Whitepapers and Guides / Power and Cooling for AI Racks
Guide
Power and Cooling for AI Racks
Power and cooling, not the chip, are increasingly the bottleneck for AI. A planning primer on what to confirm with your facility before you order hardware.
Why it matters
Rack power density has moved from roughly 10 kW in conventional deployments to well over 100 kW for current dense GPU racks. At that level, liquid cooling is increasingly a default requirement rather than an option, and the facility often decides what you can deploy.
What to confirm with your facility
- Power available per rack, and the redundancy you need
- The cooling method on offer: air, rear-door heat exchangers or direct-to-chip liquid
- Floor loading and weight limits for dense racks
- Power distribution: PDUs, connectors and circuit capacity
- Space for cabling, especially for high-bandwidth networking
- Monitoring for power, temperature and coolant
- Lead times for any electrical or mechanical works
Planning tips
- Start with the facility, then choose the hardware that fits it
- Verify the real power draw of your target configuration with the supplier
- Plan for the next hardware generation, which will likely draw more
- Consider energy efficiency, because power is an ongoing cost
- Ask early about colocation options if you cannot build your own room
How Embedica helps
Our data centre work begins with a site and power and cooling readiness assessment, before any hardware is ordered, so the build you specify is one your site can actually run.
Next steps
Related: Data Centre Infrastructure · Energy-Efficient Computing · Enclosures and Chassis
General guidance only; requirements depend on your workload, site and suppliers. Company and product names are trademarks of their respective owners.
Want this worked through for your own workload?
