Underwriting the AI Revolution: A Guide to Data Center Investment

The explosive growth of artificial intelligence has ignited a shadow boom in a once-niche asset class: the data center. This is not just another cyclical real estate trend. It is a secular demand shock. The voracious computational needs of AI have created an unprecedented, and insatiable, demand for digital infrastructure, transforming data centers into a primary target for institutional capital.

But underwriting an AI-era data center is fundamentally different from analyzing a traditional warehouse or office building. The old rules of real estate investment do not apply. The core value drivers have shifted from location and square footage to power and kilowatts. For analysts and investors, mastering this new underwriting paradigm is the only way to capitalize on the opportunity without absorbing the risks that come with it.

What Is an AI Data Center, and Why Does It Need Different Underwriting?

An AI data center is a high-performance computing facility built to accommodate densely packed GPU and accelerator hardware, with power densities five to twenty times higher than traditional data centers and cooling systems engineered for liquid-based heat rejection rather than air handling. The underwriting framework that works for traditional real estate (square footage, lease terms, exit cap rate) breaks down because the asset's value is driven by kilowatt capacity, not floor area.

The shift from traditional storage to AI-era supercomputing changes four things at once:

• Power density: A traditional data center might operate at 5 to 10 kilowatts per rack. An AI-ready facility must accommodate 50, 80, or even 100+ kW per rack.
• Cooling architecture: Air-cooled HVAC systems cannot handle AI-era heat loads. Direct-to-chip liquid cooling, immersion cooling, or rear-door heat exchangers are required.
• Lease structure: Leases are written in dollars per kilowatt per month, often with take-or-pay terms over 15- to 30-year horizons.
• Capital intensity: Build cost per MW has climbed to roughly $11.3 million globally in 2026, with US tier-1 markets running materially higher.

A real estate pro forma built around PSF rent, traditional opex ratios, and a 7-year hold with a 6 percent exit cap will produce a number for an AI data center. The number will be wrong, often by an order of magnitude.

Why Has Power, Not Square Footage, Become the Primary Underwriting Metric?

Power has replaced location as the primary value driver because AI workloads consume so much electricity that the practical constraint on a data center is no longer the building footprint or proximity to fiber, but the ability to secure committed power capacity from the local utility.

In traditional real estate, value is driven by proximity to population centers, transportation hubs, and tenant demand. In the world of AI data centers, value is driven by proximity to power. The single most critical factor for a data center development is securing access to massive, reliable, and scalable power from the electrical grid.

Your due diligence must focus on:

• Power availability: Can you secure a binding commitment from the local utility for 200, 500, or 1,000+ megawatts? This is the primary barrier to entry and the most valuable asset in the deal.
• Interconnection queue position: In many US markets, interconnection queues stretch 4 to 7 years. A project that needs 500 MW with no queue position is not a developable site; it is an option on a developable site.
• Redundancy: What is the reliability of the grid? Are there multiple substations and redundant power feeds to ensure uptime? Tier III and Tier IV facilities require N+1 or 2N redundancy minimums.
• Cost of power: The long-term, all-in cost of electricity is a primary operating expense and directly impacts tenant demand and profitability. Power cost variance of 2 cents per kWh across regions translates into hundreds of millions in NPV difference over a 20-year hold.
• On-site generation potential: 2025 and 2026 have seen rapid expansion of behind-the-meter power solutions including natural gas peakers, battery storage, and on-site solar. Sites with viable on-site generation now command meaningful premiums.

How Do You Model Revenue for an AI Data Center?

Data center leases for AI tenants are structured around contracted power capacity, measured in kilowatts or megawatts, and priced in dollars per kilowatt per month. Revenue, utilization, and lease-up all need to be modeled in power terms, not in square footage.

Your financial model must be built around this reality:

• Revenue: Instead of "Rent PSF," your primary revenue driver is "dollars per kW per month." Typical 2026 rates for new contracts run 110 to 175 dollars per kW per month in major US markets, with significant variance by power density and location.
• Utilization: Lease-up is measured by contracted kilowatts, not occupied square feet. A facility can be 100 percent leased on a power basis while less than 60 percent of the floor area is occupied.
• Lease length: AI tenant leases routinely run 15 to 25 years (some 30+), with take-or-pay structures that obligate the tenant to pay for committed capacity whether or not they consume it.
• Escalators: CPI-linked escalators are standard. Some 2026 contracts have begun incorporating power-cost pass-through escalators tied directly to the utility's tariff.
• Pass-throughs: Understand how the variable cost of electricity is passed through to the tenant. In a triple-net structure with power pass-through, the developer absorbs less commodity risk. In a hybrid structure, the developer carries it. This is a critical point of negotiation and a critical input in the model.

