So, I was asked to explain everything that's happening in AI and do it in 15 minutes and not talk too fast. So, I'll probably manage two of those, possibly all three. You can see which it is. And the challenge, I think the fundamental challenge in trying to understand AI is to try to break it apart, because once you say, 'What do you think about AI?' this becomes an impossible question. And I've just tried breaking it apart into three questions: so, talk about capital, talk about deployment, and then to talk about how this might actually change things.
So, first looking at capital, because of course that's what drives all of the deployment and all of the investment that's happening at the moment. As most people in this room will know, one way you can think about tech is that it moves in platform shifts. Every 15 years or so, there's a new thing and everything gets built around that. And so, we had mainframes and then PCs and then the web and then smartphones and now we have generative AI. And when you have these new platform shifts, a bunch of things happen. So, this is a new thing. All the new innovation and company creation and product creation and investment happens around the new thing. The old thing's still there. Airlines still run on mainframes. But everything new happens on the new thing. And inside technology, that means everything changes. The old gatekeepers fall away or they become irrelevant. New gatekeepers emerge as new value. There are new markets. There are new businesses created. Outside of technology, people kind of look at this and scratch their head and say, 'Okay, is this kind of like a new tool I get from Google or is this a fundamental threat to our entire industry?' And you kind of have to check back in 10 years and find out.
This is just a brief illustration of what that meant in tech for Microsoft. When the tech industry was Windows PCs, then Microsoft dominated the industry. Then when smartphones became the center of the industry, Microsoft became irrelevant. And for something 10 or 15 years, the Ballmer winter, no one paid any attention to Microsoft. And then charts like that get you the quotes like these that both Google and Meta say the risk of missing this for them is much bigger than the risk of over-investing, because in the end all they've done is built a bunch of empty data centers that they'll fill up in 5 years anyway. But the biggest risk of all is to miss this thing for them.
Which drives charts like this. Last year, those big four platform companies spent about $400 billion building infrastructure. This year, they've said they'll spend something like $650 billion, maybe more, maybe 700. We'll find out. OpenAI, of course, is trying to break in. Half a trillion dollars, $1.5 trillion dollars, a trillion dollars a year, none of which has actually gone into the ground yet. But we'll find out. Maybe, maybe not. But meanwhile, Nvidia can't keep up. Revenue number keeps going up and to the right as everybody takes all of that money and sends it straight off to Taiwan. Indeed, you can see it — no, no, I'm not joking. You can see it in the trade statistics. This is Taiwanese versus PRC imports to the USA. Like that is purely the AI capex boom on the one side and then Trump's tariffs on the other side that have caused that switch.
Now, of course, this amount of capital is quite hard to spend all at once. So, there are backlogs in getting the chips. There are backlogs in getting power deployment. There are some communities that are not sure they want to have a giant data center built in the neighborhood. So, there's a kind of a pig in the python problem which gets you to Kevin Scott at Microsoft and everybody else in the industry saying it's actually not been possible to build any infrastructure anything like as quickly as they would have liked.
Meanwhile, we have incredible revenue growth. So, for some reason these companies have decided to take the last 4 weeks of revenue and multiply that by 12 and call that their annualized revenue, which seems a little bit overcomplicated. So, this is just like the monthly number. But doing two or three billion dollars a month after two or three years is unprecedented. And all of these companies also say that effectively the main constraint on growth is capacity. If they had twice as many chips, they'd have twice as much revenue. And that sort of sits in a kind of a broader conversation that drives these numbers up because we have explosive revenue growth, explosive demand growth, and therefore more and more demand for capacity. And of course, that's also driven by agents and reasoning.
But then the next model is always bigger and more expensive. So, you're kind of trying to chase, and meanwhile the efficiency of the data centers keeps getting better. So, you're kind of chasing the efficiency gain on the one hand versus demand versus the next big model. So, there's a question here of, well, how much longer can these numbers keep going up? One of them is until you run out of money. So, until recently, these companies were paying for this investment out of free cash flow, which basically means the advertising industry.
This year, Meta thinks it will spend over half of its entire revenue building infrastructure. Alphabet and Microsoft not far behind. So, like they can't double those numbers again because there just isn't enough money. Which means that they go to the debt markets and they start doing kind of weird complicated off-balance sheet vehicles. And meanwhile, Oracle is trying to buy their way in as well without having anything like as much cash flow. So, there's an awful lot of stuff going on on Wall Street and in capital markets and in construction.
