The AI Debt Pile: Who Pays When Growth Slows?

The risk has migrated beyond technology shares

The AI investment boom is increasingly a question about who ultimately finances expensive buildings, processors, electricity and grid connections, not simply who makes the most impressive software. The crucial risk transfer is from technology companies to lenders, infrastructure funds, pension portfolios, utilities and the households paying regulated electricity tariffs. A data centre can be profitable at one level of occupancy and interest rates and financially fragile at another. The borrower may be a specialist landlord rather than the technology firm creating the demand, and the long life of a physical building may exceed the useful commercial life of the chips it houses. This mismatch matters when public markets begin doubting whether extraordinary capital spending can generate enough cash. The Bank of England (opens in a new tab) has warned that the use of debt to finance AI infrastructure is accelerating and could eventually transmit a technology shock into broader lending conditions. That warning is not a forecast of imminent defaults, but it identifies a financial channel that cannot be monitored by looking only at share prices.

Three shocks between October 28 and November 1

This week's events exposed three parts of the mechanism. First, late-October corporate earnings raised doubts about cash returns on enormous AI investment, with Alphabet's results drawing particular scrutiny and investors punishing technology valuations. Second, the October 30 US equity decline, including a sharp Intel fall, illustrated that a loss of confidence in projected growth can rapidly affect broader benchmarks. Third, October 31 reporting on US electricity bills showed the politics of data centres entering the November 3 midterm campaign, with households angry about rate increases and candidates considering tougher conditions on projects. A fourth event was the publication of market previews showing the midterms and the next earnings cycle as simultaneous tests for both AI regulation and equity prices. These developments are linked but not identical. Earnings interrogate revenue; stock declines reprice future profits; electricity bills distribute infrastructure costs; and election outcomes determine who can impose the terms. No single event proves an AI bubble has burst.

British infrastructure is a financing question

Britain has an opportunity to host data processing, software and supporting engineering. But the country also has finite transmission capacity, expensive electricity and a long queue of would-be connections. A September parliamentary answer, referencing information collected by the National Energy System Operator in 2025, reported 72.8 gigawatts of capacity allocated to data centres in the British connection queue across the transmission and distribution systems. This figure is a queue measure, not a forecast of 72.8 gigawatts of operating demand or a count of completed sites. Some applications may be delayed, duplicate ambitions or never reach construction. That distinction prevents a dramatic number from becoming a false statement about today's grid consumption. Nevertheless, projects reserving potential capacity can slow other demand connections if queue rules are weak. Manufacturers, housing developers and clean-energy investors may face delays that impose real costs even when the AI schemes never become active. Regulation of the queue is therefore a matter of industrial strategy and distributional fairness.

The debt claim needs precision

Claims that British data centre firms carry tens of billions in debt need careful definition. Debt can refer to borrowings by UK-domiciled operating companies, loans secured on British assets, global bonds issued by a multinational landlord, or the share of an international financing package assigned to UK facilities. These totals are not interchangeable and may overlap. Without a current borrower-by-borrower schedule the number cannot be treated as a verified UK industry aggregate. What is established is the structural risk highlighted by the Bank of England: data centre and AI-chip expansion increasingly uses external borrowing, and the assets securing it may depreciate more quickly than creditors expected. The Bank's July 2026 financial-stability report discusses estimates of more than two trillion dollars of funding needs for AI chips over five years, explicitly attributing that estimate to JPMorgan. It is a projection of global chip funding needs, not the debt currently owed by British data centre firms. Exaggerating that distinction would make a serious systemic warning less credible.

Who actually lends the money

Financing channels include bank construction loans, private credit, infrastructure debt, listed corporate bonds, equipment financing, lease obligations and equity from investors seeking stable long-term returns. The lender faces the risk that the project takes longer to connect, costs more to power, loses a major tenant or cannot refinance at the expected rate. A large technology company may sign long-term capacity agreements that reduce vacancy risk, but those agreements require examination: termination provisions, guarantees, escalation clauses and counterparty ratings determine how much risk genuinely transfers. When a data centre is sold to an infrastructure vehicle, its original builder may realise cash while future investors inherit the occupancy and financing exposure. The public description of a deal as long-term contracted infrastructure can conceal a business that depends on continued expensive model training. No lawful evidence suggests every such structure is imprudent. The policy challenge is whether banks, pension trustees and regulators stress-test assumptions that differ from a conventional warehouse or logistics property.

