Hosting NVIDIA DGX systems requires meeting several data center requirements beyond a standard server rack, starting with power, since a single DGX H100 draws around 10.2 kilowatts and a DGX B200 draws considerably more, requiring dedicated high amperage circuits, often three phase, with redundant power distribution units for reliability. Cooling capacity must match that density, which for DGX H100 can often still be met with well provisioned air cooling and hot aisle containment, while DGX B200 and future systems increasingly require liquid cooling infrastructure including a coolant distribution unit and facility water loops. Physical considerations matter too, since DGX chassis are heavy, often 130 to 150 kilograms or more, requiring reinforced raised floors or slab floors rated for that point load, along with adequate rack depth and service clearance. Networking infrastructure needs high bandwidth InfiniBand or RoCE Ethernet connectivity for multi node clusters, and facilities should also provide fire suppression appropriate for electronic equipment, controlled access, and stable temperature and humidity within NVIDIA's specified operating ranges. Skipping any of these requirements risks thermal throttling, unplanned downtime, or safety issues rather than just suboptimal performance. Nanobase AI conducts full data center readiness assessments before any DGX installation to confirm power, cooling, floor loading, and networking meet NVIDIA's specifications.
DGX systems are pre-integrated, the facility around them is not
NVIDIA ships DGX systems as fully integrated, pre-validated appliances, which removes the hardware assembly risk that comes with building a custom GPU server, but it does not remove the facility burden. A DGX chassis still has to plug into real electrical circuits, real cooling capacity, and a real floor, and NVIDIA's specifications for each of those are stricter than most general-purpose data center rooms are provisioned for. Treating a DGX purchase as "just another rack unit" is the most common planning mistake enterprises make before their first delivery date.
The requirements scale meaningfully between generations, so a facility that comfortably hosted DGX H100 systems cannot assume the same is true for DGX B200 without re-verification.
Facility checklist by category
| Category | DGX H100 | DGX B200 |
|---|---|---|
| Power draw (approx.) | About 10.2 kW per system | Considerably higher; confirm against current OEM spec sheet |
| Circuit type | Dedicated high-amperage, often three-phase | Dedicated high-amperage, often three-phase, higher headroom needed |
| Cooling | Often still viable with well-provisioned air cooling and hot-aisle containment | Increasingly requires liquid cooling infrastructure and a CDU |
| Physical weight | Roughly 130–150+ kg per chassis | Similar or greater; verify against reinforced floor rating |
| Floor requirements | Raised or slab floor rated for concentrated point load | Same, with larger safety margin recommended |
| Networking | High-bandwidth InfiniBand or RoCE Ethernet for multi-node clusters | Same, with higher per-node bandwidth expectations |
| Environmental | Stable temperature and humidity within NVIDIA's specified operating range | Same, tighter thermal margin at higher power density |
Always confirm exact figures against the current DGX datasheet for the specific model and configuration being purchased, since specifications are refined between hardware revisions.
The requirements that get missed most often
Fire suppression appropriate for electronic equipment, rather than a standard sprinkler system, is frequently overlooked in facilities that have not previously hosted dense compute, and retrofitting it after equipment arrives is far more disruptive than planning for it up front. Controlled physical access is a related requirement that matters both for security and for warranty or support terms in some enterprise agreements. Networking is another area where organizations under-plan: multi-node DGX clusters depend on InfiniBand or RoCE Ethernet fabric with enough bandwidth and low enough latency to support distributed training or high-throughput inference, and retrofitting fabric after a cluster is racked is disruptive compared to designing it in from the start.
Floor loading is the most physically unforgiving requirement on the list. A DGX chassis at 130 to 150 kilograms or more, concentrated in a single rack footprint, can exceed what a standard office or lightly-built server room floor is rated for, and discovering this after delivery is a costly problem to fix.
A pre-installation validation sequence
- Confirm exact power draw, circuit type, and redundancy (N+1 or 2N) requirements against the specific DGX model's current datasheet.
- Verify cooling method compatibility: air cooling with hot-aisle containment for H100-class systems, liquid cooling readiness for B200-class systems.
- Get a structural engineer to confirm floor load rating against the chassis weight, especially in any non-purpose-built facility.
- Confirm InfiniBand or RoCE fabric design and cabling plan before systems arrive if deploying more than one node.
- Validate fire suppression, access control, and environmental monitoring meet both NVIDIA's specification and internal compliance requirements.
Frequently asked questions
Can DGX H100 run on air cooling alone?
In many cases yes, with well-provisioned air cooling and hot-aisle containment, DGX H100 systems can operate within normal thermal limits, though this should be confirmed against the specific facility's cooling capacity rather than assumed.
Does DGX B200 always require liquid cooling?
Liquid cooling infrastructure is increasingly required or strongly recommended for DGX B200 given its higher power density; confirm the exact requirement against NVIDIA's current specification for the configuration being purchased, since this has evolved across Blackwell-generation products.
What networking does a multi-node DGX cluster need?
High-bandwidth InfiniBand or RoCE Ethernet connectivity between nodes is standard for multi-node DGX clusters, sized to avoid becoming a bottleneck for distributed training or high-concurrency inference workloads.
What happens if we skip a facility requirement like floor loading?
Skipping structural, power, or cooling requirements risks thermal throttling, unplanned downtime, or genuine safety issues, not just suboptimal performance, which is why a full readiness assessment before delivery matters more for DGX than for lighter-weight equipment.
How Nanobase AI helps
Nanobase AI, a Silicon Valley enterprise AI engineering company, conducts full data center readiness assessments before any DGX installation, confirming power, cooling, floor loading, and networking against NVIDIA's current specifications for the exact model being deployed. We coordinate with facility engineers and structural consultants so delivery dates do not outrun site readiness. See our on-premise LLM deployment guide or explore GPU infrastructure solutions.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.