AI’s trillion-dollar bet is now riding on borrowed money

Hyperscalers are on track to spend nearly $1.1 trillion on AI data centres through 2027, but a June 2026 working paper by Wharton finance professor Jessica Wachter finds that the AI sector must lift its own productivity by a factor of 2.7 by 2030 just to break even. If the projected gains fail to appear, Wachter and her coauthor conclude in their research paper that the buildout could become the largest misallocation of capital in history.

More than half of the $2.9 trillion in planned data‑centre investment between 2025 and 2028 will need external capital, with private credit expected to supply about $800 billion. That structure spreads exposure through pension funds, insurers and utility ratepayers.

Private‑credit funds are pouring at least $800 billion into hyperscaler data centres over the next four years — that single figure marks the moment the AI infrastructure boom stopped being a corporate‑cash story and became a debt story. The financing now flows through opaque special‑purpose vehicles, off‑balance‑sheet leases and securitisations, ending up in pension portfolios and utility rate‑recovery plans. A concrete example is taking shape in Louisiana, where Meta‘s Hyperion project has ballooned from a $10 billion bet into a $50 billion campus backed by a joint venture with Blue Owl Capital and a chain of entities that insulate the tech giant from the full cost. Ratepayers, meanwhile, are being asked to underwrite seven new gas‑fired power plants. What began as a bet on AI scaling has turned into a financing experiment with stakes far beyond Silicon Valley.

The break‑even math that few are pricing

Wharton finance professor Jessica Wachter published a working paper in June 2026 that calibrates the AI investment boom through 2027. Her model infers that the AI sector must raise its own productivity by a factor of roughly 2.7 by 2030 to break even, after accounting for capital costs and depreciation.

Wachter cautions that because the required productivity leap is not yet visible in aggregate statistics, the spending could prove a bubble. If the gains do not materialise, Wachter and her coauthor conclude in their research paper that the buildout would represent the largest misallocation of capital in history. Yet the emergence of that productivity remains hypothetical. A survey of 6,000 senior executives in the US, UK, Germany and Australia found that 90 per cent reported no AI‑driven productivity increase over the past three years — even as they forecast an average boost of 1.45 per cent over the next three, with US executives anticipating 2.25 per cent.

The gap between spending and revenue is already wide. Former SEC chair Gary Gensler notes that AI companies are on track to invest about $750 billion this year while total AI revenues hover between $150 billion and $200 billion. “The challenge is that the spending does not have commensurate revenues yet,” he says. Alphabet‘s Q2 2026 quarterly figures show the strain: capital expenditure of $44.9 billion pushed free cash flow to a deficit of roughly $5.9 billion — its first negative read since the 2004 IPO.

The shortfall that worries Wachter appears starkly when projected capex is set against the required productivity gain and today’s revenue.

Morgan Stanley calculates that of the $2.9 trillion hyperscalers will spend on data centres between 2025 and 2028, more than half must come from outside capital. Private credit alone is expected to supply about $800 billion. Debt issuance linked to AI infrastructure is expected to near $500 billion in 2026, with high‑yield AI data‑centre bonds now trading at spreads above June 2022 levels.

The financing chain is not abstract. Meta’s Hyperion campus in Louisiana, first announced in late 2024 as a $10 billion, two‑gigawatt project, had grown to a $30 billion undertaking by late 2025. Meta then sold an 80 per cent stake to Blue Owl Capital, forming the Beignet joint venture. A landlord entity, Laidley LLC, now owns the site and leases it back to Meta subsidiary Pelican Leap under a series of four‑year contracts. “The leases match GPU lifetimes,” Columbia Business School finance professor Stijn Van Nieuwerburgh has warned, “leaving investors with empty buildings if Meta terminates early.”

Meanwhile, utility Entergy Louisiana is planning seven additional gas‑fired power plants, raising total planned capacity to about 7.5 gigawatts — roughly six times the electricity used by New Orleans. Consumer advocates question whether all costs will be covered if Meta reduces its power purchases. Logan Burke, executive director of the Alliance for Affordable Energy, has warned that ratepayers could inherit the bill for surplus generation. Some developers are already building their own power grids to bypass utility delays, but Louisiana’s approach ties households to decades of cost recovery.

How the debt entered the plumbing

The shift from cash‑funded expansion to heavy borrowing is the structural change that turns a tech boom into a systemic risk. Until recently, hyperscalers financed data centres with free cash flow. That cushion is gone. Oracle‘s fiscal 2026 capital expenditure reached roughly 175 per cent of operating cash flow, driving free cash flow deeply negative. Across the group, the aggregate free‑cash position sits near zero.

Bank for International Settlements economists (in BIS Bulletin No. 120 and March 2026 quarterly research) have documented how hyperscalers have moved into on‑ and off‑balance‑sheet borrowing, using lease‑backed special‑purpose vehicles that sit outside bank‑style capital requirements. Private‑credit originations to AI‑sector firms exceeded $40 billion in 2025, and the BIS warns that “shadow borrowing” is concentrating exposure in lightly regulated funds. Because many of these vehicles lack capital buffers, a downturn in AI demand could send losses directly into retirement portfolios and insurance balance sheets.

