"Is AI a bubble?" is being asked in two very different rooms. In one, investors are asking whether the companies building AI infrastructure are priced for a future that may arrive late or smaller than promised. In the other, people who run businesses are asking whether it is still sensible to put AI into their operations — or whether they should wait until the dust settles.
Those are not the same question, and treating them as one leads to bad decisions in both directions: companies that freeze because of headlines about valuations, and companies that buy whatever is being sold because the technology is "obviously the future". This piece separates the two, sets out what the bubble case actually rests on, and then deals with the question a business can act on: how to build so that a correction, if it comes, costs you as little as possible.
Nothing here is investment advice. It is about building systems, not buying shares.
What the bubble case rests on
The worry is not that AI does nothing. It is that the money going into it has run ahead of the money coming out, and that the way it is being financed would make a slowdown painful.
The scale of the spending. Alphabet, Amazon, Meta and Microsoft together planned around $725 billion of capital expenditure for 2026, up 77% from about $410 billion in 2025, most of it on data centres, chips and the power to run them. That spending has to be repaid by revenue from AI products that, in many cases, are still finding their customers.
How it is paid for. Early in the boom the largest companies funded this from their own cash flow. Increasingly they borrow. In February 2026 Alphabet sold a 100-year bond as part of a wider debt raise while it expanded its AI data centres — the first century bond from a technology company since Motorola's in 1997. Debt is not a problem in itself, but it turns "this may take longer to pay off" into "this has to be paid off on a schedule".
Money that goes round in a circle. In September 2025 Nvidia announced plans to invest $100 billion in OpenAI to help finance data-centre build-outs that run on Nvidia's chips. When a supplier finances its own customers, sales figures can overstate how much independent demand exists. Analysts have pointed out the resemblance to the vendor financing and "round-tripping" that made the dot-com collapse worse.
Hardware that ages quickly. Large cloud providers generally write their servers off over five and a half to six years. Critics, the investor Michael Burry most prominently, argue that AI chips are superseded far sooner, which would mean today's profits are flattered by depreciation that is spread too thinly. Defenders reply that older chips are still rented out and still earning. Nobody knows yet who is right, because the hardware has not been around long enough to find out.
Even people at the centre of the boom concede part of the case. In August 2025 OpenAI's chief executive, Sam Altman, said investors were "overexcited" about AI and compared the moment to the dot-com era — while also calling AI the most important thing to happen in a very long time.
What the other side says
The counter-argument is not that prices are reasonable. It is that this boom is built differently from the last one. The companies doing most of the spending are among the most profitable in the world, the products are earning real revenue, and valuations, measured against expected earnings, sit well below where internet stocks were in 2000.
The dot-com comparison is worth taking seriously, because it cuts both ways. Between the late 1990s and 2002, network builders put some $30 billion into laying 90 million miles of fibre-optic cable, and by 2002 the vast majority of it lay dark. Part of the reason was a technical leap: wavelength-division multiplexing let a single strand carry many times more traffic than before, so the industry suddenly needed far less fibre than it had laid. Investors lost fortunes. And yet that cable was later lit, and it carried the internet that the next twenty years were built on.
A bubble can be right about the technology and wrong about the price and the timing. For a business deciding what to build, that distinction is most of what matters.
The evidence that matters if you run a business
If you are not an investor, the most useful data in this debate is not about valuations. It is about whether AI is producing results inside companies — and here the picture is sobering.
A National Bureau of Economic Research survey of nearly 6,000 executives in the US, UK, Germany and Australia, published in February 2026, found that about 69% of firms use AI in some form, yet 89% reported no change in productivity over the previous three years. A smaller, explicitly preliminary study from MIT's NANDA project in 2025 reported that about 95% of the organisations it examined saw no measurable return from generative AI pilots. Treat that figure with care — the sample was small — but note the reason its authors gave: the failures came less from the quality of the models than from how organisations tried to use them.
That matches how AI projects usually go wrong. The model is rarely the reason one disappoints. The reasons are everything around it: a pilot that was never connected to the systems where work actually happens, a task chosen because it was impressive in a demo rather than because it was expensive to do by hand, no way to tell whether the system was right, and nobody responsible for what it did. A correction in the stock market changes none of that. Neither does a boom.
What a correction would actually change for you
Suppose the pessimists are right and the market corrects sharply. For a business using AI, three things could move.
The price of using models. This is the one most people get backwards. The cost of AI has been falling fast: Epoch AI found that the price of reaching a fixed level of model performance fell by between 9 and 900 times per year, depending on the task, though it cautions that the fastest drops may not last. When DeepSeek released a cheaply trained, openly available model in January 2025, Nvidia lost close to $600 billion of market value in a day — bad news for chip investors, good news for anyone paying for AI by the request. A correction could push prices either way. Providers under pressure to show profits may raise prices or end generous free tiers; a glut of data-centre capacity would push them down. A system that only makes sense at today's prices is a fragile system.
Who you are buying from. In any shake-out, some providers are acquired, some change their terms and some disappear. Providers retire older models on their own schedules even in good times. If your process depends on one vendor's model behaving exactly as it does today, that is a risk regardless of the market.
Your own budget. When sentiment turns, the projects that survive are the ones that can show what they return. Pilots justified by enthusiasm are the first to be cut.
Building so the answer doesn't matter
None of this argues for waiting. It argues for building in a way that does not depend on who wins the argument about valuations. In practice that comes down to a handful of decisions, most of which cost little if made at the start and a great deal if retrofitted.
Put the model behind your own interface. Your workflow should call a component you own — one that knows which model to use for which task, what to send it and what to do when it fails — not a particular vendor's API directly. Changing provider then becomes a configuration change and a round of testing, not a rewrite.
Measure the cost of a finished task, not the price of a request. What matters is what it costs to process one invoice, answer one customer or qualify one lead — including retries, human review of uncertain cases and the infrastructure around the model. Know that number, and know how much the model's price could rise before the task stops paying for itself.
Keep the data and the record. Your documents, your prompts, your evaluation cases and the log of what the system did belong in your own storage. They are what let you test a replacement model against the old one on your real work instead of trusting a benchmark, and they are the part of the system that keeps its value whatever happens to any provider.
Design for the model being unavailable or wrong. A well-built system already routes uncertain cases to a person and keeps working, more slowly, when a provider has an outage. The same design carries you through a provider changing its terms or leaving the market.
Choose work you can measure. The same MIT study found the strongest returns in back-office functions rather than in the sales and marketing tools that attract most of the budget. That fits: documents, reconciliation, routing and data entry are work whose before and after can be counted. A project that can show its return survives a budget review in any market.
The short version
Whether AI stocks are overpriced is a question for investors, and serious people disagree about it. Whether a particular AI system is worth building is a question about one task, one process and one set of numbers — and it can be answered without predicting the market.
If the pessimists are right, the businesses that come through best will be the ones whose AI systems were never tied to a single vendor, a single price or a single round of enthusiasm. If the optimists are right, those same businesses lose nothing for having been careful. Either way, the model was never the system.