About this article: AI-assisted writing, reviewed against the cited sources. This is source research, not a lab validation of the procedures. Check the stated versions and your own recovery requirements before making changes.
A dense GPU rack is a heater with a deadline. Every watt the servers draw becomes heat that has to leave the room, and the electrical feed has to deliver that wattage at the voltage and power factor the site actually has. The useful number is not the GPU slide. It is the system maximum on the server datasheet, multiplied by how many of those servers you bolt into one rack, then checked against the inlet air the server is willing to breathe.
An AI writing system produced this article for NetLeverage's editorial studio. Nothing here was metered on a rack. The sums are arithmetic on published nameplates. They are not a wiring design, and they are not a substitute for the qualified person who signs the electrical installation.
Why it matters
GPU thermal design power is the card. The rack breaker sees the whole node: CPUs, switches on the board, network adapters, drives, fans and the loss in the power supplies. Size the feed from the GPU line and the breaker trips, or the room overheats, while the GPUs are still inside their own budget.
NVIDIA's H100 datasheet lists maximum thermal design power for the SXM module as up to 700 W, configurable, and for the H100 NVL as 350–400 W, configurable. "Up to" and "configurable" mean the card is not a fixed 700 W load. A node full of SXM modules can be set lower. A PCIe NVL card is a different budget entirely. Adding "eight times 700" is only valid for eight SXM modules set at that ceiling.
The DGX H100 datasheet is the system view of eight H100 GPUs: 640 GB of GPU memory in total, and system power usage of about 10.2 kW maximum. Eight times 700 W is 5.6 kW. The system figure is about 4.6 kW above that GPU sum. Those 4.6 kW are still heat, and they still come from the same feed. A rack plan that multiplies GPU TDP and stops there has missed almost half of this node's published maximum.
From nameplate to current
Real power on a three-phase feed is
so the line current is divided by . is the line-to-line voltage. The power factor is the share of the current that does real work. This article uses the relation as electrical arithmetic. It does not open a site's switchboard schedule, and 400 V below is an example voltage, not a claim about every Australian board.
Take the DGX H100 maximum as 10.2 kW, power factor 1, line-to-line voltage 400 V:
per phase, for one node, at unity power factor, assuming a balanced three-phase aggregate load. The DGX H100 user guide specifies six 200–240 V AC power supplies rated 3300 W at 16 A in a 4+2 redundancy arrangement; it does not describe each PSU as a three-phase input. Actual phase current depends on how qualified designers distribute those single-phase inputs and redundant feeds. The aggregate formula is not a plug, circuit-breaker or cable specification. Four such nodes, if each can reach that maximum at the same time, are 40.8 kW and about 59 A per phase on the same example. Drop the power factor to 0.95 and the current rises by , to about 15.5 A for one node and about 62 A for four. The GPUs did not change. The feed did.
That 40.8 kW is four system nameplates. It does not include a top-of-rack switch, a management switch, or storage in the same cabinet. Add those from their own nameplates. It also assumes the four nodes can all sit at the published maximum together. The datasheet's tilde on "~10.2 kW max" is NVIDIA's wording. This article did not meter a DGX, and an idle node will not draw the maximum. Size the protection and the cooling plant for the maximum you are willing to let the node reach. Size the energy bill from a measurement you have not yet got from a nameplate.
The H100 SXM ceiling matters when someone changes the configured power. If the modules are configured below 700 W, the GPU share falls and the system maximum on the DGX sheet is no longer a simple 10.2 kW you can scale in your head. Read the configured limit on the node you have. The datasheet's "configurable" is the warning.
The heat is the same watts
In steady state a running server does not store electrical energy. The watts in leave as heat. 10.2 kW at the input is 10.2 kW the cooling system has to take away, for that node, before you add the rest of the room. Four nodes at that ceiling are 40.8 kW of heat, plus switches and lights and the loss in the cooling plant itself. This article does not apply a power-usage-effectiveness multiplier. That factor was not taken from a primary document opened for this article. Ask the plant for the IT heat you added up, and let the plant's own overhead be a separate line the facilities team owns.
