DEPLOYMENT GUIDE / 01

From a rack order
to a facility brief.

Start with the systems you intend to deploy. Understand the questions that turn compute capacity into a site requirement.

START WITH THE WORKLOAD

Different workloads.
Specific requirements.

Inference & serving

Define model size, concurrency, latency objectives, geography, and growth. An inference deployment can range from a small enterprise installation to a substantial cluster.

Training & fine-tuning

Define the compute configuration, scale-out fabric, storage needs, and operating schedule. Match the facility to the selected architecture.

Workload labels alone do not establish cooling or power requirements. Start with the applicable NVIDIA reference architecture and the OEM’s configuration-specific requirements.

A FIRST LOOK AT YOUR POWER ENVELOPE

Make the starting
assumptions visible.

Enter networking, storage, and management load separately. Defaults are illustrative inputs, not a recommended configuration. A zero additional load means those systems are not yet included.

ILLUSTRATIVE FACILITY LOAD

0.192MW
Compute systems0.160 MW
Total IT load0.160 MW
Cooling basisConfirm with your OEM

An early planning estimate. Electrical peaks, redundancy, distribution ratings, environmental conditions, and expansion need separate engineering review.

Illustrative 20 kW per unit. Define whether a unit is a server or rack. Replace with your own complete-system specification.

RA families do not imply a universal power figure. The overhead factor is an assumption, not a guaranteed PUE or electrical service rating.

Need a detailed power & cooling breakdown? Open the advanced planner

ADVANCED DEPLOYMENT PLANNER

Every load.
Every assumption.

Separate the hardware demand from the infrastructure around it. Build a transparent power and cooling brief—not a promise that a site is ready.

Runs on your device. No account required. Nothing is sent when you calculate or download.

01 / Compute boundary

Illustrative starting values; replace with your configuration. No OEM sizing is implied.

One unit means one complete rack or one server—not both. System mode already includes GPU power and internal gear. Component mode ignores the complete-unit AC value.

02 / IT loads & heat capture

All AC loads are reference-load assumptions, not necessarily measured peaks. Operating fractions are fractions of electrical load—not GPU utilization. Blank liquid capture means unknown; enter 0 only for a confirmed all-air boundary.

Compute systems

Complete compute units, including internal networking and local storage already in the specification.

External networking

Scale-out fabric, external management switches and optics. Exclude in-rack gear counted in compute.

External storage

Storage servers, controllers and drives. Count storage-network switches here OR under networking, not both.

Supporting IT

Separate management, login, orchestration and security servers.

03 / Facility assumption

IT × factor includes an allowance for cooling electricity, distribution losses and other facility loads. Do not add those loads again. This is not a guaranteed PUE, especially at partial load.

YOUR PLANNING BRIEF

Review the inputs, then calculate. Unknowns will stay visible.

Calculation method & limits

System mode: quantity × complete-unit AC kW. No separate GPU addition.

Component mode: quantity × (GPU count × GPU DC watts ÷ 1,000 + host DC kW) ÷ PSU efficiency fraction.

IT total: compute + external network + external storage + supporting IT. Operating scenarios apply the entered power fraction to each category.

Thermal model: approximately all IT electrical input becomes heat. Liquid = category load × capture share; residual air = category load − liquid. Capture shares are assumed constant across scenarios. Facility equipment heat is not included in these IT thermal totals.

Facility estimate: IT × overhead factor. It is neither annual energy nor a utility reservation. Operating and reference cases are not confidence bounds.

Engineering still required: electrical transients, diversity, power factor, voltage, UPS/generator ratings, N+1/2N capacity, cooling redundancy, and OEM temperature, flow, pressure and coolant-quality requirements. Installed redundant capacity is not operating demand.

Water: coolant circulation and site water consumption are different. Neither is inferred from MW or a liquid-cooling label here. CDU, pump and chiller selection requires a mechanical design.

Equipment boundary example: NVIDIA NVL72 components. Hybrid cooling example: Dell XE9640 cooling requirements. These examples do not establish a universal liquid-capture percentage.

FIND YOUR STARTING POINT

Scale changes the conversation.

Deployment bandTotal IT loadFirst planning focus
Initial deploymentBelow 250 kWRack suitability and available services
Enterprise250 kW–under 1 MWDedicated capacity and growth
Dedicated cluster1–under 5 MWHall fit and cooling infrastructure
Large deployment5–under 20 MWPhased delivery and mechanical capacity
Campus20–100 MWUtility commitments and development
Multi-phase campusAbove 100 MWLong-term expansion planning

LandFactory planning categories, not industry certifications. Cooling method and rack density are assessed separately.

PUBLIC REFERENCE LIBRARY

Begin with the source.

Explore manufacturer material for the exact architecture you are considering. These documents have different scopes and generations; they are not interchangeable bills of materials or facility designs. Verify the current revision, supported scale and cooling option with the OEM.

Manufacturer references are provided for education. Product guides and reference architectures have different scopes. Inclusion does not imply an affiliation, certification, or endorsement. Sources reviewed September 5, 2026. Numeric presets use specific NVIDIA hardware specifications, not complete factory configurations. RTX PRO and HGX families require OEM-specific input.

LET’S BUILD THE CONNECTION

Let’s turn your requirements into a plan.

Prepare your deployment brief