AI's Bottlenecks Are Obvious: Watts and Dollars
Every boom runs into physics and finance. AI is no exception. The constraint is simple: not enough cheap, reliable electricity, and too much easy credit chasing scarce compute.
That mix can inflate asset prices, pull forward demand, and invite engineering tricks that look clever-right up until cash flow doesn't show up on time. If you work in finance, your edge is not more AI hype. It's underwriting the bottlenecks with discipline.
The two hard limits
First, energy. Training and inference need massive, steady power. New data centers are queuing for grid connections, competing for transformers, and negotiating complex permits and long-lead equipment. Even with capital in hand, delays are common.
Second, credit. Capital is pouring into GPU leasing, vendor financing, and off-balance sheet SPVs. Structures get creative when demand looks limitless. That's exactly when covenants, milestones, and counterparty quality matter most.
Why this inflates the AI trade
When power is scarce, capacity gets pre-sold years ahead. When money is cheap, builders promise scale that depends on gear deliveries, substation upgrades, and chip roadmaps. The result: valuations based on future exabytes and petaflops, not present cash generation.
We've seen the rhyme before. Tools that extend credit and shift timing risk feel innovative until the real economy (in this case, electrons and lead times) calls the bluff.
Where the cracks can show
- Delayed energization pushes revenue out while interest accrues.
- Utilization misses if inference demand is overestimated or customers migrate to cheaper models.
- Opex shock from higher all-in $/kWh erodes unit margins even when racks are full.
- Residual value risk on specialized chips if a new generation compresses resale prices.
What finance leaders should do now
- Underwrite power first: confirm interconnection queue position, substation scope, utility upgrade timelines, and curtailment terms. Require step-in rights and liquidated damages tied to energization.
- Tie funding to physics: milestone draws on site readiness, MW energized, PUE achieved, and contracted take-or-pay capacity-not on purchase orders alone.
- Price true TCO per token/training run: include capex, depreciation (assume short useful lives), power, cooling, networking, and downtime. Stress test under lower utilization.
- Hedge the inputs: layer PPAs with collars, shape risk coverage, and demand-response revenues where feasible. Match hedge tenor to debt maturities.
- Diversify vendor exposure: don't rely on one chip supplier, interconnect standard, or cooling approach. Pre-negotiate buyback or resale programs.
- Demand real offtake: favor contracts with credible counterparties, cash-backed commitments, or capacity reservations that survive model shifts.
- Protect the downside: maintenance covenants, DSRA sizing tied to energization risk, and change-of-control protections if facilities are flipped.
- Back picks-and-shovels: transformers, switchgear, modular data halls, and grid-adjacent land often deliver steadier cash than speculative compute.
Metrics that actually matter
- MW contracted vs. MW energized; months-to-energization by site.
- PUE at design, at load, and under heat stress; uptime track record.
- All-in $/kWh (energy, demand charges, transmission, hedges, plus expected curtailment).
- Utilization profile by hour and tenant; share of capacity under take-or-pay.
- Chip delivery schedule, acceptance criteria, and penalties for slippage.
- Gross margin per token/inference; CAC payback for AI services tied to the capacity.
- DSCR under delayed revenue; cash conversion cycle including deposits and prepayments.
- Residual value assumptions and tested secondary markets for prior-gen hardware.
How to structure smarter
Keep recourse limited to energized assets. Use springing covenants that tighten if energization dates slip. Align incentives with shared savings: lower PUE, lower coupon. And cap distributions until stabilization-when utilization, PUE, and tenant payments have proven out.
If you fund growth equity, insist on a capital plan that pairs grid access (or on-site generation) with realistic chip delivery and build schedules. Cash burns fast when any one of those three stalls.
Energy diligence: don't outsource it
Review utility letters, interconnection studies, and transformer lead times. Ask for historical curtailment data and seasonal pricing curves. Validate that backup generation and thermal systems are permitted and insurable.
For context on data center load growth and efficiency trends, see the International Energy Agency's research on data centers and networks here.
Where returns still look compelling
- Brownfield expansions with existing substations and proven tenants.
- Long-duration PPAs connected to capacity markets that compensate availability.
- Cooling retrofits and power upgrades with short payback and contracted demand.
- Portable/modular data halls tied to firmed mobile generation for interim capacity.
A sober framework for the "AI trade"
Model your base case on energized MW and contracted dollars, not logos or headlines. Then add delays, higher power costs, and a step-down in chip resale values. If IRR survives that, you likely have a deal-if it doesn't, you just avoided a problem.
Bubbles form when capital forgets the bottlenecks. Your job is to price them, fund against them, and get paid for taking them.
Level up your team's AI finance fluency
If your analysts need faster context on AI workload economics and vendor risk, point them to practical tool roundups and role-specific training. Start with AI tools for finance here, explore finance-focused AI courses by job here, or consult Research for deeper studies and methodology.
Your membership also unlocks:
