Usable kW, not nameplate
Ask what the breaker is rated for and what you can actually draw continuously. A cabinet sold as 40 kW on a circuit that derates to 80% is a 32 kW cabinet.
NVIDIA colocation: H100, H200, B300 and DGX-Ready data centers
The power, cooling and rack-density numbers for each NVIDIA system, taken from NVIDIA’s own documentation and what to check with a facility before you sign.
QuoteColo is not an NVIDIA partner and does not certify facilities. DGX-Ready is NVIDIA’s own program; we place into certified and non-certified facilities both, because most deployments do not require the certification.
5,000+
Clients helped since 2004
No Cost
to Buyers
500+
Worldwide providers
Quotes
in 48 h
Power and cooling, by NVIDIA system
Every figure NVIDIA publishes is a configurable maximum, not a sustained draw.
NVIDIA does not publish typical or sustained power for any of these parts, and operators can cap them lower.
| System | GPU, max | System / rack, max | Cooling, per NVIDIA |
|---|---|---|---|
| H100 SXM5 | up to 700 W | HGX system power not published by NVIDIA | OEM-dependent |
| H100 NVL (PCIe) | up to 400 W | — | Passive, air |
| H200 SXM5 | up to 700 W | HGX system power not published by NVIDIA | OEM-dependent |
| H200 NVL (PCIe) | up to 600 W | — | Dual-slot, air |
| DGX H100 / H200 | — | 10.2 kW max · 14.3 kW on the CTS variant | Air · 1,105 CFM |
| DGX B300 | up to 1,000 W | ~14.3 kW max | Air · 1,550 CFM |
| DGX B300 | up to 1,400 W | 14.5 kW or 15 kW | Air · 1,500 CFM |
| GB200 NVL72 | not published | ~120 kW per rack | Liquid, direct-to-chip. Networking and storage air |
| GB300 NVL72 | not published | up to 142 kW per rack | Liquid, MGX-based |
| Vera Rubin NVL72 | not published | Rack total not published | Liquid · 45°C warm water |
| L40S | 350 W | — | Passive, air |
| RTX PRO 6000 Blackwell, Server Edition | up to 600 W | — | Air (dual-slot) or liquid (single-slot) |
Verified against NVIDIA documentation, 20 October 2026.
Air or liquid, and how much power per rack
Cooling is decided by the system, not by the GPU generation. Eight-GPU nodes, including B300, are air-cooled in NVIDIA’s own documentation. Liquid becomes mandatory at the rack-scale NVL72 systems.
| What you are deploying | Power | Cooling | Rack planning |
|---|---|---|---|
| One PCIe card L40S, H200 NVL, RTX PRO 6000 |
0.35–0.6 kW per card | Air | Fits a standard cabinet |
| One 8-GPU node DGX H200 / B300 |
10.2–15 kW max | Air | A single high-density cabinet |
| Four 8-GPU nodes | ~41–60 kW | Air | NVIDIA’s guidance is four DGX systems per rack; more than four is not recommended. Peak demand is 20.4 kW per three-phase circuit. |
| GB200 NVL72 | ~120 kW per rack | Liquid | One rack is one system. Networking and storage stay air-cooled |
| GB300 NVL72 | up to 142 kW per rack | Liquid | Same, at higher density |
What changed, generation by generation
Six months ago most NVIDIA colocation inquiries specified H100. The question on the first call now is whether a facility can take Blackwell.
| Generation | SKU | What buyers ask for | What changed for the facility |
|---|---|---|---|
| 2023–24 | H100 | Training clusters | Density moved past what most retail colocation was built for |
| 2024–25 | H200, HGX H200 | More memory, same power envelope | Nothing. Same facility requirements as H100 |
| 2025–26 | B300, DGX B300 | AI factories, reasoning models | More power per node. Still air-cooled at the node |
| 2025–26 | GB200 NVL72 | Rack-scale clusters | The rack becomes the system. Liquid, and few facilities qualify |
| 2026+ | B300, GB300, DGX B300 | Reasoning AI, long-context inference | 288 GB HBM3e per GPU, 800G networking. Rack-scale stays liquid |
What NVIDIA colocation costs per kW
Your deployment size decides which market you are shopping in. These are planning bands for space and power — bandwidth, cross-connects, IPs, setup and remote hands are quoted separately.
