Strategic summary

Artificial intelligence is transforming the data center from a specialized real-estate asset containing standardized IT equipment into a tightly integrated compute, electrical-power, cooling, communications, and industrial-infrastructure system. AFCOM’s State of the Data Center 2026 Study reports a remarkable increase in average rack density, from 16 kilowatts per rack in 2025 to 27 kilowatts in 2026, while the average facility represented in its survey grew from approximately 32 megawatts to 38 megawatts. Seventy-four percent of respondents plan to deploy AI-capable infrastructure, 72% expect AI workloads to increase capacity requirements, and 69% are actively taking steps to increase rack density.[1]

The strategic implication is that data-center value is migrating toward four scarce resources: deliverable electrical power, high-capacity cooling, suitable land with infrastructure access, and the organizational ability to finance, build, commission, secure, and operate increasingly complex facilities. The most valuable site is no longer necessarily the one closest to a major population center. It may be the site that can obtain several hundred megawatts of reliable power, connect to multiple fiber routes, secure water or water-minimizing cooling, obtain permits, and retain community acceptance.

The financial opportunity is correspondingly large. JLL projects that nearly 100 gigawatts of new data-center capacity could be added between 2026 and 2030, approximately doubling global capacity and requiring as much as $3 trillion of total investment. JLL divides that prospective investment into approximately $1.2 trillion of data-center real-estate asset creation and an additional $1 trillion to $2 trillion of tenant-funded IT equipment.[2] Its midyear 2026 North American analysis reports 25 gigawatts of absorption in the first half of the year, approximately 1% vacancy, 66 gigawatts under construction, and 95% of the construction pipeline already committed.[8]

Those figures describe an infrastructure supercycle, but not a risk-free one. Grid constraints, equipment shortages, rising construction costs, technological obsolescence, speculative power requests, uncertain AI utilization, tenant-credit differences, environmental regulation, and growing community resistance could separate durable assets from expensive stranded capacity. Investors should therefore value a data-center project according to verified power delivery, tenant quality, cooling compatibility, commissioning readiness, expansion rights, and community durability—not merely announced megawatts or projected AI demand.

The original AFCOM report, written by Bill Kleyman with statistical analysis by Kristin Letourneau, Ph.D., was produced in February 2026 and is available here: AFCOM, ⁠State of the Data Center 2026 Study⁠.[1]

From digital real estate to industrial infrastructure

For much of the cloud-computing era, the data-center investment thesis rested on familiar variables: land, fiber connectivity, utility service, building reliability, tenant credit, lease duration, and occupancy. AI does not eliminate those variables, but it changes their relative importance.

Traditional enterprise computing distributed modest power loads across large numbers of general-purpose servers. AI training and increasingly AI inference concentrate far more electrical and thermal load into accelerator-dense clusters. These systems depend on high-bandwidth networking, synchronized computing, substantial electrical distribution, and cooling capable of removing heat close to the processors.

The result is a change in the data center’s economic identity. It remains a real-estate asset, but it increasingly resembles a privately developed utility and industrial plant. Its major systems include substations, switchgear, transformers, generators, batteries, uninterruptible power supplies, pumps, heat exchangers, cooling towers or dry coolers, water-treatment systems, and sophisticated control platforms. These systems must be designed as an integrated architecture rather than as building services added after the computing requirement is established.

This transition explains why power availability can matter more than land cost. A low-cost site without firm power is not a functioning data-center asset. A site with power promised only after several years of utility upgrades may possess option value, but it should not be valued as commissioned capacity. Similarly, an announced 500-megawatt campus may represent a phased development opportunity rather than 500 megawatts of financeable, near-term supply.

Investors should distinguish among at least four measures:

* Announced capacity.
* Utility-approved or contracted capacity.
* Energized building capacity.
* Revenue-producing critical IT load.

Confusing these stages can overstate both industry supply and the value of an individual project.

Larger facilities and the economics of scale

AFCOM reports that the average facility size represented in its survey increased from approximately 32 megawatts to 38 megawatts in one year.[1] The increase reflects a broader shift toward campuses designed for phased growth and high-density computing.

Scale can create several financial advantages. Larger campuses may distribute land, security, network, administration, and some electrical costs across more revenue-producing capacity. They can standardize designs, negotiate equipment purchases, support dedicated substations or generation, and offer large tenants contiguous expansion capacity.