Meanwhile, back to the models. We've got lots of models. There are an awful lot of benchmarks. In fact, there are probably hundreds of benchmarks and you really need to use ChatGPT to understand what all the benchmarks mean. But I think this is actually a better way of thinking about it, which is the models keep getting better and there are lots of them. And it used to be that OpenAI was ahead and now OpenAI is one of sort of half a dozen companies that can make models. And they're all kind of the same. If you go into video creation or coding, then you can see differences. But for kind of most general purpose model purposes, these things are commodities. They're very expensive commodities, but there doesn't seem to be a barrier to entry or winner-takes-all effect.
Meanwhile, of course, more stuff means more tokens. So, we have reasoning and then we have agentic. And so, we keep driving the frontier forward of what you can do with this stuff, what the capabilities are, of course, and how much money you need to spend. And so, one question one kind of looks at here is to ask, 'Well, where is this all going to settle down? Are we going to go to a world where these companies collectively spend a trillion dollars a year on capex and have half the profits?' Is this going to look different from the traditional software industry where there was no capital, but you had winner-takes-all effects that say you had monopolies? Is this going to look more like kind of commodity infrastructure where there's half a dozen of these things and they're very expensive to build, but because they're all the same, they end up selling at marginal cost? Are they going to be able to capture value further up the stack? Are they going to be like Windows or like Apple? Are they going to work more like TSMC where they make the thing, but they don't have any control over what people build on it?
All of that gets us to two sort of very, very basic questions here. The models keep getting better. The models seem to be commodities, but we don't know what we're going to do with them. We don't really know what the product is. We don't really know where the value capture is going to be. And that looks very much like the internet in like 1997 or mobile in 2005 or 2010. You knew this is a big deal. You didn't really know how it's going to work.
So, that takes me to talking about deployment. William Goldman, famous Hollywood scriptwriter, said, 'Nobody knows anything.' The early internet was full of all these ideas and companies and acronyms. The same with the mobile internet. Many of these ideas were very exciting. Everything on this screen is something that didn't work. And so, we should just sort of presume a whole bunch of stuff that we see in generative AI today will be equivalent. There will be a whole bunch of stuff that people are working on now that won't work. And then there'll be something else that does work. Of course, you could also look at that and say, 'But Benedict, everybody's using this.' So, we've got 900 million weekly active users, although only at OpenAI, only about 5% of them are paying.
But then if you look at the usage, you discover that something like 80% of ChatGPT users sent less than 1,000 messages last year, which probably means they used it once or twice a week. If you look at survey data, you see the same thing. Many more people are using this monthly or weekly than they're using this daily. And the same thing in the enterprise. If you say, 'Do you use ChatGPT?' everyone will say, 'Yes. Yes. Yes. Obviously, I do.' But if you ask people how much they use it, a much smaller number of people found something to do with it every day than they're using it maybe once a week.
And I think a kind of a useful analogy here is to think about spreadsheets. This is the first successful software spreadsheet, VisiCalc. If you show this to an accountant, it was life-changing. If you show it to a lawyer, they would say, 'Well, that's very clever, but I don't do that. I might use it next week for my timesheet.' Except that this Mac — not a Mac, sorry, an Apple II — cost over $15,000 adjusted for inflation. So, it was kind of an expensive thing to use. But an awful lot of people today are lawyers rather than accountants. They look at this thing and they think, 'Well, that's great and I might use it next week.' What else am I supposed to do with it?
And so, there's a sort of structural question here. How do you work out what to do with the new thing? To begin with, generally you absorb it. Use it for the problems you already know about, which is a line we've heard in both of the two previous sessions. But what you have to do is work out what new things you can do with this, what innovation you can have. And then, of course, sometimes someone will come along and disrupt you.
This is pretty much what's happened so far. This is what big companies are doing with this. Lots of people doing kind of analysis and a little bit of analytics and product innovation and customer service, much less new revenue because it's much harder to work out what the new revenue would be than to find some cost savings and to do some analytics in the middle. And the place where the usage is most concentrated, as you can see here, unsurprisingly, is in tech. So, again, even in tech, more people are experimenting with this than using it every day. But then there's a bunch of other industries where no one can really work out what they're supposed to be doing with this stuff.
And if you kind of systematize this again, what's really happening is there's a small number of things where this works really, really well, which is software development, marketing, customer support, and then kind of automating back-office point solutions. Everyone else is kind of scratching their head. Again, this is data from OpenAI that says the same thing. A quarter of their user base is in tech, which is not representative of the overall population.
