Data Centers: Power Pricing, PUE, and the 2024-2026 Hyperscale Re-Pricing

Key Takeaways

The 2024-2026 AI Demand Re-Pricing

A data center is a power lease with a building wrapped around it. The cap rate follows the credit of the tenant and the stability of the power; the lease structure follows the way the kilowatt is delivered; the capex stack follows the electrical and cooling spec, not the square footage. The institutional reader who underwrites the asset class as if it were specialty industrial will lose every deal in 2026 to the underwriter who underwrites it as the digital-infrastructure asset it actually is — sitting at the intersection of commercial real estate and wholesale power, and repricing faster than any other institutional asset class in the cycle.

The 2024-2026 inflection is structural, not cyclical. US data-center capacity sat at roughly 12 GW in 2020. By the end of 2025 the operational base had reached approximately 25 GW. Goldman Sachs Research forecasts global data-center power demand to increase 165% by 2030 versus 2023, reaching 1,350 TWh, with the US specifically reaching 95 GW of installed capacity (+197% from 2025) and consuming 11% of total US electricity by 2030 versus 6% in 2025. Dell'Oro tracks global data-center capex to $1.7 trillion through 2030. The AFCOM 2025 State of the Data Center survey reports 6× construction increase planned over the next three years and average rack densities up 70% year over year (16 to 27 kW). The hyperscaler Big Five — Amazon, Google, Meta, Microsoft, Oracle — have committed roughly $1 trillion of capex over 2024-2026 per JLL's 2026 Global Data Center Outlook, against 41 GW of planned owner-occupied delivery and 62 GW of leased capacity through 2030.

The institutional consequence is that data centers are no longer a specialty industrial sub-segment. They are the dominant institutional infrastructure asset class. Blackstone bought QTS for $10B in 2022. KKR and Global Infrastructure Partners took CyrusOne private at roughly $15B in the same vintage. Brookfield and Ontario Teachers' bought Compass Datacenters for $5.5B in 2023, securitized six hyperscale facilities for $830M in 2026, and announced a multi-billion-dollar KKR follow-on. DigitalBridge crossed $119B of AUM by March 2026 and announced a $4B sale to SoftBank closing in H2 2026. Stonepeak launched the Montera Infrastructure platform in 2025 with $1.5B of equity, its fourth North American data-center investment after CoreSite and Cologix. Blackstone's BXDC IPO in 2026 underwrote hyperscaler credit at a 5.75-7.0% gross asset yield. Every major infrastructure capital pool now has a staked platform. This is not a sector being entered; it is a sector being consolidated.

What follows is the institutional underwriter's read of the asset class in 2026: the capex stack at the 2026 benchmark, the ISO-by-ISO power-cost stack, the behind-the-meter and nuclear-PPA architecture that explains site selection, the PUE benchmark table translated into operating-cost math, the hyperscale vs colocation vs retail framework with cap rates for each segment, the institutional ownership map, the interconnection queue as the binding constraint, and a 50 MW Atlanta hyperscale build-to-suit (BTS) worked end to end with the cap-rate, debt-structure, and PPA-renegotiation sensitivities the institutional analyst needs to land on a recommended price. The article does not chase volume keywords; it serves the cohort the rest of the open web is not serving — the REPE associate, the institutional-platform analyst, the debt broker, and the infrastructure-fund principal who has been assigned a data-center deal and needs the asset class read at full institutional depth.

The Capex Stack — $11-20M/MW in 2026

Underwriting starts with the basis. The 2026 institutional benchmark for a hyperscale shell-and- core build — standard cloud-compute load, 1.20 PUE design, air-cooled or hybrid air-evaporative cooling — runs $11-14M per IT MW per JLL and Cushman & Wakefield's data-center pricing outlook. AI-optimized facilities — liquid cooling (single-phase or two-phase immersion, rear-door heat exchangers, or direct-to-chip), higher-voltage distribution to handle 120-240 kW racks, and densified mechanical and electrical plant — run $15-20M/MW. The capex spread between a standard shell-and-core and a GPU-dense AI training campus is the difference between a $700M and a $1.0B project on a 50 MW IT load.

The decomposition by line is the institutional analyst's first table. Electrical systems are typically 40-45% of the budget — medium-voltage switchgear, paralleling switchgear, UPS modules and batteries, standby generators with N+1 or 2N redundancy, transformers, and the distribution to the white space. Cooling and mechanical run 15-25% — CRACs, CRAH, chillers, air-handling units, evaporative cooling towers, plus rising liquid-cooling spend. Structural shell is 10-15% — the building itself, including raised floor, cable tray, and security shell. Site work, land development, and utility-feeder construction take 8-12%. Fit-out (white space build, cable management, fire suppression, BMS) accounts for the balance, with redundancy upgrades to Tier IV (concurrent maintainability plus fault tolerance) adding roughly 40% to the electrical and mechanical line items versus Tier III. The Uptime Institute 2025 Global Data Center Survey reports that hyperscale facilities increasingly target Tier III concurrent maintainability with application-layer redundancy rather than Tier IV at the facility layer, on the view that software- level distributed redundancy is cheaper than electrical-mechanical fault tolerance.