Scale also amplifies execution risk. A large campus can require extensive transmission upgrades, substations, gas supply, water infrastructure, and multiple construction phases. Its economics may depend on assumptions about future AI demand, tenant renewals, equipment density, electricity prices, and financing costs. The amount of capital at risk before stabilization can be enormous.

JLL estimates that global data-center capacity could reach approximately 200 gigawatts by 2030, supported by a 14% compound annual growth rate. It also estimates that construction costs have been rising at approximately 7% annually.[2] If construction costs continue increasing while interest rates, equipment lead times, and utility requirements remain uncertain, developers may face higher total development costs even when lease rates rise.

This creates an important distinction between nominal rent growth and risk-adjusted return. JLL reports that North American data-center rents have increased nearly 70% since 2020, averaging approximately 9% annual growth.[8] Yet the benefit to owners depends on whether rent increases exceed growth in energy, construction, financing, insurance, maintenance, and tenant-improvement costs.

The strongest projects are likely to combine contracted demand with phased capital deployment. Preleasing, take-or-pay structures, credit support, deposits, and milestone-based construction can reduce speculative exposure. Conversely, a project built around an unproven AI tenant, uncertain financing, or nonbinding capacity reservation may require a substantially higher risk premium.

Prefabrication moves into power and cooling

AFCOM finds that 37% of respondents expect to rely primarily on conventional construction, while 30% favor a hybrid model combining conventional construction with prefabricated components. Among organizations exploring modular construction, 77% are considering prefabricated power modules and 69% cooling modules.[1]

This is financially important because power and cooling equipment frequently sit on the critical path. Factory-built electrical rooms, skids, cooling modules, pump assemblies, and preintegrated control systems can reduce field labor, improve repeatability, and allow portions of construction to proceed concurrently.
 Prefabrication does not eliminate project risk. Modules must still be transported, installed, connected, tested, commissioned, and integrated with site-specific systems. A standardized module may also create vendor concentration or lock the operator into a design that becomes less competitive as accelerator platforms change.

Nevertheless, modularization can improve capital efficiency when combined with repeatable campus designs and disciplined configuration management. McKinsey estimates that smarter design, standardization, and construction methods could reduce data-center capital spending by 10% to 20% per facility and reduce projected global spending through 2030 by as much as $250 billion.[6]

For investors, the beneficiaries extend beyond traditional data-center landlords. The opportunity encompasses transformer and switchgear manufacturers, modular-power suppliers, cooling companies, engineering firms, commissioning specialists, electrical contractors, control-system vendors, and manufacturers of pumps, heat exchangers, busways, backup systems, and energy-storage equipment.

The corresponding risk is that many of these suppliers are being valued on unusually strong order backlogs. Investors must distinguish durable capacity constraints from temporary shortage pricing. New manufacturing capacity, design substitution, changing accelerator architectures, or slower data-center construction could reduce margins even if long-term demand remains strong.

Rack density changes the physical and financial model

AFCOM’s most dramatic finding is the increase in reported average rack density from 16 kilowatts in 2025 to 27 kilowatts in 2026. That represents a 69% year-over-year increase and nearly four times the seven-kilowatt average reported in 2021.[1]

A rack drawing 27 kilowatts continuously consumes approximately 236,500 kilowatt-hours per year before accounting for cooling and other facility overhead. At $0.07 per kilowatt-hour, its annual IT electricity cost would be approximately $16,600. At $0.12 per kilowatt-hour, the cost would approach $28,400. A one-megawatt continuous IT load consumes 8.76 million kilowatt-hours annually, making relatively small differences in electricity price, power-usage effectiveness, or utilization financially material.

Higher rack density does not necessarily increase total facility demand if computing efficiency improves and workloads are consolidated. In practice, however, AI demand is growing rapidly enough that improvements in performance per watt are being overwhelmed by the quantity of computing deployed.

AFCOM reports that 69% of respondents expect density to continue increasing and that 69% are already taking action. The leading responses include liquid cooling, airflow improvements, and better operational visibility.[1]

The investment implication is that not every existing data center is equally capable of being converted into an AI facility. A building designed for five- to ten-kilowatt racks may face limitations in utility capacity, floor loading, busway rating, cooling-water distribution, ceiling height, piping routes, heat rejection, backup generation, and physical space for new equipment.

Some legacy assets can be retrofitted economically. Others may remain useful for conventional cloud, storage, networking, enterprise computing, or lower-density inference. Still others could become functionally obsolete if the cost of modernization exceeds the value of the resulting capacity.

