The AI Bubble: The Party, the Bill, and the Person Nobody Wants to Sit Next to
The AI bubble is already touching the ground.
This summer, somewhere in the south east of England, a household was told not to use a hosepipe.
Not far away, a building was drawing up to nineteen million litres of water a day. Drinking quality water, because under current Environment Agency rules that is the only kind a data centre is allowed to be supplied with. It did not stop. It could not really be asked to stop. The building had been designated critical national infrastructure, which leaves water companies with very little power to turn the taps down.
Nobody voted for that trade. It was not debated on a doorstep or put in a manifesto. It emerged, quietly and lawfully, from a designation made for national security reasons, and it landed on a garden in South London in August 2026.
That is the AI bubble touching the ground. Not a chart of Nvidia's share price. A tap.
Hold the image, because everything else here is the explanation for it. The capital expenditure, the circular deals, the redundancies announced for productivity gains that have not yet arrived. All of it eventually resolves into physical objects sitting in fields, drinking, on the strength of a demand forecast that somebody wrote down and nobody independently checked.
A word on what I am and what I am not. I am not an economist. I have no view worth having on where the S&P 500 goes next, and I would be suspicious of a lawyer who claimed one. What follows is not my forecast. It is a reading of what the people who are qualified have actually said, in their own words, followed by the part I do know something about, which is what their arguments mean for a contract, a board minute or a consultation pack. Take the economics from them. Take the legal consequences from me.
In short. In July 2026 the Bank of England warned that a burst AI bubble could reduce UK GDP by up to 2.2 per cent. Economists disagree sharply on whether current valuations constitute a bubble: Jeremy Grantham calls it the biggest investment bubble in American history, Howard Marks says valuations are high but not yet manic, and David McWilliams argues the greater danger is that AI succeeds and triggers debt-destroying deflation. For UK businesses the legal exposure does not depend on which is right. It sits in long compute commitments signed at peak pricing, circular financing arrangements that inflate apparent demand, board minutes that record only the upside case, redundancy business cases resting on productivity gains that have not yet materialised, and data centre projects whose water and planning risk has not been allocated in contract.
Here is the rest of what the record shows on that tap, because this argument is usually conducted on vibes.
More than three quarters of UK data centres sit in the water stressed south and east of England. Thames Water estimates a large data centre might use between 4 and 19 million litres a day, comparable to the daily demand of around 50,000 households. Water UK, the trade body for the water companies, told MPs in a written briefing that the AI growth zone plans are "fatally flawed", that projections explicitly exclude data centres, and that the relevant policy documents make not a single mention of water. There appears to be an assumption, it said, that the country will always have enough water for its economic needs, and nothing could be further from the truth. A House of Lords report in May reported that the Environment Agency projected England facing a five billion litre shortfall by 2055.
One side of this argument treats data centres as the new aircraft carriers. Foundational to defence, cyber capability and economic sovereignty. On that view the toothpaste is out of the tube, the only live question is whether Britain has any say in how AI gets built, and local planning objections are a rounding error against national security.
And this position has serious economists behind it, not just lobbyists. The Irish economist David McWilliams, who is nobody's idea of a tech industry cheerleader, has argued plainly that you cannot live in a digital age without data centres and that Ireland needs more of them, not fewer. His point is that the objection is often incoherent: we want the digital economy, the jobs and the tax receipts, and then we object to the physical building that produces them. If you use the cloud, you have a data centre. You simply have it somewhere you cannot see. Sounds similar to the arguments we had around wind farms.
The other side asks a simpler question. At what cost, to whom, and who was asked?
I am not going to pretend I can settle that. What I can tell you is where it becomes a legal problem, and it already has.
The legal bit. This is where the next three years of judicial review sits, and a Commons committee is already taking written evidence. If data centres are critical national infrastructure, what does that do to planning law, abstraction licensing and local consent? Should permissions require independently verified economic impact assessments rather than developer projections? Should there be mandatory reporting of water consumption, source and peak demand, as several submissions to that inquiry have recommended?
