The first question used to be fiber and land. Now it is electrons, and the answer arrives with a date attached.
JLL's 2026 Global Data Center Outlook, summarized in a JLL Spark post by Carolyn Trickett and Ajey Kaushal, puts numbers on the squeeze: nearly 100 GW of new capacity between 2026 and 2030, AI workloads at roughly half of all usage by the end of the decade, and demand still outrunning that supply. In primary markets, grid connection waits now exceed four years. Capital cannot buy a queue position, so developers have started designing around the queue instead of standing in it.
Power: make it, contract for it, or move the load
The menu JLL describes is a portfolio, not a single choice. "Bring your own power" and behind-the-meter generation come first where interconnection is the binding constraint. Long-term renewable PPAs, co-located solar and wind, and battery storage carry the low-carbon share of the load. Natural-gas-backed microgrids, fuel cells and small modular reactors cover resilience where grid scaling lags demand.
The nuclear number is the one that moved: the SMR pipeline with data center operators grew from 25 GW at the end of 2024 to 45 GW by the time of the IEA's April 2026 report, Key Questions on Energy and AI, as reported by Capacity. That is a pipeline of interest and signed intent, not operating reactors. The same reporting notes that many US projects pursuing onsite natural gas remain in early stages, based on satellite tracking — a useful corrective to announcement-driven capacity math.
Software gets its own line because it changes when power is drawn, not just how much. Batteries plus load-shifting logic let operators time the heaviest training runs to hours when grid power is cleaner or simply available. That turns a training schedule into something closer to a dispatchable asset, which is a different discipline from signing a PPA.
The consequence JLL draws is that site selection stops being a fiber-and-land exercise and becomes an integrated power-portfolio decision. Anyone still underwriting sites on land cost and latency alone is solving last decade's problem.
Cooling stops being a facilities detail
Cooling can account for up to 40% of facility energy use, and conventional air cooling is hitting physical limits at AI rack densities. JLL's layered answer: advanced liquid cooling, which it says can cut cooling energy use by up to one-third while improving power usage effectiveness; hybrid systems and rear-door heat exchangers that convert existing air-cooled halls rather than stranding them; and AI-driven DCIM that adjusts airflow, liquid flow and setpoints against real-time workload placement.
Read the one-third carefully. It applies to cooling energy, not the whole facility bill, and it is a project-dependent range rather than a guarantee. The retrofit path is the more interesting half: a large installed base of air-cooled capacity already exists, and converting it is usually faster than permitting a new building.
Robert Dunn, CEO of Start Campus, told Capacity that the meaningful advances are liquid cooling, waste heat recovery and reuse, zero-water cooling and grid-interactive facility design, and argued that sustainability and performance are no longer trade-offs for AI infrastructure. That is an operator's claim and worth testing site by site, but the direction matches the physics — dense racks push you toward liquid anyway, and once you are plumbing coolant, you may as well recover the heat.
Compressing the schedule
Modular and prefabricated infrastructure cuts construction timelines by 30–50% and allows capital to be deployed in phases, per JLL. Schedule is the product here: a hall that energizes a year earlier earns revenue a year earlier, which usually outweighs a higher cost per megawatt.
Two other moves matter. Edge and distributed capacity — micro data centers and edge colocation — support low-latency inference as inference overtakes training in volume. And sustainable construction, meaning low-carbon materials, adaptive reuse of industrial assets and integrated waterless cooling, does double duty: it cuts embodied carbon and it reduces the community friction that stalls permits.
Operating a campus you cannot staff linearly
Larger footprints spread across more sites break the old staffing model. JLL points to AI-driven infrastructure management that combines DCIM, telemetry and machine learning for energy optimization, predictive maintenance and capacity planning; workload-aware software that reads GPU cluster behavior to consolidate jobs and route them to efficient locations; and edge fleet orchestration that runs thousands of distributed sites remotely without headcount growing in step with square footage.
The unglamorous dependency is instrumentation. None of this works on a site with bad metering, and retrofitting telemetry is harder to fund than a liquid cooling loop.
What the numbers do and do not settle
The demand backdrop, per the IEA as reported by Capacity: data center electricity demand rose 17% in 2025, AI-focused facilities grew faster and outpaced growth in global electricity demand by 3%, consumption is set to double by 2030 with power use from AI-enabled data centers tripling, and capital expenditure at five large technology companies topped US$400 billion in 2025 with another 75% expected in 2026. Technology companies signed roughly 40% of corporate renewable PPAs in 2025.
JLL frames all this as a $3+ trillion investment supercycle, a $500 billion technology opportunity across power, cooling, infrastructure and management, and a venture-scalable slice of about $50 billion. Read that as a well-informed commercial projection from a firm whose venture arm invests in the category — the market sizing and the pitch are the same document.
Two cautions. The IEA's finding that power consumption per AI task is falling at an unprecedented rate sits alongside the tripling forecast; both can be true, and the gap between them is where the real uncertainty lives. And the binding constraints are physical and administrative, not financial: strained supply chains for gas turbines, transformers and advanced chips, plus planning and regulatory systems that hold up connections.
The item most likely to be underestimated is the least technical. JLL notes that meeting capacity goals depends on public sentiment and community acceptance, and that early engagement is handled in-house by developers and owners today. A four-year interconnection queue is a scheduling problem. A blocked project is a permanent one.
Practical upshot: before the next site goes under contract, the useful questions are the interconnection date with a fallback generation plan, the rack density the cooling design can actually support, the modular path that lets you energize in phases, and the telemetry that makes any of it optimizable. That is an infrastructure project with a queue position, not a real estate asset with a yield.