The power crunch: How energy constraints reshape datacentre strategy

The rapid adoption of artificial intelligence (AI) is colliding with a hard physical limit. Notably, that limit is not necessarily compute capacity or silicon availability, but electricity.

For decades, technology leaders treated power as a basic utility. You plugged in your servers and the grid delivered. Yet in an AI-driven world, power is fast becoming the defining constraint on digital growth. 

Consider the fundamental difference between standard cloud workloads and generative AI. 

A standard internet search requires a fraction of the energy needed to process a single query through a large language model. Multiply that single interaction by billions of users worldwide and the true scale of the problem becomes clear. The computational density of AI workloads is creating highly concentrated energy demands that regional electricity grids were not built to handle.

We are already witnessing the consequences of this shift. Gartner predicts that by 2026, power shortages will delay more than 30% of all datacentre expansions. By 2028, datacentres will account for 10% of all US electricity demand. This is forcing a strategic reset en masse for IT and infrastructure & operations (I&O) leaders.

The emergence of the tech energy company

The scale of the challenge is rather staggering. Some proposed hyperscale datacentres will consume up to 5GW of power – enough to power roughly 3.5 million households. When you concentrate that much demand into a few square miles, existing commercial grids can buckle. More recently, this is combining with increasing friction with local communities who fear grid instability and rising utility costs.

Grid operators work on 20-plus year planning cycles, yet excitement for AI evolves and transforms month-to-month. Because the public grid cannot keep up, major technology companies are taking matters into their own hands. Constellation and Microsoft have committed to restart the Three Mile Island nuclear plant by 2028, whilst Amazon has unveiled plans to invest $20 billion into a datacentre campus near a Pennsylvania power plant.

These organisations are securing "behind the meter" power because they must. On-site power generation is transitioning from a redundancy to a primary source of power. Consequently, IT and I&O leaders increasingly evaluate alternative energy sources to guarantee future capacity. Interest is growing in small modular reactors that promise safer and more flexible nuclear power deployment, whilst natural gas microgrids are being utilised to bridge the immediate gap while renewable technologies mature.

Cooling the AI engine

Generating power is only one part of the equation. Organisations must also radically optimise how they consume it. New construction facilities will house server racks that generate unprecedented levels of heat. Traditional air cooling physically fails at these new high-density thresholds and is no longer viable for high performance computing.

Liquid cooling technologies are now the standard for AI clusters, with systems that circulate liquid directly to (and away from) high heat components for better thermal performance. Implementing advanced techniques like rear door heat exchangers significantly reduces operational costs. It represents a fundamental shift in facility engineering.

Ironically, while AI is driving the need for more cooling, it can also provide the solution. AI algorithms can analyse real time sensor data for workload and temperature to predict future cooling needs dynamically. This intelligent automation can result in up to a 40% reduction in cooling energy consumption.

The role of sustainable software optimisation

The conversation around datacentre energy often ignores the software layer entirely. However, developers play a crucial role in managing this power crisis. Training large AI models is incredibly energy intensive. The industry must shift its focus toward writing more efficient code and right sizing models for specific tasks.

By optimising algorithms to require fewer computational cycles, organisations can significantly reduce the underlying power draw of their applications. Choosing the most efficient software tool for the job is increasingly a vital energy decision.

The edge computing advantage

The industry has spent years centralising data processing to achieve economies of scale, yet power constraints are reversing that trend. We have reached a tipping point where the immense size of centralised facilities is becoming a liability.

Edge computing offers a strategic alternative. IT leaders and I&O leaders need to distinguish between training AI models and actually running them. While model training requires massive, centralised power clusters, day-to-day inferencing can often happen locally. 

By deploying more localised edge data centres, organisations can move data processing closer to the source. This decentralisation spreads the electrical demand across multiple local power grids instead of overwhelming a single region. These purpose-built facilities also require less energy to transmit data and can often utilise highly efficient free cooling techniques.

Navigating the energy transition

The organisations that succeed in the AI economy will be those that recognise energy as a strategic variable rather than a facilities afterthought. Treating power availability as a given is a recipe for stalled growth. Those who fail to adapt their site selection and architecture design risk giving up competitive ground.

The next phase of IT infrastructure will be defined by how well organisations can secure reliable power and intelligently distribute workloads. IT and I&O leaders must bring their facility managers and software developers to the same table. Those that get this alignment right will lead the next generation of digital innovation.

Gartner analysts will further explore data centre power priorities and challenges at the Gartner IT Infrastructure, Operations & Cloud Strategies Conference, taking place in London, from 16-17 November 2026.