Wiki/Operators

Colocation REITs: What the Operators Actually Disclose

The framing question for this wiki is whether AI-era hyperscaler economics actually break from traditional datacenter REIT economics, or just look like they do. The traditional colocation REITs — Equinix, Digital Realty, Iron Mountain, plus the privatized cohort (QTS, CyrusOne, Aligned, Stack, Compass, Vantage) — sit at the center of this question because they (a) build a meaningful share of hyperscale capacity, (b) compete with hyperscaler self-build, and (c) disclose far more about project economics than the hyperscalers themselves do. Their 10-Qs and earnings transcripts are an underused primary source for grounding the AI capex debate.

This page collects the numbers operators put on the record in 2025 and early 2026.

1. Yield on Cost: The Number Both Anchors Care About

Both public REITs disclose yield-on-cost explicitly. The numbers have been remarkably stable through the AI boom, which is the first surprise.

Digital Realty (DLR) ended 2025 with 769 MW under construction at an 11.9% expected stabilized yield, on $10 billion of gross development value (DLR Q4 2025 earnings call, Motley Fool transcript). That figure is essentially flat against the 11–12% range DLR has guided to for the past several years. In Q4 they delivered ~90 MW (75% pre-leased) and started another ~135 MW. Full-year development capex was ~$3 billion (Digital Realty Q4 2025 slides via Investing.com).

Equinix (EQIX) reports a different but related metric: cash-on-cash return on gross PP&E for "stabilized" assets (180+ sites at >85% utilization). In Q3 2025, 180 stabilized assets at 82% utilization generated a 26% cash-on-cash return on gross PP&E (constant currency); by Q4 2025, 187 stabilized sites at 82% produced 27% (Equinix Q3 2025 earnings presentation, PDF; Equinix Q4 2025 transcript, Motley Fool). The 26–27% figure is cash on cash on legacy gross book — i.e. on capex spent years ago at lower per-MW prices — not a forward yield on new builds.

The reconciling intuition: DLR's wholesale/hyperscale model targets ~12% on fresh capital today. Equinix's retail interconnection model converts to ~25%+ once an asset has fully ramped over a 5–7 year window. Neither operator is reporting that AI demand has materially moved the yield-on-cost target upward — pricing and cost have inflated roughly in tandem.

DLR's CEO has said explicitly that they are "pushing escalators of minimum 3%, as high as 4% or just above that, CPI-linked on renewals and new deals" (DLR Q1 2025 earnings via Markets Daily).

2. Lease Economics: Term, Escalators, Churn, $/kW

Lease term. Hyperscale leases in the >1 MW segment run typically 10–15 years; DLR's CEO has confirmed "hyperscale contracts are typically 15 years with escalators certainly 3% or maybe even higher" (DLR Q1 2025 highlights). The Blue Owl/Meta Hyperion structure (see §5) is a notable outlier — only a four-year initial term with extension options.

Escalators. DLR disclosed in Q1 2025 that over 85% of bookings included fixed rent escalators of at least 4% or CPI-linkage — a meaningful step up from the 2–3% norm pre-2022 inflation. Equinix has stated they see "no dilution in pricing" on new contracts (Equinix Q3 2025 transcript, Motley Fool).

Churn. Genuinely low. DLR reported total churn of 1% in Q2 2025, with "negligible" churn in the >1 MW segment (DLR Q2 2025 transcript, Motley Fool). Iron Mountain runs at 96.9% occupancy across 488 MW operating capacity (DCD Q4 2025 colo results).

$/kW pricing. Wholesale rates breached $215/kW/month in Northern Virginia (Ashburn) in Q2 2025 — the highest on record (Datacenters.com pricing report). Annualized, that is ~$2,580/kW/year, implying a 100 MW campus generates ~$258M of rent — which against $1.0–1.1B of build cost (see §3) cleanly produces the ~11–12% yield DLR discloses. The numbers tie out.

