The demand for infinite compute

The AI boom isn’t just about chatbots. In this video, I explain why cloud companies and chipmakers are exploding in value: we’re moving into an economy where computation becomes a fundamental input – like steel, electricity or oil. If that’s true, our demand for compute could approach infinity.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Azeem Azhar?

  So today I want to talk to you about something that may surprise you. If we think that the huge investments in artificial intelligence, data centers and infrastructure, and the exceptional profits at companies like Amazon, Microsoft or Nvidia are just about generative AI, we're missing something important. All of this matters, generative AI, but in a different way. The real story is a broader shift in the economy towards computation. So let me take you through this shift and how it connects to what we're seeing in the news and how Gen AI fits in that picture. We saw the results from many of the tech companies, many of the big tech firms,

• 0:45Azeem Azhar?

  and they showed really staggering growth, particularly in their cloud businesses. Companies like Amazon and Google and Microsoft are known as hyperscalers because outside of the businesses we most think of them running, they also run computing capability for enterprises all over the world. And those cloud businesses are growing really rapidly. Analysts reckon that Amazon's AWS grew more than 20% to about $33 billion in revenues. Google Cloud, growing faster, it's the smallest of the three, at about $15 billion in revenues. And Microsoft's Azure grew by some 40%.

• 1:29Azeem Azhar?

  And we reckon a large part of that, as it was with Amazon and Google, AI workloads. That is, running their infrastructure to serve AI companies like OpenAI, like Anthropic, but more importantly, enterprises all over the world who are increasingly building AI services internally. It's not just about the data centers and the hyperscalers. Chip companies are seeing enormous orders. I'm not sure many people would have predicted that Anthropic, a three-year-old startup, would buy a million AI processing units called TPUs from Google.

• 2:09Azeem Azhar?

  And on that subject of Google, they said that 150 of their enterprise customers were using more than 1 trillion tokens from their LLM systems, which is, sounds like quite a lot. It's even more than I managed to use. But those tokens are actually pretty cheap, so the dollar value is not as big as it might sound. Set against all of this, the tech giants will collectively invest, as we know, roughly $400 billion this year in the hardware to deliver AI. They'll invest more next year, and they'll invest more the year after that. I think it's telling that Microsoft's boss, Satya Nadella, is talking about planet-scale token factories. So is this all crazy talk?

• 2:55Azeem Azhar?

  Is this exuberance, is this excitement running ahead of where we are? Are we building capacity that we will never use? A famous episode of The Simpsons when the town of Springfield built a monorail. Is this the monorail for the global economy? I think to make sense of that, let's think about what computing really is. I mean, computing is a way of systematically processing information using tools. We call those tools computers. And processing information is a really valuable thing to do. The computers we use today are of a particular type. They, they work in binary on these chips. But there are other ways of doing computing.

• 3:38Azeem Azhar?

  Ternary computing, where you don't have bits of 0 and 1. You have trits of minus 1, 0 and 1. And of course, quantum computing with its fuzzy states. But ultimately, being machines, they can compute faster and more consistently than humans. And computing has proved to be really, really useful. And we've shown that we actually just love having access to computing capabilities. Let's compare it to something else that I really like, which is donuts. So I love donuts, but there's a marginal, diminishing marginal utility with donuts. The first one is great. I'm not sure I will ever eat number four in a single sitting.

• 4:20Azeem Azhar?

  And I can pretty much testify that I have never got as far as four donuts in a single sitting. And the same is probably true for you, but that isn't true with computing. What happens with computing is we have an insatiable demand for it, and it also has a positive price elasticity, which means that if the price comes down, we buy more and more of it. We use more and more of it. And what happens with compute is that consumption continues as capabilities expand. Now, we've got some empirical data for this, some real evidence. So a few years ago, I went through the process of estimating the amount of computing capacity in the world every year from 2023 back to the 1950s.

• 5:04Azeem Azhar?

