Underwriting the AI Revolution: A Guide to Data Center Investment

U.S. data-center electricity demand is expected to grow rapidly, but published estimates are scenario-based and should not be reduced to one universal GW figure. Lawrence Berkeley National Laboratory's 2025 update gives a 649 TWh reference case for U.S. data centers in 2030, with a compounded-uncertainty range of 521 to 843 TWh. The IEA's 2026 outlook projects global data-center electricity consumption to rise from 485 TWh in 2025 to about 950 TWh in 2030, with AI-focused consumption growing faster than the sector overall. Use the specific study, geography, energy measure, utilization assumption, and scenario when underwriting demand: LBNL's 2025 update and the IEA's 2026 outlook.

AI data center underwriting is different from a warehouse or office model because the project combines real estate, utility delivery, specialized equipment, long-lived infrastructure, and concentrated customer contracts. Square footage remains relevant, but it is not a sufficient denominator for a capacity-led product. For the broader mechanics of an auditable real estate pro forma, keep the property model separate from the infrastructure schedules below. A stabilized property acquisition can begin with TILT's Commercial Acquisition Model, while power, cooling, commissioning, and contract schedules remain project-specific.

What Should an AI Data Center Underwriting Model Measure?

Underwriting an AI-era data center requires an infrastructure-style model alongside the real estate model. The key variables are delivered electrical capacity and timing, IT load and rack-density assumptions, cooling and heat rejection, connectivity, tenant commitments, power and operating-cost pass-throughs, reinvestment, financing, and residual or stranded-capacity risk.

For each project, connect the physical design to the cash flow model:

• Load: Define whether each MW refers to utility, critical, or IT load. Reconcile contracted capacity, energized capacity, usable IT load, rack count, and commissioning dates.
• Design: Tie rack density to the selected server configuration, electrical distribution, cooling architecture, heat rejection, water, controls, and redundancy plan.
• Contract: Model the actual rent unit, ramp, minimum commitment, power reimbursement, escalator, ancillary charges, renewal rights, termination rights, guarantees, and remedies.
• Capital: Build the cost and reinvestment schedules from the project design and estimate. Do not substitute a global cost-per-MW or revenue-percentage default.

How Should You Underwrite Power Delivery and Site Selection?

For AI-oriented projects, delivered electrical capacity and the date it becomes usable may be the primary gating variables, but location still matters. Underwrite utility service, interconnection and network-upgrade scope, fiber, water or heat rejection, permitting, labor, taxes, and customer demand together.

Require evidence for the project's exact service path:

• Utility and tariff: Obtain the load letter or service agreement, tariff, deposits, milestones, and conditions for service.
• Studies and upgrades: Identify transmission or distribution studies, construction agreements where applicable, network-upgrade scope, cost allocation, and responsibility for delays.
• Schedule: Model study milestones, equipment lead times, permitting, construction, testing, commissioning, and energization as a probability-weighted schedule.
• Reliability: Analyze the actual topology, outage assumptions, maintenance plan, and contractual service obligations. Tier labels do not replace this project-specific analysis.
• Location: Confirm fiber and network access, water or heat-rejection constraints, taxes, labor, zoning, environmental approvals, and customer demand.

How Do You Model Revenue for an AI Data Center?

Model the contractual revenue unit actually used in the lease. It may be dollars per kilowatt per month, dollars per square foot, a fixed facility charge, consumption-based charges, or a combination.

Your financial model must be built around this reality:

• Capacity: Reconcile contracted utility capacity, critical IT load, rentable area, ramp schedule, power reimbursement, ancillary charges, escalators, and pass-throughs.
• Lease-up: For capacity-led products, track contracted and energized IT load alongside area, suites, racks, revenue, and operating capacity. Different denominators can produce different utilization percentages.
• Lease terms: Verify the executed term, renewal rights, termination rights, guarantees, minimum commitments, remedies, and escalators. Treat take-or-pay as a defined contractual provision, not a universal feature.
• Power costs: Model the tariff, PPA settlement, escalation, reconciliation, and the party responsible for each cost rather than assuming a standard triple-net pass-through.

As one disclosed operator example, Digital Realty's Q2 2026 Americas bookings were reported at a blended $182 per kW per month, with different rates for 0 to 1 MW and greater-than-1 MW bookings. That is a REIT leasing statistic, not an AI-only market index. Use the executed rent schedule and comparable product set instead: Digital Realty's Q2 2026 disclosure.

