NVIDIA H100, H200, B200, B300 and DGX-ready Data Centers

Real ranges (April 2026): $150–$250 per kW/month for power-only retail colo, plus $1,200–$3,500/rack/month for space and standard cross-connects. An 8× H200 HGX node (10.2 kW) typically lands at $2,500–$5,000/month all-in for power, space, and basic network in mid-tier markets. Prime markets (Ashburn, Santa Clara) run 20–40% higher. Liquid-cooled space adds $1,000–$2,000/rack premium. We send 3–5 actual quotes, not generic ranges, based on your specs.

No, not strictly. H100 and H200 (700W TDP per GPU, ~10.2 kW per HGX node) can be air-cooled in facilities with rear-door heat exchangers or proper hot/cold aisle containment. Liquid is preferred for performance and density but not mandatory. For B200 (1,000W) and B300 (1,100–1,400W), direct-to-chip liquid cooling IS mandatory — both for thermal reasons and to maintain NVIDIA’s warranty terms.

Air-cooled H100/H200 deployments in power-slotted facilities: 2–6 weeks. Liquid-cooled B200/B300 in prime markets (Ashburn, Santa Clara): 3–6 months. Plan-B markets (Phoenix, Atlanta, Reno, Columbus): often 4–8 weeks even for liquid. GB200 NVL72 hyperscale racks: 3–6 months minimum. We track live capacity: the difference between a ‘qualified’ provider and one with actual power on the floor right now is what saves you 60+ days.

For steady inference and long-running training, yes, typically 40–60% lower 24-month TCO if you own the hardware. AWS p5e at ~$4.98/GPU-hr × 8 GPUs × 80% utilization ≈ $700K over 2 years. Same workload on owned H200s in colo: ~$370–400K all-in (hardware + power + space + network). Cloud still wins for spiky, short-burst training. Hybrid (train cloud, infer colo) is the sweet spot most teams land on.

HGX H200 is NVIDIA’s reference 8-GPU baseboard that OEMs (Supermicro, Dell, Lenovo, ASUS) build into servers. DGX H200 is NVIDIA’s own complete system. Both pull ~10.2 kW per 8-GPU node, both prefer liquid but tolerate air, both want NVLink-friendly rack topology. H200 NVL is the PCIe variant: lower power (600W), air-cooled, fits standard enterprise racks at 15–25 kW. Different deployment profiles, much easier to place.

NVIDIA’s DGX-Ready Colocation program certifies facilities that meet NVIDIA’s standards for electrical, cooling, and operational quality for DGX systems. Members include Equinix, Digital Realty, DataBank, EcoDataCenter, and others. You don’t strictly NEED a DGX-Ready facility unless NVIDIA support or warranty terms require it for your specific deployment, but it’s a useful filter for AI-grade infrastructure quality. Many non-DGX-Ready providers host HGX-based systems (the OEM equivalents) just fine and often cheaper.

Most prime-market providers won’t take less than a full rack or a 100 kW commitment. Mid-tier and regional providers will take a single 10–15 kW cabinet on 12-month terms. We specifically maintain relationships with operators who accept small-footprint GPU deployments — a major reason single-node AI startups come to us instead of cold-calling Equinix.

400 Gbps InfiniBand (NDR) is standard for H200 multi-node training. 800 Gbps (NDR800) for B200/B300. Storage and management traffic should live on a separate 100/200G Ethernet fabric. Cross-connect and leaf-switch placement matter — NVLink scale-out has cable-length limits (~20–30m), so you need contiguous racks, not whatever’s spare.

Practical planning numbers: H100/H200 HGX: 30–40 kW/rack. H200 NVL (PCIe): 15–25 kW/rack. B200: 60–80 kW/rack. B300: up to 100 kW/rack. GB200 NVL72: 120 kW per rack-scale unit. Always provision for peak draw, not average. Training workloads sustain near-max for hours.

No. We’re free to you. Providers share their existing sales commission with us, the same way they’d pay an in-house salesperson. You pay exactly what you’d pay going direct, and we save you the weeks of sales calls.

Most providers don’t publish pricing, and the ones that do quote list prices that the real bill rarely matches (cross-connect fees, power overage tiers, install fees). ChatGPT’s training data doesn’t know who has 30 kW of liquid-ready capacity in Phoenix this week. We do because we’re calling them.

Tell us your ideal markets AND your acceptable Plan B. In 2026, the price difference between Ashburn and Phoenix for 30 kW of liquid-cooled space can be 20–40%. If your latency requirements allow, flexibility translates directly to dollars and weeks saved.

Yes. Most of our active conversations in 2026 involve B200 deployments. GB200 NVL72 is hyperscale-grade — fewer facilities qualify, lead times are longer (3–6 months minimum), and most deals are off-market. We have relationships with the operators that take these. Send specs and we’ll be straight about timelines.

Send us the SKU + quantity. We’ll back into the kW for you and confirm before quoting. Common gotcha: people spec ‘nameplate’ wattage from sales sheets, but real workload draw is often 70–85% of that. We size for sustained peak, which is what providers actually bill against.