Are we in an AI bubble? (these are the 5 warning signs)

0:00Are we in an AI bubble? As I speak, hundreds of billions of dollars are being poured into artificial intelligence. Many say it's the investment the technology deserves, the push towards an abundance of intelligence that could change humanity forever. Others are calling it a bubble, a repeating economic story of excess, untethered belief, and ultimately collapse.

0:31In the 1600s, a bubble led to tulips trading for more than houses. For the price of one bulb. And in the 1800s, bubbles led to thousands of miles of unused railway tracks being built. All of these ended in disaster, with money down the drain, investors losing their shirts, and shocks to the economy. The narratives seduce us, but ultimately that rhetoric turns into reality and prices collapse. To some extent, the current AI wave resembles past manias. And so there's a key uncertainty. Is this sustainable growth or are we leading to a bubble? It's complex because bubbles look a lot like booms while you're in them.

1:11A boom also involves rapid investment and optimism. But eventually fundamentals like cash flows and productivity catch up. So I spent hundreds of hours figuring this out. I dug through data, I talked to investors and economists, I made a number of models, I compared what's happening today to the bubbles of history. So all of this led me to a five-gauge framework that helps us understand AI bubble risk in much the way that a pilot uses a number of different gauges to figure out if the plane is flying safely. These gauges can give us a sense of the path we're on right now and help us understand what we need to look out for in the future.

1:54A key gauge is the investment intensity. So what is the scale of capital expenditure relative to GDP? And is that getting unhinged? When one sector attracts an enormous amount of capital, it bends the entire economy. Capital moves there, labour moves there, supply chains move there. And one problem is that any kind of reversal transmits quickly. If assets are short-lived, that can really compress the payback window. If we look historically, the US build-out of the railways in the 19th century had a number of bubbles. And at one point, the annual capital expenditure in building that railway network and the trains approached 3.6 % of GDP, right before one of the busts.

2:39In the late 1990s, investment in telecoms infrastructure, as we started to digitize telecoms and communications, reached about 1 % of US GDP level. And it left behind many, many conduits of dark fiber that we still use today. When we look at artificial intelligence in 2025, we're going to see about$400 billion go into build-out data centers. Not all of that is in the US, but roughly speaking, given the majority is, that's about 1 % of US GDP. And it seems to be rising. Maybe it'll get towards 1.5 % by 2030. Now, hardware like GPUs doesn't have the same lifespan as iron rails on a railway. They depreciate over six years.

3:20They can get used in their really high intensity uses for about three years before they have to do gentler tasks, shall we say. At the same time, American GDP growth is becoming noticeably dependent on these investments in data centers. So where does this gauge lie? My verdict is that it's on the boundary of green and amber, probably just into the amber. But if you squint, you might be able to argue that it's a green. Monetization level is an important gauge. It's the ratio of Gen AI revenues relative to the capital deployed that year to build out the infrastructure. It matters because investment must start to earn its keep.

3:57And if coverage is persistently low, this could signal rising fragility. In the case of the railways, their revenues covered about two-thirds of the annual investment at peak. In the case of the telecoms bubble about 25 years ago, new revenues covered around a quarter, maybe a bit less. But both of these strained as growth slowed. If we look at Gen.ai today, we think that about$60 billion will be spent on it in 2025. Compared to that$400 billion of capex, that gives us coverage maybe around 15%, maybe a little less, maybe a little bit more. And the indirect gains are very likely to be undercounted.

4:34There are some analysts and investment banks who have a much higher number for Gen AI revenues in 2025 than we do. But given our number, about 60 billion in revenues, 15 % coverage, I'll say this gauge is in the amber. Revenue trajectory is also important to keep an eye on. So this gauge measures the speed and breadth of revenue growth and whether that momentum is accelerating or slowing down. Its strong momentum drives revenues up, it absorbs depreciation, it repays capex, and stagnation has proceeded past busts. Rail revenue growth was weak into the 1870s crash. Telecoms was growing faster in the 1990s, but not enough compared to the scale of leverage that they had taken on.