And commercially, if your model depends on the proposition that AI infrastructure is sacred and non-negotiable, you have taken a position in a live political argument. Those rarely end cleanly. They almost never end quickly.
Which brings us to the money, and to the party.
Three kinds of people at the AI party
The first are dancing. The second are watching the door. The third wandered over to the bar, found the tab, and have gone very quiet.
I spend my working life with the third group. Founders, general counsel, finance directors, operations people. Nobody in that room asks me whether AI is real. They ask something narrower and far more useful. If the money stops, what happens to my contracts, my headcount, my covenants and my board?
The rest of this is an attempt at an answer. Not to time the market, which nobody can do, but to separate the risks already sitting in your business from the ones still living in a slide deck. I have checked every figure below against its primary source. Where a widely repeated claim did not survive checking, I say so. There is a great deal of confidently sourced nonsense on this topic, some of it in newsletters that look exactly like this one.
One word on my own position. It has not changed. AI is not the risk. AI without oversight is. That turns out to apply to the technology and to the balance sheets built around it.
Camp one: the Bank of England and the case that this is a bubble
The central bankers, and their language has changed.
In October 2025 the Bank of England's Financial Policy Committee wrote that "the risk of a sharp market correction has increased", and called the risk of spillovers to Britain's financial system material. Read that as the plain English it is. Central bankers are trained to speak in a register somewhere between a shipping forecast and a hostage video, and that sentence is, by their standards, shouting.
In the July 2026 Financial Stability Report, Governor Andrew Bailey put a number on it. An AI bubble bursting could take up to 2.2 percentage points off UK GDP.
That is the only figure here describing what happens to Britain rather than what happens to Nvidia. A 2.2 point hit is not a bad quarter for people who own tech shares. It is a recession, delivered to a country whose exposure runs through pension funds and index trackers held by people who have never knowingly bought an AI stock in their lives.
Bailey described what he called a triple whammy. "There's a triple whammy," he told the Treasury Select Committee, and it is routinely misreported as three flavours of the same overvaluation. It is not. The three are stretched AI valuations, cyber risk arising from the speed of AI development itself, and the use of AI by financial firms in automated trading. Nobody writes about the second and the third. They are the two that turn a market correction into an operational event on a Tuesday morning.
JeremyGrantham, who called dot com and housing, went on The Diary of a CEO in June. His words, not mine: this is "the biggest investment bubble in American history".
He was precise about that word American. He thinks Japan in 1989 holds the world record, trading at 65 times earnings at the peak, then falling for two decades and taking thirty five years to recover. He puts US shares now at 35 to 40 times. He thinks the fall could be around 70 per cent. His advice to retail savers ran to three words: don't own US stocks.
The part of Grantham worth more than the headline is his theory of why this happens at all. "The great bubbles always occur around the very most important ideas," he said, listing railways in the 1840s, the Nifty Fifty, the internet, and now AI. The bigger the idea, the bigger the burst.
Which produces a conclusion that people keep refusing to hold in their heads at the same time. On Grantham's account, believing AI is world changing and believing AI is a bubble are not opposing positions. They are the same position. The railways genuinely did change everything, and a great many people were ruined buying them.
Grantham has also been early before, and says so himself. In 1998 and 1999 his firm lost half its assets under management calling dot com two years too soon. He knows this. It is rather the point of listening to him: he will tell you what, and he will not tell you when.
What this means if you run something. The window is now and the next few years, and the direction of travel is from growth at any cost to show me cash flow. If you are raising, stress test the model against three things happening at once: compute prices falling on oversupply, customers deferring AI budgets, and investors wanting positive free cash flow inside eighteen months. If you are buying, think hard before signing long compute commitments into a market that may be about to get cheaper.
The legal bit. Directors approving large AI capital programmes on thin near term returns should be documenting the downside case, not just the upside one. And it is worth being honest about why they often do not. Approving the spend and being wrong is survivable, because everybody was wrong together. Refusing the spend and being wrong is career ending, because you were wrong alone. Nobody has ever been fired for buying compute. That asymmetry, not any forecast, is what most AI capex is actually priced on.