Investors should therefore examine the density distribution within a property rather than accepting a portfolio-wide average. Critical questions include how many racks can actually operate at the advertised density, whether cooling and electrical systems can sustain that load simultaneously, and whether the capacity is available under normal and contingency conditions.

Liquid cooling reaches an adoption threshold

Thirty-nine percent of AFCOM respondents say their cooling systems fail to meet all operational requirements, compared with 34% in the previous year. Thirty-six percent report that they have implemented liquid cooling, up from 19%, and another 28% plan adoption within 12 to 24 months.[1]

Respondents estimate an average transition threshold of approximately 50 kilowatts per rack, although answers range from below 20 kilowatts to above 80 kilowatts. The variation is technically reasonable because the practical limit of air cooling depends on airflow volume, pressure, temperature, humidity, rack configuration, server design, and containment.

The report finds no dominant liquid-cooling architecture. Rear-door heat exchangers and two-phase immersion are each cited by 37% of respondents. Single-phase direct-to-chip cooling is cited by 26%, while smaller shares select hybrid direct-to-chip and rear-door configurations, two-phase direct-to-chip, or single-phase immersion.[1]

These technologies should not be treated as interchangeable. Direct-to-chip cooling removes heat from processors and accelerators but may leave memory, storage, power supplies, and networking components dependent on air. Rear-door heat exchangers can extend the life of existing air-based facilities. Immersion can support high heat removal but changes equipment-service procedures, materials requirements, fluid management, warranty relationships, and operational practices.

For landlords, liquid cooling also changes the division of responsibility between building and tenant. Lease documents must define ownership and maintenance of cooling-distribution units, manifolds, piping, water treatment, leak detection, heat exchangers, and tenant-installed systems. Operators must establish acceptable fluid chemistry, pressure, temperature, redundancy, and connection standards.

A failure to resolve these boundaries can produce disputes over capital expense, operating expense, water damage, equipment warranties, and responsibility for downtime. The most valuable AI-ready facilities will therefore offer more than nominal liquid-cooling capability; they will provide a standardized and contractually defined thermal-service architecture.

Power becomes the controlling asset

The International Energy Agency estimates that global data-center electricity consumption will rise from approximately 485 terawatt-hours in 2025 to approximately 950 terawatt-hours in 2030. AI-focused data-center consumption is expected to triple during that period.[3]

The United States faces a particularly concentrated impact. Lawrence Berkeley National Laboratory’s June 2026 update estimates a reference case of approximately 649 terawatt-hours of U.S. data-center electricity consumption in 2030, with a broad uncertainty range of 521 to 843 terawatt-hours. That would represent approximately 9.5% to 15.3% of total U.S. electricity consumption, with 11.8% as the central estimate.[4]

These figures explain why power is becoming the industry’s primary gating factor. The constraint is not simply aggregate annual energy. Data centers require deliverable capacity at specific locations, at the correct voltage, with appropriate redundancy, power quality, transmission support, and commissioning schedules.

A data center may also exhibit a different load profile from conventional commercial customers. AI training can create concentrated, high-utilization demand, while some accelerator workloads can introduce rapidly changing loads. Utilities must plan generation, transmission, substations, protection systems, voltage control, and reserve requirements around actual operating characteristics rather than nameplate capacity alone.

Power scarcity is changing site selection and capital allocation. JLL reports that emerging or “frontier” North American markets account for 77% of capacity under development as developers move toward energy-rich and comparatively buildable regions, including parts of Texas, Ohio, Louisiana, and the Carolinas.[8]

This geographic expansion creates opportunities, but it also increases community, transmission, fuel, water, and market-basis risks. A nominally energy-rich location may still lack transmission capacity or firm gas transportation. A project relying on a future transmission line, pipeline expansion, or new generating plant carries development dependencies that must be reflected in valuation and financing.

On-site generation and BESS reshape capital structure

AFCOM reports that 25% of respondents currently generate power on-site, compared with 19% in the preceding year. Another 23% plan implementation within 12 months, and 17% are exploring the possibility.[1]

Behind-the-meter generation can reduce dependence on delayed utility expansion and provide greater control over project schedules. It can also improve resiliency when coordinated with the grid, energy storage, and multiple generating units.