How Should You Build the Cost and Reinvestment Schedule?

Data-center cost per MW is not a portable global average. It depends on whether MW means utility, critical, or IT load; land and shell; electrical and mechanical plant; redundancy; cooling architecture; tenant improvements; utility upgrades; financing; location; and delivery date.

Build the capex schedule from the actual design:

• Site and utility: Land, site work, utility and substation scope, network upgrades, deposits, and financing carry.
• Facility: Shell, electrical distribution, UPS, backup generation, cooling, heat rejection, controls, network, security, tenant improvements, commissioning, soft costs, and contingency.
• Reinvestment: Add separately reasoned cases for hardware interfaces, power distribution, cooling, controls, and tenant improvements. Do not use a universal percentage-of-revenue reserve.
• Denominator: State whether the estimate is per utility, critical, or IT MW and what scope, geography, redundancy, and delivery date it covers. Digital Realty and Blackstone disclosed approximately $7 billion of development cost for approximately 500 MW of potential IT load in one multi-campus transaction. That is not a market-wide benchmark. See the disclosed transaction estimate.

The economic value is heavily influenced by delivered power, electrical and mechanical infrastructure, cooling, connectivity, and tenant commitments, not just gross building area. The shell is not automatically generic: reuse depends on structure, floor loading, clear height, fire protection, electrical topology, cooling plant, water constraints, and permitting.

What Cooling Systems Do AI Workloads Require?

High-density AI deployments may require direct-to-chip liquid cooling, rear-door heat exchangers, immersion, or a hybrid design, depending on the selected servers, rack density, facility temperatures, redundancy requirements, and heat-rejection plant.

Do not infer the cooling architecture from a kW-per-rack label alone. Require a thermal design, coolant-distribution and CDU scope, water and heat-rejection assumptions, failure modes, maintenance plan, and commissioning evidence. Microsoft's 2025 environmental report documents its transition toward chip-level liquid cooling, while ASHRAE's data-center resources provide the relevant engineering context.

The design cases should distinguish:

• Direct-to-chip: Coolant reaches the processor cold plate, with building-wide distribution and CDU scope to be documented.
• Rear-door heat exchange: Liquid-assisted heat rejection at the rack boundary, with limits and compatibility established by the selected equipment.
• Immersion or hybrid: A different server, rack, maintenance, fluid, and commissioning basis that must be priced and tested as designed.

What Does PUE Tell You About Operating Costs?

PUE is total facility energy divided by IT-equipment energy. It is a boundary- and measurement-dependent efficiency metric, not a complete measure of computing efficiency or a universal AI-facility target. DOE's 2024 design guide gives 1.6 as an average-data-center reference and notes that highly efficient facilities can be below 1.1. Google reported a 2025 fleet-wide PUE of 1.09, while Microsoft reported a FY25 global average of 1.17. Those are operator-specific results, not a universal benchmark. Use measured annual energy and the same boundary when comparing assets: DOE's design guide, Google's efficiency reporting, and Microsoft's efficiency reporting.

For a transparent hypothetical, assume a constant 100 MW IT load, a 0.1 PUE improvement, and an energy price of $0.07 per kWh. The reduction is 10 MW, producing approximately $6.13 million of annual energy savings before demand charges, taxes, losses, utilization changes, or the question of whether the stated 100 MW means IT load or facility load. Replace that illustration with the project's tariff and measured boundary.

What Are the Tenant Credit and Concentration Risks?

Tenant concentration is a deal-specific credit and re-leasing risk. A tenant's parent credit, guarantee scope, contracted capacity, ramp, remedies, and ability to reuse the infrastructure must be tested in the executed documents rather than inferred from a tenant category.

Review:

• Contractual support: Identify the tenant entity, guarantor, guarantee limits, security, minimum commitments, termination rights, remedies, and credit triggers.
• Concentration: Show exposure by tenant, parent, project, power block, lease expiry, and sponsor portfolio. Multiple buildings can still represent one customer.
• Re-leasing: Test whether the building, electrical topology, cooling, network, and permits can serve another workload without material reconfiguration.
• Operating performance: Connect service levels, commissioning evidence, outage obligations, maintenance rights, and pass-through mechanics to cash flow.

