Power requirements of AI servers: what modern data centres need to know
Artificial Intelligence is rapidly transforming data centres. Where traditional server racks once operated at around 5–10 kW, modern AI environments are pushing far beyond that, often reaching 30 kW, 60 kW or even over 100 kW per rack.
This shift is not just about compute. It fundamentally changes how power is distributed, monitored and managed within the data centre.
Understanding the power requirements of AI servers is therefore essential for ensuring uptime, efficiency and scalability.
HPC & AI
Power requirements of AI servers | Data centre power guide.
How much power do AI servers use?
AI servers consume significantly more power than traditional IT equipment, primarily due to the use of GPUs and high-performance accelerators.
Typical ranges include:
• Traditional servers: 300–800 W per server
• GPU servers: 2–10 kW per server
• AI racks: 20–100+ kW per rack
Modern AI platforms, including systems from NVIDIA, AMD and GPU-based servers from manufacturers such as Supermicro, are driving these increases.
Depending on configuration, a single system can draw several kilowatts continuously, resulting in extremely high rack densities.
Why AI changes power distribution completely
AI workloads introduce a fundamentally different power profile compared to conventional IT environments:
• Higher power density
• Three-phase power as standard
• Dynamic and fluctuating loads
• Strong dependency between power and cooling
As a result, traditional power distribution strategies are no longer sufficient.
Key challenges in AI server environments
Overload risk
High-density racks significantly increase the risk of exceeding circuit limits, potentially leading to downtime or hardware damage.
Phase imbalance
Uneven load distribution across phases reduces efficiency and puts unnecessary strain on infrastructure components.
Lack of visibility
Many data centres still rely on rack-level or PDU-level monitoring. For AI workloads, this is no longer enough.
Outlet-level monitoring is becoming essential to gain accurate insight into power usage.
Complex rack configurations
AI infrastructure often combines different systems within a single rack, such as:
• Fast data refresh rates for real-time monitoring
• Integration with DCIM and monitoring platforms
In addition, modular designs allow data centres to adapt functionality as requirements evolve.
The role of monitoring and software
As power demands increase, visibility becomes critical.
Modern data centres rely on monitoring solutions that offer:
• Real-time data insights
• Historical analysis
• Alerts and threshold management
• Integration with DCIM platforms
These capabilities help optimise energy usage and support compliance with regulations such as CSRD and broader energy efficiency requirements.
For broader industry guidance, organisations such as the Uptime Institute and ASHRAE provide valuable best practices for power and cooling design.
Conclusion
AI is redefining data centre infrastructure. What was once primarily a compute challenge is now equally a power challenge.
Without the right approach to power distribution, monitoring and control, AI deployments cannot scale safely or efficiently.
By understanding the power requirements of AI servers and designing infrastructure accordingly, data centres can support high-density workloads while maintaining reliability and control.
European Commission – Energy Efficiency: https://energy.ec.europa.eu/topics/energy-efficiency_en