Pension funds can be hit twice

Pension schemes may own technology equities directly, hold diversified bonds containing major AI issuers, invest in private infrastructure vehicles or employ asset managers that finance construction. A market decline would affect different investors in different ways. Defined-contribution savers bear movements in their account values; defined-benefit liabilities are affected by interest rates, and the net impact depends on assets, hedging and funding ratios. It would be misleading to say all UK pensions are heavily exposed without portfolio-level disclosures. The broader risk is correlated exposure: the same savings account may own the tech company's shares, the developer's bonds and a utility raising money to serve new data-centre demand. In a boom those positions appear diversified because the legal issuers differ. In a downturn they may all depend on one underlying assumption: enough users and companies will pay profitable prices for artificial intelligence. Trustees should demand a look-through assessment rather than treating every instrument's name as a distinct economic bet.

Electricity makes the story political

The energy-price-volatility driver (opens in a new tab) connects remote investment choices to household welfare. Developers can pay the full incremental cost of connections and new generation, or some costs can be socialised through regulated network charges. The distribution matters more than the slogan that data centres use power. The International Energy Agency's 2026 update projects global data-centre electricity consumption rising from roughly 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030 under its central outlook. That is a global estimate with considerable project uncertainty, not a UK tariff forecast. In the United States, reporting ahead of the midterms found residential electricity prices up by more than ten per cent since 2024, with local political disputes over planned utility investment. Those bill increases have multiple causes, so it would be wrong to assign every dollar to AI. The important test is how regulators allocate grid upgrades, who pays for dedicated substations and whether new loads bring their own generation or flexibility.

The political feedback loop

American midterm candidates are discovering that the same AI boom supporting capital markets can make local voters angry. A data centre may offer construction employment, property taxes and some permanent jobs, but its ongoing headcount often falls short of the promises associated with a large factory. Nearby residents may face land-use conflicts, water demand and competition for electricity infrastructure. Governors and utility commissions therefore have incentives to demand more transparency and a larger developer share of system costs. Tougher rules could slow new building, which would reduce near-term investment spending and pressure vendors supplying chips and electrical equipment. But well-designed rules could also improve the economics of completed projects by eliminating speculative applications and forcing realistic power planning. The inflationary-pressure driver (opens in a new tab) captures one possible channel: faster electricity and infrastructure price growth can squeeze household purchasing power even if the AI sector eventually delivers productivity gains.

Equity valuations and credit risks interact

The market's worry is not necessarily that artificial intelligence has no economic value. It is that prices of shares, corporate debt and infrastructure projects may assume very rapid monetisation with little room for delay. The October 30 selloff followed earnings that investors judged insufficient relative to spending expectations, and market analysts are also watching the cost of financing. A growing number of firms seeking credit can push up borrowing spreads, particularly for companies with weak free cash flow. If refinancing becomes expensive, operators may postpone new sites, cancel equipment orders or seek additional equity. Such actions hurt existing shareholders even before a formal default. Conversely, well-capitalised customers with durable earnings may gain bargaining power as speculative developers retrench. Distinguishing the infrastructure layer from the platform layer is vital. A broad AI label can cover cash-rich hyperscalers, highly leveraged equipment users and landlords whose valuations depend on tenants they cannot easily replace.

Banks are a transmission channel, not the whole system

The Bank of England said in July that the outstanding stock of AI-related debt remained modest at that point but was changing quickly. That is a measured assessment, not a declaration of systemic collapse. A banking stress test should ask how many borrowers depend on the same customers, which loans have construction risk, how collateral would be valued if accelerators become obsolete, and whether the capital market can absorb a wave of refinancing. Private credit adds complexity because disclosure and liquidity may differ from listed bonds. Banks can also be affected indirectly if losses at asset managers and investment funds tighten financing for ordinary companies. When the Federal Reserve (opens in a new tab) assesses monetary conditions, a technology selloff can influence risk appetite and wealth even without a bank solvency crisis. The central question is how a repricing of AI cash flows affects the ability of unrelated businesses to borrow. That spillover is the meaningful public-interest risk.

