Is the AI Bubble Bursting? A 2027 Software Buyer's Guide

Open any business paper this autumn and someone is calling the top. US AI-related capital spending cleared $400 billion in 2025 and keeps climbing, a handful of chip, cloud and model companies are selling capacity to each other in deals that look suspiciously circular, and a widely cited MIT study found that around 95% of corporate generative-AI pilots delivered no measurable return. Put those next to each other and "bubble" is a fair word to reach for.

The problem is what people do with that word. Some founders are using it as a reason to freeze every software decision until the dust settles. Others are ignoring it entirely and signing three-year commitments at peak hype. Both are mistakes, and they are opposite mistakes, which tells you the headline is not actually a plan.

This guide is not a market forecast. We have no idea whether valuations correct in 2027 or 2029, and neither does anyone selling you a prediction. What we can do is separate the part of the bubble that affects you from the part that does not, and give you a way to keep investing in software that holds up whichever way the market goes.

What people actually mean by "the AI bubble"

"Bubble" is doing three different jobs in these conversations, and they are worth pulling apart because only some of them are your problem.

• A capital bubble. Investors have poured enormous sums into infrastructure, chips and model companies on the bet that demand will catch up. If that bet is early, some of those investors lose money and some companies fold. This is a financial-markets story.
• A valuation bubble. Public and private AI companies trade at multiples that assume near-flawless execution for years. A correction here hits shareholders and acquirers.
• A usefulness bubble. The claim that the technology itself does not work, or does not pay back. This is the one most relevant to you, and it is the weakest of the three.

The MIT figure gets quoted as proof of the third. Read the actual report and it says something more useful: the pilots failed mostly on integration and a "learning gap," not on model quality. Generic tools bolted onto a workflow they do not understand stall. The same study found the best returns came from unglamorous back-office automation, not the sales-and-marketing toys that soaked up most of the budgets. That is not "AI does not work." That is "most companies deployed it badly," which is a solvable problem and, frankly, the story of every technology wave.

The bubble is in the capital, not in the capability

Here is the distinction that should drive your decisions. The dot-com crash wiped out Pets.com and a few hundred billion dollars of paper wealth. It did not make the internet less useful. The companies that kept building sensible things on top of it, and the ones that waited for the panic to buy talent and infrastructure cheaply, came out of 2001 and 2002 in a stronger position than the ones that either torched their cash or hid under the desk.

The same shape applies now. If there is a correction, it lands on the people who funded the data centres and bought the stock. The models you already use do not get worse the morning after. If anything, a shakeout tends to push prices down and force the survivors to compete on reliability and real outcomes instead of demos, which is good for a buyer.

So the question "is the bubble bursting?" is the wrong one for a software budget. The right ones are narrower: which of my vendors would actually disappear in a downturn, and what am I building that keeps paying back even if the hype evaporates entirely?

What a correction would, and would not, change for you

Picture the pessimistic scenario: funding tightens through 2027, a few well-known AI startups fail, and model prices stop falling.

What changes: some vendors raise prices or disappear, free tiers get stingier, and the "we will figure out revenue later" startups get forced to charge what their product costs. You will want portability and an exit plan for anything critical.

What does not change: the software that automates your invoicing still saves the hours it saved. The agent that resolves a third of your support tickets still resolves them. A correction does not reach into your systems and undo work that already pays for itself. Value that was real stays real.

That asymmetry is the whole point. Spending tied to a concrete, measured outcome is defensible in any market. Spending tied to a narrative ("we are an AI company now") is what gets cut first when the narrative turns, and it deserves to be.

Decisions that look smart whichever way the market goes

You do not need to predict the macro to make good calls. These hold up in a boom and in a bust:

1. Fund outcomes, not AI. Approve the invoice-automation project because it saves 40 hours a month, not because it has "AI" in the deck. If you cannot name the outcome and roughly size it, you are buying a story. Our guide to measuring AI ROI walks through how to put a number on it before you commit.
2. Keep an exit from every critical vendor. Own your data, prefer tools that export cleanly, and avoid wiring a core process so tightly to one provider that you cannot leave. This is the lesson from AI vendor lock-in, and it matters most precisely when funding gets tight.
3. Prefer boring, back-office wins first. The MIT data and plain experience agree: document processing, reconciliation, triage and the like pay back faster and more reliably than flashy customer-facing bets.
4. Build the thin layer that is yours, buy the commodity underneath. Models and infrastructure are commodities trending cheaper. The workflow, data and judgement wrapped around them are your advantage. Our build versus buy breakdown covers where that line sits.
5. Scope small, prove it, then scale. A contained project that ships and pays back beats a cathedral that gets cancelled. This is also how you avoid being one of the 40% of agentic projects that get scrapped by 2027.

Notice none of these require a view on whether Nvidia is overvalued. They are just good software discipline, which is the point: the companies that survive a bubble are usually the ones that were not relying on it.

How to invest through the noise

If the bubble talk has you tempted to pause everything, run a quick test on your current and planned spend. For each line, ask: does this pay back on its own, inside a year, measured in hours or euros I can point to? Spend that passes keeps going regardless of the market. Spend that fails that test was never really an AI decision, it was a hope, and the market correcting just saves you the trouble of cancelling it later.

The 5% of pilots that worked in the MIT study were not the ones with the biggest budgets or the newest models. They were the ones aimed at a specific, measurable job and wired properly into the business. That is entirely within your control, bubble or no bubble, and it is the same approach that produces software worth owning in any decade. If you want help sorting the defensible projects from the hopeful ones before you plan 2027, that is exactly the kind of work we do, and it starts with the outcome, not the hype.