GPU POWER DENSITY
An 8-GPU H100 cabinet draws 20-25 kW (peaks to 30 kW) vs CPU cabinet at 5-8 kW. Costs: power circuit ($200-600/month per 30A), per-kWh ($0.08-0.25/kWh), power density premium ($500-2,000/month per cabinet over 15 kW).
For 32 GPU H100 deployment (4 cabinets, 100 kW): $1,600 circuits + $8,640 energy + $4,000 density premium = $14,240/month. Power is the single largest colocation cost, 3-5x space.
| Component | CPU Cabinet (8 kW) | GPU Cabinet (25 kW) | Multiplier | Annual Cost 32-GPU |
|---|---|---|---|---|
| Space | $400-800/mo | $400-800/mo | 1x | $4,800-9,600 |
| Power circuit | $200-400/mo | $800-1,200/mo | 3x | $9,600-14,400 |
| Energy | $0.08-0.15/kWh | $0.10-0.25/kWh | 1.5-2x | $86,400-180,000 |
| Density premium | $0 | $500-2,000/mo | N/A | $6,000-24,000 |
| Cross-connects | $200-400/mo | $800-1,600/mo | 4x | $9,600-19,200 |
| Total monthly | $1,900-3,500 | $11,300-21,100 | 3-6x | $135,600-253,200 |
INTERCONNECT COSTS
InfiniBand fabric for 8-node cluster (64 GPUs): 8 HDR/NDR switches ($15-40K each), 64 transceivers ($200-800 each), 64 cables ($100-500 each). Total interconnect hardware: $150-400K. Cross-connect fees: $1,600-6,400/month.
DGX H100 4-node cluster colo costs: power $8-12K, space $1.6-3.2K, cross-connects $0.8-1.6K, remote hands $0.4-0.8K = $10.8-17.6K/month. Maintenance: $8-20K/year for 4-hour support.
TCO: COLOCATION VS CLOUD VS ON-PREM
3-year TCO for 32-GPU H100: On-prem $2.13M ($1.85/GPU-hr at 85% util), Colo $1.83-2.07M ($1.59-1.80/GPU-hr), Cloud on-demand $4.10-4.92M ($3.56-4.28/GPU-hr), Cloud reserved $2.74-3.28M ($2.38-2.85/GPU-hr).
Colocation beats cloud at >60% utilization. Below that, cloud elasticity is cheaper. Emerging hybrid colo with on-demand burst in same facility at $3.00-4.50/hr with <1 ms latency to colo baseline. Hidden cost: power capacity lead times of 12-26 weeks in major metros.