Table 1. The 2026 hyperscale capex stack. JLL 2026 outlook puts the standard shell-and-core average at +6% versus 2025. AI-optimized GPU campuses run 30-50% above shell-and-core on the electrical and cooling lines. Tier IV adds roughly 40% to electrical and mechanical. Land is reported separately and varies more by market than by MW.

Land is reported separately because it scales with site selection and utility access, not with IT load. The 50 MW Atlanta BTS in our worked example uses a $30M land basis on approximately 50-75 acres adjacent to a Georgia Power transmission substation with confirmed 60 MW utility delivery. Northern Virginia hyperscale land is materially more expensive on a per-acre basis but is gated by grid availability rather than by land cost. The institutional discipline is to underwrite the deal on $/MW basis (all-in development cost divided by IT MW) and report land separately as a discrete risk concentration tied to entitlement, environmental, and utility-availability execution.

Power as the Deal — the ISO-by-ISO Stack

A data center is the largest concentrated electrical load any institutional CRE underwriter will ever model. A 50 MW IT load at 1.20 PUE consumes 60 MW continuously — the equivalent of 50,000 American homes — with a load factor approaching 100%. At wholesale power rates of $50-90/MWh, the annual energy spend is $25-50M. Under triple-net hyperscale leases the tenant pays that bill directly to the utility or under the PPA architecture; under turnkey or wholesale colo structures the landlord buys the energy and resells it at a markup. Either way, the wholesale power-cost stack determines site selection, not submarket office or industrial economics. JLL summarizes the institutional discipline in one line in their 2026 outlook: "speed to power is the primary site-selection criterion."

The six ISOs / RTOs that matter to the institutional underwriter, with the wholesale baseline, capacity-payment overlay, interconnection-queue posture, and the resulting 2026 site-selection implication for each.

Table 2. The ISO-by-ISO wholesale power stack and the 2026 site-selection implication. PJM is the premium / supply-constrained market; ERCOT is volume; MISO/SPP are cost; NYISO/CAISO are edge. The Southeast vertically integrated markets (Georgia Power, Duke, TVA) have become the institutional spillover for hyperscale BTS as Northern Virginia tightens. EIA data via the Today in Energy ERCOT analysis.

Three market-level details deserve specific institutional attention. First, the Northern Virginia pricing premium. Ashburn wholesale rates hit $215/kW/month at Q2 2025 per Cushman & Wakefield — a record high — on +30.1% year-over-year lease-rate growth. CBRE reports Northern Virginia +1,102 MW of net absorption alone in H2 2025, with primary-market vacancy across the US at a record 1.4%. The premium is not a temporary pricing anomaly; it is the value of grid-connected MW in the one corridor with the deepest fiber backbone and the largest hyperscaler footprint, and it is pricing through the supply-constraint per JLL.

Second, the ERCOT institutional pivot. The EIA forecasts ERCOT-North wholesale prices to rise +45% in 2026 alone and +79% by 2027 under the high-demand scenario, with data centers driving 46% of ERCOT's projected 87 to 145 GW load growth between 2025 and 2031. Texas Senate Bill 6 (signed June 2025) reallocates a larger share of grid-connection cost to large customers — data centers and crypto miners — which materially changes the pro-forma OpEx assumption for ERCOT-sited deals underwritten before mid-2025. Institutional underwriters reading any Texas deal need the SB 6 cost- allocation in the model.

Third, the Southeast spillover. Atlanta, Charlotte, Nashville, and the Carolinas have absorbed the institutional hyperscale demand that cannot get a Northern Virginia interconnection in 18 months. Georgia Power's substation pipeline, the Tennessee Valley Authority's industrial-load programs, and Duke Energy's expansion in the Carolinas all support 30-50 MW deliveries on 18-30 month timelines. The capex math is similar; the wholesale rate is 25-35% below NoVa; the cap-rate spread is 50-100 basis points wider than Ashburn for now, with consensus that the spread compresses as Northern Virginia stops adding deliverable MW. The worked example in this article sits in this market.

Behind-the-Meter, Front-of-Meter, and Nuclear PPAs

The wholesale rate is only half of the power story. The other half is the contractual architecture that delivers it. Three structures dominate institutional underwriting in 2026, and the nuclear-PPA case studies of 2024-2026 are the defining freshness wedge of the cycle.

