Everyone wants to talk about large language models, reasoning capabilities, and multimodal intelligence. Nobody wants to talk about data centre cooling costs.

But the boring infrastructure stuff is increasingly what determines who can actually deploy AI at scale.

The Power Problem

Training and running large AI models consumes enormous amounts of electricity. A single GPU cluster for model training can draw megawatts of power—enough to run a small town.

Data centre operators are scrambling to secure power capacity. In some regions, new AI deployments are delayed not by software or hardware, but by the time required to bring additional electrical capacity online.

This isn’t a temporary constraint. As models get larger and inference demand grows, power requirements increase. The hyperscalers are signing long-term power purchase agreements and, in some cases, investing in dedicated energy generation.

For enterprises running AI workloads in colocation facilities, power availability is becoming a procurement consideration. “Can this facility support our planned AI infrastructure?” isn’t a question IT teams were asking five years ago.

The Cooling Reality

Every watt of power consumed by AI hardware becomes heat that must be removed. Traditional air cooling struggles with the thermal density of modern AI chips.

Data centres are investing in liquid cooling systems, but retrofitting existing facilities is expensive and disruptive. New builds incorporate advanced cooling from the start, but they take years to plan and construct.

The result: a physical constraint on how quickly AI infrastructure can expand. You can order GPUs today; you may not have a place to put them that can actually keep them running at optimal temperature.

The Talent Gap

Not the AI researcher talent gap—that’s well documented. The infrastructure talent gap.

Running AI systems at scale requires engineers who understand both traditional IT infrastructure and the specific demands of AI workloads. Storage architects who know how to feed training data to GPU clusters fast enough. Network engineers who can handle the bandwidth requirements of distributed training. Operations teams who can manage novel hardware with less mature tooling.

This hybrid skillset is rare. Traditional data centre operators are upskilling, but the learning curve is steep. Companies are competing fiercely for people who can actually run AI infrastructure in production.

Supply Chain Fragility

Advanced AI chips are manufactured by a handful of foundries, primarily in Taiwan. The geopolitical implications of this concentration are increasingly discussed, but the practical implications are already apparent: long lead times, allocation constraints, and pricing power resting with suppliers.

Companies planning AI deployments a year out are discovering that hardware delivery timelines don’t match their project schedules. The second-tier hardware market—used or previous-generation equipment—is becoming more active as organisations seek alternatives to 18-month waits for new GPUs.

What This Means

For most organisations, these constraints are indirect—they affect the cloud providers and infrastructure companies that enterprises depend on. But the effects ripple outward.

Cloud AI service pricing reflects infrastructure scarcity. Availability of compute capacity in preferred regions isn’t guaranteed. Enterprise AI projects may stall waiting for resources that don’t exist yet.

The companies navigating this well are:

Planning infrastructure earlier. Not treating compute as infinitely elastic, but forecasting AI workload growth and securing capacity proactively.

Diversifying providers. Not betting everything on one cloud provider’s ability to deliver capacity.

Considering edge and hybrid deployments. Not everything needs to run in hyperscale data centres. Smaller models running on local infrastructure might avoid the bottleneck.

Investing in efficiency. Model optimisation, quantisation, and efficient inference reduce resource requirements without sacrificing capability.

The Unsexy Truth

The AI revolution is constrained by physics and supply chains as much as algorithms. Progress on foundation models outpaces the infrastructure to run them.

This isn’t pessimism—it’s planning reality. The organisations that account for infrastructure constraints in their AI strategies will be better positioned than those that assume compute appears on demand.

Sometimes the limiting factor isn’t intelligence. It’s electricity.