  I think I started in 1958. So essentially, I counted the number of computers of different types, and I estimated how much computing each one could do, and totaled that up to give a sense of the stock of computing in every year, as I said, since 1958. Now, the counting isn't exactly apples for apples, but it's directionally robust enough. Now, the normal yardstick I use is to start in 1972. It's a wonderful year, it's the year I was born. And I estimate that the total stock of computing power available globally between 1972 and 2023 increased by 11 orders of magnitude, which is roughly a factor of 100 billion.

• 5:48Azeem Azhar?

  In other words, we consumed a lot more, astronomically more compute in 2023 over an astronomically large computing infrastructure in 2023 compared to that wonderful year of 1972. It's a combination of there being many more computers, not just mainframes and mini-computers, but also laptops, supercomputers, and of course, our smartphones. And those computers themselves were much more powerful, thanks to Moore's Law, making them powerful, more powerful, every year. It turns out that in that 50 or so year, every single year, we saw 62% growth in the total computational stock on the planet.

• 6:33Azeem Azhar?

  62% compounded for 50 years is a big astronomical number.

• 6:39Azeem Azhar?

  So we've really shown that we like having this stuff around, even though lots of the time it sits idle. And the truth is, there's a ton of it out there. I mean, I reckon, and I may be wrong by an order of magnitude, that there's about 10 to the 22 flops of compute stock in the world as of 2023, where a flop is a floating point operation. So it's like an instruction that the computer is carrying out. When we think about these new AI factories, these gargantuan data centers with chips optimized to serve AI models, they're going to be filled with incredibly powerful chips. I mean, think of the Nvidia ones, it's a $5 trillion company, it shipped about 7 million

• 7:20Azeem Azhar?

  of these GPUs this year, and they probably added several zeta flops to the stock of computing. What's a zeta flop? A zeta flop is 10 to the 21. So, you know, meaningful addition to the existing stock of computing. And the thing is, all of this compute is actually being used. I mean, if you listen to analysts and you talk to the companies, as I do, you hear that they are running short of compute. So this isn't like the monorail in Springfield, which was never used. It's not like the dark fibre during the telecoms bubble that wasn't used for more than a decade. As Gavin Baker, the CEO of Atrides, which is an investment management firm, says, there are no dark GPUs anywhere.

• 8:04Azeem Azhar?

  And that's the same message that I hear when I talk to people in the industry. Utilisation rates are exceptionally high. What's happening is that the economy is moving into a computational fabric alongside the physical elements of the real economy. We're doing more and more useful work in silicon, work that we might have had to do physically in previous years and previous generations. We've always used information in our economies, there's a historical progression from tally sticks to the breakthrough of double-entry bookkeeping several hundred years ago. And computational systems emerged at the turn of the 20th century, they were initially mechanical, and since 1938, electronic.

• 8:50Azeem Azhar?

  But what they do to our information system is they allow the processing of that information to take place out of our skulls and in these machines. And the machines go faster and they do it more consistently and they benefit from manufacturing economics. And I think that that shift is what really drives our desire to use computation. And since nineteen sixties, our economies have really depended on computing. It may not feel like that if you were like me, born in the seventies, it didn't feel particularly digital, but the economies really, really depended on the ability for computers to

• 9:30Azeem Azhar?

  do the information processing from a supply chain and inventory to payroll processing. I mean, even things like fast fashion fundamentally depend on databases, on being able to log customer preferences, on being able to keep a large number of SKUs, on being able to update prices and update inventories really, really rapidly. And now we're getting to a stage where every senior exec has lived in that world of computing. Many of them grew up with the internet alongside them rather than growing into it. And that additional knowledge, combined with the greater capability of computing, is

• 10:09Azeem Azhar?

  creating a demand for companies where really they want to do things increasingly in silicone. I mean, that is a major shift in the last ten or fifteen years. Modeling, simulating drug discovery. I mean, thinking about how to optimize a portfolio of products rather than running that optimization month after month after month by manufacturing them and seeing how well they sell. Or it could be planning and optimizing the routes that your logistics vehicles are going to take. Far cheaper to do that in silicon than to drive thousands of vehicles around the country. A quick note, if you want to support us in bringing more of these conversations to the world, please consider subscribing to the show.

• 10:54Azeem Azhar?

