Summary
- Iron Mountain and Structure Research forecast annual global data centre demand approaching 90GW by 2030.
- The report says AI inference will overtake training in 2026 and represent 80% of AI critical IT load by 2030.
- European demand is expected to concentrate around London, Frankfurt, Paris, and faster-growing hubs including Madrid, Barcelona, Berlin, Düsseldorf, and Lisbon.
Iron Mountain Data Centers and Structure Research have forecast that annual global data centre demand could reach nearly 90GW by 2030, leaving supply as much as 500% short of demand if delivery cannot keep pace.
The forecast, set out in Iron Mountain’s Top 4 AI Predictions report, points to a rapid reshaping of digital infrastructure as AI moves from model training into large-scale inference. The report says AI colocation revenue for training and inference is forecast to grow at a 77% compound annual growth rate from 2025 to 2030, reaching US$134bn by the end of the decade.
The same analysis expects AI to account for around 44% of total global data centre colocation market revenue by 2030. It says demand for data centre capacity will outstrip supply from 2027 to 2030, with global annual demand reaching nearly 90GW.
The European figures are substantial. Iron Mountain and Structure Research forecast London at 2.7GW, Frankfurt at 2.68GW, and Paris at 2GW by 2030, with fast growth also expected in Oslo, Barcelona, Madrid, Zaragoza, Lisbon, Berlin, and Düsseldorf.
Inference changes the geography
AI training has driven much of the early hyperscale campus discussion. Training clusters can be placed in remote locations where power and land are available, provided network requirements and the operating model work. The next phase of AI infrastructure is less geographically forgiving.
Iron Mountain and Structure Research expect inference capacity to overtake training capacity in 2026. By 2030, inference is forecast to account for 80% of total AI critical IT load, reversing the training-heavy structure seen in the early phase of generative AI infrastructure.
Inference is the infrastructure behind live AI services. It has to respond to user requests, enterprise workflows, and application traffic at scale. That pushes some capacity closer to population centres, enterprise markets, and data sovereignty jurisdictions, rather than only towards remote power-rich sites.
European pressure will not simply move away from the established hubs. London, Frankfurt, and Paris are already constrained by land, grid access, planning politics, and build costs. If inference demand grows around those same markets, developers will need capacity in or near already stretched geographies.
Secondary hubs could benefit, although only where power, fibre, latency, and regulation align. Madrid, Barcelona, Lisbon, Berlin, Düsseldorf, Oslo, and Zaragoza all have different combinations of available land, grid questions, climate, renewable procurement, and customer demand. Forecast demand is not the same as deliverable supply.
Supply risk moves into the capital plan
The report also reframes the financial risk around AI infrastructure. Hyperscaler capital expenditure is projected at US$375bn this year, with roughly half going towards infrastructure needs such as self-build and leased data centre capacity, and half towards servers, CPUs, TPUs, and GPUs.
That split is critical because chips alone do not create usable AI capacity. GPUs require powered, cooled, secure, and connected facilities. High-density clusters need more sophisticated electrical distribution, cooling systems, controls, commissioning, and operational procedures than conventional enterprise deployments.
If demand rises faster than facilities can be delivered, the pressure will not only appear in rents. It can influence where AI services are deployed, how much latency users accept, what customers pay for inference, and whether companies can scale AI products without infrastructure bottlenecks.
The report also invokes a version of Jevons’ Paradox: falling AI costs may increase overall consumption rather than reduce it. More efficient models and cheaper inference can make AI more widely used, increasing the volume of queries, tokens, and workloads that need to be served.
That effect would deepen the strain on European power systems already supporting electrification of transport, heating, industry, and manufacturing. Data centres are joining that queue with loads that can resemble heavy industry rather than conventional IT.
The forecast is not destiny. Capacity announcements frequently exceed what is eventually built, and demand forecasts can move with model efficiency, regulation, power prices, and customer adoption. The direction is clear enough: AI is converting data centre planning into power-system planning.
The strongest markets will not be the ones with the loudest demand story. They will be the ones that can bring together grid access, land, cooling, planning, supply chains, and capital quickly enough to turn forecast load into energised capacity.
