Aligning AI Acceleration with Cloud Capacity

Iron Mountain Data Centers and Structure Research have been closely examining the major trends shaping cloud and AI delivery. In our latest industry outlook white paper, “Will AI Eat the Cloud?”, we explore these emerging dynamics in depth. These findings suggest that the trends will keep both us - and the broader industry - extremely busy over the coming years.

Here are a few of the conclusions we came to:

• AI infrastructure will account for almost half of global colocation revenue by 2030
• Inference centers will overtake training centers in the next year
• Artificial Intelligence is getting exponentially cheaper
• The integration of training with inference is defining a new AI regional zone architecture
• 2 GW hubs will be the new normal worldwide by 2030

The data center paradox: AI is set to dominate infrastructure, not cloud revenue, by 2030

The long-term vision and investment behind the growth of AI would stress most capital markets, but a significant amount of this investment is being funded through the hyperscalers’ balance sheets and several private equity firms.

The hyperscalers’ total annual capital expenditures are projected to reach $375 billion this year, a 36% increase from $275 billion in 2024, which itself saw a 64% increase from 2023. The top four hyperscalers account for approximately 90% of this spend, with commitments continuing to rise as we approach year-end.

AI services are on a powerful trajectory, mirroring the rapid rise of the cloud itself. ChatGPT, for example, reached 1 million users in just 5 days, a feat that took Netflix five years. The steepness of this curve has huge implications for the data center industry. Demand for data center capacity is projected to outstrip supply from 2027-2030, with annual global demand reaching nearly 90 GW by 2030, exceeding supply by as much as 500%.

AI revenues, meanwhile, still have a very long way to go to compete with traditional cloud revenue. Microsoft Azure data indicates that AI currently has an $11-12 billion run-rate (around 6% of total Azure revenues), with quarterly growth of 10-11% and an overall Azure AI service stream growth of 35% as of Q1 2025. Goldman Sachs forecasts that Generative AI alone will account for 10-15% ($200-300 billion) of the total $2 trillion cloud revenue by 2030.

Even by 2030, that revenue level stands in stark contrast to the infrastructure investment required to enable it. Infrastructure precedes revenue, and AI will account for half of infrastructure (measured by power) long before its revenues match those of the traditional cloud. By 2030, while still accounting for only 10-15% of the cloud´s global $2 trillion revenue, however, AI is expected to account for approximately 44% of the total global data center colocation market revenue.

Inference overtakes training

The nature of the infrastructure is also evolving. As AI applications and inference workloads accelerate, we are seeing a significant shift in the balance between training and inferencing.

In terms of critical IT capacity, AI inference is projected to grow at a 79% CAGR to 2030, far outpacing AI training’s 25% CAGR over the same period. Inference capacity is expected to overtake training capacity next year, and by 2030, AI inference will account for 80% of total AI critical IT load capacity. This is a complete reversal of the 2023 position, an exceptionally quick turnaround, which should already be a key element in our pipeline planning.

Cheaper intelligence and what it means

Fortunately, while demand for high-performance GPUs is unprecedented, and training model costs fluctuate, the average cost of artificial intelligence is rapidly decreasing. The cost of the cheapest Large Language Model (LLM) with a minimum Massive Multitask Language Understanding (MMLU) score of 83 has decreased approximately tenfold annually from 2022 (GPT-3) to 2024 (Llama 3.2 3b).

In line with the Jevons paradox, this reduction in AI training and inference costs will not reduce consumption; instead, it will fuel an exponential increase in AI demand, in the same way that increased efficiency has driven market development for everything from cars to memory hardware and network switches.

The new AI regional availability zone architecture

We are also looking at a new Availability Zone (AZ) architecture, with new AI infrastructure deployed adjacent to cloud infrastructure. This includes Extension Availability Zones with much larger capacities than original AZs, optimized for both latency-sensitive inference workloads and less latency-sensitive satellite inference + training workloads. In these redesigned availability zones, the seamless interdependence of cloud and AI infrastructure highlights their mutual reliance. Retrofitting global cloud regions to become AI factories requires not only GPUs but also significant storage and regular compute for agents, demonstrating that existing cloud infrastructure is more critical than ever.

2 GW hubs - the new normal by 2030

The shift from latency-insensitive training centers (location agnostic and does not need to be near existing cloud markets) to latency-sensitive inference centers (needs to be near existing cloud markets) will place immense development pressure on existing densely populated "eyeball-rich" hubs, as well as requiring the creation of new hubs.

Northern and central Virginia will continue to lead, with over 10 GW of capacity by 2030. This will be followed by Dallas (2.8 GW) and Phoenix (2.7 GW) will see massive absorption, matching capacity levels in Tokyo (2.8 GW), London, and Frankfurt (2.7 GW), followed by newcomer Abeline (2.4 GW). Sydney (2.4 GW), Atlanta (2.3 GW), Johor (2.2 GW) Chicago (2.1 GW), Mumbai (2.1 GW) which, together with Paris, are all on track to surpass the new 2 GW “mega hub” benchmark.

Take a deeper dive

These are just a few observations. Our white paper offers a wider analysis showing how AI will accelerate cloud growth by increasing standard compute demand. Rather than being displaced by AI, the cloud will serve as the essential foundation—delivering the vast storage capacity and interconnected infrastructure that next-generation AI services depend on.

AI is fuelling demand for a new layer of accelerated data infrastructure that is supercharging the industry, changing data center locations, introducing new compute and cooling architectures, and igniting aggressive investment and power and land banking strategies. It´s an exciting time. Only the most ambitious, far-sighted, and deep-pocketed operators will be in a position to keep in step with demand for both advanced AI infrastructure and the expanding cloud it is built on.

You can download the full paper here