  And now we have AI moving into firms. There was a fascinating. Fascinating piece of research from Wharton Business School this week, which is part of a regular survey they do. And showed that eighty percent of people who responded, senior execs in American companies with more than a thousand employees, with more than fifty million dollars of revenue, claim to use AI, generative AI every week. Forty seven percent of them claim to use it daily. But most interestingly, fifty eight percent of them said they were using some kind of agent or agent workflow in their business. And that's really, really interesting. It's really interesting when we think about what this means for the usage of compute.

• 11:38Azeem Azhar?

  The reason is that for many of us using generative AI today, there's so much we want to do with it. We've got a creative backlog, we've got a work backlog, and it's really fun to be able to use Claude or G P T five to help us work through that backlog. But ultimately, we're constrained in that query and response modality by having to live life, having other things to do, wanting to hang around with family, wanting to get some sleep, wanting to do some exercise. So the tasks run in them when when we're not using them, they're not running. But in an agentic world, the tasks run the whole time because the agents run the whole time. They are working while we are sleeping.

• 12:17Azeem Azhar?

  They are the the robots in the dark factory, the other software robots in the data center continuing to work for us. And this shift towards agentic workflows, which is is taking time, but is getting more and more mature, indicates an increase in the amount of compute that enterprises will want to use. So it's also important to understand that compute is about computation, that is systematically processing information according to some kind of algorithm or set of steps. It's not just about large language models. Large language models are the propeller engine of nineteen thirties, nineteen forties aviation.

• 12:58Azeem Azhar?

  And like propeller engines on planes, it's quite possible they will be replaced by something even snazzier. So let's tackle this idea of infinite demand. What do I mean by infinite demand? I don't mean that every atom in the universe will be enlisted for computing, but I'm sure a science fiction writer out there has got a story on just that subject. What I mean is is practically unlimited, that at each turn of the screw.

• 13:27Azeem Azhar?

  Better chips, more sensors, more data, improved algorithms. We find new things that we want to compute in business, in science, in research, in entertainment. Better materials, better routings, more efficient stock portfolios. And of course, that unending array of consumer uses. We might think that optimizations, think about DeepSeek will help. That optimizations like DeepSeek, which reduce the amount of computing you need to get the output that you want, would actually be a break on the amount of compute that is demanded in the global economy. But as we saw, once you make the output cheaper, you increase the demand for it.

• 14:11Azeem Azhar?

  So Jevons paradox, it's that positive elasticity of demand that I talked about. Because as the outputs become cheaper, you widen up the range of use cases that were previously too expensive to attempt. They weren't economically feasible and a software optimization made them more cheap. They crossed the break even point. And so we then created new tasks for them. And I think that's really important to understand, that of course optimizations and efficiencies are really, really desirable. But there is something peculiar about the act of computation. The act of processing information, the act of processing that information to help us make better decisions or invent new things or tackle difficult problems.

• 14:56Azeem Azhar?

  That means that there's an unending array of things that we'd want to point this capability at. So when I look at the market today and there's this febrile activity around the chip companies and the infrastructure companies. It's very easy to get lost in that noise and to say, how can this possibly be true? To say, this is, this is just, you know, a lot of kids running after a soccer ball on a pitch. This is the dot com again. But I ask myself, in 10 years, or in 20 years, or in 30 years, will humanity be doing more computations or will we be doing fewer computations? I mean, the answer is almost certainly that we will be doing more, much more.

• 15:36Azeem Azhar?

  Much more. That 62% annualized growth in computational stock that I calculated years ago, it's held for decades and it's even being accelerated by this AI build out. And that's one reason why this is so different from every other infrastructure boom. We're never going to get full on compute the way that we get full on donuts. We'll just want more and more. Every economic transformation we've seen has been powered by our ability to better process information. Whether it was double entry bookkeeping or writing or the telegraph, the ability to process and act on information is critical.

• 16:17Azeem Azhar?

  Now, after four donuts, you stop eating, but after four trillion trillion computations, you discover a new industry. And that means we're not yet full and quite likely we'll never get full. Thanks for listening all the way to the end. If you want to know when the next conversation is released, just hit subscribe wherever you're listening. That's all for now and I'll catch you next time.