Behind-the-meter (BTM). A co-located load at the generation source, electrically connected before the utility meter, bypassing the transmission grid. The original Talen Energy / Amazon Web Services architecture at the Susquehanna nuclear plant in Pennsylvania was a BTM arrangement — up to 960 MW (later increased to 1,920 MW) of nuclear baseload delivered to a co-located AWS data center without flowing through PJM's transmission system. The economic appeal to the hyperscaler is straightforward: avoid the 4-7 year PJM queue, lock in clean baseload from a 24/7 generator with high capacity factor, eliminate transmission losses (4-7% line loss on the PJM system), and isolate the load from PJM capacity-market pricing volatility.

The institutional issue with BTM is the cost-allocation precedent. FERC rejected the Talen-Amazon Interconnection Service Agreement (ISA) in November 2024 on the view that the BTM structure shifted grid-cost burden to the broader PJM ratebase — existing utility customers would effectively subsidize the hyperscaler's grid-connection avoidance. The Talen-Amazon deal was restructured in June 2025 as a front-of-meter retail arrangement (grid-connected, retail electricity sale at a fixed price, no FERC ISA approval required), with transmission reconfigurations completing in Spring 2026. The deal terms are largely preserved — 17 years, up to 1,920 MW, $18B in total contract value — but the architecture is now standard retail rather than BTM. The Utility Dive coverage walks the FERC decision and the restructure.

Front-of-meter (FOM) retail PPA. The new default. The Microsoft / Constellation Energy deal at the Three Mile Island Unit 1 (renamed the Crane Clean Energy Center) is the archetype: a 20-year PPA, 835 MW of nuclear baseload, restart targeted for 2028 pending NRC approval (review tracking to 2027), 100% of output contracted to Microsoft to match Microsoft's PJM-region data-center load. Constellation explicitly characterized the arrangement as a non-co- located retail PPA — the power flows through the PJM grid like any other retail electricity sale — which keeps it outside the FERC ISA framework that complicated Susquehanna. The deal is reportedly priced at a meaningful premium to the PJM wholesale baseline (multiple energy-press estimates put the all-in price around $100/MWh) but the premium is the cost of 20-year baseload certainty against PJM volatility.

Standard utility retail and on-site generation. The bulk of US data-center load still runs on standard utility retail — Dominion Energy's data-center service rate in Virginia, Georgia Power's industrial tariff, the ERCOT retail electricity provider (REP) market in Texas. On-site generation — primarily natural-gas turbines or reciprocating engines, with standby diesel for emergency — is increasingly used as bridge power while utility interconnection waits and, per JLL's 2026 outlook, is becoming a permanent fixture in supply- constrained corridors. Crusoe, CoreWeave, and other GPU-cloud operators are building gas-fired capacity onsite as a load-following resource that can be permitted faster than utility transmission upgrades.

PUE and the OpEx Pass-Through Math

Power Usage Effectiveness (PUE) is total facility energy divided by IT load energy. PUE 1.0 is the theoretical limit — every watt drawn from the grid goes to compute. Real facilities consume additional energy for cooling, lighting, security, distribution losses, and the energy required to operate the UPS systems themselves. The Uptime Institute 2025 Global Data Center Survey reports an industry weighted-average PUE of 1.54, stable for six consecutive years — meaning the average facility draws 54 watts of non-IT energy per 100 watts of IT load.

Table 3. PUE benchmark by facility class. Uptime Institute publishes the industry weighted average; AFCOM reports liquid-cooling adoption nearly doubling year over year (19% to 36% per the 2025 State of the Data Center). The hyperscale-versus-industry gap is the institutional opportunity — and the underlying engineering question that determines who eats the marginal power cost.

Translate PUE to the institutional OpEx math. A 50 MW IT-load facility operating at 1.20 PUE consumes 60 MW total. The same IT load at 1.54 PUE consumes 77 MW — 17 MW of incremental non-IT draw, every hour, every day, every year. At a blended wholesale rate of $70/MWh and 8,760 hours, the PUE gap costs $10.4M per year. Across a 15-year hyperscale lease the cumulative spread is $156M nominal. Under a hyperscale triple-net lease, the tenant pays the energy bill directly to the utility or under the PPA architecture — the landlord is indifferent to PUE on a pure cash-flow basis (the building's NOI does not vary with the cooling efficiency, only the tenant's operating cost does). Under a turnkey colocation structure, the landlord buys the energy and resells the white-space capacity at a per-kW-per-month rate — the PUE gap is the landlord's margin loss. Under a wholesale colo / power-included hybrid, the spread is split per the PPA pass- through and metering arrangement.

The institutional implication is that PUE is not an engineering trivia metric for the underwriter. It is a determinant of who bears the power-cost risk and how the lease rate is structured. The AFCOM 2025 report flags liquid-cooling adoption rising from 19% to 36% year over year as rack densities push past 27 kW average and GPU racks hit 120-240 kW; at those densities, air cooling cannot remove the heat and the PUE penalty for staying on air becomes severe. The cooling-spec capex decision — spend $3-5M/MW extra to put in liquid cooling and design to 1.10 PUE versus stay air-cooled at 1.30 PUE — is a capex-versus-pass-through decision that institutional underwriters now model line by line for AI-optimized GPU builds.