Section 172 has no exemption for exciting technology. If the project goes wrong and the board minutes contain nothing but enthusiasm, that gap will be read aloud to somebody later, slowly, by a barrister who is enjoying themselves.
Camp two: Howard Marks and the case that this is not yet mania
Howard Marks, Robin Greenwood and Jason Furman sit here, more or less. Markets are noisy. Enthusiasm is high but has not reached critical mass. Some segments are overvalued while the underlying capability is genuinely worth something.
Marks has been the most quotable. Valuations, he told CNBC, are "high but not crazy", adding the line that ought to be pinned above every trading desk: "Expensive and going down tomorrow are not synonymous." He declines to apply the bubble label because he cannot find the psychology. "To me, the main ingredient in bubbles is psychological excess," he said. "I don't detect that level of mania at this time."
That is a genuinely different claim from Grantham's, and I want to be careful not to flatten it. Grantham is measuring price. Marks is measuring people. They can both be looking at the same market and reporting accurately.
Marks is not sanguine, mind. His December memo set out the concentration plainly: AI related stocks accounting for around 75 per cent of the S&P 500's gains and roughly 90 per cent of its capital expenditure, alongside startups valued at $50 billion with no product, and circular transactions he flags as a red flag rather than a footnote.
Citi's numbers are the cleanest version of the balanced argument. Its Bear Market Checklist tracks 18 red flags across earnings forecasts, fund flows, valuations, capex, investor sentiment and equity issuance. In June it hit 10 of 18 globally, the frothiest reading since 2008, with the United States at 11.5 and Europe at 5.
The same checklist reached 17.5 of 18 in March 2000, and 13 before the financial crisis. Citi remains constructive on equities to year end. Its warning is about direction, not level: once the count reaches double digits, it has historically risen faster.
Both halves of that matter. We are frothier than at any point since 2008 and still meaningfully short of 2000. I mention it because the number travels without its context, and a reader who has only met one half of it will make a different decision from one who has met both.
There is a UK version worth borrowing. Not that AI is a bubble, but that there is a provider bubble. Too many vendors chasing the same contracts, real productivity gains underneath, and a shakeout coming for the middle of the market.
What this means if you run something. Winners solve a costed problem: cost per task, reliability, integration, compliance. Losers monetise the buildout. If your niche has suddenly acquired eleven competitors with near identical decks, you are in the provider bubble, and consolidation will find you before regulation does.
Camp three: the technology is real, the financing is not
Roughly the FinancialTimes and Economist position, and the most useful frame for anyone signing contracts.
AI is not a mirage. The financial architecture around it is fragile. Debt funded capex. Circular deals, where a hardware supplier invests in an AI lab which then commits to spend with the same supplier, inflating what looks like demand. Cheerful assumptions about how quickly capability converts into revenue.
MartinWolf's framing is the one I keep returning to, because it removes the false choice everyone else is stuck in. AI is a general purpose technology whose economic impact will take years to arrive, and the hard question is not whether it works but who captures the profits. Overinvestment, destructive competition, bankruptcies and consolidation are, on his account, standard features of that process rather than evidence against the technology. JohnPlender, in the same paper, has been blunter: the euphoria meets the classic bubble criteria, and it is a new telling of an old story.
The interesting thing about circular deals is that they are not really deception. Everyone can see the loop. It is disclosed, filed and reported on. Marks flags it in his memo; Nvidia's own estimated exposure to such deals has been put at around 15 per cent of 2025 sales.
They work because of what a purchase order normally means. Ordinarily, committing to spend billions is expensive and painful, which is exactly why we treat it as evidence. It is a costly signal. You do not put your balance sheet behind demand you do not believe in. Circular financing does not fake the signal. It makes it cheap to send, while leaving it looking identical to the expensive version. The market keeps reading the signal correctly, and the signal has stopped meaning what it used to mean.
That is why disclosure alone does not fix this, and why "but it was all in the accounts" will be a thin defence later.
Oracle is the case study, and it has moved fast enough that most commentary on it is out of date.
