Is AI Another Dot Com Bubble Waiting to Burst?

Everyone has an opinion on this one, and most of them are loud. Half the internet is convinced we are living through 1999 all over again, that the whole thing is a giant house of cards, and that when it falls it will take the pension funds down with it. The other half thinks anyone who doubts AI is a dinosaur who will be out of a job by Christmas. Both camps are exhausting, and both are partly right, which is the annoying bit.

So let us actually sit down and think about it properly. Not with hype, and not with doom. Just a proper look at where the money is going, where the value actually is, and whether the thing everyone is betting the farm on is real or just a very expensive group hallucination.

First, the money question, because that is what people actually mean

When someone asks if AI is a bubble, they are almost never asking a philosophical question about intelligence. They are asking a financial one. Is the amount of money being poured into this stuff justified by the amount of money coming back out? And right now the honest answer is: nobody knows, and the gap is enormous.

The numbers being thrown around for 2026 are genuinely difficult to picture. Depending on whose estimate you trust, the big cloud and chip players are spending somewhere in the region of 700 billion dollars in a single year building data centres, buying chips, and wiring up the electricity to run them. That is not a startup burning through a seed round. That is closer to the annual budget of a mid sized country, spent on machines that need replacing every few years.

Now hold that next to the revenue. The actual paid for, money in the bank revenue from AI products is real and growing fast, but it is nowhere near 700 billion. It is a fraction of it. The bet the industry is making is that the revenue curve eventually races up to meet the spending curve, the way broadband eventually justified all the fibre that got laid in the late nineties. The fear is that it does not, or that it takes so long that a lot of very confident companies run out of runway first.

The case that it is a bubble

Let us give the pessimists a proper hearing, because they are not idiots. Their argument goes roughly like this.

The spending is being justified by projections, not by profits. A huge chunk of the capital going into AI is essentially one group of companies buying chips from another group of companies, and some of that money is circular. Company A invests in Company B, Company B uses that money to buy compute from Company A, and everyone books it as growth. It looks like a thriving economy until you notice the same banknote is being counted several times.

The unit economics of running these models are brutal. Every clever chatbot answer costs real money in electricity and hardware, and for a lot of products the price the customer pays does not cover the cost of serving them. In the dot com days you could at least run a website cheaply once it was built. A large AI model is the opposite. It is expensive to build and then expensive again every single time someone uses it. That is a nasty shape for a business.

And the hype has clearly detached from reality in places. When a company can add the letters AI to its pitch deck and watch its valuation double overnight, you are looking at a market that is pricing a story, not a product. That is exactly the smell of 1999.

The case that it is not a bubble, or at least not that kind of bubble

Now the other side, because it is stronger than the doomers like to admit.

The dot com comparison flatters the pessimists in one important way. In 1999, most of the companies with silly valuations had no revenue and no product, just a domain name and a plan. The big AI players today are, for the most part, some of the most profitable companies in human history. They are funding this spend out of enormous existing cash flows from search, advertising, cloud, and software that people already pay for every day. When Pets.com collapsed, it had nothing underneath it. If AI spending got cut in half tomorrow, the companies doing the spending would still be printing money from their day jobs.

There is also the small matter that the technology actually works. This is not Theranos. Hundreds of millions of people use these tools every week and come back the next day, which is the single hardest thing for any product to achieve. A bubble built on something nobody wants pops instantly. A bubble built on something everybody wants can deflate slowly and still leave an enormous industry standing, exactly as the internet did after 2000.

So the sensible position is uncomfortable and boring: there is almost certainly a bubble in parts of the AI market, and there is almost certainly a real and durable industry underneath it, and the pop, when it comes, will separate the two painfully. Both things are true at once. The internet was a real revolution and the dot com crash was real and brutal. Those facts never contradicted each other.

Are we actually using AI correctly?

Here is where I get a bit grumpy, because I think the answer is mostly no, and the reason is that we have fallen in love with the wrong use cases.

The instinct across the whole industry has been to take AI, which is a genuinely new kind of tool, and cram it into shapes we already understand. We have bolted a chatbot onto everything. Your bank has one, your fridge company has one, your tax software has one, and nine times out of ten it is worse than the button it replaced. We took the most flexible technology in a generation and mostly used it to answer emails slightly faster and to summarise documents nobody was going to read anyway.

Think about it like this. When electricity arrived, the first thing factories did was rip out their big central steam engine and drop an equally big electric motor in exactly the same spot. It worked, but it was pointless. The real gains only came decades later when someone realised you could give every single machine its own small motor and redesign the entire factory around that idea. We are in the big central motor phase of AI. We are using a revolutionary thing to do the old thing slightly better, instead of asking what the old thing even was.

The correct question is not how do I add AI to my product. It is what could I never do before that is suddenly possible now. Those are completely different questions, and almost everyone is asking the first one.