Hyperscale vs Colocation vs Retail

The institutional taxonomy of data-center assets is three segments, distinguished by size, tenant universe, lease structure, term, rate, and cap rate. The segment framework is the article's most referenced table because the rest of the asset class collapses out of it — debt-market access, securitization viability, cap-rate stack, exit-buyer pool, and operational complexity all sort by segment.

Table 4. The three-segment institutional framework. Hyperscale dominates new institutional deal flow on size, tenant credit, and cap-rate compression. Wholesale colo is the long-duration institutional book (Equinix, Digital Realty, CoreSite). Powered shell trades wider because the tenant takes the fit-out risk. Retail trades wider still on tenant churn and operational intensity. Cap rates per JLL 2026 outlook and Cushman & Wakefield pricing data; the BXDC 2026 IPO underwrites the hyperscale band at 5.75-7.0% gross asset yield.

Hyperscale (>50 MW, BTS or powered shell). A single tenant, typically Big Five (Amazon, Google, Microsoft, Meta, Oracle) plus the IG enterprise contingent (Apple, Tesla, Nvidia direct deployment, increasingly sovereign and federal). Lease structure is triple-net or absolute net, 15-year initial term with two or three 5-year extensions at FMV or fixed strike, escalators fixed (2-3% annual) or CPI-linked with floor and cap. Rate structure is per-kW-per-month wholesale, typically $150-220/kW/month in 2026 Atlanta hyperscale benchmark, with NoVa premium pricing the upper end and Tier II Sun Belt markets the lower end. Deal sizes $200-300M+ per facility. Cap rate band 5.5-6.5% on stabilized hyperscale per JLL Q1 2026 and the BXDC IPO underwriting. The CMBS market now accepts hyperscale BTS as collateral at securitization-grade cap rates per the Brookfield- Compass $830M 2026 ABS securitization (six hyperscale facilities).

Wholesale colocation (1-50 MW, turnkey or wholesale). Multi-tenant. The Equinix / Digital Realty / CoreSite / Cologix institutional book. Lease structure is turnkey (landlord delivers white space ready for tenant racks; tenant pays per-kW-per-month wholesale rate) or wholesale (tenant takes a private suite or cage at landlord rate). Term 5-10 years with options. Credit mix is enterprise, SaaS, financial services, federal — broader than hyperscale but durable. Rates $150-300/kW/month at the wholesale tier. Cap rate band 6.0-7.0% stabilized per JLL and CBRE Q1 2026. Interconnection density — the cross-connect ecosystem inside the facility — is the operational moat; Equinix's IBX network commands a margin premium on cross-connect revenue that translates to cap-rate compression on the trophy assets.

Powered shell. The asset is the electrical and mechanical capacity plus the structural shell, delivered for a hyperscaler or wholesale colo operator to fit out the white space. Landlord delivers the kW of utility capacity, the structural building, and the basic distribution; tenant builds out the IT racks, CRAH, UPS top-up, and white-space fit. Lease term often longer (10-20 years) to amortize the shell capex against a longer hold; rate structure quoted on a per- kW-shell basis at roughly 35-45% of the corresponding turnkey rate. Cap rate 7.0-8.0% because the landlord has not delivered the full fit and the tenant bears more capex risk.

Retail / cabinet (<1 MW, cabinet-by-cabinet). The legacy CenturyLink / Lumen portfolio, the small-MSA single-facility operators, the network-operator hosting facilities. Lease structure cabinet-by-cabinet at $1,500-2,500/cabinet/month, term 1-3 years often with month-to- month structure, tenant churn 10-15% annual. Cap rate 7-8% reflecting operational intensity, churn, and broader tenant credit dispersion. Not the institutional core for new deal flow but a real portfolio book at the trade-press tier and an opportunistic acquisition target for the platforms that can roll it up.

The 2026 Institutional Ownership Map

Name the names. The 2026 institutional landscape sorts into three pools: public REITs, PE / infrastructure- fund-owned platforms, and the GP / sponsor universe that surrounds the platforms.

Public REITs. Digital Realty (DLR) — the largest publicly traded data-center REIT by network footprint, 300+ data centers across 50 metros globally, 2026 core FFO guidance $7.90-8.00 per share, projected revenue $6.6B, market cap approximately $68B (May 2026). Per the DLR 10-K filings the portfolio runs across hyperscale, wholesale colo, and interconnection-dense IBX-equivalent assets. Equinix (EQIX) — the interconnection-density leader, 270+ data centers across 77 markets and 6 continents, 2026 AFFO guidance $4.20-4.28B (+9-11% YoY), market cap approximately $103B and enterprise value $124B (May 2026), the EQIX 10-K walks the interconnection-density model. Iron Mountain (IRM) — the hybrid records-management plus data-center REIT, smaller exposure but worth the cross-reference as a third institutional public option. The combined public-REIT market cap in the cluster sits at roughly $200B+.