Air cooling moves that heat in the air stream across the server. Liquid cooling moves a large share of it in the coolant. The NVIDIA Blackwell datasheet, opened 10 November 2026, describes the GB200 NVL72 as 36 Grace CPUs and 72 Blackwell GPUs in one NVLink-connected, liquid-cooled, rack-scale design. It says liquid cooling increases compute density and reduces floor space. The same datasheet's comparison of that rack with H100 air-cooled infrastructure is marked projected performance, subject to change. The pages extracted here do not print a kilowatt figure for the NVL72 rack. Do not invent one. A liquid rack still rejects the same class of heat; it rejects it to a coolant loop, and the loop still ends at a heat exchanger that the building has to feed. Density goes up. The building's heat rejection does not become optional.
What air the server will accept
Thermal requirements must come from the exact equipment maker and the applicable licensed environmental standard. A general data-centre class is not evidence that a particular server is certified for it. ASHRAE includes a high-density-server class as well as its other equipment classes; the NVIDIA datasheet used here does not establish which ASHRAE class applies to this installation. Ask the equipment maker and the facility designer to identify that class and the correct revision before approving an operating envelope. This article does not reproduce ASHRAE's protected limits, tables or notes.
The independently published NVIDIA DGX H100 datasheet specifies an operating temperature range of 5–30 °C. That is a vendor limit for the stated system, not a recommended room set point and not an ASHRAE classification. An inlet at 31 °C exceeds that vendor maximum regardless of a general room label. The facility design needs operating margin, the equipment maker's altitude and humidity limits, and a plan for loss of cooling. Those are engineering inputs rather than numbers to infer from GPU wattage.
A useful acceptance record identifies the server model and configuration, the vendor document revision, inlet sensor locations and uncertainty, the permitted environment, the alarm set points approved by the designer, and the response when cooling capacity falls. Liquid-cooled systems additionally need the maker's coolant, pressure, flow and water-quality requirements. A temperature recorded in a corridor is not a measurement of any of those conditions. No electrical or cooling installation is approved by this article.
Inlet temperature is the air entering the server, not the hot aisle behind it and not a sensor in the corridor. Blanking panels, a missing floor tile, and a rack intaking its own exhaust all raise the inlet while the room average still looks acceptable. The arithmetic above does not detect that. A thermometer at the server intake does.
What to choose and why
| Question | Use this number | Do not use |
|---|---|---|
| How hard can one H100 SXM be driven? | Datasheet maximum TDP, up to 700 W, configurable | A single "700 W GPU" assumed for every card, including the 350–400 W NVL |
| How hard can an eight-GPU DGX H100 be driven? | About 10.2 kW maximum system power | 8 × 700 W = 5.6 kW |
| What current is that at the board? | at the site's voltage and power factor | 14.7 A copied from the 400 V, unity-power-factor example |
| What must the cooling remove? | The same watts, as heat, plus the other loads in the rack | A GPU-only heat figure |
| What air should the front of the server see? | The exact vendor limits and the equipment class confirmed by the facility designer; the DGX sheet specifies 5–30 °C | The warm end of an allowable class the server does not list |
| A 72-GPU liquid rack | The vendor's rack power, when you have it, and a coolant loop sized to that heat | A guessed kilowatt figure. The Blackwell datasheet opened here does not print one |
Our recommendation, not a vendor's: write the system maximum, the configured GPU power, the site voltage and the inlet limit on one page before you order a rack. If any of those four is missing, the current and the cooling load are not known yet.
A worked rack, labelled as arithmetic
One DGX H100-class node at the published maximum, example feed 400 V line-to-line, power factor 1:
| One node | Four nodes, all at that maximum | |
|---|---|---|
| Published system power | ~10.2 kW | 40.8 kW |
| GPU share if eight SXM modules are at 700 W | 5.6 kW | 22.4 kW |
| Rest of the node, by subtraction | ~4.6 kW | ~18.4 kW |
| Line current at 400 V, pf 1 | ~14.7 A | ~59 A |
| Heat to remove from these nodes | ~10.2 kW | 40.8 kW |
The subtraction is this article's arithmetic on the two NVIDIA figures. It is not a line item NVIDIA published. Four nodes are an illustration of density, not a rack SKU. A rack that also holds a 400 Gbit/s switch adds that switch's nameplate on top. A node configured below 700 W per GPU, or a node of H100 NVL cards at 350–400 W, is a different row.
