Bring your own GPU servers. We provide up to 17.3kW per rack on 208V 3-phase power, air-cooled hot/cold aisle cooling, and 50 Mbps of bandwidth per kW (about 16 TB/month) — on Grant County PUD hydro power at roughly 4.6¢/kWh, among the cheapest commercial electricity in the United States. Owner-operated, with a direct line to the people who run the network. We do not rent servers or GPUs.
Customer-owned servers only. These are the numbers GPU buyers ask for first.
| Power per rack | Up to 17.3kW — 208V 3-phase, 60A circuit, 48A continuous |
|---|---|
| Single-phase | 30A single-phase PDUs available |
| PDUs | APC and Raritan |
| Cooling | Air-cooled, hot/cold aisle. Liquid cooling not offered today; being evaluated. |
| Bandwidth | 50 Mbps per kW included (about 16 TB/month per kW) on a 1 Gbps connection |
| Network | Grant County PUD fiber, static IPs, VLAN isolation, cross-campus L2 |
| Pricing | From $200/kW/mo all-in. $0 setup. |
| 1 × 8-GPU node | H100/H200-class, about 10kW each, plus networking and storage |
|---|---|
| 1 × 8-GPU node | B200-class, about 14kW — fits within a single 17.3kW rack |
| 2–3 × 4-GPU servers | About 5–6kW each |
| 4–8 × GPU workstations | RTX-class servers at roughly 1.5–3kW each |
A single 8-GPU server draws 3–6kW. A growing multi-server cluster can pull 25–50kW or more. At those power densities, where you plug in matters more than almost anything else.
Grant County PUD delivers Columbia River hydroelectric power at rates consistently ranked among the lowest commercial electricity costs in the US. The same reason the hyperscalers built here — and we're passing that advantage directly to you.
208V, 3-phase power available at Mattawa campuses for high-density GPU workloads. We configure circuits around your hardware — not the other way around. Custom amperage and phase configurations available; contact us with your specs.
You're not working through a procurement process and a sales team. Call the owner directly at 509-906-1250. Custom power configurations, non-standard circuits, or unusual hardware — if it's feasible, we'll work it out with you directly.
Any GPU-dense, power-hungry AI workload benefits from cheap hydro power. Here's what our customers are running or planning to run.
Running LLM inference 24/7? Power cost dominates your TCO. Colocate your inference hardware in Grant County and your per-token cost drops dramatically versus Seattle, Portland, or California facilities. Single GPU server or a multi-server rack — we work with you on the right footprint.
Training runs that go for days or weeks are exactly where cheap power matters most. Colocate your training hardware — GPU clusters, storage arrays, and interconnects — in a facility that doesn't penalize you for running at full power draw around the clock.
Pair your GPU compute with our S3-compatible object storage on the same local network. Your training data and model checkpoints live next to your compute — sub-millisecond access, simple flat pricing, no cloud storage bills eating into your budget.
Need your AI workloads distributed for resilience? Split your cluster across two campuses — sub-millisecond cross-campus VLAN connectivity means your nodes can communicate as if they're on the same rack, while getting power diversity across different grid feeds.
Power, space, and connectivity — all in. Real bandwidth included, not nickel-and-dimed by the gigabyte: 50 Mbps per kW included (about 16 TB/month per kW) on a 1 Gbps connection, with no per-GB fees.
Every AI deployment is different. Tell us your GPU count, power draw estimate, and timeline — we'll give you a real number, not a quote that requires three more calls to finalize.
Get a Custom Quote →See how much you'll pay to colocate GPUs at Columbia vs. cloud pricing. All-in power cost, no surprise bandwidth charges.
Mattawa West is live now with 5kW available for smaller AI/GPU deployments. Mattawa East — a half mile away on the same fiber backbone — is open with 2MW of capacity and high-density power circuits. Contact us for details. Dedicated builds: for larger-scale operations, with no fixed timeline yet.
Yes, if you own the hardware and it fits our rack power and cooling envelope. Racks deliver up to 17.3kW on 208V 3-phase power, air-cooled in a hot/cold aisle layout. An 8-GPU H100/H200-class node typically draws about 10kW, so a rack usually holds one such node plus networking and storage. Check your server's maximum draw and airflow requirements; direct liquid-cooled platforms are not supported today.
GPU and AI colocation starts at $200 per kW per month all-in (power, cooling, space and bandwidth). Pricing is based on a 12-month term; month-to-month is $210 per kW. A 10kW deployment is $2,000 per month, and a full rack reserved at 16kW is $3,000 per month on a 12-month commitment. There is no setup fee.
No. Columbia Colocation does not rent servers or GPUs. You bring your own hardware and we provide space, power, cooling and network.
Racks are fed by 208V 3-phase 60A circuits (48A continuous, about 17.3kW) using APC and Raritan PDUs. 30A single-phase PDUs are available for smaller or older equipment.
Not today. Racks are air-cooled in a hot/cold aisle layout. We are evaluating liquid cooling for future deployments, so tell us your requirements.
Your connection is 1 Gbps. Each kW includes 50 Mbps of sustained bandwidth, about 16 TB/month if used continuously, with no per-GB fees. Usage is measured at the 95th percentile. Overage is $2 per Mbps. Fiber is Grant County PUD with static public IPs and VLAN isolation.
Grant County PUD supplies Columbia River hydroelectric power at roughly 4.6 cents per kWh, among the lowest commercial rates in the United States. For GPU servers that run around the clock, power is the largest operating cost.
GPU count, power draw estimate, interconnect requirements, timeline, redundancy needs — the more you tell us, the faster we can give you a real number. No sales scripts, no runaround.
Mattawa East (2MW, open) is taking customers now. Multi-rack and custom-room builds (4–8 racks, up to 300–400kW) are possible with a longer-term commitment. For a dedicated location, we have 2MW and 2.5MW shells available; contact us.
Related pages