PE / infrastructure-fund-owned platforms. QTS Realty (Blackstone, $10B 2022 close at a 21% premium); Switch (DigitalBridge and IFM, $11B 2022 close); Compass Datacenters (Brookfield Infrastructure and Ontario Teachers' Pension Plan, $5.5B 2023 close, $830M ABS securitization in 2026, KKR follow-on multi-billion-dollar investment announced); CyrusOne (KKR and Global Infrastructure Partners, $15B-equivalent 2022 take-private); Cologix (Stonepeak); CoreSite (Stonepeak); Montera Infrastructure (Stonepeak, $1.5B equity 2025 vintage, BTS 100+ MW projects in Tier I/II markets). Combined enterprise value of the PE-owned platform book sits in the $45-55B range. Blackstone's BXDC IPO in 2026 underwriting hyperscaler credit at a 5.75-7.0% gross asset yield is the institutional cap-rate benchmark.

Infrastructure-fund AUM. DigitalBridge at $119B AUM (March 2026); to be acquired by SoftBank for approximately $4B (announced December 2025, closing H2 2026); DBP III closed at $11.7B in commitments. Blackstone Infrastructure Partners, Brookfield Infrastructure, KKR Global Infrastructure, Stonepeak — each with a data-center sleeve in the multi-billion-dollar range. Aggregate digital-infrastructure-targeting AUM at the major platforms exceeds $300B. The institutional thesis is consensus and the capital is staked; the institutional underwriter's job is now to underwrite specific deals against the cap-rate stack the platforms have set, not to defend the asset class as an allocation.

NAREIT publishes the public-REIT sector data; data-center REITs returned -14% total return in 2025 (the worst-performing REIT sector) despite operational FFO +21.3% year over year and NOI +7.2% year over year, a public-versus-private dislocation walked in detail in the closing section of this article.

The Interconnection Queue Is the Binding Constraint

Land selection is now grid selection. The institutional underwriter who treats site selection as a real-estate exercise — submarket vacancy, comp lease rates, demographics — will lose every deal in 2026 to the underwriter who treats it as a grid-availability exercise.

The defining data point is Northern Virginia. Dominion Energy is managing approximately 70 GW of large-load interconnection requests in its service territory, with another 4 GW of expected data- center load by 2028, against substation capacity of approximately 2.1 GW in the constrained Loudoun County corridor. The Utility Dive coverage of PJM's data-center problem walks the systemic queue across the broader PJM footprint: 5-7 year wait, 30 GW of incremental data-center requests, 32 GW of expected total load growth to 2030. The Dominion interconnection requirements PDF is the operating document; institutional underwriters should read it the way they read a lease abstract.

The ERCOT load forecast is the same story in a different ISO. Texas projects 87 GW of system load in 2025 growing to 145 GW by 2031, with data centers accounting for 46% of the incremental load. EIA forecasts ERCOT-North wholesale prices +45% in 2026 and +79% by 2027 under the high-demand scenario. The MISO and SPP footprints have lighter load growth and shorter queues but smaller fiber backbones, which constrains hyperscale tenant interest.

The geographic implication: the 2026 institutional site-selection map is migrating out of Northern Virginia even as Virginia's existing inventory leads the world. CBRE reports that Southern California, Austin-San Antonio, and Central Washington each surpassed New York Tri-State in 2025 inventory. Atlanta, Charlotte, Nashville, Phoenix, Reno, Columbus, San Antonio, and Indianapolis are absorbing institutional hyperscale deal flow. The 50 MW Atlanta worked example in this article sits in the Southeast spillover — Georgia Power confirmed 60 MW delivery to a transmission substation, 24-month construction timeline, no PJM queue dependency.

Worked Example — 50 MW Atlanta Hyperscale BTS

The deal. A 50 MW IT-load hyperscale build-to-suit in the Atlanta MSA, anchor tenant a Big Five hyperscaler (BBB+ / A-rated parent corporation, system-level guarantee), BTS structure with the landlord developing the facility to tenant-spec, 15-year initial lease term with two 5-year extension options at FMV, triple-net structure with tenant pass-through of utilities and property taxes. The site is approximately 50-75 acres adjacent to a Georgia Power transmission substation with confirmed 60 MW utility delivery, 24-month construction timeline, projected 1.20 PUE design using free-cooling and evaporative-cooling hybrid. The lease rate is $1,800/kW/month wholesale (the 2026 Atlanta hyperscale benchmark).