And I think it's kind of just worth looking at software development and thinking at a minimum this is like the new AWS. And so, AWS was probably several orders of magnitude change in how quick and easy it was to build software. And even if nothing else happened, this is again several orders of magnitude change in how quick and easy it is to build software. What does that mean? Well, I think I might disagree with some of the panelists in the previous session. Very obviously you can make way more software way, way more quickly, much more cheaply. You can also do a bunch of things that you couldn't do before because LLMs can do things that software couldn't do before. And there'll be a bunch of legacy apps with big complex expert systems that will get replaced by LLMs.
That doesn't mean that an accountant in your accounts payable system is going to say, 'Yeah, I don't like Stripe. I'm going to build my own this weekend using Anthropic.' Probably not a great way of thinking about how software evolves.
Some people in this room may be familiar with charts like this. This is a market map of every company in the marketing technology industry. We do not suffer from a shortage of software. The hard part is not writing the software. The hard part is working out what it's supposed to be doing. It's working out the business and the use case and the tooling and the legacy system and the way to get people to use it.
Even when it works, it still takes time. Something like only about a third of enterprise workflows have moved into the cloud so far after 25 years. So, it takes time to build the new thing. It also takes time to make people use it. I find this story hilarious because Accenture is going to punish people for not using AI, which mainly means they haven't worked out good use cases for AI.
And here we have Steve Jobs giving us kind of the fundamental problem that it's not the user's job to work out what to do with a new thing. It's the entrepreneur's job. It's the software company's job to work out what you should do with the new thing. And so, that kind of captures a bunch of questions. Like you've got these transformative new capabilities that you can build stuff with. But you've got a probabilistic system that will always get things wrong and we don't have any way to fix that. And you've got this kind of blank screen question. So, how do you build product around that? Well, you kind of build software. You think about use cases and data and tooling and user experience and go to market. And you build an actual software company around that software in order to find those solutions.
How would you work out what those would be? Well, one answer might be as a big company you go to Accenture, even if they can't necessarily work out what to do with this themselves. Another is that you do pilots. And so, this is day two and every big company now is doing a bunch of pilots. Some of those pilots go into production, some of them don't, because that's what pilots are. And again, that's concentrated in software development and in marketing.
And then every startup in the world is out there trying to find those use cases and build those use cases and unbundle Excel, Salesforce, Photoshop, ChatGPT, and turn those into single-purpose use cases and go out and get people to use them.
Now, sometimes when you automate something with a new piece of software, then all sorts of other stuff changes, too. So, this is a chart about barcodes. In the mid-1970s, the US grocery industry deployed barcodes. And if you didn't need to do a stocktake every Friday night with a piece of paper, then suddenly it became possible to run your business in a completely different way. You could have five or 10 times more products because you could keep track of them all.
And that takes me to the last section, which is to talk about what all of this is going to change. So, I'm going to kind of repeat this slide again. A lot of conversations I've had with big companies are kind of along the lines of, 'Well, is that it? We've deployed a bunch of stuff now. We've all got ChatGPT. And then half of us didn't use it. And we've got image generation and we're doing natural language search and we've got a SKU tagging project and some review summarization. Is that it? That seems like there should be more than that. There should be something else.'
The problem is, of course, working out how a fundamentally transformative technology is going to change everything is impossible. I mean, you can guess. But imagine asking this question about the internet in 1997. Like you'd have got some of it, but you would not have predicted all the things that we're doing today.
However, I think there are again a couple of building blocks you can use to look for those kinds of answers. And that's to think about automation and to think about price elasticity. And then to think about levels of abstraction. I also used to look at the last wave of machine learning and say this gives you infinite interns. So, you want to listen to every call coming into the call center and tell me if the customer is angry. You want to look at every picture in the CMS and tell me if it's a man or a woman or a cat or a dog or a pair of shoes or a coat or whatever it is that you're trying to sell.
Now, if you kind of systematize that across generative AI, you can kind of build like a classic consulting slide and you can say, 'Well, there's perception and then there's document creation and there's some analysis and then there's action.' So, I should say agentic here. The tech industry kind of suffers from a sort of Tourette's syndrome. Only instead of shouting it in inappropriate moments, people will say metaverse. Or now it's said agentic. So, I should probably try and say agentic some more just to make sure we're on brand.
Now, some of this again is about absorbing again. So, every marketing company is now using this technology to create massively more assets. But there's a kind of price elasticity question here around automation. If you make it massively cheaper and easier to do something, do you do the same amount of work for less money with fewer people or do you do more work with more people? Was having a building full of people doing something really boring your defensibility? And then more interestingly, what is it that becomes possible if you don't need to have millions of people to do that?