How Has Power Procurement Changed in 2026?

Power procurement should be underwritten source by source. Utility service, tariff and transmission, physical or virtual PPAs, on-site generation, storage, backup generation, and curtailment rights provide different combinations of energy, capacity, attributes, and reliability support.

For each source, confirm the contract and model settlement, basis, congestion, fuel, interconnection, environmental, and pass-through risks. Most data centers still prefer grid service, while the IEA reports that U.S. developers are pursuing on-site gas in response to slow grid connections. That is an emerging development pattern, not a universal 2026 default.

Hyperscalers are also pursuing long-term arrangements involving renewable, nuclear, storage, and other firm-power technologies. Treat announced projects as options until the required permits, financing, interconnection, fuel or power contract, and operating date are verified. Do not include future nuclear capacity in a 2026 base case without a contractual and delivery basis. The IEA's 2026 analysis provides the appropriate context.

The model should separate:

• Firm service: Delivered utility capacity, tariff, upgrades, reliability, and energization milestones.
• Contractual supply: PPA energy, capacity, attributes, settlement, term, credit, and basis exposure.
• Local resources: Generation, storage, backup, fuel, emissions, permits, maintenance, and operating constraints.
• Tenant economics: Which costs are passed through, which remain with the landlord, and how reconciliations and escalations work.

What Are the Common AI Data Center Underwriting Mistakes?

The damaging errors are denominator errors, unsupported defaults, and failure to connect legal and engineering evidence to the financial model.

1. Using the wrong revenue unit. Model the contractual unit actually used: capacity, area, fixed facility charge, consumption, or a combination.
2. Using a generic design. Tie load, rack density, cooling, power distribution, heat rejection, water, and commissioning to an equipment schedule and engineering scope.
3. Assuming energization. Replace national queue defaults with utility milestones, upgrade scope, equipment lead times, and explicit delay cases.
4. Hiding pass-through risk. Model tariff, PPA settlement, escalation, basis, reconciliation, and the party responsible for each cost.
5. Using a generic obsolescence reserve. Identify the failure mode: rack power, floor loading, busway, switchgear, coolant distribution, heat rejection, water, network, controls, code, or tenant-improvement turnover. Size the reserve from engineering scope and scenarios.
6. Confusing a market reference with a return forecast. Public lease, yield, cap-rate, and securitization disclosures describe particular assets or portfolios. They do not replace the project's own cost, ramp, financing, and exit assumptions.

How Should You Underwrite Returns and Financing?

Returns must be underwritten from the asset's actual cost, rent, ramp, power delivery, operating costs, capex, debt, tax, exit, and tenant-credit assumptions. Public disclosures provide transaction-specific reference points, not a universal 2026 return table.

For context, Blackstone's prospectus describes a target initial gross asset yield of 5.75 to 7.00 percent or higher for its intended portfolio, while a June 2026 Digital Realty transaction disclosed an expected stabilized capitalization rate above 6.5 percent. Yield, cap rate, levered IRR, and development return are different measures. Neither disclosure establishes a market-wide cap-rate or IRR range. Review the Blackstone prospectus and the Digital Realty transaction disclosure as examples, not comps without adjustment.

Stabilized data-center cash flows have also been used in secured revenue-note and asset-backed transactions. That evidence supports the existence of transactions, not automatic eligibility, lower-cost debt, a broad 144A market, or compressed exit cap rates. Review collateral leases, tenant concentration, ratings, reserves, liquidity support, covenants, amortization, legal isolation, power obligations, and servicing. See the Aligned 2026-1 materials and Centersquare 2025-1 notes for transaction-specific evidence.

Why Should You Underwrite These as Infrastructure?

These assets are best framed as infrastructure-heavy real estate: the building matters, but the investment case also depends on delivered power, specialized electrical and thermal systems, connectivity, customer contracts, and the ability to reinvest as the design envelope changes.

The model should show the evidence behind each major assumption: utility milestones, engineering scope, executed lease terms, operating-cost pass-throughs, tenant credit, financing documents, reinvestment cases, and residual or stranded-capacity risk. That is the difference between a real estate model with a data-center label and an underwriting framework that can be tested at investment committee.

Frequently Asked Questions