The lesson of Archegos was that the real danger is the borrowing no one can see whole. It is taking shape again, far larger, in the boom around artificial intelligence.

In the spring of 2021 a fund that most people had never heard of blew a hole in the balance sheets of the world's biggest banks. Archegos Capital Management, the family office of New York investor Bill Hwang, had borrowed tens of billions of dollars from at least nine banks to build enormous, concentrated bets on a handful of shares. Each bank could see the slice of risk it had financed. None could see the whole. When the shares turned, the unwinding cost the banks more than $10bn between them in a matter of days. Credit Suisse took the worst of it — around $5.5bn, and coordinated penalties and corrective measures from regulators in Britain, the United States and Switzerland, including the largest fine the Bank of England's regulator had ever imposed. Within two years it had ceased to exist as an independent bank.

The striking thing about Archegos was not the greed or the leverage. It was the blindness. The danger was never any single loan. It was the concentration no one could see, because the borrowing was scattered across lenders — and because, at the bank that lost most, the executives at the top only grasped the size of the exposure in the days before it detonated.

Five years on, that same blindness is taking shape again — on a far larger scale, and around the most fashionable bet in finance.

This week Anthropic, one of the leading artificial-intelligence labs, filed confidentially for a stock-market flotation. OpenAI and others are moving towards the same destination. As the AI wave reaches the public markets, ordinary investors will be able to own a piece of it. At some point they will likely join the tech stocks that have come to dominate stock markets and indices in which our pensions and savings are invested. But the shares are only the visible tip. Underneath sits a great deal of borrowing — and, as with Archegos, it is arranged in a way that makes it hard to see whole.

The bet beneath it all is simple: that AI will make companies across the global economy far more productive. That belief is now priced into businesses well beyond the technology firms themselves, and it remains largely unproven. JPMorgan's own analysts have worked out that the industry would need to generate something like $650bn of revenue a year, indefinitely, merely to earn a modest return on what is being spent. Markets are pricing a hope. That is no great matter — until people begin to borrow against the hope.

The issue is not whether AI succeeds or fails. It is what happens if the gains arrive much later, prove much smaller, or accrue to fewer companies than investors expect. If that happens, the first pressure point may not be the AI companies themselves but the debt raised to build the infrastructure around them — the data centres, power capacity, chip supply chains and financing structures that have been built on the assumption that demand will continue to rise.

That borrowing is already taking shape — just not where most people are looking. Much of it does not sit with the famous names themselves. OpenAI, the company at the centre of the boom, carries relatively little debt on its own balance sheet, yet the infrastructure being built around it requires vast sums of capital. The financing sits instead with data-centre operators, infrastructure providers, suppliers and, increasingly, private-credit funds that now lend where banks have retreated. As a result, the exposure is dispersed across the financial system rather than concentrated in the companies that attract the headlines. On a smaller scale, optimism is folded into ordinary company loans, where cost savings from AI that have not yet been made are counted towards the earnings a loan is judged against.

We have seen this shape before, as it happens. It is the shape of the late-1990s telecoms boom — suppliers lending to their own customers, debt piled on infrastructure built for demand that had to keep rising — and that ended not in a stock-market dip but in a wave of defaults.

Here is where Archegos returns. The financial system measures risk one borrower and one asset class or sector at a time. It has no category for a single, shared bet that runs across hundreds of unrelated companies, many of them not labelled "AI" or "tech". Much of the lending now sits in funds whose loans carry no market price, only the manager's own valuation; and the banks that finance those funds rarely look through to what the money is ultimately backing. So, as in 2021, each lender sees its own slice and reassures itself that the slice looks safe. No one sees the whole. The difference is one of scale: this time the concentrated bet is not one fund's, but a productivity thesis increasingly embedded across the market — in equity valuations, corporate earnings forecasts, infrastructure debt, private credit and the assumptions banks make about borrowers' future cashflow.

Regulators are not blind to the boom. The Bank of England, the European Central Bank and the IMF have all warned that valuations look stretched and that a fall could spread well beyond the AI companies. They are building the right tools, too — stress tests, monitoring of who is lending to whom. The trouble is that the tools are aimed at an ordinary recession, cover only part of the market, and oblige no one to act. None of them yet asks the question that matters: what happens to all this borrowing if AI simply disappoints?

Forcing banks to hold more capital against AI lending is, for now, politically out of the question. Washington is championing the build; Britain and Europe have staked their growth on attracting exactly this investment; and any country acting alone would merely push the lending elsewhere while keeping the risk, which finds its way home through banks that operate everywhere. Indeed, the tide runs the other way: on both sides of the Atlantic regulators are easing bank capital rules in the name of growth, thinning the buffers that would absorb a shock — even as this new concentration builds.

But the first thing needed is not new rules — it is better sight. Regulators could extend their stress tests to model not just a downturn but an AI disappointment; require banks to look through the funds they finance to what those funds are really lending against; and treat exposure to the AI bet as something to be measured across the system, rather than lost between categories. None of that tells anyone where to lend or what to charge. It would simply let the authorities see how large the exposure is, where it is concentrated and how it might spread — so that, if the bet sours, they are ready for it rather than surprised by it.

Banks have their own version of the same task, and it too is about sight, not retreat. Treat AI as one exposure running across the book — the chipmaker, the data-centre developer, the power supplier and the funds lending to them all — and aggregate it, rather than let it hide inside a dozen industry codes; look through the funds they finance to what those funds actually hold; stress the book against an AI disappointment rather than a generic downturn; and put someone senior genuinely in charge of the whole picture.

Selling the risk on, as banks are busily doing now, is not the same as seeing it — a danger passed to someone less able to bear it has not left the system, only moved. The same applies to a quieter habit — folding projected AI savings into the earnings a loan is measured against. The discipline here is not to refuse the loan but to see it clearly: once efficiencies not yet realised inflate the EBITDA, the leverage on the page is lower than the leverage in fact, and a covenant that looks comfortable rests on savings that may never arrive. The honest figure is the loan measured on today's earnings, before the AI adjustment.

Time will tell whether AI is a bubble. That is not a question regulators can answer, and it is not their job to try. Their job is to understand the risks building on the back of the AI bet, and to be ready for the strain that would pass through the system if it turned. The lesson of Archegos was not that someone took too much risk. It was that, almost until the end, no one could see how much. We should not have to learn it twice.

Martyn Hopper is founder of Martyn Hopper & Partners, a financial-services regulatory practice. He was previously head of enforcement at the Financial Services Authority and then a partner at Herbert Smith and Linklaters for over 20 years. He has acted in major prime-brokerage and counterparty-risk enforcement cases.