This is what happened in the UK around this in the 19th century. This is the population and then the number of steam engines, the motive power of all of the steam engines that were in the UK. And when you got that many steam engines, you can have a train that goes from London to Scotland in 10 hours. And you couldn't have done that by hiring 10,000 horses and tying them to the front of the train. Or 10,000 people and tying them to the front of the train.
When you have this new technology, it unlocks things that just weren't possible before. You don't just automate the old thing. I think one way to think about that is to go back and look at the internet itself. One of the things that the internet did was it split apart the product from the physical asset. So, if you were in the music business, suddenly you didn't have to make little pieces of plastic anymore and that destroyed your company. If you were in the newspaper business, you didn't have to manufacture these newspapers every day and put them on trucks. So, you kind of separated apart what the product was that people were coming to you for from the way you were doing it.
And I think there's kind of a similar point here as we look at generative AI. Again, if you separate out the need to have all of that stuff being done by people bored out of their minds sitting in warehouses somewhere and can they get a machine to do that, what new things become possible? One way to think about this is to think about e-commerce and advertising and marketing. I always kind of look at images like this, which of course I made with ChatGPT, and think the thing is Amazon and Google and Meta don't really know what's in those boxes. They just know that there is a SKU and people who bought this SKU bought that SKU. But they don't know why. They don't know what those products mean.
And if we move to generative AI, you start asking completely different kinds of question. And so, there's the old question, which is here's a coat, where can I buy it? You can do that on Amazon now or Net-a-Porter. But then you push to more interesting questions. And what really happens is if you ask the kind of question that you would just never ever have thought of asking, you can change the level of abstraction and push to different places.
The enterprise side of this again is once you have all of this data and you can automate that, what are different things that you could never have done before with people, no matter how many people you might have had in the past?
Meanwhile, of course, what happens to the people? A couple of people mentioned layoffs. The one really obvious thing that got automated in the 20th century was accounting, because you have adding machines and then mainframes and then cloud computers and then machine learning. And every decade we got more accountants. Because it turned out that the job wasn't adding up the numbers. The job was something else.
And so, you get this kind of question of what is it that the customer is buying? Why do they want that? Do they just want the slide or the spreadsheet or the SKU or the answer to the question? Or were they coming to you for something else? For kind of experience, authenticity, judgment, preference, curation, suggestion. Do they want something other than just automating the answer?
Now, I'm kind of over time, so I'll go through a few concluding comments very quickly. Firstly, it seems to me all AI questions have one of two answers. The answer is either how did it work the last 10 times or no one knows. And the great part about the last 10 times is the answer to those questions is also no one knows.
But we've been here a lot before. This is an IBM ad from 1950 that says an IBM electronic calculator is like having 150 extra engineers, which is also the pitch for every AI software company today. Also, of course, they're all white male engineers with receding hairlines. All the ads from this period look like this. Once you see it, you can't unsee it.
But this is a US government report from the '50s that talked about automation. And this is one of the things that they talk about: ready-mix concrete, coal mining, and electronically controlled elevators. What the hell is an electronically controlled elevator? Well, this is what they're talking about. This is the elevator in my apartment building uptown. We have an attended elevator. There's no button. It's a streetcar. You get into the elevator and you say hello to the doorman and he drives you to your floor. And you have to make sure to give him a good tip, otherwise he might not come quickly when you ring the button. So, you give him a good tip every Christmas.
This is a perfect case of automation because this is how many people worked as elevator attendants in America in the 20th century. The first half is you build lots of skyscrapers. The second half is Otis invented the automatic elevator. You press the button, the elevator comes. And when it worked, you just don't see it anymore. We don't think of that as being automation. When you got into the elevator to come here this evening, you didn't say, 'Ah, I'm going to use an automatic elevator today.' It's just a lift.
This is a classic quote about AI from 1970: AI is whatever doesn't work yet. It seems to me that machine learning isn't AI anymore. And I've heard conversations multiple times in the last two or three months where people say, 'Oh, that's not AI. That's just image recognition. That's not AI. That's just sentiment analysis.' And you think, 'Wait a minute. 15 years ago, that was witchcraft.' 10 years ago, AI made that possible. Now, it's not AI anymore because once it works, it's not AI. It's just software.
And that's probably how this will work as well. In 10 years' time, it will just be software, and we'll all be excited about something else. With that, I will say thank you.