The safe order
1. Read the system power and the operating temperature on the server datasheet, and the configured GPU power on the node. The H100 sheet's "up to 700 W (configurable)" and the DGX sheet's "~10.2 kW max" and "5–30 °C" are the three fields that drive the rest.
2. Count servers, then count everything else in the cabinet. Multiply system maxima. Add switches and storage from their own sheets. Do not multiply GPU TDP and call it a rack.
3. Convert to current at the voltage and power factor the board uses. The 400 V, unity-power-factor lines above are examples so the formula is visible. A 415 V board, or a power factor of 0.9, changes the amps. A qualified person checks the protective device, the cable, and the diversity the installation rules require. This article stops at the arithmetic.
4. Hand the same kilowatts to cooling as heat. State whether the path is air across the server or liquid to a rack manifold. For the NVL72, the Blackwell datasheet says the rack is liquid-cooled. The heat still has to leave the building.
5. Compare measured inlet conditions with the approved equipment envelope. Record the vendor limits and applicable licensed standard, including altitude, humidity and change-rate constraints confirmed for the actual system. Do not copy a generic class into the design or treat an allowable maximum as a set point.
6. Meter the running node before you believe the nameplate as an average. The maximum is for protection and for cooling capacity. The average is for the bill. They are not the same measurement, and this draft did not take either one on a live rack.
Risks and limits
- Configurable TDP is a range. H100 SXM is up to 700 W. H100 NVL is 350–400 W. Using 700 W for an NVL card overstates that GPU budget.
- The DGX figure is "~10.2 kW max" on the datasheet opened 10 November 2026, for eight H100 GPUs, 640 GB total, four NVSwitches. The tilde is theirs. It was not confirmed with a clamp meter here.
- GPU TDP summed is not system power. 5.6 kW against about 10.2 kW is the gap on these two sheets.
- The 400 V examples are not a site. Current scales with and with . Protection, cable sizing and discrimination are outside this arithmetic.
- Steady-state heat equals electrical input for the IT load. It does not include cooling-plant overhead, and no PUE figure is applied here.
- ASHRAE material is referenced, not reproduced. Obtain the applicable standard and equipment classification through the facility designer; no permission to reproduce protected material is assumed.
- The DGX operating range is 5–30 °C. It does not establish an ASHRAE class or a recommended room set point.
- The Blackwell datasheet describes a liquid-cooled 72-GPU rack and does not, in the extract used here, state its kilowatts. A 25× performance sentence on that sheet is marked projected and subject to change. It is not a power measurement.
- No NetLeverage facility, power capacity or cooling plant is described.
Sources
- NVIDIA, NVIDIA H100 Tensor Core GPU datasheet PDF — https://dam-cdn.nvd.orangelogic.com/AssetLink/u5hh6fv4r7564i4y673484y3m20083nj.pdf (opened 10 November 2026). SXM maximum TDP up to 700 W, configurable; H100 NVL 350–400 W, configurable.
- NVIDIA, NVIDIA DGX H100 datasheet PDF — https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/nvidia-dgx-h100-datasheet.pdf (opened 10 November 2026). Eight H100 GPUs, 640 GB total, system power usage ~10.2 kW max, operating temperature 5–30 °C.
- NVIDIA, NVIDIA Blackwell datasheet PDF — https://dam-cdn.nvd.orangelogic.com/AssetLink/y441155802qub41q118b2852i557jem5.pdf (opened 10 November 2026). GB200 NVL72: 36 Grace CPUs and 72 Blackwell GPUs, liquid-cooled, rack-scale. Rack kilowatts not printed in the pages used. Performance comparisons marked projected, subject to change.
- ASHRAE TC 9.9, equipment thermal guidelines reference card, fifth edition — https://www.ashrae.org/File%20Library/Technical%20Resources/Bookstore/Supplemental%20Files/Therm-Gdlns-5th-R-E-RefCard.pdf. Consult the licensed source for applicable classes and limits; protected tabular values and notes are not reproduced.
- NVIDIA DGX H100 User Guide, system and power-supply specifications — https://docs.nvidia.com/dgx/dgxh100-user-guide/introduction-to-dgxh100.html.
This article was originally published on .