Barclays is insuring its AI loans so that it can lend more

Posted on: 11 November 2026

In the 1980s a corner of the Lloyd's market believed it had found a way to make catastrophe risk disappear. The syndicates writing London Market Excess of Loss business, LMX to those inside it, were groups of private investors known as Names who pledged their personal wealth against the risks an underwriter accepted on their behalf. They insured one another against losses above a given threshold, then passed part of that cover on to other syndicates, which passed part of it on again, so that most of the risk never left the market at all. Every syndicate looked at its own book and found it protected. On 6 July 1988 the Piper Alpha platform exploded in the North Sea, killing 167 men. The direct insured loss was around $1.5bn, but as the claims travelled from contract to contract the losses inside the spiral were later estimated at roughly ten times that figure, and some syndicates discovered that they had covered the same platform several times over. Since the Names carried unlimited liability, many of them took heavy losses on their own savings and property. What emerged in the analyses that followed was that even the people working in that market had not fully understood what the spiral was doing. Each contract made sense on its own terms. The problem was the circuit, which nobody was measuring.

The circuit is being drawn again, a short ride on the Docklands Light Railway from Lime Street. Bloomberg reported at the end of October that Barclays has been sounding out insurers about covering the credit risk on more than $10bn of loans that finance artificial intelligence infrastructure, meaning the data centres built for the hyperscalers, the handful of companies such as Microsoft and Meta that buy computing capacity on an industrial scale. On some facilities the bank would like insurers to take more than half of the risk. HSBC is in preliminary talks with insurers over project finance loans, some of them tied to the same buildout, although the bank describes this as ordinary portfolio management and says its exposure to data centres is small. In May the Financial Times reported that JPMorgan, Morgan Stanley and Japan's SMBC were looking for ways to pass on parts of this debt because they were approaching their internal limits.

Read quickly, this looks like prudence. The motive reported by people close to the Barclays talks points elsewhere, however, because the aim is not to hold less of this risk but to free regulatory capital so that the bank can write more of it. Regulatory capital is the cushion of the bank's own funds that supervisors require it to hold against its loans. If part of a loan's risk is transferred to an insurer, and the Prudential Regulation Authority, which supervises British banks, accepts the protection, the cushion required shrinks and the bank has room again. That room does not stay empty for long, since the hyperscalers' appetite for financing keeps growing and a bank that steps back simply hands the mandate to a rival that does not. The capital released flows straight back into the sector the risk has just left.

This is the part I find most interesting. A concentration limit is one of the oldest rules in banking, born of the plainest lesson of the trade, which is that a lender who lends too much to one borrower ends up depending on that borrower more than the borrower depends on him. It was designed as a brake. Once a market exists that will buy whatever sits above the limit, it stops braking and becomes a meter, telling the bank how much protection it must buy in order to keep lending. It no longer measures how much risk there is in the system, only how much of it remains on one particular balance sheet. Those are different quantities, and they can move in opposite directions for years without anyone noticing, because every individual balance sheet complies with the rules.

There is a detail, though, that makes the story less comfortable. A survey published in May by Willis Towers Watson, one of the largest insurance brokers in the world, found that several providers of credit insurance were approaching their sector limits on data centre exposure or had already reached them. The second link in the chain has its own number that it should not exceed, and it is touching it. When that capacity is full the demand will not vanish. It will look for a third link, in private credit funds and risk transfer structures where supervisors see less. In its May report the Financial Stability Board, which coordinates financial regulators across the largest economies, noted that artificial intelligence already accounted for more than a third of private credit deals in 2025, against an average of 17 per cent over the previous five years.

The second lesson comes from 2008. AIG, the American insurance group, had sold European banks protection whose sole purpose was to reduce the capital their supervisors required against roughly $250bn of loans and securities. It was the same exchange banks are seeking now, capital relief in return for a premium. Those contracts themselves produced no losses. AIG went down because of a different part of its book, the protection it had written on securities tied to American subprime mortgages, and when it did the capital relief enjoyed by the European banks nearly disappeared with it overnight. A Congressional oversight panel later estimated that the rescue, which grew to $182bn, spared those banks from having to raise as much as $16bn of capital at the worst possible moment. Practitioners call this wrong-way risk, the case in which the guarantor weakens at precisely the moment the guarantee is needed.

So it is worth asking what the insurers writing cover on data centre loans hold in their own portfolios. A large insurer invests for a living in corporate bonds and, increasingly, in private credit, and some of the largest corporate bond issues of the past two years have come from the hyperscalers themselves. In a world where demand for computing disappoints, the data centre loan sours at the same moment as the assets the insurer would use to pay the claim lose value. The cover is good for as long as nobody needs it.

For anyone who sits on a board or manages a family's capital the practical conclusion is a single one. A hedged exposure is not a reduced exposure, because the risk does not leave, it changes shape and becomes risk on the counterparty. Before treating it as closed it is worth asking who has taken it and, above all, whether that party's own assets would suffer in the same scenario in which it would be called on to pay. If the answer is yes, the protection is an opinion about the future signed by someone betting on the same outcome. The same reasoning applies to anyone holding units in private credit funds, which are the natural third link in this chain.

The signal to watch will not be the first default but the price of protection. As long as insurers accept the risk at modest premiums the meter keeps turning. Something is already moving, since on 22 October the cost of insuring Meta's debt against default reached a record, the day after the company raised its forecast for AI spending. On the day premiums jump or capacity runs out, the banks will have to choose between keeping the risk on their own books and lending less. Either way, investment in data centres will slow, and for a reason that has nothing to do with artificial intelligence.

After Piper Alpha, Lloyd's introduced Realistic Disaster Scenarios, which oblige syndicates to show what a defined catastrophe would cost them once their reinsurance is taken into account. Nothing comparable yet tells anyone what a fall in demand for computing would cost the chain of banks, insurers and funds now passing AI risk from one to another.