AI Eats the World | Benedict Evans

[music] [music] Right. I didn't know I was going to get

Right. I didn't know I was going to get dry ice as well as pyrochnics and lasers. I must be a really big deal now.

lasers. I must be a really big deal now.

Um, so I was asked to explain everything about AI and only take 15 minutes and not talk very and not talk too fast. So

I I'll see what I can do. I think this is always the place to start. The tech

industry moves in platform shifts. We

always kind of know this chart and we know that generative AI is the new platform shift, but I think it's just worth going back and thinking how many lessons do we learn from the last five times that absolutely everything changed

in completely different ways. What is it that happens in a platform shift? What

should we be thinking about now? All of

the innovation and the investment and the company creation moves to the new thing. Inside technology, the new

thing. Inside technology, the new gatekeepers emerge. Old gatekeepers fall

gatekeepers emerge. Old gatekeepers fall away. There is new value capture. There

away. There is new value capture. There

are new and bigger markets and old things get destroyed. Outside of

technology, is this a new tool? Is this new revenue?

Is this an existential threat? How much

do we have to care about it? That's what

every big company is kind of looking at and scratching their head about now inside technology. Um, platform shifts

inside technology. Um, platform shifts mean that companies gain dominance and lose dominance. This is Microsoft's

lose dominance. This is Microsoft's share of computing. When the PC was the center of the tech industry, that was Microsoft. They dominated it. When we

Microsoft. They dominated it. When we

moved to smartphones, Microsoft became irrelevant for a decade because they were no longer a player in the thing that set the agenda. Going back a generation, there were a lot of people trying to make PCs before Apple and

before m IBM and Microsoft. And none of those companies succeeded. In fact, the same thing for web browsers, for search, for social, for smartphones. The first

companies are very, very rarely the companies that end up getting all the value. Part of that is because we don't

value. Part of that is because we don't really know how the new thing is going to work. Um, we all use the web now, but

to work. Um, we all use the web now, but it was not clear in the mid '90s how this was going to work. And there were all sorts of other companies and acronyms and ideas and business models that ended up not being the thing. The

same thing for mobile. I used to come to Helsinki to talk to a company called Nokia. Some of you may remember it. um

Nokia. Some of you may remember it. um

and talk about IM mode and Java and WAP and none of those things for the future.

The same thing with generative AI now.

We should presume that half the stuff we're working on is not going to be the answer and isn't going to work. Um that

means we get a lot of noise and a lot of hype and a lot of nonsense that you just kind of have to sweep aside and ignore.

And you also generally of course get bubbles. Um and in bubbles people draw

bubbles. Um and in bubbles people draw straight lines on log scale charts and they say say things like you just don't understand how exponential growth works.

Um, and they always say this is different and the problem is they're always right. It always is different.

always right. It always is different.

The.com bubble was different. This is

clearly different to the.com bubble. But

it doesn't mean it can't be a bubble.

Um, but when you sweep aside all the noise and all the dust is settled, the world has changed and it becomes part of the way we all live. Um, online dating now is about 60% of new relationships in the USA. This has just become part of

the USA. This has just become part of the way the world works. And on the enterprise side, you get an order of magnitude more tools with each generation. When we went from PCs to

generation. When we went from PCs to SAS, we went from dozens of apps to hundreds of apps. we should expect the same thing to happen with LLMs probably.

Um, however, there is one way that this is all different um from all those other platform shifts, which is with all the other ones, you knew what the physical limits were. You knew what the next

limits were. You knew what the next phone could and couldn't do. You knew it couldn't fly. You knew that Telos

couldn't fly. You knew that Telos wouldn't give everybody fiber next year in 1997. With AI, we don't know why

in 1997. With AI, we don't know why these models work so well. So, we don't know how much better they're going to get. All we have is this sort of vibes

get. All we have is this sort of vibes based forecasting where people say what they feel might happen. And so it might be that this is only as big as the internet or PCs or smartphones, which seems pretty big to me. Or it might be

much bigger than that. It might be a much more fundamental shift. We just

don't know. We'll have to come back in 10 years and find out. Um but if this is only as big as the internet, that feels like enough to get excited about. And so

we ask, well, how is this going to work?