It is not free power. On-site generation increases initial capital requirements and introduces fuel-price exposure, emissions permitting, maintenance, spare-parts requirements, operating personnel, and equipment-replacement obligations. A natural-gas turbine may reduce grid dependence but create dependence on firm gas transportation and local pipeline capacity. Reciprocating engines may offer modularity and faster response but introduce different maintenance, noise, and emissions profiles.

Battery energy storage systems, or BESS, are another major theme. AFCOM reports that BESS and solar each received 44% when respondents were asked which non-fossil technologies were gaining the most traction.[1]

Technically, BESS is not an energy source and is not inherently non-fossil. It stores electricity, and its environmental profile depends on the charging source. Its value lies in flexibility: ride-through, peak management, renewable integration, grid services, load smoothing, and potentially reduced use of diesel generators.

The financial case depends on the service stack. A battery justified only as emergency backup may have a different return profile from one earning revenue or reducing costs through demand-charge management, energy arbitrage, capacity programs, and ancillary services. Investors must also account for degradation, augmentation, fire protection, insurance, warranty limits, interconnection rights, and replacement capital.

AFCOM appropriately concludes that on-site generation is normally most valuable as part of an integrated architecture rather than as a permanent rejection of the grid. The likely winning design combines grid service, local generation, energy storage, demand management, and, where technically and economically feasible, renewable or nuclear supply.

The report discusses FERC Order No. 2023 as part of the interconnection environment. The order does replace the former serial process for many generating projects with a first-ready, first-served cluster-study process and stronger commercial-readiness requirements.[7] It is important, however, to recognize that Order No. 2023 primarily reforms generator and storage interconnection to the transmission system. It does not independently solve utility-service delays for large data-center loads.

Nuclear power is a strategic option, not a near-term baseline

Thirty-two percent of AFCOM respondents identify small modular nuclear reactors as a non-fossil technology gaining traction.[1] That percentage measures industry interest, not actual deployments.

The potential attraction is clear. Advanced nuclear systems could provide high-capacity-factor, carbon-free generation suitable for large campuses with relatively constant demand. Their smaller unit sizes could theoretically align more closely with phased data-center growth than conventional gigawatt-scale reactors.

The obstacles are equally important: licensing, technology maturity, project duration, construction execution, fuel arrangements, waste management, security, financing, and public acceptance. Nuclear projects must be evaluated as long-duration energy infrastructure, not as equipment that can simply be ordered with a data-center building.

For most projects developed during the present construction cycle, nuclear power is more likely to appear through long-term power-purchase arrangements, support for existing nuclear generation, or future development partnerships than through a reactor operating directly behind the meter. Investors should assign value according to licensed, financed, contractually supported milestones—not announcements or memoranda of understanding.

Cloud repatriation is workload optimization, not the end of cloud

AFCOM reports that 67% of respondents experienced some movement of workloads from public cloud to on-premises or colocation environments in 2026. That compares with 80% in 2025, 71% in 2024, 83% in 2023, and 59% in 2022. Among respondents experiencing migration, 81% report higher facility power demand and 36% describe the effect as significant.[1]

These figures should not be interpreted as proof that cloud computing is contracting. The survey appears to measure whether an organization moved any workloads, not the net proportion of computing leaving public cloud.

An enterprise can expand its cloud spending while repatriating selected workloads. Persistent, predictable, data-intensive, latency-sensitive, or heavily utilized applications may be less expensive or more controllable in owned or colocated infrastructure. Bursty workloads and applications needing rapid geographic deployment may remain better suited to hyperscale cloud.

AI strengthens this portfolio approach. Training, fine-tuning, retrieval, inference, storage, and data preparation have different economic and latency characteristics. The infrastructure decision increasingly depends on data gravity, accelerator utilization, egress expense, compliance, model sensitivity, and the value of operational control.

For investors, this supports both hyperscale cloud growth and continued demand for colocation, neocloud, sovereign-cloud, and enterprise infrastructure. The important variable is not “cloud versus on-premises,” but which operator can provide the lowest risk-adjusted cost for a defined workload.

DCIM evolves toward an operational intelligence layer

AFCOM reports that DCIM adoption is highest in security, asset management, cable management, power and environmental monitoring, and change management. Fifteen percent already use AI or machine learning in DCIM, while 66% plan integration within three years.[1]

Potential applications include predictive maintenance, thermal optimization, capacity forecasting, anomaly detection, energy scheduling, and digital-twin analysis. These capabilities become increasingly valuable as facilities grow too complex for operators to manage through disconnected monitoring systems.

