Table of Contents
Most GPU power numbers you’ll read are TDP — the rated maximum. Real sustained AI inference often hits TDP within a few percent. For 24/7 operation that’s a different cooling and power problem than gaming or training-burst workloads.
RTX 5090 sustained AI draw: ~550-575 W. RTX 6000 Pro: ~600 W. Cold-aisle target: 30 °C max. N+1 power feeds. Don't run consumer GPUs in a closet without proper datacenter cooling — first thermal event ends the deployment.
Real power draw
| GPU | Rated TDP | Sustained AI inference | Burst peak |
|---|---|---|---|
| RTX 3050 | 130 W | ~125 W | ~140 W |
| RTX 3060 12 GB | 170 W | ~165 W | ~175 W |
| RTX 4060 | 115 W | ~110 W | ~125 W |
| RTX 5060 Ti | 180 W | ~175 W | ~190 W |
| RTX 3090 | 350 W | ~340 W | ~420 W (bursts) |
| RTX 4090 | 450 W | ~430 W | ~480 W |
| RTX 5080 | 360 W | ~345 W | ~380 W |
| RTX 5090 | 575 W | ~555 W | ~620 W |
| RTX 6000 Pro | 600 W | ~580 W | ~620 W |
| A100 80 GB SXM4 | 500 W | ~480 W | ~520 W |
Cooling envelope
For sustained 100% load:
- Cold-aisle temperature: 25-30 °C target. Above 30 °C, cards thermal-throttle.
- Airflow: 4U chassis with 3+ rear-mounted blowers. Front-to-back airflow.
- Inter-card spacing: 2-slot gap between adjacent GPUs in a multi-GPU build.
- Power supply: 1.5× peak power for headroom. 5090 server needs 1200W+ PSU.
Rack layout
- Power per rack: a single 8× 5090 server draws ~5 kW. Most racks deliver 6-10 kW max.
- Cabling: 12VHPWR (16-pin) on Blackwell — verify proper seating, watch for melting on misaligned connectors.
- UPS: 5-minute UPS for graceful shutdown on power loss.
- N+1 PSU: dual PSU with automatic failover.
Verdict
Power and cooling decide GPU lifespan more than anything else. A 5090 in a poorly-ventilated closet thermal-throttles within hours; in a proper datacenter chassis it runs at full speed for years.
Bottom line
This is why dedicated datacenter hosting beats DIY for sustained AI loads. Power, cooling, and 24/7 ops are real specialised problems.
