Yes, the AI boom has a balance sheet problem
The equity story looks unstoppable. The balance sheet says otherwise.
AI is no longer a pure software scale play. It's becoming one of the largest debt-funded infrastructure builds in modern history-data centers, power, networks, and chips-stacked on top of complex financing. That shift is pulling private credit into the core of tech, tightening links between a small set of firms, and pushing more risk off balance sheet.
AI stops looking like software
For two decades, big tech scaled software with low marginal cost and rising cash piles. That pattern broke when training and serving frontier models demanded concrete, copper, and silicon at a pace internal cash can't match.
UBS estimates roughly $125 billion of AI data center and project finance issuance in 2025, up from about $15 billion a year earlier. McKinsey pegs total data center spend near $7 trillion by 2030. The majors-Amazon, Google, Meta, Microsoft, Oracle-issued about $121 billion in new debt this year, with at least another $100 billion expected next year. Returns now hinge on utilization, timing, and the cost of capital. That's more utility or telco than software, and credit is starting to price it that way.
Oracle shows how fast sentiment can turn
Oracle rallied hard on AI optimism and OpenAI ties, then snapped back. A revenue miss took the stock down double digits in a day, and the real tell showed up in credit.
Fresh investment-grade bonds-about $18 billion issued in September-sold off, with paper losses exceeding $1 billion. CDS widened to levels last seen during the financial crisis. That's what happens when an equity growth story rests on infrastructure timing and leverage.
The web of circular financing
Nvidia sells the picks and shovels, and it's winning-so far. But even its success leans on others' willingness and ability to keep spending.
To grease demand, Nvidia is investing in customers, extending financing, and backing data center deals that loop back into chip purchases. OpenAI sits at the center too-customer of Oracle, Amazon, Microsoft, and CoreWeave; investor and partner in parts of the chain; committed to hundreds of billions in compute over time while generating around ten billion in annual revenue with large losses. Profitability is still years out.
CoreWeave is another pressure point: no profits, roughly $14 billion in debt, tens of billions in lease obligations, and heavy customer concentration-about 70% of revenue from Microsoft. Money circulates inside a small group of firms, amplifying growth on the way up and risk on the way down. It also obscures where losses land if demand slows or timelines slip.
When debt goes off the books
As borrowing climbs, tech is turning to special-purpose vehicles. Meta, xAI, Google, and others have used SPVs to finance data centers and chip purchases, then lease assets back. The result: cleaner corporate balance sheets, preserved ratings, less transparency.
Private credit is filling the gap. Morgan Stanley estimates private lenders could supply more than half of the $1.5 trillion needed for data centers through 2028. Disclosure is thin, links to banks and insurers are growing, and real-time mapping is hard.
Securitization adds another layer. Data center cash flows are being bundled into ABS-digital infrastructure now makes up about $82 billion of the US ABS market, up roughly ninefold in under five years. Some loans are even backed by GPUs. As new chip generations arrive, older models lose value. If collateral prices fall, lenders can force sales into a soft market, driving prices lower and tightening credit at the worst time.
What finance teams should track now
- Utilization and take-or-pay: Track contracted vs. actual AI workload utilization by cohort and vintage. Watch renegotiations.
- Power and timing: PPA commitments, interconnect queues, and delay penalties. Every quarter of slippage matters for cash burn.
- All-in cost of capital: On- and off-balance-sheet debt, SPVs, leases, vendor financing, and securitizations. Blend the rate and the covenants.
- Customer concentration: Exposure to OpenAI, Microsoft, Nvidia-backed entities, and a handful of foundation model buyers.
- Vendor dependence: Nvidia allocations, prepay terms, resale restrictions, and chip depreciation schedules.
- Residual value risk: Haircuts and advance rates on GPU-backed loans; sensitivity to next-gen chip launches.
- ABS performance: Triggers, DSCR tests, step-ups, cash traps, and servicer incentives across data center ABS deals.
- Private credit linkages: Co-investors, warehouse lines, and fund-level leverage that can transmit shocks.
- Cross-default pathways: Keep a live map of guarantees, comfort letters, purchase obligations, and revenue-sharing agreements.
- Refi walls: Maturity stacks from 2026-2029 under different rate scenarios.
Three stress paths to model
- Demand slip: Model training slows, inference ramps later than planned. Utilization drops 10-20%, hitting DSCR and covenant headroom.
- Collateral shock: New GPU class cuts residuals on prior gens by 30-50%. Margin calls and forced sales widen spreads and reduce fresh capacity.
- Funding tightening: Private credit pullback raises spreads 200-400 bps. SPV issuance slows; sponsors must consolidate or inject equity.
Positioning and practical moves
- Equity: Treat AI infra more like utilities. Pay for contracted, high-visibility cash flows; discount expansion promises without firm PPAs and customer commitments.
- Credit: Prefer senior secured with strong collateral controls and step-up covenants. Demand transparent look-through to SPVs and leases.
- Treasury: Hedge rate risk across construction windows; align swap tenors with ramp schedules, not just lease terms.
- Risk: Build GPU collateral haircuts that anticipate obsolescence, not just resale comps. Stress liquidity against two missed allocation cycles.
- Governance: Require board visibility into off-balance-sheet exposure and circular financing ties with key counterparties.
The bottom line
The AI economy is now bound together by leverage, timeline bets, and intricate financing. Spreads have widened, stocks are volatile, and transparency is scarce where it matters most.
Growth is real, but so is the capital intensity. Treat this like infrastructure. Price the timing risk. Demand clarity on who is on the hook when the music slows.
Related resource
If you're upgrading team skills for AI use cases in finance, this curated list can help: AI tools for finance.
Your membership also unlocks:
