AI Data Centers: The $200 Billion Lending Puzzle No One Can See Clearly
AI has turned data centers into the new trophy asset. At least $200 billion was borrowed last year to fund sites, shells, and the energy hookups needed to keep racks humming. Forecasts peg total AI infrastructure spend at $3 to $5 trillion over the next few years, with hundreds of billions more in fresh debt on deck.
On paper, it sounds simple: buy land, build, secure electricity, lease to blue-chip tenants. In practice, lenders can't easily tell how much exposure they actually have to AI. The financing stack is a tangle, and the risk is layered in ways many balance sheets don't fully reflect.
Why lenders can't see their true exposure
Capital stacks now blend private credit, SPVs, corporate bonds, high-yield debt, and asset-backed securities. Some loans look like traditional CRE; others look like construction or project finance that need a takeout at maturity. Exposure gets duplicated across vehicles, syndications, and securitizations.
That makes concentration risk hard to track. One institution can be in the senior, the mezz, the bond index, and an ABS pool-thinking they're diversified when they're staring at the same assets from four sides.
The refinancing trap
Banks are planning ahead by structuring loans so a $10 billion campus can be sliced into $1 billion pieces at refi. Smart idea-if demand holds, rates cooperate, and tenants expand on schedule. If not, maturities stack up fast.
There's a bigger concern: leverage is rising across AI-linked borrowers. The Bank for International Settlements has flagged higher debt loads, which means any stumble can echo through credit markets.
What to underwrite now (developers, lenders, and builders)
- Grid and interconnection: Queue position, substation scope, transformer lead times, curtailment risk, and any clawback provisions from the utility.
- Electricity costs and certainty: Hedged pricing, PPAs or equivalent, passthrough mechanics to tenants, and exposure to congestion pricing.
- Permitting and politics: Local moratoria, data center caps, water-use restrictions, and environmental review timing.
- Cooling strategy and resilience: Water sourcing, reuse plans, dry/hybrid systems, heat-recovery options, and backup generation availability.
- Tenant credit and lease quality: Parent guarantees, step-in rights, pre-leasing thresholds, termination rights, renewal options, and density commitments.
- Revenue coverage: Minimum rent vs. usage-based components, escalators, and DSCR tested against realistic energy and maintenance costs.
- Tech readiness: Liquid cooling capacity, floor loading, busway flexibility, and upgrade paths for higher rack density.
- Schedule risk: GPU and switchgear lead times, commissioning windows, and liquidated damages that actually bite.
- Exit options: Breakup-friendly parceling, zoning flexibility, and eligibility for ABS/CMBS or insurance takeout.
- Capital reserves and covenants: True-to-life contingency, capex reserves for retrofits, and triggers that force early course corrections.
- Fiber and rights-of-way: Dark fiber availability, redundancy, easements, and cross-connect commitments.
- Insurance and outages: Business interruption for grid events, cyber incidents, and equipment failures.
Why some investors are sidestepping direct data center risk
More sophisticated capital is moving upstream: transmission upgrades, substations, generation and storage, and long-duration grid assets. Those get paid whether a specific AI training cluster hits targets or not.
Energy demand from data centers is real and growing. For context, see the IEA's work on the sector's electricity needs: IEA commentary on data centers and AI.
Stress-test these scenarios before you wire funds
- Interconnection delayed 12-24 months despite signed agreements.
- Electricity prices double; passthroughs don't fully cover timing gaps.
- GPU deliveries slip two quarters; lease rent commencement moves right.
- Anchor tenant slows expansion or rightsizes capacity after efficiency gains.
- Local rules cap new connections or restrict water usage mid-project.
- Rates stay higher-for-longer; takeout markets thin when your debt matures.
Practical moves you can make this quarter
- Lock utility commitments with penalties for delay and clear curtailment language.
- Structure mini-perm with baked-in conversion to term if coverage and leasing hurdles are met.
- Demand parent-level support or LC backstops from hyperscale tenants; limit termination options.
- Design for modular phasing so cash flow turns on in smaller, financeable chunks.
- Pre-buy long-lead gear (transformers, switchgear) with assignment rights to lenders.
- Install metering and contracts that separate base rent from usage, with transparent escalators.
- Underwrite a second life: can the shell pivot to colocation, high-performance compute, or industrial use?
The fork in the road
This buildout could mint a decade of growth-or leave a trail of distressed shells and stranded interconnects. The difference comes down to underwriting discipline, honest timelines, and exit optionality. If AI profits lag or if we overshoot on capacity, the correction will move fast.
Map your exposure across entities, stress the ugly cases, and fund what still pencils. If your team needs to level up on AI's impact across roles, here's a practical place to start: Complete AI Training: Courses by Job.
Your membership also unlocks:
