Is AI a bubble?

Generative AI has been moving at a ridiculous pace. Since ChatGPT landed in late 2022, we've gone from "look what this thing can do" to billions in investment, companies rebuilding products around models, and executives being asked what their AI strategy is. Which makes the obvious question worth asking: Are we watching a technological revolution, or another bubble? Probably both. The mistake is assuming that if the market is a bubble, the technology must be one too.

// why it looks like a bubble

The pattern is familiar. A new technology arrives. Everyone gets excited. Money follows the excitement. Valuations detach from revenue. Suddenly every company has an AI strategy, whether it needs one or not. We've seen this movie before with dot-coms, crypto, NFTs and the metaverse. AI has some of the same symptoms. Companies are being valued on what they might eventually become rather than what they're earning today. Start-ups are raising enormous amounts of money while still figuring out what their actual product is. That's usually where things get uncomfortable. Then there's the cost. Training and running frontier models requires extraordinary amounts of compute, hardware and electricity. The models are getting more efficient, but "more efficient" doesn't mean "cheap". The economics still have to work. And impressive demos aren't the same thing as profitable businesses. A model that can write an email in three seconds is interesting. A system that reliably saves a company £10m a year is a business. We're still figuring out how much of the former becomes the latter.

The interesting question isn't whether AI is useful. It obviously is. The question is whether the value being created is large enough to justify the enormous amount of capital being spent to create it. Some companies will prove that it is. Some won't. That's what a correction looks like.

// the bubble isn't the technology

The internet was a bubble. The internet wasn't a mistake. The dot-com crash destroyed huge amounts of paper wealth and killed thousands of companies. It also left behind infrastructure that eventually became the foundation of the modern economy. AI can follow the same path. The market can dramatically overestimate the value of individual companies while simultaneously underestimating the long-term importance of the underlying technology. Both things can be true. In fact, they're probably going to be.

// people are already getting value

The strongest argument against "AI is just hype" is that people are actually using it. Developers are writing code faster. Support teams are automating repetitive conversations. Marketing teams are generating and testing content. Researchers are processing documents and datasets. Lawyers and analysts are using models to accelerate research. The gains aren't theoretical anymore. They're not necessarily revolutionary either. A 20% improvement in a boring workflow is still a 20% improvement. And boring improvements compound.

// the technology is moving too fast

One of the stranger things about the current AI cycle is how quickly the baseline keeps moving. Better reasoning. Better multimodal capabilities. Smaller models. Lower inference costs. Faster hardware. Longer context windows. Better tooling. Every time the economics start looking difficult, the technology changes the equation again. That doesn't guarantee that the current valuations are justified. It does make the underlying technology difficult to dismiss.

// what probably happens next

I don't think the most likely outcome is: AI 🚀 → AI 💥 → AI disappears It's closer to: AI 🚀 ↓ too much money ↓ too many companies ↓ reality arrives ↓ consolidation ↓ useful products survive ↓ AI becomes boring And honestly, the last step is the interesting one. The biggest technologies eventually stop feeling like technology. Nobody talks about whether their company should "adopt the internet" anymore. It's just there. AI is likely heading in the same direction.

// ROI becomes the filter

The first phase was about experimentation. Companies wanted to understand what AI could do. The next phase is going to be much less forgiving. The questions become: How much does it cost? How accurate is it? How much time does it save? How much revenue does it create? Can we trust it with real work? What happens when it gets something wrong? That's a much better environment for good products. It's also a terrible environment for companies whose entire pitch is "we have AI".

// smaller models, sharper tools

I don't think the future is necessarily one enormous model doing everything. There will be plenty of those. But there will also be smaller, specialised systems that are cheaper, faster and easier to control. A model trained for one domain doesn't need to know everything. It needs to know enough about the thing you actually care about. That can mean better accuracy, lower costs, stronger privacy and fewer compliance headaches. The future of AI probably looks less like one giant brain and more like a collection of increasingly useful tools.

// follow the picks and shovels

There's another lesson from previous technology cycles. You don't necessarily want to own the company making the most exciting application. Sometimes you want to own the thing everyone needs to build the applications. Cloud infrastructure. Chips. Data platforms. Model hosting. Security. Observability. MLOps. The companies building the foundations can benefit regardless of which individual AI application wins. The gold rush has always been good for shovel sellers.

// the uncomfortable bits

AI still has plenty of problems that don't disappear because the models get better. Copyright. Privacy. Hallucinations. Bias. Security. Deepfakes. Regulation. Accountability. The technical capability to generate an answer is not the same thing as having a system you should trust with an important decision. That's why governance isn't going to be a side project. As AI moves into production, knowing when not to trust the model becomes almost as important as knowing how to use it.

// what businesses should actually do

Don't adopt AI because everyone else is. Find an expensive, repetitive or frustrating problem. Measure it. Try AI. Measure it again. If it works, keep it. If it doesn't, remove it. That's the strategy. Not: "We need an AI strategy." But: "We have a problem. Can AI solve it better?" The technology is powerful. It is not magic.

// my prediction

Some AI companies are going to disappear. Some valuations are going to look ridiculous in hindsight. Some investors are going to lose a lot of money. There will probably be a correction. And none of that means AI was a bubble. The useful distinction is between an AI bubble and a bubble in AI companies. The second is almost inevitable. The first is much harder to argue. AI is already producing measurable productivity gains, changing software development, reshaping knowledge work and creating entirely new products. The hype will eventually become boring. That's usually what happens when something becomes real. The winners won't necessarily be the companies with the biggest models or the loudest announcements. They'll be the ones that turn increasingly capable technology into something people actually need. The bubble can burst. AI probably won't.