Capex. $14M/MW basis on the $11-14M shell-and-core band (Atlanta sits at the higher end of the shell-and-core range because of tier III spec and 60 MW utility-feeder cost) on a 50 MW IT load = $700M total project cost. Decomposed: land $30M, vertical construction $580M (structural shell, electrical and mechanical, fit-out, soft costs, contingency), $90M electrical and cooling spec uplift for hyperscale-grade reliability and the 1.20 PUE design.

Construction debt. A 65% loan-to-cost construction facility at $455M, priced at SOFR plus 225 basis points, 36-month draw period (24-month construction plus 12-month tail for completion and stabilization), interest-only with capitalized interest, converting at certificate of occupancy to a 10-year fixed CMBS perm at 5.30% / 30-year amortization / 1.40× debt service coverage. Sponsor equity $245M plus the developer fee and carried interest. The construction-to-perm structure walked in the construction loans, draw schedules, and interest reserves article is the operating reference for sizing the interest reserve and the S-curve draw schedule on this deal.

Lease economics, Year 1 stabilized.

• Gross rent: 50,000 kW × $1,800 / month × 12 = $90.0M

• Tenant pays all utilities and property taxes (NNN pass-through), so no landlord opex line for energy or taxes

• Landlord ground-lease payment (if applicable) and asset-management fee: $5.0M

• Reserves for life-safety, capex, and structural maintenance: $5.0M

• NOI = $90.0M − $5.0M − $5.0M = $80.0M

Exit value at stabilization. $80.0M NOI capitalized at the 2026 Atlanta hyperscale cap rate of 6.0% (sitting between the NoVa hyperscale tightness of 5.5% and the rising-rate scenario at 6.5%) implies an exit value of $1.33B. The basis is $700M; the unlevered development value-creation is roughly $630M, or roughly 90% on cost. Carried over the 24-month construction plus the stabilization period, the unlevered IRR is in the high-teens; the levered IRR (65% LTC construction, 60% perm, sponsor promotes at IC-standard waterfalls) underwrites at 13-15% over the 15-year primary lease term per institutional standards.

Table 5. The 50 MW Atlanta hyperscale BTS worked example. Inputs disclosed; institutional reader can reproduce every dollar number on the page from the inputs shown. The cap-rate, PPA-reset, renewal, and PUE-drift sensitivities are walked in the next section.

Sensitivity — PPA Reset, Renewal, PUE Drift, Cap Rate

The institutional IC discussion turns on four sensitivities. Each is modeled explicitly because each has a material effect on the levered IRR and on the exit-value distribution at year 15.

1. PPA renegotiation at year 8 — the mid-lease power-cost reset. Hyperscale BTS leases at the Atlanta benchmark assume a stable power-cost regime under the tenant's PPA or utility-tariff arrangement. If the underlying PPA is a 10-year fixed-price contract that resets at year 8 (Constellation, Vistra, Talen, and similar PPAs frequently include mid-term reset mechanisms), the post-reset wholesale rate may step up materially — institutional underwriters typically model a 30-60% step-up in the reset year as the base case for ERCOT-sited deals (per EIA's +79% by 2027 high-demand forecast) and 15-25% for PJM and the Southeast. The lease is NNN so the tenant absorbs the step-up; the underwriting question is whether the tenant's BBB+ credit continues to support the higher all-in cost or whether the credit deteriorates against capacity- constrained power and the lease becomes a renegotiation conversation. The nuclear-PPA architecture described above — Microsoft-TMI, Talen-Amazon, Vistra long-dated contracts — is the institutional answer: lock in 17-20 year power-cost stability and remove the reset risk entirely. Deals where the tenant has signed a long-dated nuclear or natural-gas PPA price tighter than deals where the tenant is exposed to PJM or ERCOT wholesale variability.

2. Tenant renewal at year 15 — the BBB+ counterparty probability. The 15- year primary lease has two 5-year FMV extension options. Institutional underwriting standard for a Big Five hyperscaler with a long-dated infrastructure investment in a facility built to their spec is a 75-85% extension probability per institutional STNL underwriting consensus — the counterparty has the credit, the operating fit, and the relocation cost is substantial. Build the sensitivity at three nodes: (a) tenant renews at FMV with NOI continuing at 2.5% annual escalators through the option period (the base case, $1.33B exit holds); (b) tenant declines to renew and the asset re-tenants at a 12-month downtime plus a 10-15% market-rate haircut on the new tenant lease (the asset becomes a 12-15 year residual lease on a stabilized basis, exit cap widens 25-50 bps to 6.25-6.50%, exit value $1.20-1.25B); (c) the deep distress case where the asset cannot re-tenant on hyperscale terms and is repositioned to wholesale colo at $200-250/kW/month and 50% IT load utilization (exit cap 7-8%, exit value $0.7-0.9B). The institutional IC weight in 2026 is 75% (a), 15% (b), and 10% (c); the blended exit is roughly $1.25-1.30B.

