Everything About AI
NextFin News - The Bank of England has begun questioning lenders about data center projects being financed before a tenant has signed a lease to pay rent, a supervisory inquiry into the riskiest edge of the artificial-intelligence infrastructure lending boom. The central bank's outreach to market participants focuses on loans known in the real estate industry as speculative financing, where construction starts without an occupier committed to the completed project, according to people with knowledge of the matter.
The move marks a shift from general warnings about AI-related financial risk to hands-on fact-finding inside bank loan books. It comes as private credit is estimated to finance roughly $700 billion in data center capital spending between 2026 and 2028, a figure the Bank of England has cited, and as nearly half of this year's data center bond issuance lands in junk-rated territory.
The Situation: A Regulator Moves From Warning to Inquiry
The Bank of England's questioning is part of a wider effort to understand lending to the data center industry, the people said. The inquiry has not been described as a formal investigation, and no rule changes or enforcement actions have been announced. Its significance lies in what it targets: the specific underwriting practice of lending against a project's future prospects rather than a signed lease from a creditworthy occupier.
Speculative development is not new to commercial real estate. Office towers, warehouses, and industrial parks have long been built on the bet that tenants will arrive. What is different in data centers is the concentration of the demand side. A handful of hyperscalers — the giant cloud and AI companies — account for most of the projected need, and their capital spending plans are moving faster than the power grid, water supply, and construction pipeline can absorb. When a lender finances a facility before a lease is signed, it is effectively underwriting the borrower's ability to find one of those few hyperscalers before debt service comes due.
The scale of the build-out helps explain why supervisors are paying attention. The estimated cost of building the data centers needed for the AI boom has been put at around $3 trillion, a figure that no single technology company, however large, can fund from its own cash. That gap is being filled by debt: AI data center and project financing deals reached about $125 billion so far in 2026, up from roughly $15 billion in the same period of 2024, according to a UBS report cited in market coverage. Commercial mortgage-backed securities tied to data centers have also accelerated, with about $17 billion issued since the start of 2025 — more than triple the amount sold in the previous two years — and data centers now representing roughly 8% of new commercial property bond deals.
The risk is not abstract. In 2026 alone, about $80 billion in data center bonds have been raised, and nearly $37 billion of that — close to half — is rated below investment grade. When debt is stacked against a project that has no tenant, no operating history, and a completion date that depends on grid connections and equipment delivery, a missed assumption does not just hurt a bondholder. It can leave lenders holding a highly specialized building that is expensive to repurpose.
Why Speculative Lending Is the Weak Link in the AI Build-Out
The first-order risk of speculative data center lending is straightforward: if no tenant signs a lease, the project generates no rent, and the loan defaults. But the mechanism that matters to a regulator runs deeper, through three channels that connect a single project to the wider financial system.
Channel one is asset specificity. A data center is not a generic warehouse. It is designed around power density, cooling capacity, and the rack layout a particular generation of chips requires. If the intended occupier walks away, the building cannot easily be converted to another use. Commercial mortgage-backed securities investors are already being forced to widen their underwriting to account for power availability, grid constraints, cooling needs, and technology-driven obsolescence — factors that had little to do with traditional property lending. A speculative project that never secures a tenant is the extreme version of this problem: the collateral exists, but its income-generating purpose does not.
Channel two is distribution and concentration. Large construction loans are rarely held by one bank. They are arranged, then sliced and sold to other lenders and investors so the arranger can free capital for the next deal. That distribution process itself can become a risk signal. Two data center campuses linked to Oracle's expansion were financed through approximately $38 billion in construction facilities arranged by JPMorgan Chase and Mitsubishi UFJ Financial Group — about $23.25 billion for a Texas facility and $14.75 billion in Wisconsin, both tied to developer Vantage Data Centers. The placement of that debt stretched across more than two dozen banks and investors, took months to complete, and left some lenders still seeking to place less than $1 billion well into 2026. When distribution takes months, it means the market is being asked to absorb more exposure to the same narrow group of borrowers: hyperscalers, data center developers, and AI tenants.
Channel three is the leverage-valuation loop. This is the channel that Andrew Bailey, the Bank of England governor and chairman of the Financial Stability Board, has flagged at the global level. In his August 2026 letter to G20 finance ministers and central bank governors, Bailey warned that markets "remain vulnerable to a potentially disorderly correction that could spread across borders," pointing to fragilities in sovereign debt markets, vulnerabilities in private credit, and stretched asset valuations. The letter also cautioned that the increased use of leverage in bond and equity markets is interacting with high valuations, market concentration, and AI-related optimism in a way that could amplify a future correction. Speculative construction lending sits at the intersection of all three: high valuations for AI capacity, concentrated demand from a few tenants, and leverage layered through banks, private credit funds, and bond markets.
Markets remain vulnerable to a potentially disorderly correction that could spread across borders.
— Andrew Bailey, Bank of England Governor and Chairman of the Financial Stability Board, in his August 2026 letter to G20 finance ministers and central bank governors
The practical consequence is that a correction in AI demand expectations would not hit only equity holders. It would travel first to the most junior, least secured lenders — the ones who financed construction before a lease existed — and then, through distribution chains, to a broad set of banks and funds that thought they owned diversified credit.
