Global AI data center delivery: Some big questions & answers

Global AI delivery: 9 key questions & answers for IT infrastructure buyers

Push is coming to shove in the world of accelerated infrastructure delivery. According to BloombergNEF, over 23 GW of capacity is being built worldwide, with over $100 billion of leases signed in the last six months. Capex estimates for 2026 leapt by 56% between August 2025 and February 2026, to over $800 billion. This raises many questions for infrastructure buyers.

Here are a few of the key ones, along with topline answers taken from the latest Iron Mountain Data Centers paper “Building at Scale: How We Deliver Accelerated Infrastructure”

1. Can operator supply keep up with AI demand?

Supply is up, but demand is even greater. About 20% of total data center capacity is already used for AI, and estimates suggest AI will drive $75 billion in demand by 2028, increasing its market share to 35%. This is on top of growing cloud and enterprise demand. To put this in perspective, McKinsey notes that to avoid a deficit, at least twice the capacity built since 2000 must be constructed in less than a quarter of that time. So the short answer to this question is no, construction cannot keep up with demand; what we are seeing is a best effort in a booming market.

2. What sort of data centers does the market need?

Put simply, everyone from enterprises, service providers, clouds and neoclouds to hyperscalers is buying facilities designed around GPUs rather than CPUs. While traditional cloud storage remains relevant, AI workloads demand exponentially greater power, new types of power distribution, and advanced cooling. With such widespread demand, there is no single type or scale: facilities range from vast superclusters like Stargate, Hyperion, or Project Rainer, to a huge range of medium-sized inference data centers.

3. What will the next generation of AI facilities look like?

Categories are emerging, but for buyers the key distinction is between training and inference facilities. Training facilities are mega-scale (100 MW - 1 GW+) and can be remote, while inference facilities are medium to large (5-100 MW) and must be located close to end-users to maintain low latency (<50 ms). Inference is forecast to grow at a 79% CAGR to 2030, eventually becoming the dominant form of AI infrastructure. So, to a great extent, next generation infrastructure will mean inference infrastructure.

4. Where are data center operators building fastest?

The burgeoning inference infrastructure market is being located closer to end users, built in and around existing metropolitan availability zones. Northern Virginia will remain by far the largest hub, expected to reach 8.5 GW by 2030. Other major global hubs include London, Frankfurt, Paris, Tokyo, and Sydney, with fast-growing emerging hubs in cities such as Johor and Mumbai. Distribution will be key, driven by power availability. According to Structure Research, by 2030 there will be more than 25 Gigawatt-plus hubs worldwide (there are just 5 today) with total capacity around the 60 GW mark.

5. How are data center operators scaling up to meet demand?

The scale and spread of demand involves significant capital commitment; Iron Mountain Data Centers, for instance, is partnering in emerging markets like India and doubling its global development pipeline to 2.6 GW over the coming years. Proactive site selection and managing grid constraints through Behind the Meter (BtM) energy solutions is also becoming critical to deal with and address grid constraints and support the energy transition. In line with cloud business models, the more forward-looking operators are extending their capabilities deep into the energy sector via new grid collaborations and technologies.

6. How has data hall design changed to address AI customer needs?

Design has shifted toward large-scale campus-style developments and liquid cooling architectures. GPU racks require 10 times more power density than CPUs, with densities reaching 200 kW/rack. This necessitates upgraded UPS controls and flexible power redundancy models, some of which also move away from traditional 2N standards for certain AI workloads. Look for quick design and construction refresh rates as higher densities will require regular power and cooling technology refreshes.

7. How can data center operators accelerate delivery while reducing risk?

This is a tough balance, where experience counts. Capital commitment, specialized functional teams and strong supply chains are vital. Successful delivery depends on booking equipment early, standardizing components, and leveraging global platforms to fill talent or component gaps. A solid track record, good accessible senior talent, and a strong local market presence and reputation continue to be the key factors.

8. Is building at scale undermining data center operator standards and sustainability?

Related to risk reduction is the question of maintaining standards and supporting customer GHG and broader sustainability targets. This varies across the market, but while some operators will undoubtedly deploy fossil fuels as a bridging solution, standards can assist building at scale, and sustainable build certifications like BREEAM and LEED are more important than ever to protect the AI provider's right to operate. PUE and WUE will be non-negotiable with possible gains from closed-loop cooling. New granular GHG protocols will also drive local-grid renewables with 100% clean power tracked hour-by-hour becoming mainstream.

9. What are the key features to look for in an AI infrastructure provider?

Broadly speaking, the infrastructure provider market is responding very well to the AI-driven surge in demand and new design criteria. The best operators offer a mix of ambitious investment, strategic site selection focused on advanced power and cooling solutions, standards-based design for agility, and efficient delivery through diverse supply chains. Probably the most important characteristic in the current market is high responsiveness to changing customer needs. Specifications change and procurement needs to adapt quickly to address this. Flexible design will be particularly critical as the GPU technology cycle shortens.

Get more details

Want to know more? You can find more detailed answers to these and other questions in the new and comprehensive IMDC paper “Building at scale: How we deliver accelerated infrastructure”