Artificial intelligence was proposed as a tool for reducing data-centre energy use in April 2017, as connected devices and digital services placed greater demands on servers and cooling systems.
Enterprise provider Aegis Data argued that automation, monitoring and prediction could help operators match power and cooling to real requirements. The company also warned that AI was still immature and needed rigorous research before facilities could depend on it.
That balance between opportunity and operational caution remains central to sustainable computing.
Cooling Was a Major Efficiency Target
Servers turn electricity into computation and heat. Data centres therefore consume additional energy moving that heat away from equipment and maintaining reliable temperatures.
Operators traditionally tracked power usage effectiveness, or PUE. The metric divides total facility energy by the energy used by IT equipment, with a theoretical ideal of 1.0.
A lower PUE indicated less overhead, but it did not describe whether servers were doing useful work. An efficient building full of underused machines could still waste substantial electricity.
Natural Conditions Could Reduce Demand
A Rackspace facility opened near West Crawley in 2015 demonstrated how design choices could improve performance. It reported a PUE of 1.15 and used outside air for cooling, taking advantage of the British climate.
The design was contributed to the Open Compute Project, allowing others to learn from it. That approach showed that sustainability was not solely an AI problem. Location, airflow, equipment layout and open engineering all mattered.
Software could optimise a facility, but it could not compensate indefinitely for poor physical design.
DeepMind Offered an Early Example
Google had applied DeepMind technology to its data-centre cooling systems and reported a 15 percent improvement in overall energy efficiency. Sensors provided information about temperature, power and equipment, while models recommended more efficient control settings.
The example made a compelling case for machine learning. Cooling involved many interacting variables, and conditions changed continually. A model could find combinations that fixed rules or human operators might miss.
Reported savings still needed context, including the facility baseline, measurement period and additional computing required by the optimisation system.
Automation Introduced New Operational Risks
Data centres supported critical services, so an efficiency experiment could not compromise uptime or damage hardware. A faulty recommendation might save energy briefly while creating heat or pressure conditions outside safe limits.
Aegis therefore stressed testing before AI replaced human interaction. Operators needed limits, fallback controls and the ability to understand why a system changed a setting.
The safest path was gradual: begin with recommendations, validate them against engineering knowledge and expand autonomy only when performance was demonstrated across seasons and unusual events.
Better Metrics Extended Beyond PUE
Professor Ian Bitterlin of the British Computer Society had criticised treating PUE as a general measure of data-centre quality. He argued that improving server effectiveness was the only way to improve overall effectiveness.
That critique highlighted the danger of optimising a convenient metric. Facilities also needed to consider hardware utilisation, workload efficiency, carbon intensity, water consumption and equipment lifespan.
AI could support those goals, but only if its objective reflected them. A system told to minimise electricity cost might shift work to a time when power was cheaper but more carbon-intensive.
Always-On Monitoring Could Find Waste
Continuous analysis could identify unusual consumption, predict equipment failure and adjust capacity before demand changed. It could also reveal servers that were powered but doing little work.
Those capabilities reduced the burden of manually interpreting thousands of sensor readings. They also created a new dependency on accurate sensors, clean data and secure control systems.
Cybersecurity was part of sustainability: a compromised management platform could disrupt operations or manipulate settings, erasing efficiency gains.
AI Was One Part of a Larger System
IBM and other providers were also experimenting with cognitive technologies in their facilities and services. The industry’s direction was clear: data-centre operation would become more software-driven.
Yet the most effective programme combined several measures. Efficient processors, workload scheduling, renewable electricity, heat reuse and careful cooling design could work alongside predictive models.
Operators also needed transparent reporting. Efficiency claims were more meaningful when accompanied by absolute energy use and emissions.
The 2017 Caution Was Well Placed
AI offered a practical way to understand complex facilities and make more precise adjustments. Google’s early result showed that significant savings were possible without waiting for entirely new buildings.
The technology did not eliminate engineering discipline. Models had to be tested, bounded and monitored, and their goals had to reflect the full environmental footprint.
The lasting lesson was that greener data centres would come from combining intelligent control with sound physical design and better measurement. AI could help operators see and reduce waste, but sustainability depended on the system around the model.