The Regulatory Response: Mapping Before Acting
The Bank of England's questioning fits a pattern of escalating supervisory attention rather than a sudden intervention. In its Financial Stability Report in July 2026, the central bank set out a forward-looking framework for monitoring AI's macrofinancial implications. Later in the year, it began considering whether to include artificial intelligence in its next system-wide exploratory scenario — the stress test that measures the resilience of the UK financial sector against hypothetical dangers. Separately, the Bank of England has been running a year-long deep-dive into the roughly $16 trillion private markets ecosystem, driven by concerns about the sector's ability to weather a downturn.
That private-markets focus matters because much of the riskiest data center finance sits outside the traditional banking system. Private credit is estimated to fund about $700 billion of data center capital spending over 2026 to 2028. Private credit funds do not mark their assets to market daily the way public bond funds do, and their valuations can lag a deteriorating reality. A supervisor asking lenders what they know about speculative data center exposure is, in effect, trying to illuminate a corner of the system where prices are not being set transparently.
The timing also reflects a shift in market discipline. Earlier in the AI debt boom, bond investors often underwrote data center financing on the strength of the tenant — a hyperscaler's credit rating — rather than the structure itself. By mid-2026, that was changing: bond investors began pricing data center debt more like project finance, where the asset's own cash flow matters. That is a healthier market development, but it also means speculative projects without tenants face a steeper climb. Commercial real estate services firm CBRE has warned AI upstarts that they need a serious partner to guarantee lease payments, or they will face an uphill battle securing new sites. Several operators have said they will only enter serious talks when a partner with strong credit ratings stands behind the tenant's obligations.
For the Bank of England, the question is not whether AI data centers will be built. Many will be, and demand for computing capacity is real. The question is how much of the financing assumes demand that has not yet been contracted — and who holds the loss if it never arrives.
The Counter-Thesis: Why This May Not Be 2008
The strongest argument against reading the Bank of England's inquiry as a warning of imminent trouble is that today's data center lending does not resemble the loose underwriting that preceded the 2008 financial crisis. The largest data center financings are tied to tenants or sponsors with investment-grade balance sheets, and the biggest borrowers are the hyperscalers themselves, whose cash flows dwarf the debt service on their facilities. Demand for AI computing capacity is not a speculative bet on a housing price index; it is backed by visible capital spending commitments from some of the world's most profitable companies.
There is also a supply-side reality that supports the speculative builders: powered land with grid connections is scarce, and a developer who can secure power capacity and permits holds a valuable asset even before a tenant signs. In that view, a "speculative" data center project is less like an empty office tower and more like pre-positioned infrastructure in a market where the bottleneck is electricity, not demand. If anything, the risk may be underbuilding, not overbuilding.
That counter-thesis has force in the near term. But it rests on two assumptions that the Bank of England's inquiry is designed to test. First, it assumes that hyperscaler capital spending continues at its current pace — a pace that is already outstripping earnings and free cash flow for several of the largest players, pushing some to borrow. Second, it assumes that the specialized nature of the assets protects lenders. In fact, specificity cuts both ways: it protects a facility with a signed tenant, but it magnifies the loss on one without. A project financed on the expectation that a hyperscaler will arrive is only as strong as that expectation.
The falsifying signal is concrete: if the hyperscalers collectively slow their announced capital spending — a visible, quarterly-reported number — while speculative project announcements continue to rise, the mismatch between contracted demand and financed supply would confirm the regulator's concern. Conversely, if capital spending holds and lease signings keep pace with groundbreaking, the speculative share of the book would prove manageable.
What Comes Next: Three Horizons for the Risk
Short term, the Bank of England's inquiry is information-gathering, not a tightening of rules. Markets are unlikely to react sharply to questions alone. The immediate effect will be on lenders' internal reporting: expect more granular disclosure of data center exposure, and more caution around loans without pre-leased capacity. Banks that arranged large speculative packages may find distribution taking longer, as the Oracle-Vantage example already showed.
Medium term, the risk is concentrated in the projects that break ground in 2026 and 2027 without signed tenants and come online into a market where hyperscaler spending has normalized. The exposed parties are the lenders at the junior end of the capital structure — private credit funds and the banks still holding undistributed slices — and the borrowers who counted on refinancing completion loans with long-term, tenant-backed debt that may not be available on the same terms.
Long term, the structural question is whether AI computing demand proves durable enough to absorb the capacity being financed today. This is where the cyclical-versus-structural call has to be made, and the two forces point in different directions. The cyclical leg is the capital-spending wave: if hyperscaler budgets slow as earnings catch up with outlays, the demand for speculative capacity would ease and the risk would look like a credit cycle that reverts once leverage is worked down. The structural leg is the physical constraint: power, water, and permitting bottlenecks do not self-correct on a quarterly cycle, and a facility that has secured grid capacity holds scarcity value regardless of the credit cycle. The Bank of England's concern is that the cyclical leg is being financed as if it were structural — that lenders are pricing today's scarcity premium as though it will last through the life of a 10-year loan.
The base case is that supervisory attention slows the riskiest lending at the margin without derailing the build-out: lenders will demand more pre-leasing before committing construction capital, and pricing will widen for projects without tenant backstops. The upside case is that demand continues to outrun supply, speculative projects find tenants, and the inquiry is remembered as prudent housekeeping. The downside case is a synchronized slowdown in hyperscaler spending and a repricing of AI valuations, which would hit speculative projects first and travel through the distribution chains that spread the exposure across banks and private credit.
The Bank of England is not trying to stop the AI infrastructure boom. It is asking a simpler, sharper question: when the music slows, who is left holding the buildings that no one signed for? The answer matters less for the hyperscalers — who can fund their own capacity — than for the lenders who bet they would.
Explore more exclusive insights at nextfin.ai.