Blue Owl walked away from a $10 billion Oracle data centre financing in December 2025 and the stock fell. That was the warning shot. What followed was worse. Oracle's capital expenditure surged to $55.7 billion in fiscal 2026 from $21.2 billion the year before, leaving free cash flow at negative $23.7 billion. It has $638 billion in remaining performance obligations, of which around $300 billion is reportedly attributable to a single customer. In June it announced plans to raise roughly $40 billion more, having already taken $43 billion in debt and $5 billion in equity in the prior year. The stock had its worst week since the dot com bust in 2001 and has fallen roughly half from its September peak.
None of that is a story about AI failing to work. It is a story about who is holding the paper when the capacity arrives.
Alphabet is the quieter signal. Its free cash flow turned negative for the first time since it went public. Moody's has warned that heavy capital spending relative to revenue will lead to "declining, and in some cases negative, free cash flow", and will hurt leverage ratios where the spending is debt financed.
While we are here, a word about a figure you have seen everywhere. The claim that 95 per cent of enterprise AI pilots never reach production is a mangling of a study that measured pilots showing no measurable return in the profit and loss account. That is a different finding, and a more interesting one. If you are going to cite it, cite what it said.
The legal bit, and this is where I earn my fee. Three questions I now ask on any AI heavy commercial deal.
First, is this strategic partnership arm's length, or is it circular? If your counterparty's revenue and your counterparty's investment run in a loop with the same third party, the demand you are underwriting may not exist.
Second, who bears the risk if a data centre project is delayed, descoped or becomes uneconomic? In a great many agreements that risk is currently allocated by silence, which means it sits wherever the litigation eventually puts it.
Third, if you are raising on an AI narrative, does your disclosure overstate organic demand or understate capex risk? Prospectus liability does not care how good the technology is.
Camp four: what AI is actually doing to jobs
Here I have to correct received wisdom, including a version of it I have repeated myself at conferences this summer.
The story everyone is telling is that AI is eating jobs at an accelerating rate. Through May, the story held. AI was cited in 38,579 US job cuts in that month alone, 40 per cent of the total, up from 7 per cent in January. It was the highest monthly figure since tracking began.
Then July happened. US employers announced 33,429 job cuts, the lowest monthly total in two years. Cuts through July stand at 477,033, down 41 per cent on the 806,383 announced in the same period of 2025. Hiring plans are up 25 per cent on last year. AI has been cited in 112,713 cuts this year, about 24 per cent of the total, and it has led all reasons for five consecutive months.
Andy Challenger, whose firm sells outplacement services and therefore has every commercial incentive to say the opposite, put it this way: "while AI is shifting the labor market, it is not dismantling it."
Both things are true at once. AI is the most cited reason for redundancy in the United States, and redundancies overall have fallen off a cliff. If you had only the May data, you wrote a very confident article that has aged badly. I nearly did.
Now the wrinkle, and I want to flag clearly that this next part is my own reading rather than anybody's finding.
Firms are cutting in anticipation of AI's effects rather than in response to demonstrated productivity gains. That much is documented, and the scepticism is not mine alone: Glassdoor's chief economist has publicly cautioned against taking corporate AI attributions at face value, and researchers at the OxfordInternetInstitute have used the word scapegoating. Some firms are using AI as a respectable narrative for cuts that were coming anyway on cost pressure, National Insurance and offshoring.
Read that as an economist and it is a forecasting error. Read it as a marketer, which is closer to my own instinct here, and it is something else. If you cut two hundred roles and attribute it to AI, you have told the market you are serious, disciplined and early, at a moment when being seen to be early is worth more than being right. The productivity gain is the stated reason. The announcement is the actual product. The redundancy is an advertisement, and the people in it are the media spend.
That is a theory, not a finding, and I would like to be wrong about it. The numbers make it hard.
The legal bit. If a redundancy is justified by a productivity gain that has not yet materialised, the business case in the consultation pack is doing an enormous amount of work. Tribunals will start asking for the evidence behind it. "We are becoming an AI first organisation" is a strategy statement, not a redundancy justification, and I would not want to defend it under cross examination.
