Retrofitting existing power systems for AI clusters
- By Victor Avelar, Stuart Sheehan & Allegia (Gia) Wiryawan
- 16 Dec 2025
- 4 min read
The explosion of AI workloads is fundamentally reshaping data center design. Modern AI clusters consume megawatts of power and push legacy infrastructure to the breaking point. Most data centers were designed for workloads with rack densities ten times lower than today's requirements. Retrofitting existing power systems isn't optional – it's a strategic imperative.
As organizations scale AI capabilities, they must ensure power systems can handle unprecedented density, synchronicity, and variability without compromising reliability or sustainability.
AI workloads stress power systems differently than traditional IT applications. Beyond training large language models, inference, fine-tuning, and emerging "long thinking" operations all strain infrastructure in unique ways.
Detailed power profiles remain under development. Because AI research evolves rapidly, power system design is a moving target. Until empirical data solidifies, data centers must plan for worst-case powerscenarios to ensure readiness.
Five attributes define how AI workloads challenge infrastructure:
- Synchronous computationsimultaneous peaks across thousands of accelerators
- Thermal power designintense heat generation and cooling demands
- Peak powershort, high-intensity consumption bursts
- AI cluster sizescaling effects that magnify load and distribution challenges
- Network latencycoordination delays that worsen load synchronization
Combined with limited load diversity, these factors make legacy power paths vulnerable to overloads and safety hazards.
Number 1
Conduct comprehensive load studiesPerform facility-wide load assessments to identify spare capacity, vulnerabilities, and upstream power limitations. These establish the foundation for retrofit planning and “headroom” design.
Number 2
Deploy power quality metering and monitoringUse electrical power management systems and advanced metering software to track real-time conditions, detect harmonics, and manage load transitions proactively.
Number 3
Mitigate overloads and step loads- Software-based power control: Limit peak consumption through accelerator management
- Energy storage: Use batteries or supercapacitors to smooth idle-to-peak transitions
- UPS optimization: Upgrade controls and batteries for frequent power fluctuations
- Infrastructure scaling: Retrofit UPSs, switchgear, and generators for AI peak loads
Number 4
Increase distribution block sizesUpgrade PDUs and RPPs with higher-rated breakers (800+ A) to handle density increases. Optimize rack symmetry and use validated reference designs that consider redundancy, voltage, and efficiency.
Number 5
Limit arc flash hazards- High-impedance PDU transformers and current-limiting breakers reduce fault current
- Line reactors add impedance where necessary to ensure compliance and safety
Number 6
Transition to higher-voltage systemsMigrate from 120/208 V to 240/415 V to improve power efficiency and reduce conductor mass. Consolidate transformers and standardize designs for multi-rack clusters.
Number 7
Expand rack PDU capacityDeploy high-capacity PDUs (125+ A) or more) and regional variations suited to local electrical codes. Use extension kits or OCP busbar systems for compact distribution in constrained spaces.
Number 8
Manage harmonics and cooling dependenciesPlace CDU pumps on dedicated UPS with true double-conversion technology. Install active harmonic filters to stabilize waveforms and protect downstream systems.
Designing for what's next
Retrofitting for AI is an iterative process. Facilities adopting validated reference designs, continuous monitoring, and scalable electrical architectures will remain adaptable as AI models evolve.
By combining robust electrical engineering with proactive monitoring and safety measures, data centers can meet the extreme demands of AI clusters while future-proofing operations. The question isn't whether to retrofit – but how effectively and urgently to implement necessary changes.
To explore the full spectrum of challenges and retrofit strategies for AI-ready data centers, download the complete white paper: "Retrofitting Existing Power Systems for AI Clusters". This in-depth guide from Schneider Electric offers practical insights, validated reference designs, and actionable recommendations to help you future-proof your infrastructure and support the next generation of AI workloads with confidence.
Latest in AI & Technology
A cybersecurity blueprint: Securing critical infrastructure through zones of influence
Bridging data and AI for success: Why it matters for businesses
Transforming data center services: AI-driven condition-based maintenance
The looming power crunch: Solutions for data center expansion in an energy-constrained world