Lease structure. Turnkey (full M&E, cooling, generators) is priced $/kW. Powered shell (slab + power + roof) is priced $/sq ft and shifts capex to the tenant. Recent investor sentiment skews 42% turnkey hyperscale / 28% powered shell (National Real Estate Advisors) — but powered shell is gaining share as hyperscalers want more control over AI-specific M&E.

3. Build Cost per MW

The pre-AI rule of thumb was $7–12M/MW for turnkey colocation. The current numbers from JLL and Cushman & Wakefield:

Year	Global avg $/MW (turnkey colo)	Source
2020	$7.7M	JLL 2026 Global DC Outlook
2025	$10.7M	JLL
2026E	$11.3M	JLL
AI-optimized	$20M+	JLL

So traditional colo build cost grew ~7% CAGR 2020–2025, but AI-ready facilities are roughly 2x because of (a) liquid-cooling plant, (b) higher power density per square foot (which counterintuitively reduces sq ft but increases dollar per MW because the M&E is denser), and (c) longer lead times on transformers, switchgear, and chillers.

Powered shell, by contrast, is $4–8M/MW (Landgate). The gap between powered shell and turnkey is the M&E + cooling capex that hyperscalers are now increasingly willing to self-fund.

4. AI Workload Differences: How Much Actually Changes?

The "AI changes everything" narrative emphasizes liquid cooling, 100 kW+ racks, and longer build cycles. The disclosures support some of this but not all of it.

Rack density. Traditional enterprise colo averaged 6–10 kW/rack. AI training now drives 60–132 kW/rack standard, with 300 kW achievable. Aligned's DeltaFlow liquid cooling system targets up to 300 kW per rack (Aligned via IntuitionLabs). Liquid cooling held 46% of the cooling market by 2024 and is the standard for new AI builds (Tom's Hardware DC cooling 2025).

Footprint compression. Air-cooled 40 kW racks need 25 racks per MW and ~2,500 sq ft. Liquid-cooled 100 kW racks need only 10 racks and ~1,000 sq ft per MW. This is why AI capex per MW is up but capex per sq ft is up even more — the dollar density of the building rises faster than its electrical footprint.

Cooling capex. Air cooling for 1 MW runs $1.5–2M. Liquid cooling for 1 MW runs $3–4M — roughly 2x (Introl liquid vs air analysis).

Build cycles. Turnkey deployments still take 12–24 months. Power interconnect, not construction, is the binding constraint. Iron Mountain has 200 MW landing in the next 18 months and 400 MW within 24 months — these are largely pre-leased before they exist (DCD Iron Mountain coverage).

Yield on AI builds — the actually surprising data point. Despite AI capex per MW being 2x, the disclosed yield on cost has not risen. DLR is still guiding 11.9% on a development book that is overwhelmingly AI/hyperscale. The implication: AI pricing is also up ~2x, and operators are not capturing scarcity rents at the project IRR level — they may be capturing them in lease-up speed and pre-leased percentages instead.

5. Lease vs. Own: The Hyperion Inflection

The Meta + Blue Owl Hyperion JV is the cleanest disclosed datapoint on lease-vs-build math available, and it is unusual on several counts (Meta press release; DCD coverage).

Total cost: $27B for a gigawatt-scale campus in Louisiana — implies ~$27M/MW for an AI-ready hyperscale build, in line with the JLL $20M+ figure plus land and substation.
Capital structure: Blue Owl funds own 80%, Meta 20%. A Morgan Stanley-arranged SPV issued $27B of A+ rated debt and $2.5B of equity — anchored by PIMCO ($18B) and BlackRock ($3B).
Lease term: Only four years initial with extension options. This is the surprising part. Traditional REIT hyperscale leases run 10–15 years. A four-year term effectively converts Meta's commitment into something closer to operational lease accounting than a 15-year financial obligation.
Why it matters: This is not a colocation lease — it's structured project finance with a single tenant, an 80% off-balance-sheet equity sleeve, and a short-dated lease that gives Meta optionality if AI economics deteriorate. It is "lease" in name only.