The investment case should not assume unrestricted autonomous control. Data-center operations involve safety-critical electrical and mechanical equipment. AI may forecast a cooling failure or recommend moving workloads, but operational authority should remain bounded by deterministic controls, equipment limits, verified telemetry, change-management procedures, and human authorization.

The most credible DCIM platforms will demonstrate integration with heterogeneous equipment, reliable asset data, cybersecure interfaces, auditability, and measurable operating savings. Platforms that merely add a conversational interface to incomplete facility data may create little durable value.

Security converges across people, identity, cyber systems, and physical access

AFCOM respondents rank human threats first at 61%, followed by ransomware at 58%, AI-enabled synthetic-identity attacks at 46%, advanced persistent threats at 44%, and distributed denial-of-service attacks at 42%.[1]

These are rankings of concern, not incident probabilities. Nevertheless, the convergence is strategically important. A modern data center can be attacked through credentials, contractors, remote-management interfaces, supply-chain components, deepfake communications, physical access, or compromised administrative systems.

Eighty-four percent report changes to physical-security requirements during the previous 12 to 24 months. Expanded surveillance, vendor screening, perimeter security, and access controls are the leading responses.[1]

The financial consequences extend beyond the cost of cybersecurity. An incident can interrupt tenant operations, create contractual liability, increase insurance costs, impair financing, damage regulatory relationships, and reduce the perceived reliability of an entire portfolio. Security therefore affects asset value and cost of capital.

Investors should evaluate whether security controls extend across the complete operational environment: IT systems, operational technology, building controls, maintenance vendors, remote access, identity verification, physical entry, software supply chains, and incident recovery. Backup systems that share credentials, networks, or administrative domains with production systems may not provide true recovery independence.

Supply-chain and workforce constraints become balance-sheet risks

Sixty-six percent of AFCOM respondents report continuing supply-chain constraints, and 15% say such delays have contributed directly to outages. Eighty-seven percent report workforce gaps affecting operations.[1]

High-density facilities require synchronized delivery of GPUs, networking, memory, storage, transformers, switchgear, generators, cooling equipment, pumps, controls, and specialized labor. A late transformer or cooling component can strand otherwise completed capacity. The economic effect is delayed rent commencement, extended interest carry, liquidated damages, and lower project return.

Uptime Institute’s 2026 global survey independently identifies limited power availability, declining grid reliability, rising costs, supply-chain limitations, and staffing shortages as major operating constraints.[5] Unlike the AFCOM report, Uptime states that its 2026 survey included more than 800 data-center owners and operators, providing a useful methodological cross-check.

The workforce problem is not simply a shortage of traditional IT personnel. AFCOM identifies demand for data-center engineers, multi-skilled operators, power and energy engineers, and security specialists.[1] Liquid cooling, on-site generation, batteries, higher-voltage distribution, and AI-enabled controls require a combination of electrical, mechanical, software, cybersecurity, and operational expertise.

For investors, workforce availability should be part of market selection. A facility in a low-cost location may experience commissioning or reliability problems if skilled operators, electricians, controls engineers, and service organizations are unavailable. Partnerships with technical colleges, universities, apprenticeship programs, and equipment vendors can become part of the asset’s operating moat.

Reading the AFCOM statistics correctly

The AFCOM study is valuable, but its survey percentages should be interpreted as directional evidence rather than statistically definitive measurements of the worldwide industry.

The published report does not disclose its sample size, field dates, geographic distribution, organization types, respondent-selection method, weighting, response rate, confidence intervals, or complete questionnaire. It also does not establish that successive annual surveys contain comparable respondent populations.

This is particularly important for the reported 27-kilowatt average rack density. A larger share of hyperscale, colocation, GPU-cloud, or AI-focused respondents could raise the arithmetic mean substantially without demonstrating that the typical installed rack throughout the industry increased by 69%. A median, percentile distribution, facility weighting, and respondent count would provide a stronger basis for comparison.

The report also occasionally shifts between perception and deployment. Respondents were asked which technologies were “gaining the most traction,” but the discussion sometimes treats those results as adoption rates. Similarly, combining present renewable-energy use with future plans does not guarantee that all planned projects will be completed.

These limitations do not invalidate the report’s strategic message. Its principal findings are consistent with the IEA, Lawrence Berkeley National Laboratory, Uptime Institute, McKinsey, and JLL. They do mean that investors should not use a single AFCOM percentage as the basis for demand forecasts, engineering design, environmental projections, or asset valuation.