Um how is this new thing going to be useful? Um where is the distribution and

useful? Um where is the distribution and the value capture and the value destruction going to be as this thing gets rolled out? So, first let's look at what's happening inside the tech industry. What are people doing? And I

industry. What are people doing? And I

think these two quotes are really all you need to know if you only had one slide. The risk of underinvesting is

slide. The risk of underinvesting is bigger than the risk of overinvesting.

That's what set the agenda. Um, this

started working 3 years ago. Everyone

says this is a huge opportunity and a huge threat. We can't miss it. It needs

huge threat. We can't miss it. It needs

a lot of capex. Okay, how much capex are we talking about here? And this year, um, the big four platform companies will spend close to $400 billion building infrastructure, which is up 4x from a

couple of years ago. Um, and Nvidia can't keep up. We'll have their numbers out later today, but they're at a, uh,4 to50 billion a quarter run rate. Um, 40

to$50 billion a quarter um, last quarter. Um, the power industry can't

quarter. Um, the power industry can't keep up either. Um, if you talk to data center people today, getting access to electricity is actually a bigger problem than getting chips from Nvidia. And

there's a lot of other problems as well.

Um, and so you have all the hyperscalers saying it's basically impossible to build data centers fast enough. It's

impossible to keep up. And they all plan to spend a whole bunch more money next year. In fact, they they expect the

year. In fact, they they expect the growth rate to go up as well as just the dollar amount to go up for some of these companies. Um, what all those

companies. Um, what all those announcements might add up to is unclear. It's not clear how much of that

unclear. It's not clear how much of that is real, how much of that gets built.

Um, how much of this is a bragger instead of a megawatt or a gigawatt. Um,

but these something like three times, four times, five times more data center, $3 trillion, $5 trillion maybe, maybe maybe more. And that sounds like kind of

maybe more. And that sounds like kind of a ludicrous number. But if you annualize that, that's like 56 700 billion a year.

And that is kind of what giant global capital intensive industries spend. So

if you could actually build a business around that, that is an amount of money um that could be found. Indeed, at the moment, that money is not coming from capital markets. That's money is coming

capital markets. That's money is coming from cash flow from giant profit very profitable companies. These companies

profitable companies. These companies became much much bigger in the last 10, five and 10 years and basically all of that that growth is going out of the door to build data center. Although that

is starting to shift um Microsoft um is going to do something like $50 billion of leasing this year um as they decide their cash flow actually isn't enough um to pay for everything that they want to build. And we've seen a bunch of these

build. And we've seen a bunch of these kinds of stories. Meta has done two deals for $30 billion each. Oracle might

have to borrow 100% of revenue um in order to build the stuff that it's committed to. And of course, or OpenAI

committed to. And of course, or OpenAI wants to join the club. Um depending on which day you listen to Sam Alman, they're going to spend $1.5 trillion.

They're going to spend a trillion dollars a year. They don't actually have any of cash to buy this themselves. Um

so we get this wonderful phrase um circular revenue where the hyperscalers give lots of money to Nvidia and Nvidia gives it to OpenAI and OpenAI uses it to buy chips from Nvidia to compete with the hyperscalers but also gives it to

AMD and Broadcom to compete with Nvidia and then this all gets kind of very exciting and great fun. Um but of course if you were running these companies what would you do if you're at the top of a bubble like Nvidia has $70 billion of free cash flow and they can't spend it

fast enough so they use it to buy market share. Um, OpenAI has mind share in

share. Um, OpenAI has mind share in stock, but they have a commodity product with no differentiation and no platform and no infrastructure. So, what do you do? Well, you scramble as fast as you

do? Well, you scramble as fast as you can um to build something more sustainable. Um, and Oracle is a cash

sustainable. Um, and Oracle is a cash generative legacy business that's been declining for 30 years. So, what would you do if you were if you were Larry Ellison and wanted people to invite you to parties again? Um, so this is all very exciting and if you're on Wall

Street, you made lots of money from it.

