NVIDIA H100 Colocation: Where to Find GPU Servers

The NVIDIA H100 remains the most sought-after GPU for AI training and inference in 2026. While supply has improved since the extreme shortages of 2023-2024, finding reliable H100 colocation — facilities that can provide the power, cooling, and networking these GPUs demand — is still a significant challenge. This guide helps you find and evaluate H100 colocation providers across the United States.

What Makes H100 Colocation Different

Colocating H100 GPU servers isn't like colocating standard servers. The infrastructure requirements are fundamentally different:

• Power density: A DGX H100 system draws 10.2 kW. Four in a rack is 40+ kW — 4-8x a typical server rack.
• Cooling: At 40+ kW per rack, liquid cooling is effectively mandatory. Air cooling alone cannot dissipate the heat.
• Networking: Multi-node H100 training requires InfiniBand NDR (400 Gbps) or high-speed RoCE. Standard Ethernet is insufficient for training workloads.
• Power quality: GPUs are sensitive to power fluctuations. Clean, conditioned power with 2N redundancy protects your $300K+ investment per server.

H100 Server Configurations

NVIDIA DGX H100

The turnkey solution from NVIDIA. Each DGX H100 includes 8x H100 GPUs connected via NVSwitch, 2x Intel Xeon CPUs, 2TB system memory, and 30TB NVMe storage. Power draw: 10.2 kW. Cost: ~$300,000-400,000.

NVIDIA HGX H100

The GPU baseboard used by OEM partners (Dell, HPE, Supermicro, Lenovo) in their custom server designs. Same 8x H100 GPU configuration as DGX but with vendor-specific chassis, CPU, memory, and storage options. Power draw varies by configuration: 6-10 kW.

Custom H100 Servers

Some organizations build custom servers with 4x or 8x H100 PCIe or SXM5 GPUs. These offer more flexibility in configuration but may sacrifice some of the NVLink bandwidth of the DGX/HGX platform.

Where to Find H100 Colocation

H100 colocation is available across major US data center markets. Here's where to look:

Top Markets for H100 Colocation

• Northern Virginia: The most facilities with H100 support, but power availability is constrained and pricing is premium.
• Dallas-Fort Worth, Texas: Large market with competitive power pricing and growing GPU colocation options.
• Phoenix, Arizona: Affordable power, available capacity, and improving connectivity make Phoenix increasingly popular.
• Chicago: Central US location with strong financial sector demand driving GPU infrastructure investment.
• Silicon Valley/Bay Area: Premium pricing but proximity to AI companies and talent.

Facilities with H100 Support

H100 Colocation Pricing

H100 colocation costs vary by market, provider, and deployment size. Typical ranges in 2026:

• Per-DGX system (managed): $10,000-18,000/month including power, cooling, and basic management
• Per-rack (self-managed): $5,000-12,000/month for a 40 kW liquid-cooled rack
• Large clusters (1+ MW): Custom pricing, typically 20-40% below list rates

For a complete cost breakdown, see our GPU colocation pricing guide. For a comparison with cloud alternatives, read colocation vs cloud for AI workloads.

What to Ask H100 Colocation Providers

When evaluating providers, ask these specific questions:

• Power: What is the maximum kW per rack? Is liquid cooling included or extra? What's the PUE?
• Cooling: Do you support direct-to-chip liquid cooling? What coolant temperature can you deliver? Is the CDU infrastructure shared or dedicated?
• Networking: Can you support InfiniBand fabric across racks? What's the maximum non-blocking cluster size? Do you offer managed InfiniBand?
• Experience: How many H100/DGX deployments do you currently support? What's your largest GPU cluster?
• Timeline: How quickly can you provision a liquid-cooled rack? Is the infrastructure pre-built or built-to-order?
• Support: Do you offer GPU-specific remote hands (driver updates, health monitoring)? What's the hardware replacement SLA?

H100 vs H200: Which to Colocate?

The H200 offers 1.7x more HBM memory (141GB vs 80GB) and higher memory bandwidth, making it superior for large model inference and memory-bound training. However, H100 remains more widely available in colocation facilities and offers the best price-performance for most training workloads.

Choose H100 if: training is your primary workload and cost matters. Choose H200 if: inference on large models is critical or your training is memory-bandwidth-limited.

Getting Started

Ready to find H100 colocation? Start with our AI-ready data center directory to filter facilities by GPU type, cooling, and market. Or learn more about how to choose the right AI data center for your workload.