The fifth camp: Davis McWilliams and the deflation risk
There is a position missing from most of this debate, and it is the one that should worry a debt heavy business most.
David McWilliams set it out in the IrishTimes eleven days ago, and his argument runs in the opposite direction to everyone above. His concern is not that AI fails to deliver. It is that it delivers exactly what it promises.
His logic is short. All technological innovation is deflationary, because what drives technology is efficiency, and efficiency means getting more out of less. Falling prices follow. As prices fall, wages fall. And since most people's income comes from wages, and income is what services debt, the first casualty is solvency. Global debt is now at a record $353 trillion, with $29 trillion added last year alone.
His historical parallel is the deflation of 1870 to 1900, when rail freight rates in the US fell from about 3 cents per tonne mile to around 0.75, wheat went from $1.19 a bushel to 49 cents, and agricultural land values in Ireland and England halved in thirty years. That period was kicked off by the crash of 1873 and a slump in railway stocks. Railway building slowed. Railways carried on changing the world regardless.
His conclusion is that AI follows the same path, and that the legacy is deflation rather than inflation, with "unemployment rises as AI replaces the service jobs of the clerical middle (consultants, lawyers and administrators), leading to falling tax revenue and mass debt defaults."
I notice he put lawyers on that list. I have no serious basis for arguing with him about it.
I include this camp for a reason that has nothing to do with balance for its own sake. If you are a founder or an SME owner reading the earlier sections and quietly concluding that the safe move is to hope AI underdelivers, McWilliams is the person who takes that comfort away. There is a version of the future where the technology works completely, the productivity arrives, and the resulting deflation is what breaks the debt. Both tails are live. Planning for only one of them is not caution.
When does it turn? Three scenarios, and what to watch
Nobody can time this. You can, though, watch the right instruments rather than the wrong ones.
Scenario one, soft landing. Capex growth slows but stays positive. Valuations compress gradually. Weaker vendors exit, stronger ones consolidate. New roles appear in governance, integration and domain applications.
Watch for: hyperscalers holding positive free cash flow, stable credit spreads on AI related debt, no major data centre defaults.
Your move: differentiate and get profitable. Expect normal funding conditions with harder questions about unit economics.
Scenario two, provider shakeout. Capex growth halves over eighteen to twenty four months. Data centre projects are delayed or cancelled. Infrastructure and pure play valuations compress sharply. The broader market wobbles without breaking.
Watch for: write downs on data centre projects, widening credit spreads on AI debt, distress among operators. A survey of more than 400 data centre executives this summer found 68 per cent expecting distressed situations to increase over the next twelve to eighteen months.
Your move: capital gets scarce for AI heavy stories. There is money in helping firms govern, optimise and integrate the AI they have already bought.
Scenario three, full bust. A trigger, a default, a failed listing, a macro shock, causes sharp repricing. Hyperscalers cut capex. Revenues miss. Growth and employment take the hit.
Watch for: sustained negative free cash flow at the majors, multiple project failures, a collapse in AI listings. We have already had a wobble on that last one, with a major data centre IPO pricing below its range last month.
Your move: preserve cash, defend the core, and expect a great deal of contract renegotiation. Restructuring and insolvency will not be short of instructions.
What the law should be asking about the AI bubble
One newsletter is not going to settle any of this. The questions are worth putting on the record now, because the answers get litigated later.
Corporate and securities. Are disclosure regimes adequate for circular financing arrangements? Should directors have an explicit obligation to stress test AI capital programmes? How should prospectuses treat growth assumptions resting on demand that does not yet exist?
Employment. If AI attributed redundancies are anticipatory rather than evidenced, how should tribunals treat the business case? What stops AI led restructuring from landing disproportionately on protected groups?
Public law. If state support for AI infrastructure rests on contested multipliers, what are the grounds for challenge? Should communities have stronger rights to information where projects touch water security?
Financial regulation. How should regulators treat related party deals inside AI ecosystems, where the same money appears on three balance sheets and is counted as demand on all of them?
