other half is that the AI looks at 300 ads and turns that into three bullet points.
And that's a pretty profound change in how advertising and brand and retailing work. Just to kind of make that real, I just moved apartment, so this is very real for me. You go to Amazon and it says, you bought packing tape, so maybe you need bubble wrap and boxes. That's a straightforward kind of "if this then that" comparison. What they ought to be doing, what they'd like to be doing, is say, "Okay, maybe you're moving. Maybe you'd need light bulbs."
And they might be able to do that with an expanded level of correlation within their data set. But the more interesting question would be, say, "Okay, you're moving, so we'll show you an ad for home insurance." Remember that Amazon now has the world's third-largest advertising business. That's not something that's going to come out of their purchasing data. And so you start shifting the kinds of questions that you might be able to answer. So I'll just skip over that slide for time.
Pulling that to a more kind of enterprise use case: imagine you're looking at a dashboard and you're thinking, what do I do with AI here? The easy thing is to say, find me that metric, find me that product, find me that feature. People talk about talking to their dashboard. You push that a little bit further: make a dashboard for Black Friday and restock. Pull together for me all the metrics that I might need for this and make me a useful dashboard for this use case in this place. Generate this software for me. This is a generative interface.
But then push this one step further. What you might really want is to say, what's likely to be out of stock that's most profitable, and what can I substitute for that? And that's not the kind of question that you can ask using a SQL query. It's not the kind of question you can ask with a database. And so you start pushing through from use our existing tools better, get better use out of our existing tools, and push further to, well, what are things that we can do with this that we kind of couldn't have done before?
That we never really realized were questions before because they were so out of scope of what we could do. All of this is to say — I started this section by saying, what's our AI strategy? That's a lot of different questions. That might be a CIO question. It might be an operations question. It might be a CEO and board question. It might be a question for your system integrator, or it might be a question for Bain or McKinsey. It might be a new tool that you use, or it might be a new industry, or a fundamental existential threat to what you do.
And so I just want to pull this together with a couple of concluding observations. Firstly, it's worth remembering all the stuff that we were excited about before AI is still there. Like, e-commerce is now about 30% of US retail. Amazon sold $65 billion of ads in the last 12 months. And all the other new stuff is still there. So, Waymo has got autonomous driving working. They're doing now something like a million rides a month just in California and expanding into a bunch of other cities.
And then we have people playing with other future technologies. So on the left here we have Meta's latest AR glasses. On the right here, one of many humanoid robots that are being enabled by a combination of electric motors and machine learning that means they don't fall over. But it's also kind of worth stepping back and thinking, how many times have we been here before? This is an ad from IBM from the 1950s that said that the IBM electronic calculator will give you 150 engineers.
And that's almost — I've literally heard that pitch from AI companies. It's like, this AI software will give you 150 engineers. Of course, they're all also white men with receding hairlines, which is hopefully not the profile of all of your engineers today. It's amazing if you look at the job ads from this period, it will say, you know, General Dynamics is hiring men like you. Men at our company are building this product. You kind of see it and you think, wait, what did I just read?
And people used to talk a lot about automation back then. This is a government report from 1955, and one of their use cases they proposed for automation was electronic elevators. In case this is confusing for anybody, what on earth is an electronic elevator? This is the elevator in my apartment building in Manhattan. We have what's called an attended elevator, which means the doorman gets in and he's got an accelerator and a brake. It's basically a vertical streetcar, and if he's drunk, he could kill you.
He could just drive it up through the top of the building. In 1955, Otis launched the first automatic elevator. And so this is a chart of how many people worked as elevator attendants in the USA in the 20th century. It's a perfect bell curve of automation. To begin with, every building gets an elevator, and then every elevator gets automated. And today it's just apartment buildings in Manhattan. That's like a middle-class luxury that you've got an elevator attendant and doorman.
And the point of this is that when the automation — when something is new and exciting, it's automation or it's software or it's technology, and then when it's done, it's just disappeared and you don't notice it anymore. This is my final slide. This is a quote from an AI scientist called Larry Tesler in 1970 who said, "AI is whatever machines can't do yet, because when it works, it's not AI anymore. It's just software."
Ten years ago, image recognition was AI. Today, image recognition is just what your phone does. Today, large language models are AI. It's quite possible that in 10 years' time, that will just be what computers do, and that's the way the world always worked. And every company in this room will have deployed it and built it and created a bunch of value, and then it'll be boring and we'll move on to something else. And we'll go on to the next platform shift. And with that, I'll say thank you. [applause]
So I'm not sure. Do we have time for questions or do we want to go straight on? Sure. Okay. Maybe one or two questions. Do we have a microphone, or any questions, or did I answer all possible questions already?
