Where will the global growth of AI infrastructure be fastest?

To enable the next generation of AI-powered services and products, a new generation of advanced infrastructure needs to be built at record-breaking speed and scale. As inference overtakes training as the primary consumer of accelerated computing, location is becoming much more critical to the successful delivery of low-latency services. But, with power limitations, slower permissions, and rising demand for AI data sovereignty, where are these data centers most likely to be built?

Here are a few quick insights into the geographical drivers and constraints, along with predictions, courtesy of Structure Research, for the top five hubs in each global region by 2030. The information is taken from the latest Iron Mountain Data Centers paper “Building at scale: Where we deliver accelerated infrastructure”

How big is the infrastructure challenge?

Current estimates suggest that approximately 20% of total global data center capacity is already dedicated to AI. According to projections, AI - delivered by hyperscalers and neoclouds - will drive $75 billion in data center demand by 2028, potentially pushing AI's share of the total market to 35%. Combined with existing cloud growth, the infrastructure demand this creates is immense. As McKinsey has noted, to avoid a supply deficit, the industry would need to build at least twice the capacity constructed since 2000 in a fraction of that time.

What will the bulk of new capacity be used for?

While the headlines are still dominated by multi-gigawatt training campuses, the bulk of new capacity will be required for inference. Training facilities develop and refine large AI models, but inference is where these models are put to work in production environments, generating responses for everything from chatbots to self-driving algorithms. Inference has already overtaken training, and by next year inference infrastructure will account for over 60% of AI infrastructure, growing at a 79% compound annual growth rate (CAGR) through 2030, vastly outpacing the 25% CAGR for training facilities. Because inference is a recurring, real-time necessity, it will quickly become the dominant form of AI infrastructure and the greatest consumer of power and space.

What are the infrastructure requirements for inference v training?

The geographical strategy for AI splits into two categories based on the facility function:

• Training Hubs: Hyperscalers and neoclouds are building massive, remote "superclusters" where power is abundant and land is available. Projects like OpenAI’s Stargate in Texas, Meta’s Hyperion in Louisiana, and the Google cluster in Ohio are prime examples of this "mega-facility" approach (100 MW to 1 GW+).
• Inference Hubs: These are distributed, latency-sensitive facilities. Because inference needs to happen close to the end-user, these medium-sized facilities (5–100 MW) are being integrated into existing cloud regional architecture, often located adjacent to existing cloud on-ramps to ensure fast connectivity.

Are new data center hot spots emerging?

Yes, and fast. The delivery of AI to the end-user requires a fundamentally more distributed architecture. This, in combination with power and permitting constraints and the demand for AI data sovereignty is powering new Tier 2 markets—areas that offer power, dense connectivity, and proximity to populations. For instance, while Northern Virginia will clearly remain the world’s largest data center market, space and power constraints are pushing development into contiguous markets, such as the I-95 corridor toward Richmond. In Europe, traditional core hubs like London, Frankfurt, and Paris are being complemented by emerging southern markets like Milan and Madrid, offering dense high-speed connectivity and low-cost green energy. In Asia Pacific, traditional hubs like Singapore and Sydney are being caught up fast by new markets like Johor and Mumbai.

North America: what will the top 5 data center markets be in 2030?

1. Northern Virginia 8.5GW
2. Dallas, Texas 2.8GW
3. Phoenix, Arizona 2.7GW
4. Abilene, Texas 2.4GW
5. Atlanta, Georgia 2.3GW

North America will continue to be the undisputed leader in capacity, with expansion to a projected 27.2GW of critical IT capacity by 2030. 68% of this capacity - 18.7GW - is forecast to be built across the top five markets.

Northern Virginia (NoVA) is projected to reach 8.5GW of capacity by 2030, with spillover creating a new sizable hub known as the NoVA Extension (2.3GW by 2030). Other established hubs such as Dallas (2.8GW), Phoenix (2.7GW) and Atlanta (2.3GW) will also continue to see massive absorption. On the Training side, remote GPU “supercluster” locations are emerging in regions offering abundant, low-cost energy, like Abilene, TX (2.4GW).

Europe: what will the top 5 data center markets be in 2030?

1. London 2.7GW
2. Frankfurt 2.7GW
3. Paris 2GW
4. Milan 1.4GW
5. Madrid 1.2GW

10GW (75%) of a forecast total European capacity of 13.3GW by 2030 is expected to come from the top five markets.

While starting from a smaller base, infrastructure growth in Europe is forecast to grow even faster than North America through 2030 (24.7% CAGR versus 20.9%), with core hubs London (2.7GW), Frankfurt (2.68GW), and Paris (2GW) forecast to have the greatest capacity by 2030. Driven by power constraints, renewables availability, and data sovereignty regulations, a new generation of southern hubs are also emerging, including Milan (1.4GW), Madrid (1.2GW), and Lisbon (466MW).

Asia-Pacific: what will the top 5 data center markets be in 2030?

1. Tokyo 2.8GW
2. Sydney 2.4GW
3. Johor 2.2GW
4. Mumbai 2.15GW
5. Seoul 1.7GW

The top five markets in Asia-Pacific are forecast to account for 11.25GW (68%) of the region´s 16.5GW total by 2030.

Asia-Pacific covers a massive area and a huge number of extremely diverse markets. The current leading markets Tokyo and Sydney are forecast to maintain their position in 2030 with 2.8GW and 2.4GW respectively. Seoul (1.7GW) and Singapore (1.2 GW) will continue to grow fast, but not as fast as Johor (2.2 GW) and Mumbai (2.1 GW) which will overtake them. Johor provides a high-potential overspill market from strategically located but space-and-power-constrained Singapore, and Mumbai is the first of a large cluster of high-growth Indian hubs.

How reliable are current infrastructure predictions?

Building data centers currently takes years, not months, so the trends defining locations through 2028 are already clear. For AI inferencing, they are similar to, and built upon, the topography of the cloud. But there is still huge scope for change. These forecasts are based to a great extent on current pipelines, but with demand at such high levels, and a growing range of bottlenecks and delays, it will be interesting to see if existing markets can maintain their growth levels to the end of the decade, and if emerging hubs manage to achieve, or even exceed, their potential.

Find out more

You can find more detailed answers to these and other questions, including the forecast top 10 hubs per global region, in the new and comprehensive IMDC paper “Building at scale: Where we deliver accelerated infrastructure”