But, but what does all of this got us if we're actually out in the tech industry?

Well, the models keep getting better. Um

Google has another model again today apparently. Um there are Chinese models,

apparently. Um there are Chinese models, there are open source models, there are lots of new acronyms. Um but we don't have any apparent mo and we don't really have any clarity on what the product or the value capture is going to look like.

We still don't really know how this is going to work in 5 years time. This is

what I mean by by far more models. This

is deliberately a very simple chart. Um

there are more models every week. They

keep getting better kind of in a straight line. Um but they're all kind

straight line. Um but they're all kind of the same. Um, this is the top 10 models as a percentage of the top model on two of the different most general purpose benchmarks. And the point of

purpose benchmarks. And the point of this is a truncated axis. They're all

kind of within 5 to 10% of each other.

Um, and every week there's a new one that's in the lead and sometimes they're a bit more than 5% ahead, but they all kind of converge on the same scores.

Where they don't converge, um, is on usage where we have Open AI and then we have people who have distribution and then we have people who don't have distribution. Um, Claude benchmark

distribution. Um, Claude benchmark scores are just as good as anybody else on this chart. Um, but nobody's heard of it. Um, and indeed, you can go back and

it. Um, and indeed, you can go back and look at OpenAI and say, "Well, they've got 800 million weekly active users.

Like, everyone's using this stuff now.

It's over there. It's over. It's done.

We're done." But if you go and ask, "Well, what does a weekly active user mean, you discover that most people who are using this don't use it very much.

Most people who are using generative AI use it once a week or once a month. And

this week, they haven't thought of anything that they could do with it. Um,

you see much of the same picture and all the other data that that we can get to.

And so we get this question, right?

Well, what is it that would drive this to become part of everybody's life? Um,

is it just early? Um, is it the wrong use cases? Is it that most people don't

use cases? Is it that most people don't actually think like that and aren't able to use a general purpose tool like that?

How much, in other words, as Steve Jobs tells us, how much do you have to work out what the product should be and show it to them? Do you have to start with the experience and work back with the technology rather than spraying the technology at everybody? Um, so if

you're building a model lab, where do you compete? Do you go down the stack

you compete? Do you go down the stack and compete on capital which is what happened in chips and aircraft um and um AWS or do you go up the stack and you compete on network effects and product

um and go to market the way the software industry works. What you can't do is

industry works. What you can't do is just have a commodity model and keep spending hundreds of billions of dollars on it. In that light I think it's very

on it. In that light I think it's very interesting to look at Microsoft um which used to have no capital and sell you a $1 CD for $500 and last quarter they spent 45% of their revenue on

capex. So are they shifting away from

capex. So are they shifting away from network effects? If you're open AI, the

network effects? If you're open AI, the answer is yes. Everything tomorrow,

yesterday, straight away, please with somebody else's balance sheet. Um

everything from, you know, chips to buy to add app platforms and data sets and e-commerce integrations. Everything

e-commerce integrations. Everything right now on somebody else's balance sheet. As Jim Barksdale told us, there's

sheet. As Jim Barksdale told us, there's two ways you can make money. You can

bundle or you can unbundle. And OpenAI

is doing both of those with app platforms and browsers and social media and everything else. But you're not going to be out able to out compete every startup in the world if the product is to just build more software

on top of the stack. Um, so where is the value capture going to be in building models? Is it in the models themselves?

models? Is it in the models themselves?

Is it in having the most capital? Or is

it in building normal software companies and taking them to market the way we've done it with every other platform shift in the past? It seems like that might be the answer. And that takes me to

the answer. And that takes me to thinking about what's happening outside of the tech industry. Um, how is it that you deploy new technology? Well, what's

the pattern? How do you always do it?

