An 8 GPU HGX B200 server requires substantially more power and cooling than the previous Hopper generation, with the eight GPUs alone drawing in the range of 8 to 11.2 kilowatts depending on the specific SXM power configuration, before adding CPUs, memory, storage, and networking, which typically pushes total system draw toward 12 to 15 kilowatts per node. At that density, most data centers find air cooling alone insufficient or highly inefficient, and NVIDIA's own reference designs for B200 and the related GB200 NVL72 rack lean heavily toward direct to chip liquid cooling using cold plates and a coolant distribution unit to remove heat efficiently and keep GPUs within safe operating temperatures under sustained load. Facilities without existing liquid cooling infrastructure should budget for CDU installation, plumbing, and potentially reinforced power distribution to support this density, since retrofitting an air only data center for B200 scale deployments is a significant project rather than a minor upgrade. Some lower power B200 SKUs and smaller GPU count configurations can still run on well provisioned air cooling, but full 8 GPU deployments generally benefit from liquid cooling. Nanobase AI evaluates a client's existing facility against B200 power and cooling requirements before finalizing a deployment plan.
The jump from Hopper to Blackwell is a facilities problem, not just a spec upgrade
Teams that have run HGX H100 nodes for a year or two often assume an HGX B200 node is a drop-in replacement because it uses the same 8-GPU baseboard concept and a similar rack form factor. The power and thermal density of an HGX B200 node is high enough that most facilities cannot simply swap it into an existing H100 rack without re-checking circuit capacity and cooling headroom. The GPUs themselves are only part of the increase; higher-bandwidth NVLink switching, faster networking, and denser memory all add incremental draw on top of the GPU compute itself.
Planning for B200 at the facility level means treating it as a new class of rack density rather than a generational refresh, particularly for organizations still running mixed Hopper and Blackwell fleets during a transition period.
HGX H100 vs HGX B200: side-by-side facility profile
| Attribute | HGX H100 (8-GPU) | HGX B200 (8-GPU) |
|---|---|---|
| GPU memory per node | 640 GB HBM3 (80 GB × 8) | ~1,440 GB HBM3e (~180 GB × 8) |
| Aggregate memory bandwidth | 3.35 TB/s per GPU | 8 TB/s per GPU |
| GPU-only power draw | Up to roughly 5.6 kW (700 W × 8) | Roughly 8–11.2 kW depending on SKU |
| Estimated total node power | Roughly 8–10 kW | Roughly 12–15 kW |
| Recommended cooling | Air cooling with hot-aisle containment, widely deployed | Direct-to-chip liquid cooling strongly recommended at full 8-GPU density |
| NVLink generation | 4th generation, 900 GB/s per GPU | 5th generation, higher aggregate bandwidth per GPU |
The numbers above are node-level estimates; always confirm exact draw against the specific OEM configuration, since PSU sizing, CPU choice, and storage count all shift the total.
Why air cooling runs out of headroom at this density
Air cooling works by moving enough conditioned air across a heatsink to carry heat away before the chip throttles, and that approach scales reasonably well up to a certain watts-per-rack-unit threshold. At 12–15 kW in a single 4U to 6U chassis, the airflow required to keep junction temperatures in range becomes impractical for most raised-floor data centers, both because of the fan power needed and because hot-aisle temperatures start affecting neighboring equipment. NVIDIA's own reference designs for B200 and the related GB200 NVL72 rack lean on direct-to-chip liquid cooling, using cold plates on the GPU and CPU dies connected to a coolant distribution unit, precisely because it removes heat far more efficiently per unit of rack space than air ever can at this density.
Some lower-power B200 configurations, particularly those with fewer GPUs per node or reduced power targets, can still run on well-provisioned air cooling with hot-aisle containment. A full 8-GPU HGX B200 node at its higher power SKUs is the case where liquid cooling stops being optional in practice.
What a facility actually needs to add
- Confirm existing PDU and circuit capacity against the higher per-rack kW figure; many H100-era racks were provisioned for 8–10 kW, not 15 kW plus.
- Budget for a coolant distribution unit, plumbing, and leak detection if moving to direct-to-chip cooling, rather than assuming existing air handling is sufficient.
- Re-verify floor loading and rack stability, since liquid-cooled chassis and CDUs add weight beyond a standard air-cooled server.
- Plan power redundancy (N+1 or 2N) at the new higher baseline rather than the prior H100-era assumption.
Frequently asked questions
Can an HGX B200 node run on air cooling at all?
Lower-power B200 configurations or reduced GPU counts can run on well-provisioned air cooling with hot-aisle containment, but a full 8-GPU node at higher power SKUs generally needs direct-to-chip liquid cooling to stay within safe thermal limits under sustained load.
Do we need to re-cable an existing H100 rack for B200?
Power distribution, not just data cabling, is the more likely constraint. Confirm the rack's circuit and PDU capacity supports the higher per-node draw before assuming a physical swap is sufficient.
How does B200 power draw compare to a GB200 NVL72 rack?
An 8-GPU HGX B200 node draws roughly 12–15 kW, while a full GB200 NVL72 rack, a much larger rack-scale system with 72 GPUs and 36 Grace CPUs in a single NVLink domain, draws approximately 120 kW; they are different scales of deployment, not directly comparable per-node.
Is retrofitting liquid cooling into an existing data center realistic?
It is a real project involving plumbing, a coolant distribution unit, and often facility water loop upgrades, but it is a common and well-understood retrofit rather than a full data center rebuild for most facilities planning Blackwell-generation density.
How Nanobase AI helps
Nanobase AI, a Silicon Valley enterprise AI engineering company, assesses existing rack power and cooling capacity against HGX B200 requirements before any hardware is ordered, and designs the direct-to-chip liquid cooling retrofit when needed. We size the full stack, from PDU capacity to Kubernetes GPU Operator scheduling once nodes are online, so the facility and the software layer are ready together. See our GPU infrastructure services for the full scope.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.