The short
- Why now: Training and inference need dense racks near dependable power; long contracts de-risk cash flow.
 - Attractive traits: Pre-sold capacity, step-up pricing, and low churn when apps embed into workflows.
 - Hidden risk: Delayed feeders/substations turn metal into idle cost. Power > silicon.
 - Core idea: Infra funds price certainty; venture prices upside. AI hubs sit in the sweet overlap.
 
Deals at a glance
| Deal type | Lock-in risk | Payback logic | Tenor | Power dependency | 
|---|---|---|---|---|
| Lease-to-own (LTO) racks | Low–mid (title shifts) | Capacity pre-booked; rent ≈ 3–5 yrs payback | 5–7 yrs | High: feeder date = revenue date | 
| Capacity pre-lease (shell & core) | Low (take-or-pay) | Index-linked rates; step-ups on refresh | 7–10 yrs | High: substation energization critical | 
| Rev-share with operator | Mid (volume risk) | Share of GPU hours; upside in bursty peaks | 3–5 yrs | Mid: curtailment caps upside | 
| GPU SPV with anchors | Low (anchor SLA) | Anchor pays floor; spot fills tail | 4–6 yrs | High: site PUE + cooling limits | 
Underwriting checklist
Contracts & cash
- Take-or-pay floors, CPI-linked escalators, renewal options baked in.
 - Anchor credit quality > logo value; counterparty diversification.
 - Pipeline of tenants for backfill at expiration.
 
Power & thermals
- Firm grid tie-in letter with dates; transformer/feeder lead times.
 - Cooling headroom for 50–100 kW/rack with liquid loops.
 - PUE targets realistic in summer peaks; redundancy plans tested.
 
Build & delivery
- EPC schedule with penalties; long-lead items already procured.
 - Permits secured, especially water rights where relevant.
 - Ops team SLAs for response/repair windows.
 
Site triage: power, pipes, people
| Factor | Green-light cue | Red flag | 
|---|---|---|
| Substation | Energization date & redundancy confirmed | “Awaiting allocation” with no letter | 
| Cooling | Liquid loop ready; CRAC only for low-density bays | Only air cooling for dense racks | 
| Water | Non-potable source, closed-loop reuse | New draw in water-stressed zone | 
| Ops | 24/7 on-site with spares SLA | Remote-only crew, no spares plan | 
Anchor profiles (why they stay)
- Foundry AI (training): Needs long runs, low interruption risk; favors take-or-pay with scale discounts.
 - Enterprise AI (inference): Predictable daily cycles; values burst rights during product launches.
 - Consumer AI (apps): Spiky demand; pays for burst + routing across sites with low egress.
 
Stickiness test: If switching requires moving data, re-qualifying security, and re-plumbing storage, churn is low.
Risk box
- Power slip: Every delayed feeder month burns cash; keep a liquid reserve for 6–9 months.
 - Thermal ceiling: If racks de-rate in heat waves, revenue caps below plan.
 - Single-tenant risk: One anchor > 60% revenue? Insist on step-down covenants.
 
What to watch: Lease-to-own structures tied to confirmed power windows (substation dates + redundancy). If power slips, returns do too.