Inside the $110 billion AI economy (and who is really paying for it)

The AI economy is real and substantial. Forensic analysis of 1,000+ companies reveals where the revenue is flowing, and what it will take for the AI data center buildout to pay off.

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

• 0:00Azeem Azhar?

  The AI economy is real and it is substantial. In the 12 months to June 2026, it's $110 billion of revenues that flowed into a sector that didn't exist four years ago. It's a big number, but more than that, it's a number we've got to faster than any technology wave we've seen in history. Three times faster than the mobile phone, three times faster than the internet, three times faster than cloud computing. So not only is it a real sector, it's growing very, very quickly indeed.

• 0:35Speaker 01

  When you say real, what do you actually mean? How do we know the AI economy and its revenues are real?

• 0:42Azeem Azhar?

  One of the biggest concerns about the AI economy is whether any of it is real at all. Is the money just being handed from one company to another and back to the originating company? And that's a real concern that sometimes happens during technology waves. So we set out to trace how revenues came in at the top of the system and went all the way through from the application companies to the foundation model companies, through to the hyperscalers and the infrastructure companies that actually run the chips that serve the models. And that was a long painstaking process. We looked at over a thousand companies to try to figure out some of these numbers. We used forensic accounting to deconstruct financial models.

• 1:23Azeem Azhar?

  And we spent a lot of time ensuring that if a dollar came in at the top of the system, we only counted it once as it made its way through. So it is real, but of that $110 billion, a large part of that is being spent by companies that are loss making. So the, the big AI labs like Anthropic and OpenAI spend tens of billions of dollars a year to train their next models. And if you're a real cynic, you could say, well, that's not real revenue. But in truth, it's in the order of 20 to 25 billion of that 110 billion. Then you might say, well, what portion of that 110 billion is coming from startups funded by venture capital?

• 2:03Azeem Azhar?

  I mean, it's one thing if an excitable founder has persuaded a Silicon Valley VC to give them $10 million. It's another thing if a stalwart member of Main Street has decided to put shareholder capital and their accumulated profits into AI. And even if you take away the money being spent by companies backed by venture capital, it's only a few billion dollars. So at the end of all of that, we're, we're looking at something in the order of 70 to 75 billion dollars coming from Main Street profitable companies spending on AI like they spend on, on IT.

• 2:41Azeem Azhar?

  If you want to see our full report on the state of the AI economy, visit intelligence.exponentialview.co.

• 2:53Speaker 01

  AI infrastructure spend has been enormous. What has to happen for it to pay off?

• 2:58Azeem Azhar?

  The infrastructure story is astonishing. The big hyperscalers and new startups serving AI are undergoing an infrastructure build out of a scale that we haven't seen since the railway build out in the United States in the 19th century. The big question is, are they going to make a return from that? And what we identified in our research was that roughly at the end of 2025, the amount of revenue flowing in to Gen AI was just about covering the depreciation expense on the infrastructure spending that the hyperscalers were doing. Now, let's understand what that number means.

• 3:40Azeem Azhar?

  What that means is that the cost of building all these huge data centers is just about being met by customer revenues. That's a good start, but it's not a good business. Because of course you need to more than cover your depreciation costs. You have to cover them with a really healthy margin so that you can pay for all of your other expenses, like your staff and your marketing, and ultimately show a profit to your shareholders. What we discovered was that we've passed one milestone. It's an important milestone, but it's not the signal of long-term sustainability. And the question is, will we get to that long-term sustainability? At the moment, revenues are growing far faster than the growth of AI infrastructure expense.

• 4:25Azeem Azhar?

  So the bet would be that at some point in the next two or three years, we might get to that point. But equally, there's a lot of uncertainty in this market and it could be the case that we don't.

• 4:36Speaker 01

  Do you think we should treat tokens as units of intelligence?

• 4:39Azeem Azhar?