5:19The dot-coms were growing, but probably in the high tens of percents. When we look at Gen.AI today, we are seeing even the very big companies like Anthropic and OpenAI growing at 80%, 100%, 200 % per annum and looking at very fast growth rate into 2026. When we look at a cohort of big Gen AI startups that are not yet public and much smaller than OpenAI and Anthropic, we count about 18 to$20 billion of revenues in 2025, substantially above last year. And some of these companies are growing at 400, 500, 600, 700 % per annum. So on a conservative path, I think Gen AI could easily reach$100 billion in revenues by 2026, possibly even more.

6:07Enterprises are really early in their adoption, so too are consumers. And despite that, new data centers are running at a very high utilization as they come online. The gauge that measures revenue trajectory is definitely in the green. So this gauge looks at how far asset prices are running ahead of earnings. I use a simple proxy price to earnings ratios and some other measures. Extreme multiples can only persist if earnings outrun gravity. Overheating is a classic bubble tale of too many investors chasing a story, not fundamentals. The dot-com is a great example of this. Some of the dot-com leaders were trading at a PE over 600.

6:52It's crazy to say that. If you look today, the top large tech names, and these are the ones most exposed to AI through their data centers or their models, they're trading at a PE of around 38. It's elevated, but it's very far from dot-com territory. Their earnings are strong. Their cash generation is strong. Their order books are strong. Their revenue momentum is really, really high. So for me, this valuation level gauge still sits in the green. Another gauge that's important to track is the quality of capital that is supporting this CapEx. I mean, simply put, you want great experienced investors or you want the companies themselves to be funding this expansion.

7:36So you need to look at the resilience of the funding sources across the stack? Is this internal cash versus external debt? Who is providing that debt? Is it coming from established players from established markets? Are there complicated structures that might obfuscate risk while expanding access into those funding flows? Weak capital structures amplify drawdowns and credit squeezes can force really disorderly unwinds. Railways relied on heavy debt and retail money and they were hit hard in the panic of 1873. At the turn of this century, the telecoms companies had piled lots of cheap debt. It was very clear revenues were not keeping up and many of them defaulted as revenues lagged.

8:20In the case of the dot-coms, it was all about equity exuberance. And the moment there was any sense of uncertainty, things collapsed. Where we are today is that the big tech companies are minting cash and they have tons of it on their balance sheets because for many, many years, they've been producing products that we want, that enterprises want, and they've been doing so profitably. So they are largely able to fund this from their own balance sheets. Although over the course of the next four or five years, they will need to find external funding. There's a funding gap of about one and a half trillion dollars that will need to be filled by private credit and other types of structures.

8:59These types of things could be signs of fragility. I don't think they are now. But of course, this story isn't even. There are companies like the NeoCloud, these very young startups that are building out data centers for AI. And of course, there are some of the model companies themselves that are going to require an enormous amount of capital on an ongoing basis, but haven't had a chance to accumulate their own reserves. And so you're going to start to see more and more debt structures being used there. Perhaps they already are. And you might have some covenant issues and other types of things to be concerned with.

9:32But there is also some ballast today that didn't exist before because governments around the world have committed to AI as infrastructure. And there's more than a trillion and a half dollars of sovereign AI commitments. That doesn't mean governments are going to stump up that cash, but it means they want that cash to be stumped up through public and private initiatives and other policy levers and incentives they can pull. And also we've got really big pools of capital that are now running around in sovereign wealth funds, which are quite tech lean forward. And so they do provide some ballast and some access to capital as we continue with this buildup.

10:13Okay, so imagine I'm that pilot in the plane, I've got my five gauges. I look at them, most of them are green. One is maybe trending into amber and one is in amber, but it's trending back down to the green. It feels quite safe. It feels like we're not yet in bubble territory, but of course you have to keep watching your instruments. Tulips only left a footnote in the history books, but railways and the web rewired the world. We're still using some of that infrastructure today. Artificial intelligence looks a lot like the latter. And my gauges for now say that this is still in the boom phase rather than the bubble phase.

10:53But keep in mind that the question of whether a technology is a bubble is fundamentally a financial one. Even if there is a financial bubble brewing around artificial intelligence and it bursts, it doesn't make the technology any less revolutionary, any less meaningful or any less useful. We'll still be left with the infrastructure for an AI economy. We will be left with companies building and investing with a greater deal of financial discipline. We will be left with a lot of talent that knows how to build these models, the applications on top of them, make them safe, make them reliable. And there will still be billions of people using AI every day.

11:38We still, after all, use train tracks and websites, even though their bubbles burst a very long time ago.