Related: BlackRock/MGX/GIP acquired Aligned for $40B in 2025 (5 GW of operational and planned capacity, implying $8M per planned MW at the equity level — but the assets are mostly already built or near-built) (Data Center Frontier on Aligned deal). Blackstone took QTS private at $10B in 2021 and has since committed >$25B of follow-on infrastructure spend.

The pattern: the traditional public REIT model is being supplemented (not replaced) by infrastructure-fund private capital that can tolerate longer hold periods and more bespoke lease structures. Public REITs still set the price discovery anchor.

6. Backlog and Pre-Leased Percentages

This is where the AI demand signal shows up most cleanly in the disclosures.

Digital Realty.

Backlog (signed-but-not-commenced leases): ~$1.4B at 100% share / $852M at DLR's share, end of 2025 (TIKR backlog summary; DLR Q3 2025 supplement).
2025 bookings: $1.2B — second consecutive year over $1B.
Q4 2025 deliveries: 90 MW, 75% pre-leased.
Hyperscale fund: closed €3.225B of LP equity in DLR's inaugural closed-end fund — oversubscribed.

Equinix xScale.

21 live xScale facilities, ~415 MW leased as of Q3 2025 (Equinix Q3 2025 transcript, Motley Fool).
~90 MW of xScale delivered in 2025.
58 major projects underway globally, including 12 xScale projects.
Pre-leasing: ~90% across operational and under-construction xScale capacity as of early 2024 (Dgtl Infra xScale coverage). Specific buildings (Paris PA9x, Dublin DB5x) were 100% pre-leased at JV contribution.

Iron Mountain.

488 MW operating, 96.9% occupied.
63 MW signed in 2025 (43 MW in Q4 alone).
200 MW landing in 18 months, 400 MW in 24 months. Earlier 2024 disclosure: 185 MW under construction at 79% pre-leased.
Targeted 125 MW of leasing in 2025 with 11% constant-currency revenue growth (Seeking Alpha).

GDS Holdings (China). Spun out stabilized assets via a Shanghai-listed C-REIT IPO at 5.2% projected 2026 dividend yield, 16.9x EV/EBITDA (DCD coverage). The cap rate / yield delta vs. US REITs (~5% in China, ~6–7% implied for DLR/EQIX, vs. 11.9% on new development) shows what "stabilized" really means: the development yield converges to the cap rate as risk burns off.

What This Tells Us About the "AI Changes Everything" Narrative

Pulling the threads together:

Yield on cost has not moved. DLR has been at 11–12% before, during, and through the AI capex boom. Equinix's stabilized cash-on-cash is 26–27%. Pricing and cost have inflated together.
Build cost per MW has roughly doubled for AI-optimized turnkey ($10M → $20M+), but powered shell costs are flat ($4–8M). This means hyperscalers can self-fund the delta — and that's exactly what the Hyperion deal does.
Lease terms are getting shorter for the largest deals. Hyperion is 4 years; classic hyperscale colo is 15. This is a material change and pushes risk back onto the asset owner.
Pre-leasing is the real signal of AI demand strength. 75% (DLR Q4 deliveries), ~90% (xScale), 79% (Iron Mountain). Capacity is committed before concrete is poured. Backlogs are at record levels.
Churn is essentially zero in the >1 MW segment. Once a hyperscaler is in, they don't leave — at least not yet.
The private/infra-fund pool is now bigger than the public REIT pool by deal volume. Aligned ($40B), QTS ($10B + $25B follow-on), CyrusOne ($15B), Compass, Stack, and the Meta-Blue Owl Hyperion SPV ($27B) collectively dwarf DLR + EQIX + IRM new development. The public REITs are now price discovery and a fraction of the actual financing.

The contrarian takeaway for this wiki: at the project-economics level, very little has changed. Underwriting still targets ~12% stabilized yield on cost, leases still have escalators, churn is low, pre-leasing is high. What has changed is the scale (gigawatt campuses instead of 30 MW buildings), the capital structure (infra-fund SPVs instead of REIT balance sheets), and the tenant concentration (one hyperscaler per campus vs. dozens per facility). The economics of a single MW look familiar. The aggregation of MW does not.