Um, to start with, you absorb it. You

make it a feature. You automate the stuff you already know. And it takes a bit longer to work out new things, new products, new revenue lines, new innovations, and then maybe someone comes along and disrupts the market. Um,

most of what's been happening so far is that step one, those easy obvious places where you absorb it into what you're already doing, which is software development marketing customer support, um, a lot of automation, um, inside big companies, automating

individual point solutions. Um, but it's what you can easily see. Um, how would you work out beyond that what you would automate? Well, you go and hire some

automate? Well, you go and hire some consultants. Um, Accenture claims to be

consultants. Um, Accenture claims to be doing um over $1.5 billion dollars of new generative AI bookings every quarter. Um, and then you do a bunch of

quarter. Um, and then you do a bunch of pilots. Um, but it's kind of worth

pilots. Um, but it's kind of worth remembering that pilots and deployment takes a while. Um, this is people who already use generative AI. Um, even

though everyone in tech has been saying agentic really loudly for the last year and a half or two years, it's actually a very very small proportion that's actually got its way into production.

Um, and that's because it just kind of takes time. Something like a third of

takes time. Something like a third of big companies have got at least one generative AI product in deployment so far, but like another quarter don't plan to do anything just yet. It takes a while. Um, in fact, it always takes a

while. Um, in fact, it always takes a while. If you're in tech, um, cloud is

while. If you're in tech, um, cloud is old and boring, but it's actually only about a third of enterprise workflows.

It just takes a while to deploy this stuff. Of course, once you've deployed

stuff. Of course, once you've deployed it, then everything changes. This is a chart of barcodes. Um once the grocery industry deployed barcodes, then it became possible to manage five or ten times more skews. You could change how

the whole of the rest of your industry worked once you've done this very simple but very difficult piece of automation.

And so that gets back to kind of the question at the beginning of this section like everyone's had a bunch of AI presentations, everyone's deployed a bunch of stuff, but what is the disruption? What is the fundamental

disruption? What is the fundamental change? Is this just automation? Is it

change? Is this just automation? Is it

something else? In other words, how would you think about the innovation and the disruption rather than just the automation? and ask trying to ask that

automation? and ask trying to ask that question now is like asking what the internet will do in like 1997. Um but

you can kind of try and look for kind of really kind of core structural questions and I'm kind of posing two here. The

first is if we go back to the internet this unbundled a lot of things from physical. You didn't need physical

physical. You didn't need physical assets anymore and very often you didn't realize that was what you were you didn't realize you were a bundle until the internet changed it. So what is it that LLMs will change? The second answer

is the internet unbundled or created all sorts of new aggregation models. Whe

that's Amazon or Door Dash or Instagram.

How do LLMs change that? How could they give that aggregation, recommendation, discovery, suggestion in different ways?

Um, if AI gives you infinite interns on the first slide, well, what would you do with infinite interns? How does that change your business? Do you do the same work with fewer people? Do you do more

work with the same number of people? Um,

do you was employing lots of people the thing that made you defensible? What

becomes possible if you don't need millions of people? What was only possible if you could get five million people to work on it and now you don't need that? Now you can get it get that

need that? Now you can get it get that in a machine and you can automate it. Um

I don't know but but but lots of people will try and find out. This is what happened in the 19th century. Steam

engines gave the UK something like 250 million equivalent labor units. 250

million extra working pairs of hands to build stuff and pull stuff and make things. What will generative AI do to

things. What will generative AI do to give us that kind of equivalent in productivity and labor? Um, the other side of that question, how do we think about what this does to internet platforms? Before the internet, we had

platforms? Before the internet, we had human editors, whether that was in a shop or a TV station or newspaper or magazine. People decided what we saw on

magazine. People decided what we saw on the internet, we scaled that massively and now we have algorithms deciding what we see and those algorithms are all kind of fed by us. YouTube knows what we would like by looking at what people do.

An LLM doesn't necessarily need to do that and it can understand things at a different level and maybe it can understand more things in different ways. So maybe it can shift all of those

ways. So maybe it can shift all of those giant platforms and all of those algorithms into different places. This

is as I suggested earlier an awful lot of value to capture. The global industry ad industry ad industry is about a trillion dollars and big tech companies have about half of that. Um and they're all using this to optimize their

existing business. This is the absorb

existing business. This is the absorb phase. They're optimizing their ad

phase. They're optimizing their ad businesses. Marketers are using this to

businesses. Marketers are using this to optimize asset creation. They're all

going from making five assets to 100 assets or 300 assets for every campaign.