The lease structure has direct implications for the right side of the capital stack. Take-or-pay leases with investment-grade tenants are increasingly being securitized in 2026 through ABS and 144A bond markets, which has pulled in lower-cost debt for stabilized assets. Modeling without considering the take-out structure misses meaningful value.

What Is the CapEx Profile of an AI-Era Data Center?

The initial construction cost is only the beginning. Data centers require substantial and continuous capital expenditure to remain competitive, with cooling infrastructure being the largest single capex category and the one most prone to obsolescence.

Build cost per MW averaged roughly $11.3 million globally in 2026, up 6 percent year-over-year on rising equipment costs and tight construction labor markets. US tier-1 markets (Northern Virginia, Dallas, Phoenix) run 10 to 25 percent above the global average. The cost stack:

• Day-one cooling infrastructure: Direct-to-chip liquid cooling systems are far more complex and expensive than traditional air handlers. A 100 kW per rack design adds roughly 1.5 to 3 million dollars per MW vs a 30 kW per rack design.
• Electrical infrastructure: Switchgear, UPS systems, transformers, generators. Scales with redundancy tier.
• Building shell: Surprisingly the smallest line item. The structure is generic; the equipment inside is what matters.
• Fiber and network connectivity: Less critical than it used to be (AI workloads have lower latency sensitivity than the legacy cloud workloads that drove past investment), but still material.
• Ongoing capex: Technology evolves rapidly. Your model must include significant reserves for future upgrades to power distribution, cooling systems, and racks themselves. A reasonable industry default is 4 to 7 percent of revenue annually, with step-functions every 5 to 7 years for major equipment refreshes.
• Obsolescence reserve: Newer than the standard capex reserve. Models the risk that next-generation AI hardware (post-Blackwell, post-Rubin) requires structural building changes the current facility cannot accommodate.

Investing in AI data centers is less like buying real estate and more like investing in a private power plant or a toll road. The barriers to entry are immense, the capital requirements are staggering, and the technical expertise required is profound.

What Cooling Systems Do AI Workloads Require?

AI workloads at 50+ kW per rack cannot be cooled by traditional air handling. They require liquid cooling, either direct-to-chip, rear-door heat exchangers, or full immersion, with material implications for both day-one capex and ongoing operating costs.

The three primary cooling architectures in 2026:

• Direct-to-chip liquid cooling: Coolant lines run directly to the GPU or CPU cold plate. Most common in 2026 deployments. Requires building-wide coolant distribution infrastructure (CDU systems) and additional capex per rack but achieves high cooling efficiency.
• Rear-door heat exchangers: Air cooling supplemented by liquid-cooled heat exchanger panels on the back of each rack. Lower capex per rack than direct-to-chip but caps the achievable power density at roughly 60-80 kW.
• Immersion cooling: Servers submerged in dielectric fluid. Highest cooling efficiency but requires entirely different rack and server form factors. Niche today but growing.

The PUE (Power Usage Effectiveness) ratio is the standard benchmark for cooling efficiency. Industry-leading AI facilities target 1.15 to 1.25 PUE; older air-cooled facilities run 1.4 to 1.8. Each 0.1 PUE improvement at 100 MW saves roughly 8.7 million dollars per year at 7 cents per kWh, which is why cooling design is a primary underwriting input.

What Are the Tenant Credit and Concentration Risks?

The primary tenants for AI data centers are a small handful of hyperscale tech companies (AWS, Google, Microsoft, Meta, Oracle, plus a growing roster of AI-native companies). Concentration risk is structural to the asset class, not a deal-specific concern.

While most of these tenants are investment-grade, the concentration creates significant risk. Losing a single tenant that occupies a large block of power capacity can be catastrophic for the asset's cash flow. The mitigants in 2026 deals:

• Take-or-pay lease terms: The tenant is obligated to pay for contracted power capacity whether or not they consume it. Long-tail of contracted revenue even if the tenant's business model shifts.
• Long lease terms with cancellation penalties: 15+ year terms with substantial penalties for early termination de-risk the cash flow profile.
• Parent guarantees: Most hyperscale leases are guaranteed by the parent corporate entity, not the SPV subsidiary.
• Multi-tenant configurations: Some sponsors are intentionally avoiding single-tenant deals, accepting lower rents in exchange for diversification.