| Your deployment | Total power | Planning band | Availability |
|---|---|---|---|
| One to four 8-GPU nodes | 10–60 kW | $225–250 / kW / month | Near-term options |
| Eight to thirty nodes, or a few NVL72 racks | 100–500 kW | $200–250 / kW / month | The size providers want right now |
| Rack-scale at multi-megawatt | 1 MW+ | $185–250 / kW / month | Plan 6–12 months. The $185 low end is build-to-suit, not ready-now space. |
| Any size, needed now | — | $200–315 / kW / month | Depends on region, kW, configuration and timing |
Of the 26 North American listings in our inventory marked immediate or Q4 2026, twelve are under 500 kW.
Listings are not unique sites, and listed power is not automatically one contiguous block.
Listings are not unique sites, and listed power is not automatically one contiguous block.
Colocation vs cloud for H200 and B300
These are not the same purchase.
Cloud rents you the GPU. Colocation rents you the room — you buy the GPU yourself. Any comparison that skips that is comparing a rental to a mortgage payment.
| Cloud, on-demand | H200, $/GPU-hour | B300, $/GPU-hour |
|---|---|---|
| AWS | $7.91 | $14.24 |
| Microsoft Azure | $10.60 | not published |
| Google Cloud | $10.60 | not published |
| CoreWeave | $6.31 | $8.60 |
| Together AI | $5.99 | $8.19 |
| Nebius | $5.40 | $8.50 |
| Crusoe | $4.29 | not published |
On-demand list rates read from each provider’s published pricing page, 23 October 2026. Per-GPU figures derived from 8-GPU node prices where the provider quotes per node.
Two things in that table are worth more than the numbers. The hyperscalers have converged to the cent — Azure and Google are $84.80 and $84.81 for the same 8-GPU H200 node. And neither of them publishes a B300 on-demand rate at all.
Markets: where NVIDIA capacity actually lives in 2026
Everyone starts in the same four metros. The capacity on our list this quarter is mostly somewhere else.
| Market | On our list now | Rate | Cooling | Available |
|---|---|---|---|---|
| Denver | 750 kW | $190–225 / kW | Air and liquid | Q4 2026 |
| Pennsylvania | 500 kW | $200 / kW | Air | Power now, install Q4 2026 |
| Utah | 800 kW listed | $240 / kW | Air | 500 kW of it immediately |
| Boise | 200 kW | $300 / kW | Air | Immediate |
Space and power, $/kW/month. Listed offers from our October inventory, not closed deals. Provider names withheld.
Across North America, 26 listings on our inventory are marked immediate or Q4 2026 — twelve of them under 500 kW. Listed power is not automatically one contiguous block, and a quarter needs a year attached before it means anything.
Where everyone looks first
Ashburn, Dallas, Santa Clara and Chicago. Tight for dense capacity, with the longest waits for anything liquid-ready.
Where we also quote
Phoenix, Reno, Atlanta, Columbus, Quincy, Denver, Boise, Montreal and Toronto. Priced out of your first choice, or waiting on it — we quote a secondary market alongside it so you can weigh location against cost with real numbers.
There are still openings for 2026 if you have some flexibility on location and size. Waits for 2027 could be significantly longer — that is an outlook, not a forecast.
What to check with a facility before you sign
Cooling matched to your system
Air is fine if it supports your exact hardware and rack density. Find the available kW first, then check the cooling fit.
