AI boom to leave data centre demand far above supply

Iron Mountain has published four forecasts on how artificial intelligence will reshape global data infrastructure, pointing to a widening gap between data centre demand and supply by 2030.

Working with Structure Research, Iron Mountain projects annual global demand for data centre capacity will reach nearly 90GW by the end of the decade. Demand at that level could exceed available supply by as much as 500% as spending on servers, GPUs and new capacity continues to rise.

Across the sector, hyperscaler capital expenditure is projected to reach USD $375 billion this year, up 36% from 2024, according to figures cited by the two companies. About half of that total is expected to be spent on servers and GPUs, with the rest allocated to data centre capacity.

Supply pressure

The forecast shows how quickly infrastructure needs have changed since generative AI tools entered the mainstream. Demand for compute, storage and power has risen sharply as companies expand AI systems from development into broader commercial use.

One of the report's central findings is the shift from training large models to running them at scale for everyday use. Early investment focused largely on the computing resources needed for training, but the balance is now expected to shift towards inference, when models respond to user requests and business workloads in real time.

Structure Research and Iron Mountain expect inference capacity to overtake training capacity in 2026. By 2030, inference is forecast to account for 80% of all AI critical IT load, implying that infrastructure for live services will be four times larger than infrastructure dedicated to training.

That shift has implications for where facilities are built. Training clusters can often be concentrated in large, remote campuses with access to land and power. Inference systems serving end users may need to be closer to major population centres to reduce latency and handle growing traffic volumes.

As a result, the report points to greater demand for new capacity in and around large urban markets. That could add pressure to already constrained locations, where grid access, land availability and planning limits have become more significant barriers to expansion.

Regional hubs

The projections also suggest very large data centre hubs will emerge across every major global region. By 2030, several markets are expected to exceed 2GW of capacity, highlighting the concentration of AI infrastructure in a relatively small number of established and fast-growing locations.

In North America, Northern Virginia is projected to reach 8.5GW, maintaining its position as the largest single hub in the forecast. Dallas is expected to reach 2.8GW and Phoenix 2.7GW, reflecting continued growth in markets with access to land, network links and enterprise demand.

In Europe, London is forecast to reach 2.7GW, Frankfurt 2.68GW and Paris 2GW. The outlook also points to faster growth in Madrid, Barcelona, Berlin, Dusseldorf and Lisbon as operators and customers look beyond the region's biggest centres for additional capacity.

In Asia-Pacific, Tokyo is projected to reach 2.8GW, Sydney 2.4GW and Johor 2.2GW. Mumbai is expected to reach 2.15GW, adding to the list of markets likely to play a larger role in meeting AI and cloud demand.

Cost question

The forecasts also highlight a commercial issue for companies adopting AI tools more widely: usage costs. While the price of the cheapest large language models has been falling quickly, the report argues that lower prices are more likely to increase usage than reduce spending.

That means businesses may face a different kind of cost pressure as employees make heavier use of AI systems priced on consumption. Token-based and usage-based charging models can create new oversight challenges for finance and technology leaders, especially when AI tools are deployed across large workforces with few limits on use.

According to the outlook, companies will need to manage demand carefully and put controls around internal AI use. The issue is framed less as a question of access to models and more as a question of how organisations limit unnecessary use while directing spending towards tasks with clearer returns.

The analysis presents an industry moving from experimentation to scale, with infrastructure, geography and cost all shifting at once. Its starkest conclusion is that annual global demand will reach nearly 90GW by 2030 and exceed available supply by as much as 500%.