A necessary counterargument

There is a substantial countercase. AI has demonstrable demand, some firms earn real revenue, and demand for computing may expand across medicine, software development, research, logistics and consumer services. Investment in power and fibre can outlast individual technology generations, benefiting other industries. Efficiency improvements may increase usage even if the energy consumed per task falls. Market prices may correct without a credit crisis; equity holders usually absorb valuation losses before insured depositors or taxpayers do. The IEA's projections are scenarios, not promises that every proposed data centre will be built. Similarly, higher UK connection queues can reflect speculative applications rather than delivered demand. A rigorous analysis must reject the false choice between an inevitable AI crash and limitless growth. The right test is whether revenues and utilisation support spending after realistic electricity costs, hardware replacement, financing rates and taxation have been charged.

What it moves

Pension funds: downside exposure rises where multiple holdings depend on the same AI spending cycle; published asset allocations and funding ratios are the observable measures. Bank lending: spreads on data-centre and private-credit loans may widen if returns disappoint, with a possible knock-on reduction in credit for other firms. Electricity prices: local tariff and connection charges face upward pressure where upgrades are socialised, while dedicated developer funding could reduce household exposure. AI equities: valuations are vulnerable to earnings misses, slowing cash conversion and tougher election-year siting rules; strong cash generation can offset that risk. Corporate bond yields: weaker issuers may face rising spreads while government bonds attract defensive demand. Interest rates: the effect is mixed because falling technology investment can reduce growth while power costs can add inflation. The direction of each variable depends on the shock, so a blanket prediction of cheaper money following an AI correction would be unjustified.

Who benefits and who pays

The earliest winners are owners able to sell capacity under secure contracts, suppliers of transformers and cooling equipment, and investors who collect financing fees before a project's lifetime economics are known. Big technology customers can also benefit from competitive capacity offers and lower prices if speculative building creates excess supply. Workers may gain in construction, electrical engineering and specialised maintenance. If the cycle reverses, the loss bearers include shareholders, lenders, private-credit investors and pension savers whose funds own exposed assets. Ratepayers may remain responsible for network assets built for demand that never arrives if tariff rules allocate costs poorly. Municipalities can suffer from abandoned construction sites or unmet tax forecasts. The distribution is not fixed by technology; it is determined by contracts, regulation, capital structure and public disclosure. This is why oversight of who bears the downside should precede any decision to subsidise new developments.

The tests after the election

Watch November 3 election results, the next reports from major chip and cloud companies, utility rate cases and UK connection queue reforms. Read the cash-flow statements alongside revenue growth: high accounting earnings with persistently negative free cash flow make financing more sensitive. Track credit spreads of heavily borrowing AI issuers against broader investment-grade benchmarks. In Britain, ask the regulator how much grid expansion is allocated to new data-centre customers and how many speculative connection requests have been removed. Ask pension funds for exposures by underlying AI revenue assumption, including private vehicles. The sector may generate tremendous productivity and still leave many projects overfinanced. The central editorial test is whether the people who choose investment risk pay for that risk, rather than transferring the bill to household electricity customers or pension savers who never consented to the bet.

Sources

Bank of England: July 2026 Financial Stability Report (opens in a new tab); Reuters: Midterms and AI-Wary Stocks (opens in a new tab); AP: Technology Selloff (opens in a new tab); Reuters: Power Bills and Data Centres in Midterms (opens in a new tab); IEA: Key Questions on Energy and AI (opens in a new tab); UK Parliament: Data Centre Grid Queue (opens in a new tab); Bloomberg: AI Policy and Midterm Investors (opens in a new tab); Financial Times: Midterms Loom Over Investors (opens in a new tab); IEA: Energy Supply for AI (opens in a new tab)