  When you think about the oil industry, we know what West Texas Intermediate is. It's a constant standard for what a barrel of oil is and what the price is. And people across the market can compare one with another. And many markets have that. They have these sort of constant metrics, whether it's the horsepower of a car or the fuel efficiency, measured in exactly the same way. The challenge in AI is that it's a very, very young industry. It's only a few years old and we haven't yet figured out a way of putting a value to the intelligence that's produced. So our proposal is that you have a, a quality weighted token. The token is the thing that ChatGPT or Claude produces, but not all tokens are the same.

• 5:22Azeem Azhar?

  They get produced by models that are more intelligent and models that are less intelligent. And they can be more useful or yes, less useful. So our initiative is to figure out a quality adjusted token that can start to look like the West Texas Intermediate standard or the Brent crude standard so that investors, companies, providers can understand what they're producing, what they're selling and what it's worth.

• 5:47Speaker 01

  AI keeps getting 10 times cheaper. So why are bills going up?

• 5:51Azeem Azhar?

  AI models are on a learning curve. That means every year they get about 10 times cheaper. But it's really odd that our AI bills haven't gone down. In fact, they've gone up virtually universally. So what's going on? Well, that's called elasticity of demand. As the price of tokens comes down, companies are buying more and more of this. They're buying more of this intelligence and the overall scale of the revenue is going up. Now, my hypothesis is that this elasticity, this behaviour will continue for quite a while. Companies in America and around the world have just got started. Even as models get more efficient, they come from China, they go to open source,

• 6:30Azeem Azhar?

  the declining price of the token will be more than made up for by increasing volumes of token consumption.

• 6:38Speaker 01

  Revenues are growing fast, but AI still represents a tiny part of the wider economy. How do you reconcile that?

• 6:44Azeem Azhar?

  AI revenues have grown so quickly, three times faster than the internet, three times faster than mobile, three times faster than the cloud. So it's incredibly rapid. And yet it doesn't really feel that a lot of companies have figured out what they're actually doing. If you look at how much American companies are spending per employee per month, it's only a few dollars. The real leaders in using AI, the firms who can prove that they've grown faster because they've invested more heavily in AI are spending hundreds of dollars, sometimes thousands of dollars per employee per month. So that $110 billion revenue number has got lots and lots of room to grow

• 7:25Azeem Azhar?

  as more and more American firms through Main Street get more confident, get more practice and get more persuaded that AI is something that will change their business. So it's very much still in the foothills.

• 7:38Speaker 01

  What's the impact of low-cost open source models?

• 7:41Azeem Azhar?

  You know, we often obsess about the most powerful models from Anthropic and OpenAI and we're amazed at what they can do. But life isn't always about having the most incredible, most powerful system. Sometimes you just need an efficient, cheap version to do the thing you need to do. And what we're learning is that open source is becoming that thing for companies as they start to use AI. And you know what? It makes perfect sense. You don't always need to go off and see the most expert doctor if all you've got is a sore ankle. Just go and see your local physician, go to the pharmacy. And open source plays that role. It plays the role of the efficient, cheap models that will lift most of the workloads that most people use most of the time.

• 8:25Azeem Azhar?

  And once in a while, we'll want to find the most powerful models to answer the hardest, hardest questions.

• 8:31Speaker 01

  What's the biggest problem companies run into with AI?

• 8:33Azeem Azhar?

  I speak to about 100 CEOs a year and they've all run into the same problem with AI. Individual productivity is up, but its impact isn't visible across the firm as a whole. I know what's going wrong. A company is not simply the sum of its employees' tasks. It's a system for turning information into decisions and decisions into action. And most of those systems were designed for a world in which intelligence was scarce, expensive and exclusively human. That world is disappearing. With AI, developers can write 50% more code, but they're still waiting in the same review queue. Teams with AI can act immediately, but management won't make a decision for a month. I call that congestion. It's an organisational problem, not a technical one.

• 9:17Azeem Azhar?

  Your company most likely runs at a rhythm designed for scarce intelligence moving at human speed. So the question isn't, where do I add AI? It's now that intelligence is abundant, what should my company even look like? Which decisions should move from monthly to daily? And what things become more valuable precisely because intelligence is cheap? The winners will be the companies that redesign themselves around a new economic fact. Intelligence is no longer scarce, but the ability to act on it still is.