Um, but there was an old joke like old a couple of years ago, a joke that half of AI would be turning three bullet points into emails and the other half would be turning emails into three bullet points.

And so if we're now what we're actually doing is using AI to turn three bullet points into 300 ads, who's going to be looking at the ads? Um, and how will they be deciding what you should do?

This is what gets us to agentic commerce. I just go to chat GBT or

commerce. I just go to chat GBT or Gemini and say, "What should I buy?" And

that gets you a very different buying journey to Amazon or or Instagram or anything we've been doing in the past.

Of course, again, this is very very early. Um, if you're in tech, everyone

early. Um, if you're in tech, everyone stopped using Google, but outside of tech, even like the young people who are supposed to get this stuff are not really doing that very much. Most people

are still using search where we've already always used it. And we should remember like the web has been dying since it started. These are two covers from Wired from like 10 and 20 years

ago. Um the web is dying again now.

ago. Um the web is dying again now.

Maybe perhaps we don't know. But we

should kind of push back and think what's the sort of fundamental question that we were trying to ask here. Um this

is a shop in Tokyo that only sells one book. You don't have to go into when you

book. You don't have to go into when you go into the shop, you don't have to work out what book to buy. But on the internet, we have infinite product, infinite media, infinite retail space.

We have no idea what to buy. And we have these very imperfect systems to try and find it for us. And now we have systems that will understand that at very different levels of abstraction and be able to make very different kinds of recommendation. What will they

recommendation. What will they recommend? What will they change? How do

recommend? What will they change? How do

we split things apart around what you ask for? Why did you ask for that? Why

ask for? Why did you ask for that? Why

are you hiring those people? Why are you making this purchase? Is it about just getting the thing just pure utility? Or

is about experience, curation, recommendation, authenticity? Do you

recommendation, authenticity? Do you want to have some mood music for an hour or do you want the artist's authentic experience? Why is this being created?

experience? Why is this being created?

Um, in other words, what is our AI strategy is a lot of very different questions depending on who you're asking and why you're asking. And then just some final observations. Um, all the

stuff we were excited about before AI is still there. Like e-commerce is 30% of

still there. Like e-commerce is 30% of retail now and going up further and more outside of the USA. Whimo has autonomous cars working. Um, and um, maybe um, some

cars working. Um, and um, maybe um, some other companies will be able to do it as well. Meta is working on glasses. Some

well. Meta is working on glasses. Some

people are excited about humanoid robots, but we should kind of go back and think again. How many times have we been here before? This is an IBM ad from 1951 that said their electronic calculator will give you 150 engineers.

Like how many AI ads are basically saying exactly the same thing. This is a US government report from 1955 about automation. Um, which is what people

automation. Um, which is what people called technology then. Um, and this is uh a passage in that what was going to get automated. One thing that might get

get automated. One thing that might get automated was electronically controlled elevators. Um, this is what they're

elevators. Um, this is what they're talking about. This is an elevator in my

talking about. This is an elevator in my apartment building in Manhattan. It's a

manually operated elevator. This is what elevators were. And then technology

elevators were. And then technology meant you could automate it. Um, when's

the last time you got into an elevator and said, "Ah, I'm using an automatic elevator now. It's just a lift." This is

elevator now. It's just a lift." This is a number of people employed as elevator attendants in the USA. First, we build lots of elevators, then we automate them, then we forget that they were ever

not automated. Um, this is Larry Tesla

not automated. Um, this is Larry Tesla in 1970 saying exactly the same thing.

AI is whatever machines can't do yet because once it works, once you've automated it, you forget that anything was ever any any different. The world

has completely changed in ways that were completely different from anything else.

And then it happens again 10 years later um and then a 10 years later and then 10 years later and we just carry on automating everything and forgetting about it. And with that I will say thank

about it. And with that I will say thank you. [applause]

you. [applause]