Data centers are the hottest assignment in commercial appraisal right now, and the easiest one to get wrong. Global occupancy ended 2025 at 97%, and the hyperscalers are putting $635 to $690 billion of capex into the ground in 2026 alone. When the demand story is that loud, it's tempting to let it do your underwriting for you.
Don't. The numbers above describe the sector, not the building in front of you. A data center is still a special-use property with a power contract, a tenant, a lease, a building that may or may not be designed for the workloads it's supposed to house, and a county that may or may not want it there. Underwrite those things or you're not appraising. You're cheerleading.
Here's the checklist I work through, in the order the value actually gets decided.
1. Power and interconnection
What to check: Confirmed, deliverable megawatts. Not “near a substation,” not “power available in the area.” The interconnection agreement or utility letter, the queue position with a date attached, substation capacity, and whether any of the load is planned behind the meter.
Why it matters to value: Land near a constrained substation with a five-year queue is a different asset than land next to surplus generation. Power is the first and largest value driver; everything else on this list is commentary if the electrons can't get there.
The stats: In one major regional footprint, there are 60 GW of potential large-load power projects in planning but only 13 GW in advanced stages, with a 6 GW generation shortfall looming if it all gets built. Contracted capacity is so far behind demand that it won't come online until late 2026 into 2027. And a quarter of sites have stopped waiting on the grid: 25% of data center sites now run on-site generation, up from 19% a year ago.
My rule: when comparing sites, I want to know whether the megawatts come from the grid, from behind the meter, or from a press release. They are not the same thing.
2. PUE and WUE: the building's operating efficiency
What to check: Measured PUE (not the brochure figure), the cooling system type, the water source and WUE, and how utility costs actually flow through the lease.
Why it matters to value: PUE is the building's gas mileage. A high PUE means higher operating expense for the tenant and a shorter economic life for the building, both of which land in your income approach and your obsolescence analysis.
The stats: The weighted-average PUE across 681 operators was 1.54 in 2025, and the long-run trend is real but slow. The U.S. average PUE improved from 1.6 in 2014 to 1.4 in 2023. Water is the number nobody prices until a county does: U.S. data centers consumed 21.2 billion liters of water in 2023.
One gap I'll flag honestly: my research doesn't have current hyperscaler WUE benchmarks in citable form. Ask the operator for WUE directly. It's a ratio worth having, and its absence in public sources tells you something about how new this disclosure is.
3. Tenant credit and lease escalators
What to check: Who the tenant actually is and what their credit is worth, the lease term, the rent basis ($/kW/month), the escalator structure (fixed vs. CPI), and rollover concentration, meaning how much income expires in any 12- to 24-month window.
Why it matters to value: This is a credit-and-term story, same as any net-leased asset. The rent level tells you where you are in the cycle; the term and escalators tell you how much of that rent you get to keep.
The stats: Hyperscale wholesale leases run $100 to $150/kW/month on 10- to 15-year terms, structured like single-tenant industrial. Colocation runs $150 to $250/kW/month on 5- to 10-year terms, more like multi-tenant office. Know which one you're looking at, because the valuation math is different. The pricing power is real: requirements of 1 MW and above averaged 13% rent increases. And lenders have a bright line on concentration risk. CMBS underwriting flags lease roll of more than 30% in 24 months or 40% in 12 months. Run that test on every rent roll.
4. Capex and construction cost per MW
What to check: The all-in cost per megawatt for the specific asset, and just as important, what's included. Shell and core is one number; the number with AI fit-out is a different one. Also the capital structure: how much is construction debt, and what pre-leasing did the lender require?
Why it matters to value: Replacement cost anchors the cost approach, and the capex scale tells you about tenant staying power. Nobody walks away from a nine-figure fit-out casually.
The stats: Average construction cost in the largest global markets ran $10.7M per MW in 2025, rising to $11.3M in 2026, and AI-optimized fit-outs add up to $25M per MW on top of the shell. Quote a cost figure without saying which scope it covers and you've said nothing. On the financing side, senior construction loans typically fund 55 to 65% of costs and require about 60% pre-lease commitments, which means the lender already did a version of your tenant-credit analysis. Read their conclusions before you finalize yours.
