Half of capacity to AI by 2030 — and the inference shift ahead

The AI infrastructure buildout is not a single demand wave — it is a restructuring of what data centers are for. JLL forecasts global data center capacity rising from roughly 103 GW today to 200 GW by 2030, with AI workloads accounting for approximately half of that capacity by the end of the decade.

103 → 200 GW

global data center capacity, today to 2030

~50%

of data center capacity expected to serve AI workloads by 2030 (about 25% in 2025)

Training workloads drove the first phase: dense, high-power clusters in locations where power could be aggregated at scale. Inference — running trained models in production — is different. Latency, geographic distribution, and edge proximity matter more. JLL projects inference could overtake training as the dominant AI requirement by 2027.

2027

when inference could overtake training as the dominant AI requirement

For land and power strategy, this shift has two consequences. First, demand is broadening beyond traditional hyperscale hubs into secondary markets where power may be more available — but utility processes are less familiar with large-load data center requests. Second, the MW-per-acre math is changing as rack densities rise and liquid cooling adoption accelerates.

Hyperscaler capex plans underscore the scale of investment. The four largest cloud operators have announced combined 2026 capital expenditure approaching $725 billion — up roughly 77% year over year — with a significant share directed at AI infrastructure and the power to run it.

~$725B

combined 2026 capex plans of the four largest hyperscalers, up ~77% year over year

Construction costs are rising in parallel. JLL forecasts average global data center shell-and-core construction cost at $11.3 million per MW in 2026. When power is scarce, the land beneath it is valued not for acreage but for the megawatts it can support on a documented timeline.

$11.3M / MW

forecast average global data center construction cost in 2026

Anywhere DC screens for both training-scale and inference-scale paths — because the utility conversation differs. A 200 MW campus request triggers different studies than a 10 MW edge deployment. Readiness work must match the load.