3. PUE drift from 1.20 to 1.30. Hyperscale facilities designed to 1.20 PUE typically drift toward 1.25-1.30 PUE over a 15-year hold as compressor efficiency declines, cooling plant fouls, and IT load patterns evolve toward higher rack density (which is harder to cool at the same PUE). The capex to hold the 1.20 design through the primary lease term — periodic chiller replacements, evaporative-cooling overhauls, possible liquid-cooling retrofit if the tenant pushes rack density past 50 kW — runs $25-40M cumulatively over 15 years on a 50 MW facility, or roughly 4-6% of the original $700M basis. Under a triple-net lease the question is whether this capex is reimbursable (typically yes under hyperscale BTS leases with PUE-target covenants) or landlord-borne. The underwriting practice is to assume 60-70% reimbursement and reserve the balance against the equity return. The IRR sensitivity is roughly −50 to −100 bps if the entire cost is landlord-borne; minimal if fully reimbursable.

4. Cap-rate movement at exit. Three scenarios. (a) Northern Virginia pull. As NoVa tightens further (1.4% vacancy at year-end 2025, supply +36% but absorption +38% YoY per CBRE), the cap-rate compression migrates to spillover markets — Atlanta hyperscale could trade through 5.5% by year 15 if the Southeast becomes a recognized institutional sub-market for hyperscale trophy assets. Exit value at 5.5% on $80M NOI = $1.45B. (b) Base case. Atlanta hyperscale holds the 6.0% benchmark. Exit value = $1.33B. (c) Rising-rate / risk-off scenario. Atlanta hyperscale widens to 6.5% on a generalized cap-rate move. Exit value = $1.23B. The 100 basis points spread on cap rate translates to roughly $220M of exit value — the institutional sensitivity that matters most to the equity return on a 15-year hold.

The cap rate calculator and formula article walks the mechanical sensitivity; the institutional reading here is that the four risks above are correlated — a deteriorating power-cost regime correlates with cap-rate widening, weakens tenant renewal probability, and increases PUE drift through higher cooling-load demand. The institutional underwriter models them jointly, not as four independent univariate sensitivities.

AI Demand and the Public-Market Dislocation

2025 ended with a paradox. Data-center REITs — Digital Realty, Equinix, Iron Mountain — delivered a -14% total return for the year, the worst-performing REIT sector per NAREIT, even as the same operating companies reported FFO +21.3% year over year and NOI +7.2% year over year on a sector basis. The public-market thesis fled the sector on technology-multiple compression following DeepSeek's early-2025 reasoning-model release and a broader narrative that hyperscaler capex would pause as AI training requirements rationalized. Q4 2025 active-fund holdings rebounded to 134% of benchmark index weight per NAREIT, signaling the public-market correction is already reversing.

Private-market capital underwrote through the noise. Blackstone closed the BXDC IPO at hyperscaler credit yields of 5.75-7.0%. Brookfield-Compass securitized $830M of ABS against six hyperscale facilities. Stonepeak launched Montera with $1.5B of fresh equity. SoftBank announced the $4B DigitalBridge acquisition. KKR followed on into Compass. The institutional read of this bifurcation is that the public-market dislocation was a multiple correction in a sector whose operational and structural tailwinds are intact, not a structural break in the asset class. The operating fundamentals — CBRE's 1.4% primary-market vacancy, +30% YoY Northern Virginia lease rates, AFCOM's +70% YoY rack densities, Goldman's +165% global power-demand forecast, Dell'Oro's $1.7T capex forecast through 2030, JLL's $1T hyperscaler capex through 2026 — have continued to compound. Private-market cap rates have compressed accordingly.

For the institutional CRE allocator the implication is that data centers are the asset class where the operational tailwind is most decoupled from public-market sentiment and where the structural demand curve is least cyclical. Every other major real-estate asset class is competing for capital allocation against a sector where the demand growth is measured in 100-200% over five years and the binding constraint is grid availability, not consumer or business demand. The institutional discipline is to underwrite the specific deal against the cap-rate stack the platforms have set, on the power-cost architecture the tenant has contracted, with the PUE target the engineering can deliver, and the cap-rate sensitivity the exit-buyer pool will price.

Six Mistakes Practitioners Make

• 1. Underwriting site selection as a real-estate exercise. The institutional site-selection criterion in 2026 is grid headroom and interconnection-queue position, not submarket lease comps or demographics. Northern Virginia is the most expensive corridor in the US because it has the deepest grid and fiber backbone — not the cheapest land. Atlanta, Charlotte, and the Carolinas are absorbing institutional flow not because they have the best submarket office or industrial economics but because Georgia Power, Duke, and TVA can deliver large MW on faster timelines than Dominion. Land cost is roughly 4-5% of the project; the utility-feeder and grid-availability access is the deal.