What the trends mean for investors

The data-center industry is entering a period in which demand remains exceptionally strong but value creation will be increasingly uneven.

The probable beneficiaries include owners of power-secured data-center capacity; utilities able to invest without shifting excessive risk to other customers; manufacturers of electrical and cooling equipment; engineering and construction firms; liquid-cooling suppliers; energy-storage providers; grid-technology companies; and operators with proven commissioning and facility-management capabilities.

The principal risks include:

* Paying AI-level valuations for assets incapable of supporting AI densities.
* Treating requested or announced power as deliverable capacity.
* Financing construction before establishing tenant credit and power certainty.
* Underestimating substation, transmission, gas, water, cooling, and permitting schedules.
* Overbuilding for training demand if inference develops a different geographic and technical profile.
* Selecting cooling or electrical systems that become incompatible with future hardware.
* Assuming high current rents will permanently exceed growth in capital and operating costs.
* Underestimating community opposition involving electricity rates, water, noise, emissions, land use, and tax incentives.
* Concentrating exposure in a small number of hyperscalers, neoclouds, equipment suppliers, or fuel sources.
* Mistaking a technology announcement for commercially commissioned capacity.

The strongest assets will have verified power, flexible thermal architecture, multiple expansion paths, high-quality tenants, disciplined capital phasing, transparent community relationships, and an operating organization capable of integrating IT, electrical, mechanical, security, and energy systems.

The industry at an inflection point

The AFCOM report captures a fundamental change: compute is no longer separable from energy infrastructure. AI is forcing data-center owners to redesign the path from the electrical grid—or on-site generating plant—to the processor and from the processor’s heat output back to the environment.

This creates a large and durable investment cycle, but it also raises the cost of error. The next generation of data centers will be more capital-intensive, more technically specialized, more dependent on energy markets, and more exposed to regulatory and community scrutiny.

Power-secured land, liquid-cooling capability, modular construction, BESS, behind-the-meter generation, intelligent infrastructure management, supply-chain resilience, and specialized talent are becoming components of a single competitive system. No one component is sufficient. Cheap land without power has limited value. Power without customers can become stranded infrastructure. Accelerators without cooling cannot operate. A completed building without skilled operators and resilient supply chains cannot deliver contracted reliability.

The industry is therefore evolving from a real-estate-led model into an integrated infrastructure model. The investment winners will not necessarily be those announcing the most megawatts. They will be the organizations that convert credible demand, financeable power, adaptable engineering, disciplined construction, and community legitimacy into commissioned, revenue-producing capacity.

References

1. Kleyman, Bill, and Kristin Letourneau. State of the Data Center 2026 Study: Ten Years of Data Reveals How AI Is Reshaping Data Center Design, Power, Density, and Talent. AFCOM/Informa, February 2026. Original report⁠.
2. Jones Lang LaSalle. 2026 Global Data Center Market Outlook: Navigating AI Demand, Power Constraints and Global Opportunities in 2026. January 5, 2026. JLL report⁠.
3. International Energy Agency. Key Questions on Energy and AI. 2026. IEA executive summary⁠.
4. Smith, Sarah Josephine, Alex Hubbard, Alexander Newkirk, Mohan Ganeshalingam, Billie Holecek, Dale A. Sartor, Michael Mills, and Arman Shehabi. United States Data Center Energy Usage Report: 2025 Update. Lawrence Berkeley National Laboratory, June 2026. LBNL report page⁠.
5. Donnellan, Douglas, Andy Lawrence, Daniel Bizo, Owen Rogers, Peter Judge, John O’Brien, Jacqueline Davis, Max Smolaks, and Rose Weinschenk. Uptime Institute Global Data Center Survey 2026. Uptime Institute Intelligence, July 24, 2026. Uptime Institute report⁠.
6. McKinsey & Company. “Scaling Bigger, Faster, Cheaper Data Centers with Smarter Designs.” August 1, 2025. McKinsey analysis⁠.
7. Federal Energy Regulatory Commission. Explainer on the Interconnection Final Rule: Order No. 2023. Updated following Order No. 2023-A. FERC explainer⁠.
8. Jones Lang LaSalle. “Data Center Demand Exceeds Expectations as H1 2026 Absorption Hits Record 25 GW.” August 11, 2026. JLL North America midyear findings⁠.