PIMCO credit strategists caution that debt investors in AI‑tenanted facilities face obsolescence and refinancing risk with limited upside. Goldman Sachs credit analysts report that a basket of high‑yield AI data‑centre bonds is now trading at widening spreads, suggesting the market is beginning to price potential strain if revenues disappoint.

The Louisiana example shows how that financial chain reaches utility bills. The Louisiana Public Service Commission has already approved three gas‑fired plants totalling about 2.3 gigawatts for Meta’s campus, with seven more on a fast track. The 20‑year cost recovery schedules mean households could pay higher tariffs for decades even if Meta’s power needs fall short of projections. The debt behind the data centre ultimately lands on a ratepayer’s monthly statement.

Beyond the headline

The Money Trail

The funding chain now runs from hyperscaler finance desks through a web of private‑credit vehicles, securitisations and utility cost‑recovery mechanisms, ending in pension portfolios and household bills. The apparent diversification masks a concentrated bet on a key assumption: that frontier‑scale AI will generate enough cash to support both equity valuations and the fixed coupons attached to the infrastructure. Understanding who ultimately holds the obligations — not just who owns the data centres — reveals why a misfire would ripple far beyond Silicon Valley.

The Response Gap

Regulators have begun to flag AI‑related leverage, but policy responses lag the speed of the buildout. Utility commissions are approving multi‑gigawatt gas plants on accelerated timetables, while financial supervisors are only starting to map off‑balance‑sheet exposure in private credit. Bridging this gap could benefit from tools that treat AI infrastructure more like systemically important real‑estate and energy assets — with stress tests and disclosure standards that match the scale and concentration of the risks now accumulating in long‑dated contracts.

The Reach

A single actor — the global pension industry — plays a significant role between AI ambition and household security. Through allocations to investment‑grade tech bonds, infrastructure funds and private‑credit strategies, pension schemes are financing a large share of the data‑centre boom. If AI delivers the promised productivity, these portfolios benefit from sustained yields and asset values; if the boom stalls, the same exposure could force de‑risking and lower future payouts, tightening retirement conditions for workers who never chose to bet on hyperscaler capex.

The costs that could land close to home

With AI infrastructure debt now embedded in pension funds, utility bills and credit portfolios, three groups face distinct pressures.

• US‑based investor with APAC emerging market exposure

  Evaluate your portfolio’s exposure to private credit, infrastructure funds and hyperscaler debt. If a US AI infrastructure retrenchment tightens global credit, funds with large allocations to leveraged tech‑infrastructure could suffer — and those losses may spill into APAC bond and equity holdings. Check your manager’s stress‑test scenarios and sector‑concentration limits now.

• Western pension fund manager

  Conduct a thorough review of your fund’s direct and indirect AI‑data‑centre exposure. Ask asset managers for disclosure on private‑credit holdings, lease‑back SPVs and securitised AI‑facility paper. Compare the risk assumptions they use to those flagged by the BIS and Goldman Sachs. If the boom stalls, the fund’s ability to meet long‑term liabilities could narrow.

• Louisiana resident and utility ratepayer

  Monitor Louisiana Public Service Commission filings for Entergy’s gas‑plant cost‑recovery plans. Consumer advocates such as the Alliance for Affordable Energy track these dockets and can explain how rate structures might change if Meta scales back power purchases. The approvals being fast‑tracked now will shape your electricity bill for decades.

• Western corporate credit analyst

  Deepen your scrutiny of hyperscaler balance sheets and their off‑balance‑sheet obligations. Pay particular attention to the pace of capex relative to operating cash flow, the maturity profile of AI‑linked bonds, and the spreads on high‑yield data‑centre baskets. Rising credit‑caution signals could force a repricing of the entire sector well before revenue numbers confirm the stress.

Explainer

Hyperscaler

Large cloud‑computing companies — Alphabet, Microsoft, Amazon, Meta and Oracle — that operate massive networks of data centres and offer computing power at enormous scale. Their combined cloud‑services revenue exceeded $300 billion in 2025. They are the primary funders and tenants of the current AI data‑centre buildout.

Private credit

Loans made by non‑bank lenders, often pooled into funds, to companies that are not easily served by public bond or bank‑loan markets. The US private‑credit market has tripled in size since 2015, reaching roughly $1.7 trillion in assets. In the AI buildout, private‑credit funds finance special‑purpose vehicles that own data centres and collect lease payments from hyperscalers.

Special‑purpose vehicle (SPV)

A legal entity created for a specific financial purpose, such as owning a single data‑centre campus and issuing debt backed by its lease revenue. SPVs keep the assets and liabilities off the sponsoring company’s balance sheet. Because they are often lightly regulated, losses inside an SPV can be hard to trace and may unexpectedly reach institutional investors.

Securitization

The process of bundling cash‑flow‑generating assets — such as data‑centre leases — into bonds (ABS) or commercial mortgage‑backed securities (CMBS) that are sold to investors. In the AI infrastructure space, securitisation issuance backed by data‑centre lease payments rose by roughly 40 per cent in 2025 alone. The bonds provide fixed‑income investors an alternative way to gain AI‑related exposure but also embed the risk of tenant default or early lease termination.