• 2. Modeling PUE as a flat operating assumption. PUE drifts over a 15-year hold as compressors age, cooling plant fouls, and rack density rises. Underwriting at the commissioning PUE (1.20) without modeling the drift to 1.25-1.30 over the hold understates OpEx by 5-10% in years 10-15. The right discipline is to model the design-PUE in years 1-3, a slow drift in years 4-10, and accelerated drift in years 11-15 absent capex intervention — and to budget the maintenance capex against the lease's pass-through provisions.

• 3. Assuming the lease is the entire credit story. The lease guarantee from Amazon, Microsoft, Google, Meta, or Oracle is investment-grade and the lease document is institutional standard. But the underlying economic stability is in the PPA or utility tariff the tenant has signed against the facility — that is what determines whether the tenant's operating cost stays stable at year 8 and year 12. A 17-20 year nuclear PPA underneath the lease is materially more valuable to the credit story than a rolling 3-5 year wholesale exposure. Read the PPA architecture; price it.

• 4. Mis-sizing capex for AI-optimized builds. Standard shell-and-core at $11- 14M/MW does not buy you a GPU-dense AI training campus. The AI-optimized build at $15-20M/MW includes liquid cooling, higher-voltage distribution, denser mechanical and electrical plant. If the tenant pushes rack density past 50 kW and the lease economics require liquid cooling, budgeting at the shell-and-core basis understates the capex by $50-300M on a 50 MW facility. Confirm the rack-density target and the cooling architecture before underwriting the basis.

• 5. Ignoring Texas SB 6 and the regulatory pivot. Texas Senate Bill 6 (June 2025) reallocates a larger share of grid-connection cost to large customers — data centers and crypto miners pay more of the substation and transmission upgrade cost than under the prior regime. ERCOT deals underwritten before mid-2025 typically did not assume the SB 6 cost-allocation in the model; underwriters reading a Texas deal in 2026 need to model the new allocation explicitly. Equivalent regulatory pivots are being discussed in PJM, MISO, and SPP — assume the cost-allocation moves toward the data center in 2026-2030 across the ISO footprint.

• 6. Treating the public-market dislocation as a signal about the asset class. Data-center REITs returned -14% in 2025 even as the underlying operating businesses delivered FFO +21% and NOI +7%. The public-market move was a multiple correction in a tech-correlated sector, not a structural break in CRE fundamentals. Q4 2025 active-fund holdings rebounded to 134% of index weight. Private-market platforms underwrote through the noise at 5.75-7.0% hyperscaler gross asset yield. The institutional discipline is to underwrite specific deals against the private-market cap-rate stack, not against the public-market REIT total return.

From Hyperscale BTS to Apers

Data-center underwriting sits at the intersection of CRE and power — the institutional discipline lives in the model. The 50 MW Atlanta worked example walks every line an institutional analyst builds: the capex stack by basis component, the lease rate by kW-month, the NNN pass- through, the NOI bridge, the cap-rate exit, the construction debt and CMBS take-out, the four sensitivities the IC discussion turns on. Apers is the platform institutional CRE teams use to build and run those models — not as a one-off Excel build for each deal, but as a workflow that ingests the inputs the analyst gathers and produces the analysis the IC asks for in minutes rather than days.

Related Articles

• Industrial underwriting: clear heights, dock doors, and the 2026 supply check — the cluster anchor. Data centers diverge from the rest of industrial on every dimension; the clear-heights and dock-doors playbook is the spec-driven sibling to the power-and-PUE playbook here.

• Logistics and distribution — the e-commerce sensitivity overlay — the demand-side companion. Both data centers and e-commerce logistics are sectors where 2024- 2026 demand has structurally re-priced the asset class; both reward institutional underwriters who model the demand curve rather than backward-looking comps.

• Cold storage — specialized modeling and tenant credit — the sibling power-intensive specialty industrial article. Both cold storage and data centers are power-intensity-dominated specialty industrial, both bifurcate the underwriting on tenant credit, both require operational diligence the generalist underwriter does not bring.

• Medical office buildings — single-tenant credit vs multi-tenant WALT — the credit-tenant lease underwriting analog. Single-tenant hyperscale BTS and single-tenant credit MOB share the institutional framework: the lease document is the asset, the credit guarantor sets the cap rate, and the institutional buyer pool prices the deviations from institutional standard.

• Cap rate calculator and formula — the mechanical reference for the exit-value calculation in the worked example. The 6.0% cap applied to $80M NOI to land at $1.33B is the formula; the institutional discipline is the cap-rate band selection across hyperscale, colocation, and powered shell.

• Construction loans: draw schedules and interest reserves — the construction-to-perm debt structure underlying the worked example. The 65% LTC construction facility at SOFR+225 with a 36-month draw, capitalized interest, and CMBS conversion is the institutional pattern walked in that article.

FAQ