• 9:49Speaker 01

  You say intelligence is no longer scarce. What does that look like in everyday life?

• 9:53Azeem Azhar?

  In 1978, my family had one camera, just like this one. We took about 24 photos a year. They were expensive and rare. Today, cameras are everywhere. We take so many photos and videos, we don't give it a thought. They used to be magical, now they're ubiquitous. And we use them in ways my parents could never have imagined. With AI, useful intelligence is going to become as ordinary as a photo. Three years ago, I used ChatGPT to write novelty rap songs for a dollar.

• 10:27Azeem Azhar?

  Today, I have several AI agents working for me all the time. They build software, they research, they create products. It costs little more than a fancy coffee every day. And honestly, it already feels mundane. As intelligence becomes cheaper, we will use much more of it. But most interestingly, we'll use it in unforeseeable ways. As unforeseeable as a camera in a doorbell was to my parents back in 1978.

• 10:54Speaker 01

  Which of the frontier labs is ahead at the moment?

• 10:56Azeem Azhar?

  It doesn't matter who wins the AI race. OpenAI, Anthropic, or DeepSeek could stumble. Another lab could overtake them. But the larger story will continue. We've discovered something remarkable: a way to manufacture intelligence. And once we can make something, we can learn to make it better and more cheaply. We've seen that with solar panels, computer chips, and batteries. It's called the learning curve. The more we make, the cheaper it gets. In the summer of 2025, it costs around 27 cents for Google's best model to write a bit of code. 10 months later, DeepSeek could do it for less than half a cent. This is part of a strong trend. Every year, AI models get 10 times better.

• 11:41Azeem Azhar?

  And this trend doesn't look like it's slowing down. Nothing in history, not oil, not electricity, not computer chips, has become so cheap so quickly. And here's a really special thing about AI. Oil can't learn how to make more oil. Solar panels don't improve new panels. But AI can learn how to make AI. And that is extraordinary.

• 12:04Speaker 01

  Do you think AI is really making us stupid?

• 12:07Azeem Azhar?

  An economics professor allowed students to take their midterm exams at home. Their average score was 96%. That's miles above normal. Suspicious, he asked them to sit their final papers in person. And the class average fell by half. Only three got anywhere near their original score. And by the way, kudos to those three students. I think the rest of them turned off their brains and got AI to do their thinking. But even then, asking whether AI is making us stupid is not the right question. The right question is why do the students want to avoid the hard thinking at all? It's because the incentives in the system don't reward their thinking. They just reward the result.

• 12:50Azeem Azhar?

  If the goal of study is high test scores, of course, we'll all take the easiest route to those good scores. SAT coaching in America is an entire industry built on gaming the score while not teaching anything. In a world of AI, we must change the incentives so that they reward reasoning. The hard thinking, the late nights, the dead ends, not just a final grade. If we reward only the answer, AI will give us better answers and weaker thinkers. If we reward the reasoning, AI could help us think more deeply than ever before.

• 13:24Speaker 01

  How far behind are China's AI labs?

• 13:26Azeem Azhar?

  If you think Chinese AI labs are just playing catch up, you're mistaken. Washington's sanctions have choked the supply of AI chips to Chinese labs. That's actually forced them to develop the most efficient AI models on the planet. American labs do have eight times more computing power than their rivals. But Chinese labs have become much smarter about how they use their limited resources. In fact, they get five times more intelligence per unit of compute than the Americans do. That's a bit like having a car that can travel five times further on one gallon of gas. I call this China's efficiency moat. Their models are 10 times cheaper than equivalent American ones. And that's already winning them US customers.

• 14:10Azeem Azhar?

  Companies like Airbnb, Coinbase, Uber, and the majority of Silicon Valley startups. That doesn't sound like catch up to me. One group wants to create the most powerful intelligence at any price. The other is figuring out how to make capable intelligence at a remarkably low cost. Those are different strategies. And we should not assume that the most expensive strategy is the one that ultimately wins.

• 14:42Azeem Azhar?

  To get my analysis every week, just sign up at exponentialview.co.