GPU Colocation Pricing Guide 2026: Power, Space, and Interconnect Costs

GPU colocation pricing in 2026 is composed of three independent line items: power (metered or flat-rate per kW), space (per rack unit or full cabinet), and cross-connects (fiber or copper per port per month). A single DGX B200 cabinet in a Tier 3 facility runs approximately $3,800/month for a 30kW committed power draw, $1,200/month for a full 47U cabinet with 2-post rail, and $450/month per 100G cross-connect to the nearest IX or direct cloud on-ramp.

The industry trend in 2026 is toward power-only pricing models where the colo charges solely for delivered kW and the space is included. Equinix, Digital Realty, and CyrusOne offer these contracts at $140–$200 per kW-month for 20–50kW cabinets. Above 50kW per cabinet, the rate drops to $110–$160/kW-month because the utilization guarantees offset the overhead of dense power distribution.

A DGX H100 draws a sustained 10.2kW under full load (8 x 700W GPUs plus CPU/switch overhead). At $170/kW-month colo pricing, the monthly power cost per DGX is $1,734. The B200 DGX draws 14.4kW sustained, costing $2,448/month in power alone. A single 40kW cabinet can hold three DGX H100s but only two DGX B200s, making the per-GPU power cost $57/month for H100 and $102/month for B200 in the same colo lease.

Power metering methodology matters. Utility-grade metering (revenue-grade CTs) adds $20–$40/month per circuit but ensures accurate billing. Some colos use facility-level PUE-adjusted billing, multiplying your metered power by the facility PUE (typically 1.35–1.55). At a 1.45 PUE, a DGX H100's $1,734 metered power becomes $2,514 billed. Negotiating power-only contracts with fixed-rate kW pricing eliminates PUE passthrough uncertainty.

Most Tier 3 colocation facilities built before 2022 provision 5–8kW per cabinet as standard. GPU clusters need 30–60kW per cabinet. Retrofitting a single cabinet for 40kW costs $4,000–$8,000 in PDUs, busway taps, and branch circuit upgrades. Facilities with existing 415V 3-phase power distribution can handle 50kW per cabinet with minimal modification. Facilities limited to 208V face hard ceilings at 25kW before voltage drop becomes prohibitive.

The shift to liquid cooling in 2025–2026 changes the colo calculus. A B200 NVL rack with rear-door heat exchangers rejects 70kW per cabinet but requires chilled water loops at 18–22C entering temperature. Only 38% of Tier 3 colos in the US have the hydronic capacity for per-cabinet liquid cooling. The rest require facility-side CDU installation at $15,000–$30,000 per cabinet, a cost typically amortized over 12 months into the lease rate.

GPU colocation requires high-bandwidth connectivity to cloud providers or the public internet. A 100G Direct Connect to AWS (dedicated 10G virtual interface) costs $350/month plus $0.02/GB egress. Azure ExpressRoute at 10G is $550/month with $0.01/GB egress. Equinix Fabric virtual cross-connects at 1G are $125/month but cap at 2Gbps burst, insufficient for checkpoint replication from a GPU cluster.

For multi-node training across colocated cabinets, the intra-facility cross-connect is the hidden cost. A 400G InfiniBand link between two cabinets in the same colo costs $1,200–$2,000/month for the dark fiber patch. For clusters spread across multiple cages, these cross-connects can add $6,000–$12,000/month, equivalent to 2–4 additional DGX H100 power feeds. Some colos waive inter-cabinet cross-connects above a 200kW committed power draw.

At 128 GPUs (16 DGX H100s) running 24/7 for 3 years, colocation at $170/kW-month with $450/month cross-connects totals $1,230,000 in housing costs plus $3,840,000 in server hardware depreciation. Total TCO: $5,070,000 over 3 years. Running the same on AWS p5 spot at $3.52/GPU/hr with 85% utilization costs $3,300,000 per year or $9,900,000 over 3 years, 95% more than colocation.

The colo advantage shrinks when utilization drops. At 50% utilization (12 hours/day active), cloud spot at $3.52/GPU/hr costs $4,950,000 over 3 years versus colo at $5,070,000 fixed. The breakeven utilization point between colo and hyperscaler spot is 56% for H100s and 62% for B200s, where the higher power draw of B200s increases colo fixed costs faster than cloud hourly costs.

Remote hands charges are the most commonly underestimated colo cost. A GPU hardware swap averages 2 hours at $200/hour. At an anticipated 1 failure per 5,200 hours per DGX node, a 16-node cluster expects 28 hardware interventions per year costing $11,200. Cross-connect moves, cage expansions, and smart hands for cable management add another $8,000–$15,000 annually.

Security compliance fees apply for regulated workloads. HIPAA-compliant colo zones add $200–$500/month surcharge per cabinet. SOC 2 Type II reports are typically included but PCI-DSS-compliant cages require badge access systems and CCTV retention costs of $150–$300/month. These compliance adders total $4,000–$10,000 per year for a typical 4-cabinet GPU cluster.

Colocation makes financial sense for teams with 64+ GPUs running at over 60% utilization for more than 18 months. Below that threshold, the capital commitment and operational overhead of colo exceed the premium paid for cloud or bare metal rental. Above 256 GPUs, colocation becomes the dominant economic choice, cutting 3-year TCO by 40–50% versus hyperscaler reserved instances.

ClusterBid simplifies the comparison by listing both colocation-compatible bare metal and cloud GPU options side by side. A team that expects stable utilization above 60% can route long-running training through colo providers on the platform while using cloud spot for burst and experimental workloads, achieving a blended rate that beats either strategy alone.