Redefining AI Data Center Power System - Independent Power Paths for Compute System and Cooling System: To meet 60KW-100KW /Rack AI densities with higher reliability and maintainability, the compute power and cooling power shall be electrically decoupled: 1. Compute: 13.8kV AC - Solid State Transformer - 800V DC backbone - Rack level 48V DC - on board VRMS. 2. Cooling System: Dedicated 400V AC(480V AC) or 400V DC Busbar, fed by Group A/ Group B redundant sources( UPS-backed) via STS; serves centralized pumps skids/CDUs/dry coolers/rear-door HEX loops for multiple racks. This separation reduces stress and heat on rack DC buses, improves fault isolation and serviceability, and aligns with high-density liquid-cooling practices. This summary outlines a new approach to AI data center electrical power design: separating compute and cooling power paths to improve efficiency and reliability. I haven't gone into the full implementation details here, but if this topic resonates with your current projects, I'd be very interested in exchanging ideas. Feedback and comments from operations engineers and R&D specialists working on next-generation AIDC architectures are very welcome. Let's discuss how this can be applied in practice.
What is the mechanism for the independent loads to coordinate during and respond to a system disturbance? Can they do so effectively with such low latency requirements? Do you have a paper attempting to model this coordination that accounts for communications latency?
The Data Centre has composite or hybrid loads. Different types with entirely different characteristics are aggregated in one place.
This is an interesting architecture, especially from a utility interconnection standpoint. You are effectively creating two distinct, large-scale load types. The compute path, fronted by a Solid State Transformer (SST), is a massive power electronic load. This provides opportunities for grid-supportive features (e.g., fast frequency response, voltage support, power factor control) but also introduces challenges in power quality (harmonics) and system modeling for transient studies. How do you see the dynamic response of this entire facility during grid disturbances, like a voltage sag? Managing the ride-through coordination between the SST-fed compute path and the UPS/STS-backed cooling path will be key.