How Has Power Procurement Changed in 2026?

Power procurement for AI data centers in 2026 has moved from utility contracts as the default to a hybrid model where utility supply is supplemented by long-term PPAs, on-site generation, and storage.

Three shifts are reshaping the procurement landscape:

• Take-or-pay PPAs with renewable generators: Hyperscalers are signing 15- to 25-year PPAs directly with solar, wind, and increasingly storage developers to lock in power costs and meet sustainability commitments. The data center developer often passes through some or all of this PPA cost to the tenant.
• On-site generation: Behind-the-meter natural gas peakers, battery storage, and rooftop or adjacent solar are increasingly being incorporated into project designs. The economic and reliability case is strongest in markets where interconnection queues are blocking incremental utility power.
• Nuclear and SMR partnerships: Several hyperscalers signed nuclear or small modular reactor partnership agreements in 2024-2025 (Amazon-Talen, Microsoft-Constellation, Google-Kairos). These are still pre-operational but reshape the long-term power assumption set.

For underwriting purposes, the implication is that the model needs to handle multiple power sources, multiple contract types, and complex pass-through structures, not a single utility line item.

What Are the Common AI Data Center Underwriting Mistakes?

The damaging modeling errors involve using real estate templates instead of infrastructure templates, understating cooling capex, ignoring interconnection queue risk, missing PPA pass-through structure, and overlooking the obsolescence reserve.

1. Modeling revenue per square foot. AI data center revenue is contracted by power capacity, not floor area. A facility leased at 100 dollars per kW per month produces vastly different cash flow than the same square footage leased at 25 dollars per PSF triple net. Real estate templates produce a number; the number is meaningless.
2. Understating cooling capex. 100 kW per rack designs require day-one capex that is 1.5 to 3 million dollars per MW higher than 30 kW designs. Models that assume "modern data center" cooling without specifying the kW density routinely come in 15 to 25 percent under-budget by completion.
3. Ignoring interconnection queue risk. Power commitments without executed interconnection agreements are not power commitments. Models that assume Day-1 energization at signing of an LOI routinely miss 18 to 36 months of carry cost.
4. Missing PPA pass-through structure. Whether the tenant or the landlord bears the cost of grid power versus PPA power, with what escalation, and with what reconciliation mechanism, can shift project IRRs by 200 to 500 basis points.
5. Overlooking obsolescence reserve. AI hardware life cycles have compressed to 3 to 5 years. Building infrastructure life cycles are 15 to 25 years. Modeling without a reserve for mid-life rack-density upgrades or cooling-architecture changes overstates terminal value.
6. Treating concentration risk as a market risk rather than a structural feature. Single-tenant or two-tenant data centers are not portfolios; they are bilateral contracts with physical assets attached. Underwriting them on a real estate cap-rate basis misses the credit-equivalent exposure.

How Should You Structure Returns on Data Center Investments?

Data center investment returns should be evaluated against infrastructure benchmarks (long-duration, low-vol, high-leverage capacity) rather than real estate benchmarks, with attention to the take-out structure (sale to REIT, ABS securitization, infrastructure fund exit) that determines the exit multiple.

Target return ranges by strategy in 2026:

• Core/Stabilized (long-term lease to investment-grade tenant): 8 to 11 percent levered IRR, supported by ABS or 144A take-out
• Development (greenfield, power-secured): 14 to 19 percent levered IRR, exit to REIT or infrastructure fund
• Power arbitrage (acquire site with power, develop later): 18 to 25+ percent IRR depending on power-cost appreciation and queue position evolution
• Opportunistic (distressed or obsolete facility repositioning): 25+ percent IRR with substantial execution risk

The Blackstone data center REIT launch and the maturation of the ABS market for stabilized digital infrastructure cash flows have meaningfully compressed exit cap rates for tier-1 assets while leaving secondary-market assets and assets without power commitments trading at much wider spreads. The exit assumption is now as important as the underwriting assumption.

The Bottom Line: This Is an Infrastructure Play

Investing in AI data centers is less like buying real estate and more like investing in a private power plant or a toll road. The barriers to entry are immense, the capital requirements are staggering, and the technical expertise required is profound.

The opportunity is undeniable, but the complexity is non-negotiable. You cannot afford to analyze these assets with a generic real estate pro forma. Your financial model must be specifically engineered to handle power-based revenue, complex pass-through structures, multi-source power procurement, and heavy ongoing capital expenditure with obsolescence reserves.

Frequently Asked Questions