Georges Elhedery was asked about artificial intelligence and answered in the vocabulary of a credit officer. The technology is transformational, the HSBC chief executive told CNBC’s Dan Murphy on October 20, and the bank is adopting it across its own operations. Then he turned to the buildings that make it work, and the tone changed.
“One has to be selective,” he said. “You have to look at the quality of the tenancy agreements. You have to look at the pricing of these contracts.
Not all of them are of that level of virtuosity that you can support strongly. Some of them may be more challenging.”
The remark is unremarkable as risk management and pointed as market commentary. Europe’s largest lender, with close to $3 trillion in assets, sits on both sides of this trade: it lends to the companies building data centers and it runs them for its own computing. When the head of that institution describes the sector in terms of tenant quality, he is telling borrowers what the next credit committee will ask.
What a bank buys when it finances a data center is not the building. It is a lease, usually signed by a company with a credit rating, often structured as a special purpose vehicle that owns the shell while the tenant owns the equipment inside it. The loan is underwritten on the rent, and the rent is only as good as the tenant’s willingness to keep paying during a downturn in demand that has not happened yet.
That is where the pricing question becomes a credit question. Contracts signed in 2024 and 2025 assumed power costs, construction costs and interest rates that have all moved. A tenancy agreement struck at one set of assumptions is worth less to a lender when the cost of capital behind it has risen, even if the tenant is perfectly solvent. Elhedery’s distinction between contracts of different quality is a way of saying that the sector is no longer priced as a single asset class.
The financing environment has changed faster than the buildings. Hyperscalers raised roughly $200 billion in bond financing last year, according to Mohamed El-Erian, the Allianz chief economic adviser, who said at a conference in London this month that the figure could approach $1 trillion. Government borrowing is running at the same time, and refinancing is happening at higher rates than the debt it replaces.
That combination produces a crowding effect. When the largest technology companies issue debt, they compete for the same pool of money that a private developer needs for a single campus in Ohio. The hyperscalers pay less for it because their balance sheets carry other revenue. The developer pays more, or does not borrow at all, which is how a buildout that looks national in the aggregate becomes selective at the project level.
McKinsey has estimated that global data center spending could reach $7 trillion by 2030, a number that circulates in investor decks and board presentations. The figure is a forecast of demand, not of returns, and the gap between the two is where lending decisions are actually made. A country can need computing capacity and still contain projects that never earn back their cost of capital.
HSBC’s own positioning explains the caution. The bank has said it wants to be the leading financier of what it calls transition ecosystems, and its customer base includes data center operators and AI companies. That ambition requires it to lend into the sector while avoiding the concentration that would make a single tenant’s troubles into a capital problem.
Being selective is not a retreat from the business. It is the condition for staying in it.
For the companies raising money, the practical consequence is a longer and more intrusive process. Lenders will ask who the tenant is, how long the lease runs, whether the rent escalates, who pays for power if the grid connection slips, and what happens to the equipment if the tenant leaves. Those questions were asked in 2021 as well, but they were asked quickly, because capital was plentiful and the answers were assumed.
Smaller developers will feel the change first. A company with one campus and one tenant has no diversification to offer and no other cash flow to pledge. A company with five campuses and investment-grade tenants can borrow against the portfolio. The sorting that Elhedery describes is already happening in the debt markets, where the difference between the best and worst borrowers has widened through the year.
There is a reasonable argument that this is the healthy version of the cycle. A market that funds every project teaches nobody anything, and a market that forces sponsors to disclose tenant quality produces better information for everyone involved, including the tenants. The risk is timing: the projects that get cut are the ones furthest from completion, and the capacity they would have provided is the capacity that relieves prices later.
Elhedery did not say the boom is over, and nothing in his remarks suggested he thinks demand is fake. He said the opposite about the technology. What he declined to do is treat a sector-wide enthusiasm as a substitute for underwriting, which is the distinction a lender is paid to make and the one that tends to be forgotten in the middle of a shortage.
The question his comment leaves behind is narrow and answerable. When the next large data center financing is syndicated, how many banks will bid, and at what spread? The answer will say more about the cost of AI capacity in 2028 than any forecast published this quarter.