Are most startups just recycling old ideas with a fresh coat of AI?

A lot of them, yes. And it is worth being precise about why, because it is not simply that founders are lazy or cynical.

There is a whole category of company right now that people unkindly call a wrapper. The product is a thin layer of interface sitting on top of somebody else's model, doing something you could largely do yourself if you knew the prompt. Notetaking apps, email writers, generic chatbots for this or that industry. There are thousands of them, they mostly look identical, and most of them will not exist in two years. They are recycling ideas that already existed, note takers and email tools and support bots, and the only new ingredient is the model underneath, which they do not own and which their competitor can rent just as easily.

The brutal problem for these companies is that they have no moat. If your entire product is a clever prompt and a nice login screen, then the moment the underlying model gets better, your product either becomes redundant or gets absorbed as a free feature by the model provider itself. We have watched this happen repeatedly. A startup builds a lovely tool to do one specific thing, the base model ships an update that does that thing natively, and the startup evaporates. Founders call this getting steamrolled, and it is the defining risk of building on top of a platform that is improving faster than you are.

But let us be fair to the other side, because the wrapper sneer is a bit smug. Plenty of hugely valuable businesses in history have been thin layers on top of someone else's infrastructure. A shop is a thin layer on top of a wholesaler. A restaurant is a wrapper around ingredients you could buy yourself. The value is not always in owning the deep technology. Sometimes it is in the taste, the trust, the specific workflow, the customer relationship, the boring integration work that nobody else wants to do. A wrapper that deeply understands one narrow industry, speaks its language, and slots into its existing tools can absolutely survive, because the moat was never the model. The moat was knowing the customer better than the model provider ever will.

So the honest split is this. Most AI startups are recycling old ideas and will not make it. A minority are using the new capability to do something that genuinely could not be done before, or are wrapping it in so much specific hard won domain knowledge that the wrapper itself becomes the valuable thing. The trick, as an observer or an investor or a founder, is telling those two apart, and it is genuinely hard in the moment because they look identical on the surface.

So where is AI actually, undeniably great?

This is the part I actually enjoy, because when you strip away the hype there is a solid core of things that are simply, quietly brilliant, and they tend to share a shape. The best uses of AI are not the flashy general assistants. They are the small, specific, unglamorous solutions to problems that used to be blocked by one of four things: not enough thinking hours, not enough capital, not enough money to make it worth doing, or simply not enough brains in the room. AI knocks down all four of those walls, and that is where the real magic is hiding.

Small custom solutions that never needed to scale

Here is a category almost nobody talks about because it is not sexy and it does not make headlines. There are millions of tiny problems inside real businesses that were never worth solving, because solving them meant hiring a developer for three months to build a bespoke tool that only twelve people would ever use. The maths never worked. The cost of a custom solution was always higher than the pain of just doing the thing manually forever.

AI quietly obliterates that maths. A small firm can now build the daft little internal tool that reads their specific messy spreadsheets, or sorts their particular flavour of customer email, or drafts the one weird report their regulator demands every quarter. It does not need to scale to a million users. It does not need to be a company. It just needs to save eleven people two hours a week each, forever, and it was never economically possible before. This is genuinely new, and it is happening in thousands of unglamorous offices with zero fanfare. No press release, no funding round, just a problem that was too small to solve suddenly becoming solvable on a Tuesday afternoon.

Problems that were blocked by sheer thinking power

Some problems were never blocked by money or ambition. They were blocked by the fact that a human brain, or even a large team of them, simply cannot hold enough variables at once or read enough material in a lifetime.

Protein folding is the poster child, and rightly so. Working out the shape a protein twists itself into was a problem that could take a talented researcher years for a single structure, and there are hundreds of millions of them. It was not a money problem. You could not throw a thousand more graduate students at it and win, because it was a problem of raw pattern recognition across a search space bigger than the human mind can grip. AI cracked it to a useful degree and handed biologists a map they simply did not have before. That is not a chatbot writing your emails. That is a genuine new key to a locked door, and there is no old fashioned amount of effort that would have opened it.

The same shape shows up in materials discovery, in reading medical scans for the faint early signs a tired human eye misses on the ten thousandth image of the day, in sifting astronomical data, in modelling how a new drug might behave before anyone spends a fortune synthesising it. These are all problems where the bottleneck was never willpower or funding. It was the sheer breadth of pattern that no individual could ever hold in their head. That is the sweet spot. That is where AI is not a slightly better tool but a genuinely new capability.

Problems that were blocked by cost and capital

Then there is the huge middle ground of things that were always technically possible but economically absurd. Translating a small company's entire website and support material into forty languages used to cost a fortune, so nobody bothered and whole markets stayed closed. Giving every single customer a genuinely personal response instead of a template used to require an army of staff. Reviewing every contract a mid sized firm signs, rather than spot checking a few and praying, used to be a luxury only banks could afford.

