The future geography of AI infrastructure

Spheres of inference

While the tech refresh rate for chips appears to be speeding up, there are some parts of the accelerated computing revolution that are easier to predict. Because building takes time, data center infrastructure is one of them.

This Iron Mountain Data Centers (IMDC) ebook gives a quick overview of the factors that are driving and shaping the expanding infrastructure footprint that will support high-performance hyperscale and neocloud services around the world.

The new drivers

The adolescence of AI

While a lot of complex AI models are still being developed, the first generation of LLMs and agents are now out on the street and earning, having graduated through training to production, aka inference. By its nature, inference will require much more capacity than training, as it will be round-the-clock and cumulative. To support this coming of age, the next few years will see a wholesale shift in the investment in high-density liquid-cooled large-scale infrastructure, away from (but maintaining) training, towards less “mega” more distributed end-user-focused inference hubs.

• The scale of the challenge: According to McKinsey “to avoid a deficit, at least twice the data center capacity built since 2000 would have to be built in less than a quarter of the time.”
• Growing AI dominance: About 20% of total data center capacity is already being used for AI, and Cushman & Wakefield estimates that AI will drive $75 billion in data center demand by 2028, pushing this share up to 35% of the total market. By 2030 this figure is likely to be 50%.

Power, production & people

Power availability is the primary dictator, and limiter, of AI ́s growth, with particular pinch points in and around densely populated mature markets, struggling with ageing infrastructure and the energy transition. This puts these areas, the traditional locations for clouds and colocation, at a premium.

• To support enterprise AI growth, which is forecast to run at 84.9% over the next five years (Structure Research), AI facilities need to be integrated into existing cloud regional architecture, with latency-sensitive inference workloads embedded next to existing cloud on-ramps.
• Cloud zones are getting significantly bigger, and a new regional architecture is emerging. This includes extensions to Availability Zones with much larger capacities. These are optimized for both inference workloads, which are very latency sensitive, and satellite inference + training workloads which are not so latency sensitive.
• Where the strain on existing generation and transmission is too high, fast-growing “spillover” hubs are being formed, where power can be provided in reasonable proximity to existing cloud availability zones.
• AI data sovereignty is now a major geopolitical issue, making training as well as inference investment more geographically diverse.

To access the full content, click on "Download Resource" button below.