Market analysis confirms AI data center demand vastly outpaces available supply, creating structural capacity shortage extending through 2027–2028.
Iron Mountain and Structure Research have released forecasts projecting that artificial intelligence will create a severe imbalance between global data centre demand and supply by 2030. Annual global demand for data centre capacity is expected to reach nearly 90GW by the end of the decade, potentially exceeding available supply by as much as 500%.
The scale of required investment is already apparent. Hyperscaler capital expenditure is projected to reach $375 billion this year, a 36% increase from 2024. Approximately half will be spent on servers and GPUs, with the remainder allocated to data centre capacity.
A critical shift is underway in how AI infrastructure is deployed. Early investment focused on the computing resources needed to train large language models, but the balance is now expected to pivot decisively towards inference—running those models at scale to respond to user requests and business workloads in real time. Structure Research and Iron Mountain forecast that inference capacity will overtake training capacity in 2026. By 2030, inference is projected to account for 80% of all AI critical IT load, meaning infrastructure for live services will be four times larger than infrastructure dedicated to training.
This shift has profound implications for facility placement. Training clusters can be concentrated in large, remote campuses with access to land and power. Inference systems serving end users, by contrast, need proximity to major population centres to minimize latency and manage traffic volumes. The result is increased pressure on already constrained urban markets, where grid access, land availability and planning restrictions have become significant barriers to expansion.
The forecasts identify emerging data centre megahubs concentrated in relatively few locations globally. In North America, Northern Virginia is projected to reach 8.5GW—the largest single hub—followed by Dallas at 2.8GW and Phoenix at 2.7GW. European hubs include London (2.7GW), Frankfurt (2.68GW) and Paris (2GW), with accelerating growth expected in Madrid, Barcelona, Berlin, Düsseldorf and Lisbon. Across Asia-Pacific, Tokyo is forecast to reach 2.8GW, Sydney 2.4GW, Johor 2.2GW and Mumbai 2.15GW.
Beyond infrastructure, a second challenge is emerging around cost control. While prices for the cheapest large language models have fallen sharply, the report argues that lower prices are more likely to increase usage than reduce overall spending. Token-based and usage-based charging models create new oversight challenges for finance and technology leaders, particularly when AI tools are deployed across large workforces with minimal usage restrictions. Companies will need to establish controls around internal AI adoption, directing spending towards tasks with measurable returns rather than simply enabling broad access to models.