Installed CDUs, not planned ones
Only relevant if you are deploying NVL72 rack-scale. Ask which room, how many cabinets, at what density, and whether the CDUs are installed or on order.
Contiguous racks
Multi-rack NVLink scale-out has cable-length limits, so the racks have to be adjacent rather than wherever the facility has space.
Who sends us NVIDIA requirements
| You are | Your situation | What we look for |
|---|---|---|
| AI startup, first node | You’ve bought one HGX or DGX node, it draws 10–15 kW, and you have nowhere to put it | Mid-tier and regional providers that take a single rack without a 100 kW minimum |
| Scaling AI company | You’re running 8 to 32 nodes across 4 to 10 racks and growing | That is 100–500 kW, the size providers are actively looking for right now |
| Neocloud or GPU reseller | You’re buying at multi-megawatt scale with an expansion plan behind it | Wholesale capacity, phased delivery and build-to-suit. Plan 6–12 months |
| Research, HPC or university | You’re standing up a cluster on a grant timeline | Shorter terms than a 36-month enterprise commitment |
| Regulated AI | You’re deploying GPUs in healthcare, federal or finance | Providers filtered by current compliance scope, physical access controls and a documented audit process |
Get 2–4 qualified options in 48 hours
Free for buyers
Your price is the same as going direct.
You will get all-in price estimates by email
From 500+ providers. No sales calls, no long forms.
Trusted service since 2004
5,000+ clients helped.
NVIDIA hardware questions
What is a DGX-Ready data center, and do I need one?
DGX-Ready is NVIDIA’s own certification for colocation facilities, validating that electrical, cooling and control products meet NVIDIA’s standards for DGX deployments. Seventeen operators hold it in the Americas. Most deployments do not require it — a facility that can deliver your power and cooling qualifies whether or not it carries the badge. NVIDIA states it does not endorse the companies on that list.
Which NVIDIA systems actually require liquid cooling?
The rack-scale ones. GB200 NVL72, GB300 NVL72 and Vera Rubin NVL72 are liquid-cooled by design, with cold plates on the processors. Eight-GPU nodes are not: NVIDIA documents DGX B300 as air-cooled, at 1,500 CFM. A B300 node does not need a facility with CDUs installed.
How many DGX systems fit in one rack?
Four. NVIDIA’s own guidance is four DGX systems per rack, and says more than four is not recommended. At 10.2 to 15 kW per system that is roughly 41 to 60 kW of rack load, with peak demand of 20.4 kW per three-phase circuit. That is the number to give a facility when you ask about density.
What is the difference between HGX H200, DGX H200 and H200 NVL?
DGX is NVIDIA’s complete system, specified and supported by NVIDIA. HGX is the baseboard an OEM builds its own server around. H200 NVL is the PCIe card, up to 600 W, in a standard dual-slot air-cooled form. They are three different things to quote and three different power profiles.
Why does NVIDIA publish DGX power but not HGX power?
Because HGX is a baseboard, not a finished system. The HGX datasheet has no power row at all; NVIDIA defers to the OEM, because the chassis, PSUs and fans are the OEM’s. If a page quotes you an exact HGX wattage, it came from somewhere other than NVIDIA. For an 8-GPU Hopper system the closest NVIDIA figure is DGX H100/H200 at 10.2 kW maximum.
What rack density should I plan for B300?
For DGX B300 nodes, plan on 14.5 to 15 kW per system — NVIDIA’s own physical-specs and power-specs tables give both figures — and four per rack. For GB300 NVL72 the rack is the system, at up to 142 kW, liquid-cooled. Those are two different deployments and they shop in different markets.
Is GB300 NVL72 120 kW or 142 kW?
142 kW is the figure to plan against. NVIDIA’s rack-scale user guide covers GB200 and GB300 together and gives approximately 120 kW; the GB300-specific reference architecture says up to 142 kW. Treat 120 kW as GB200 typical and 142 kW as the GB300 ceiling, and size the room for the ceiling.