None of these were impossible before. They were just too expensive to be worth it, so they did not happen. AI drops the cost of these tasks by an order of magnitude, and when you make something ten times cheaper you do not just save money, you unlock things that were previously off the table entirely. A tiny company can now behave, in some respects, like a large one. That redistribution of capability down to the small players is, to me, one of the most genuinely exciting and underrated stories in the whole field, and it gets drowned out by people arguing about whether a chatbot is conscious.

Why on earth are Meta and Google rehiring developers?

This is one of the most telling stories of the past year, and it cuts straight through the hype, so it is worth slowing down on.

The narrative in 2025 was clean and terrifying. AI writes code now, therefore we do not need as many people who write code, therefore the layoffs. And the layoffs were real. Meta cut thousands of roles through 2026 and pivoted hard towards AI. Microsoft shed tens of thousands across the same stretch. The message from the top was blunt: the machine does the typing now.

And yet, look a little closer at the actual hiring data and something odd jumps out. Google spent the same period advertising far more engineering roles than the year before, something in the order of sixty percent more. The demand for people who can actually build and wrangle AI systems has gone, in the words of people who track this stuff, in only one direction, which is up. So we have mass layoffs and a hiring surge happening inside the same industry, sometimes inside the same company, at the same time. What is going on?

Two things, and they matter enormously for the bubble question.

The first is that the layoffs and the rehiring are not the same jobs. A lot of what got cut was the middle, the layers of routine work that AI genuinely can now assist with or automate, and a fair amount of it was simply the correction of over hiring during the pandemic boom, dressed up in fashionable AI language because saying we hired too many people in 2021 sounds worse to shareholders than saying we are becoming an AI company. What is being hired back, at eye watering salaries, is a smaller number of very senior, very capable engineers who can design the systems, fix what the AI breaks, and build the actual products. The shape of the workforce is changing, not simply shrinking. It is getting more expensive at the top and thinner in the middle.

The second thing is the more important one, and it is the quiet admission hidden inside all of this. If AI could really replace software engineers wholesale, the most sophisticated AI companies on the planet, the ones who build the models, would be the first to stop hiring them. They have every incentive and every tool. Instead they are fighting each other for engineering talent and paying record sums to get it. That tells you something the marketing will never say out loud. The tools are astonishing at assisting a skilled human and nowhere near capable of replacing one. The people closest to the technology are voting with their chequebooks, and they are voting for humans. If you want a single honest signal about how mature this technology really is, ignore the press releases and watch who the model builders are desperately trying to hire.

Where the true costs are hidden

This is the part the glossy demos never show you, and it is where a lot of the bubble risk genuinely lives, so let us dig into it properly. The sticker price of AI is not the real price, and the gap between them is where a lot of companies are going to get hurt.

The first hidden cost is inference, which is the fancy word for actually running the model every time someone uses it. Everyone obsesses over the eye watering cost of training a model, the one big upfront number that makes headlines. But training is a one off. Inference is forever. Every single question, every generated image, every summarised document costs real electricity and real hardware time, and it never stops for as long as the product exists. Plenty of AI companies are quietly losing money on every active user, subsidising the true cost to keep the growth numbers pretty, in the hope that chips get cheaper or that they can raise prices later without everyone leaving. That is a dangerous game, and it is the same game a lot of dot com companies played right before the music stopped.

The second hidden cost is the one nobody puts on a slide: the cost of being wrong. An AI that is right ninety five percent of the time sounds fantastic until you remember that the five percent arrives wrapped in exactly the same tone of total confidence as the ninety five. So now you need a human to check the output, which eats a good chunk of the efficiency you were promised. In low stakes work, a slightly wrong holiday suggestion, that is fine. In law, medicine, accounting, or engineering, the checking is not optional, and the cost of the mistakes that slip through the checking can dwarf everything you saved. A lot of the productivity gains that look so impressive in a demo quietly evaporate once you add the cost of the human who has to sit behind the machine making sure it did not invent something.

The third hidden cost is people and process. Dropping AI into a company is not plug and play, whatever the vendor tells you. Someone has to clean the data, connect the systems, retrain the staff, rewrite the workflows, and babysit the thing when it drifts. The licence fee on the invoice is often the smallest number in the whole project. The real bill is the months of expensive human effort it takes to make the shiny tool actually useful in a messy real world business, and that bill rarely appears in the excited projections about how much money AI is going to save everyone.