Dancing near the exits
The AI bubble debate is really a debate about us. We are extraordinarily good at inventing transformative things and hopeless at pricing them while they are happening. We swing between this time is different and we have seen this before, always with total confidence in whichever one we are currently saying.
I have deliberately not told you which camp is right, because I am not qualified to and because it is the wrong question for most people reading this. Bailey, Grantham, Marks, Wolf and McWilliams disagree with each other in ways that are genuinely unresolved, and anyone who tells you otherwise is selling a newsletter rather than writing one.
What I would say is this. Every one of those five positions, including the optimistic ones, produces the same instruction for a business. Build for optionality rather than prediction. Shorter commitments. Documented board decisions. Contracts that allocate the risk deliberately instead of by silence. A straight answer to what happens if your largest customer's funding halves. None of that requires you to know who is right. All of it looks sensible in every scenario above, which is the only kind of advice worth taking about a future nobody can see.
If you are a lawyer or a policymaker, the useful posture is to stop treating this as somebody else's subject. If the boom becomes a bust, the aftermath gets fought in tribunals, planning inquiries and regulatory hearings long after the trading desks have moved on.
And for the rest of us, a mixture of awe and scepticism seems about right. AI really might change everything. That is not an argument for handing your money, your job or your faith to every company that has added the letters to its pitch deck.
Because whatever happens to the valuations, the buildings are already there. If it all works, we will have paid for them in drinking water and told ourselves it was obviously worth it. If it does not, we will still have paid for them, and there will be a select committee about it.
So enjoy the revolution. Dance if you want to. Just be the person who checked where the exits are, and be ready to be extremely unpopular at the bar.
I unpack this sort of thing with in house counsel and founders every fortnight on the Beyond The Fine Print podcast, and in our Linkedin newsletter. If you are looking at an AI heavy contract or a capital programme your board has not properly stress tested, that is the work I do.
Rory O'Keeffe is the founder of RMOK Legal, a solicitor regulated by the SRA, and an SCL accredited Leading IT Lawyer.
RMOK Legal. We Look After It.
Sources. Bank of England Financial Policy Committee record (October 2025) and Financial Stability Report (July 2026); Andrew Bailey, evidence to the Treasury Select Committee; Jeremy Grantham, The Diary of a CEO (June 2026); Howard Marks, CNBC interview (October 2025) and memo "Is It a Bubble?" (December 2025); Citi Research Bear Market Checklist (June 2026); Martin Wolf and John Plender, Financial Times; Challenger, Gray & Christmas job cut reports (May and July 2026); Oracle FY2026 results and SEC filings; Moody's Ratings on hyperscaler leverage; David McWilliams, Irish Times (August 2024 and 1 August 2026); Water UK written briefing to MPs; House of Lords report (May 2026).
Commentary, not legal or investment advice.
FAQ
-
Nobody can time it. The Bank of England said in October 2025 that the risk of a sharp correction had increased, and in July 2026 modelled a hit of up to 2.2 percentage points to UK GDP. Economists including Howard Marks argue valuations are elevated but not yet at mania levels.
-
The financial exposure runs mainly through pension funds and index trackers. The commercial exposure is more direct: long compute contracts signed at peak pricing, customers deferring AI budgets, and funding conditions tightening for AI-heavy business models.
-
Section 172 of the Companies Act contains no exemption for emerging technology. Directors approving large AI programmes on thin near-term returns should record the downside case, not only the upside. Board minutes that contain nothing but enthusiasm become the weak point if the project fails.
-
It can, but the business case must be evidenced. Where redundancies are justified by productivity gains that have not yet materialised, tribunals are likely to test the reasoning. A strategy statement about becoming AI-first is not on its own a redundancy justification.
-
It describes arrangements where a supplier invests in a company that then commits to spend with that same supplier. Everyone can see the loop and it is properly disclosed. The problem is that it makes a purchase commitment cheap to signal while leaving it indistinguishable from genuine demand.
-
More than three quarters of UK data centres are in the water-stressed south and east. Under current Environment Agency rules they must be supplied with drinking-quality water. Their designation as critical national infrastructure leaves water companies with limited power to restrict supply during shortages.