Summary
- McKinsey estimates AI-related data centre demand could rise from about 44GW to 155GW by 2030.
- Global investment in data centres, excluding IT hardware, may exceed $1.7tn cumulatively through 2030.
- The analysis frames AI capacity around power, cooling, project delivery, electricity prices, and capital cost.
McKinsey has estimated that AI-related data centre demand could rise from about 44GW in 2025 to 155GW by 2030, turning the AI infrastructure boom into a test of power, cooling, land, capital, and project delivery.
The firm’s analysis of colocation data centres says total global data centre demand could rise from about 82GW to about 220GW over the same period. It also estimates that global data centre investment, excluding IT hardware, may exceed $1.7tn cumulatively through 2030.
The report separates AI training and inference demand, then examines infrastructure economics across markets including London, northern Sweden, Northern Virginia, Singapore, and China. Its comparison of a 100MW liquid-cooled AI colocation facility places electricity prices, power and cooling equipment, construction, and delivery speed at the centre of competitiveness.
Demand is splitting the market
Training and inference create different infrastructure demands. Training facilities can be concentrated in very large campuses where land, power cost, and access to accelerators dominate. Inference is likely to become more distributed as AI becomes embedded in applications that need lower latency, data locality, or customer proximity.
That split is uncomfortable for European markets. London, Frankfurt, Amsterdam, Paris, and Dublin remain major digital hubs, but each faces some combination of power constraints, planning scrutiny, environmental pressure, and land scarcity. More power-advantaged locations in the Nordics and Iberia may support larger AI campuses, although connectivity, customer depth, grid readiness, and execution capability remain decisive.
McKinsey’s analysis notes that the UK had 2.2GW of installed hyperscaler and colocation capacity in 2025, while Sweden had 0.8GW. It also says the Nordics benefit from lower-cost, low-carbon power, a cooler climate, and room to scale. London remains attractive because demand and network depth still carry weight, even where costs are higher.
The economics are not simply a property question. McKinsey models levelised infrastructure cost for a 100MW liquid-cooled AI colocation facility, excluding IT hardware. In that framing, utilities become the largest driver of cross-country cost variation, while power and cooling equipment are another major factor.
Deliverable megawatts beat paper pipeline
The report’s investment numbers sharpen the difference between announced capacity and delivered capacity. A data centre pipeline can expand quickly on paper, but facilities are built at the pace of substations, permits, equipment orders, contractors, commissioning teams, and energisation.
AI raises the stakes because the load is more concentrated and more expensive. High-density clusters require stronger electrical infrastructure, liquid-cooling readiness, carefully managed redundancy, and more intense commissioning. Retrofitting existing halls can work for some deployments, but wholesale conversion is difficult where floor loading, pipe routes, maintenance access, and electrical topology were designed for lower-density use.
Colocation providers face a strategic squeeze. Customers want faster capacity and higher density, while operators must protect resilience and long-term flexibility. A facility built too tightly around today’s AI hardware can become difficult to adapt; a facility designed too cautiously may miss current demand.
Europe’s policy environment adds another filter. Energy-efficiency reporting, water scrutiny, heat-reuse expectations, planning pressure, and critical infrastructure resilience requirements are all becoming part of the capacity equation. Markets with coordinated power planning and predictable permitting will have an advantage over those relying on fragmented queue management.
McKinsey’s analysis makes a practical point: AI infrastructure is not a software cycle. It is a construction, power, cooling, and capital cycle. The winners will be the markets and operators able to convert electricity connections, equipment procurement, and planning consent into resilient capacity before customer demand shifts elsewhere.