And then there is the cost that does not land on any single company's books at all: electricity and water. These data centres draw staggering amounts of power and, in many designs, water for cooling. That cost is real, it is rising, and increasingly it is being paid by everyone in the form of higher energy prices and strained local grids, whether they use AI or not. It is a genuine externality, and it is one of the reasons the current trajectory may not be as clean or as cheap to sustain as the growth charts assume.

The benchmark problem, or why tokenmaxxing is quietly rotting the leaderboards

This one is more technical, so let me build it up gently, because it is genuinely important and almost nobody outside the field understands it. The short version is that the scoreboards everyone uses to decide which AI is best are being gamed, and the way they are being gamed is making the models more expensive and slower for reasons that have nothing to do with being more useful to you.

Start with the basic idea. To claim your new model is the best, you run it against a set of standard tests, called benchmarks. Maths problems, coding challenges, reasoning puzzles, exam questions. The model gets a score, the score goes in a chart, the chart goes in the marketing, and investors and journalists treat the chart as gospel. So far so reasonable. The trouble is that the moment a number becomes the thing everyone is judged on, people stop trying to be good and start trying to score well, and those are not the same thing at all. There is an old law that says when a measure becomes a target, it stops being a good measure. AI benchmarks are a live demonstration of it.

There are a few ways this goes wrong, and they stack on top of each other. The first is contamination, which is the polite word for the test answers leaking into the study material. These models learn from vast sweeps of the internet, and the benchmark questions and their answers are often sitting right there on the internet too. So the model has, in effect, seen the exam paper before sitting the exam. It scores brilliantly, everyone applauds, and then it falls flat on a genuinely new problem it could not have memorised. You have measured memory and called it intelligence.

The second, and this is the one your word tokenmaxxing points at, is more subtle and more interesting. Many benchmarks reward a model for reaching the right final answer, and it turns out that models score higher when they are allowed to ramble through a long chain of reasoning before answering, generating heaps of intermediate text, what the field calls tokens, on the way. More thinking tokens, better scores. So the labs tune their models to produce these enormous verbose reasoning trails, because it pushes the leaderboard number up. Tokenmaxxing is exactly that: maximising the number of tokens the model chews through in pursuit of a better benchmark result.

Here is why that should annoy you as a user rather than impress you. Every one of those extra tokens costs money and takes time. You are paying, in cash and in seconds of your life, for the model to talk to itself at great length so that it can win a contest you are not even competing in. A model that has been tuned to top the leaderboards may be measurably slower and more expensive to run than a plainer one that gives you an equally good answer in a quarter of the words. The benchmark says the verbose one is better. Your bill and your patience say otherwise. The score and the actual usefulness have come apart, and the gap is being paid for by you.

And it compounds into a genuinely bad incentive for the whole field. If the way to look best is to burn more compute per answer, then the entire industry is being quietly pushed towards models that are more expensive to run, right at the moment when the biggest question hanging over the whole bubble is whether anyone can run these things profitably. The leaderboards are cheering the models on in exactly the wrong direction. They reward the behaviour that makes the unit economics worse. When you next see a chart showing model X has pipped model Y by two points on some benchmark, the correct response is a shrug and a question: better at what, measured by whom, and at what cost per answer to actually run.

So, bubble or not? A grown up verdict

Let me try to land this honestly rather than reaching for a tidy soundbite, because the tidy soundbites are all wrong.

Yes, there is a bubble. There is far too much money chasing far too many identical ideas, valuations in places have completely detached from any plausible revenue, a lot of the spending is circular and flattering itself, and a painful correction that wipes out a big crowd of the wrapper companies and the weakest of the model players is not just possible, it is close to inevitable. When it comes, it will be brutal and it will be loud, and a lot of people who confidently told you AI could not fail will discover they were describing a stock price, not a technology.

And also, no, it is not only a bubble, and this is the part the gleeful doom merchants keep getting wrong. Underneath the froth there is a real, working, deeply useful technology that hundreds of millions of people have already woven into their days and will not give back. There is a genuine core of problems, in science, in medicine, in the thousands of tiny unglamorous corners of ordinary businesses, that AI has actually solved or made solvable for the first time. That core does not evaporate when the share prices fall. It just gets quietly more valuable while everyone is distracted by the crash.

The dot com era is the perfect guide precisely because both halves of it were true. The crash was real and it ruined people. The internet was also real and it went on to eat the entire economy. The companies that died deserved to die. The ones that survived, the ones actually solving a real problem rather than adding an e to their name, became some of the most important businesses in the world. The same sorting is coming for AI. The bubble bursting and the technology mattering are not opposing predictions. They are the same story told at two different speeds, and the smart move is to stop shouting about which one is true and start working out which side of the line your own idea sits on.

The winners will be the ones who stopped asking how to sprinkle AI on top of what already exists, and started asking what suddenly became possible that never was before. Everyone else is just laying fibre nobody will light, and buying a very expensive front row seat for the pop.