5. Density and rack power trends
What to check: The average rack density the building was designed for, the floor loading, the cooling type (air vs. liquid), and whether there's a credible upgrade path.
Why it matters to value: Density is obsolescence risk in plain English. A building designed for yesterday's rack densities reprices the day the tenant's workloads outgrow it.
The stats: Average rack density hit 27 kW in 2026, up from 16 kW a year earlier and just 7 kW in 2021. AI training workloads demand 40 to 100+ kW per rack. Cooling is racing to catch up: 36% of operators have liquid cooling deployed and another 28% plan to adopt it within 12 to 24 months. When I tour a facility, I'm checking the cooling and the floor loading against those numbers. A 2018-vintage enterprise hall and a 2026 AI-ready facility are not the same asset class anymore, and the market is starting to price them that way.
6. Staffing risk
What to check: Whether the local labor market can actually staff the building, meaning technicians, electricians, operations people, and whether the operator has a track record of keeping them.
Why it matters to value: An unstaffed or understaffed building underperforms: more downtime, slower response, weaker renewal probability. It's an operating risk that feeds directly into your income projections and your risk premium.
The stats: 46% of operators report difficulty finding qualified candidates, and 37% struggle to retain the staff they have. That's not a footnote. Nearly half the industry can't hire. In a tight labor market, the building with the staffed, stable operations team is worth more than the identical building down the road that keeps losing its electricians. Ask about turnover, not just headcount.
7. Residual and replacement value
What to check: The expected useful life, the cost to retrofit versus rebuild, typical holding periods, and who the buyer is at the end of your hold.
Why it matters to value: The exit is where the money is made. Underwrite the building at year 10 or 15, not just year 1.
The stats: The first hyperscale data center ABS gives us real benchmarks: tenant renewal probability around 75% and an expected useful life of 32 to 34 years. But “useful life” assumes the building keeps up. Retrofitting an older facility runs $50 to $100M per 10-MW facility, which the lenders themselves call “a challenging economic equation for older properties.” And the hold-period data says this is a long game: 93% of holders expect holding periods beyond five years. Underwrite accordingly. Short-hold, quick-flip math doesn't fit this asset class.
8. Entitlement and zoning risk
What to check: The actual zoning status (not the broker's version), any moratoriums or pending ordinances, noise and water-use restrictions, setback requirements, and the planning staff's genuine posture toward the project.
Why it matters to value: A site that's perfect except the county just paused data center approvals carries a real, quantifiable discount. Entitlement risk isn't theoretical anymore.
The stats: The political fuel is measurable. U.S. residential electricity prices rose 29% from 2019 to 2025, and “why are my bills going up” is the oldest and most effective argument against any new development. On the ordinance side, I'll be straight with you: my research includes a qualitative policy memo on local data center ordinances, but no citable count of moratoriums by jurisdiction. Don't let anyone quote you a national moratorium number they can't source. Talk to planning staff before you believe the word “shovel-ready.”
The bottom line
Underwriting a data center is underwriting eight things, and power is first among equals. The sector's growth numbers, 97% occupancy and $635 to $690B in annual hyperscaler capex, tell you the tide is rising. They don't tell you whether the specific building floats.
If someone brings you a data center assignment, start with the same questions I start every one with: where do the electrons come from, what does the tenant's lease actually say, was this building designed for the workloads of 2026 or 2016, and what does the county actually allow? Answer those four and the valuation mostly writes itself. Skip them and no amount of sector enthusiasm will save the report.
Sources: JLL 2026 Global Data Center Outlook; AFCOM State of the Data Center 2026; Uptime Institute Global Data Center Survey 2025; AlixPartners 2026 Data Center Market Outlook; CRE Finance Council Data Center E-Primer (Feb 2026); Lawrence Berkeley National Lab 2024 U.S. Data Center Energy Usage Report; CRS R48762.3 Data Center Energy Infrastructure: Federal Permit Requirements (2025); E3 Understanding the Drivers of Rising Electricity Rates and the Role of Data Centers (2026); IBISWorld OD6584 Hyperscale Data Center Services in the US (Mar 2026).