How much power and cooling does a GB200 NVL72 rack require, and does that affect where I can host it?

Teams considering a GB200 NVL72 often focus on the GPUs and overlook a practical gate: the rack's power and cooling demands are far beyond what a standard data center provides, which shapes where you can host it. A GB200 NVL72 rack draws on the order of 100 kilowatts and requires liquid cooling, far more power density and heat removal than a typical air-cooled data center supports, so hosting it usually means a purpose-built facility or a managed cloud provider rather than a standard colocation rack. This guide explains the power and cooling requirements and why they push most teams toward hosted capacity instead of self-hosting.

Power density beyond a normal rack

A standard data-center rack is provisioned for a modest power draw, often in the single-digit to low tens of kilowatts. A GB200 NVL72 packs 72 GPUs and 36 CPUs into one rack, which pushes its power draw to roughly 100 kilowatts, an order of magnitude above what a conventional rack is wired and provisioned for. Confirm the exact figure against NVIDIA's specification and your configuration, since it varies, but the working point is that this is a high-density system, not a normal server rack.

That density is the first hosting constraint. A GB200 NVL72's roughly 100-kilowatt draw exceeds the power delivery of most standard data-center racks, so hosting it requires a facility engineered for high power density, not a general-purpose colocation space. Many existing data centers cannot deliver that much power to a single rack without significant electrical upgrades, which is why the power requirement alone rules out a lot of would-be hosting locations before cooling even enters the picture.

Why liquid cooling is required, not optional

The same density that drives the power draw creates a heat-removal problem that air cannot solve. Packing this much compute into one rack generates heat at a concentration that airflow, the cooling method most data centers are built around, cannot carry away fast enough.

Liquid cooling is the answer, and it is a requirement rather than an upgrade. A GB200 NVL72 is designed for liquid cooling because the heat density of 72 GPUs in one rack is beyond what air cooling can remove, so the hosting facility must support liquid cooling infrastructure, which most air-cooled data centers do not have. Retrofitting a facility for liquid cooling is a major undertaking involving coolant distribution, plumbing, and heat rejection systems. So the cooling requirement, like the power requirement, narrows hosting to facilities specifically built or upgraded for it, and the two together are why a GB200 NVL72 is not something you drop into a standard server room.

What the requirements mean for where you host

Use the frame below to see how power and cooling shape the hosting decision.

Hosting option	Can it meet ~100kW and liquid cooling	Practical read
Standard colocation rack	Usually no	Power and cooling fall short without upgrades
Self-built facility	Only if purpose-engineered	Large capital and time investment
Purpose-built AI data center	Yes	Designed for high density and liquid cooling
Managed GPU cloud	Yes, provider operates it	Access capacity without hosting it yourself

The pattern is consistent: the power and cooling requirements rule out standard hosting and make self-hosting a major infrastructure project. For most teams, the practical path is our platform, which already operates GB200 NVL72 racks in facilities built for the density, which turns a hosting problem into a rental decision. Unless you have a purpose-built data center, the requirements themselves point toward managed capacity.

Accessing hosted GB200 NVL72 on GMI

Since the power and cooling requirements make self-hosting impractical for most teams, the direct path is our managed platform, which operates GB200 NVL72 racks in suitable facilities. We are an AI-native GPU and inference cloud that runs NVIDIA hardware in its own data centers and publishes dedicated GPU list pricing.

We currently list GB200 NVL72 at from $8.00 per GPU-hour Available Now, operating the racks in facilities built for their power and cooling needs, so you access rack-scale Blackwell without engineering a 100-kilowatt, liquid-cooled environment yourself. Verify the current rate and availability on our GPU infrastructure (https://www.gmicloud.ai/en/gpus) and pricing (https://www.gmicloud.ai/en/pricing), since Blackwell pricing and stock move quickly, and confirm region since data-center location determines where the capacity physically sits. This is the value of managed capacity for a system this demanding: we handle the power delivery, liquid cooling, and facility engineering, and you rent the compute. When the workload is sustained large-model serving, our Prime Inference (https://www.gmicloud.ai/en/models/prime-inference) provides reserved GPU capacity with single-tenant isolation and warm serving on this hosted rack-scale hardware. Start a conversation in our console (https://console.gmicloud.ai) or contact our sales team.

Let the facility requirements decide the hosting model

If you plan to host a GB200 NVL72 without accounting for its power and cooling, you will run into a facility gate before the GPUs ever run: roughly 100 kilowatts of power and mandatory liquid cooling that a standard data center cannot provide. Those requirements narrow hosting to purpose-built facilities, which makes self-hosting a major capital project and managed cloud capacity the practical path for most teams. So let the facility requirements decide the model: unless you operate a data center engineered for this density, access a GB200 NVL72 through our managed infrastructure, and confirm the specifics against NVIDIA's figures and our available regions.