GPU server power: 2, 4 or 8 GPUs on 230 V 16 A, 32 A and 400 V three-phase feeds

How much power a GPU server draws with 2, 4 or 8 cards

A GPU server with two 600 W cards needs about 1.8 kW and fits one 230 V 16 A feed; a server with four 600 W cards needs about 3.4 kW and a 32 A or three-phase feed, unless the cards run at 450 W; an eight-card server needs 6 to 8 kW and a 400 V three-phase feed on each side. These are our estimates for servers with the NVIDIA RTX PRO 6000 Blackwell Server Edition or the NVIDIA H200 NVL, both rated at up to 600 W per card, next to the figures the makers publish. The RTX PRO 4500 Blackwell Server Edition at 165 W changes the picture: eight of them stay on a 16 A feed.

The cards set most of the load. NVIDIA lists the RTX PRO 6000 Server Edition at "Up to 600W (configurable)", the H200 NVL at up to 600 W and the RTX PRO 4500 Server Edition at 165 W. For the processors we take two AMD EPYC per server, sized to at least six cores per GPU: the 16-core EPYC 9135 at its default TDP of 200 W for two cards, the 32-core EPYC 9355 at 280 W for four and the 64-core EPYC 9555 at 360 W for eight. We then add 15 per cent for memory, drives, network cards, fans and conversion losses. That last figure is our assumption; a measurement on the delivered system replaces it with a real value.

Card power from NVIDIA's product pages for the three cards, read on 24 October 2026. Dell's "~8 kW" per node in the column "Est. Node Max Power (Full Load)" of its AI Factory with RTX PRO 6000 datasheet (XE7740 and XE7745, undated), read on 24 October 2026. Our estimate: cards plus two processors (2 × 200 W for two cards, 2 × 280 W for four cards and for eight RTX PRO 4500 SE, 2 × 360 W for eight 600 W cards, TDPs from AMD's EPYC 9135, 9355 and 9555 pages), plus 15 per cent.

The one published figure sits above our estimate. Dell labels its 8 kW per eight-card node as an estimated maximum at full load, and it covers Dell's own fan and switch design. For servers with two or four cards, and for any H200 NVL server, we found no published wall-power figure, so those rows are arithmetic only. Our article on how many GPUs fit in one server allows for up to about 10 kW for an eight-card server with high-TDP processors. Where a maker's figure exists for the exact configuration, we plan the feed on it; where none exists, we plan it on the rating of the power supplies the maker specifies for that configuration, until a measurement on the delivered system gives a real value.

Feed capacity: 230 V single-phase and 400 V three-phase

A single-phase feed delivers voltage times current; a three-phase feed delivers √3 × 400 V × current, which is the same as three 230 V phases of that current. We plan at 80 per cent of the rating, the margin in our guide to GPU rack power and cooling. IEC breakers are rated for continuous current at their nominal value, so the margin is our planning practice, and it leaves room for start-up peaks and for the other equipment in the rack.

Our arithmetic: 230 V × amps for single-phase, √3 × 400 V × amps for three-phase.

A three-phase feed only delivers its rating when the load is spread over the three phases. Each phase of a 32 A three-phase feed carries 32 A at 230 V, 7.36 kW, and a server's power supplies plug into outlets on individual phases. With four or eight supply cords per server, the cords can be spread so that each phase carries a similar share.

Which feed each configuration needs

Putting the two tables together gives the feed per rack position. With redundant supplies on A and B feeds, the feed on each side has to carry the whole server, so the column applies to each side.

Our estimates from the tables above. Two eight-card servers at Dell's 8 kW take 16 of the 17.74 kW we plan on one 400 V 32 A three-phase feed.

The two-card row stays well below the 2.94 kW of a 16 A feed by our estimate. Its margin depends on the rest of the configuration, so check the rating of the power supplies the maker specifies as well: if it lies above 2.94 kW, plan on it, or run the cards at the 450 W cap until a measurement on the delivered system gives a real figure. The four-card row is where the cap decides the feed: at 450 W per card the server stays on 16 A, at 600 W it needs 32 A or three-phase.

We build these servers with two, four or eight cards and choose power supplies, cords and the card power mode to fit the rack position; send us your feed rating, A/B requirement and PDU outlet type, and our team replies with the card count and power cap that fit before we quote.

Power supplies: counts, zones and redundancy by maker

The servers below use supplies of 2,400 to 3,200 W, but the rules for how many and in which arrangement differ from maker to maker. These are the rules in the documents we read.

Dell XE7740 Technical Guide Rev. A04 (April 2026) and XE7745 Technical Guide Rev. A09 (September 2026); HPE ProLiant Compute DL380a Gen12 QuickSpecs; Cisco UCS C845A M8 spec sheet Rev. A.11 (11 May 2026); Supermicro SYS-422GA-NRT datasheet; Puget Systems E282-4U configurator; read on 24 October 2026.

Dell's two guides are the most explicit. The CPU zone holds supplies 1 and 2 and the GPU zone supplies 3 to 8, and at least one supply has to be installed in each zone. For full redundancy the XE7740 guide asks for "1+1 in CPU zone and 3+3 in GPU zone", which is also how the A/B grid splits: each grid feeds one CPU-zone and three GPU-zone supplies. With eight 600 W cards, the XE7740 and the XE7745 support 4+4 A/B grid redundancy with 3,200 W supplies and not with 2,400 W ones, and 3+1 is not supported with either.

HPE asks for 2,400 or 3,200 W supplies whenever H200 NVL or RTX PRO 6000 Server Edition cards are configured, five of them for two or four double-wide cards (two for the system board, three for the GPUs) and eight for eight or ten. Supplies of different types cannot be mixed. Cisco's C845A takes up to four 3.2 kW supplies with N+1 redundancy on 100 to 240 V input, each with an IEC C20 inlet.

N+1 and A/B protect against different failures. N+1 keeps the server running when one supply fails; A/B keeps it running when a whole feed, PDU or UPS fails, which only works if the supplies on one side alone can carry the server. Puget states the consequence for its 3+1 design: redundancy works only while the server draws less than three of its four 3,200 W modules can deliver, 9.6 kW by our arithmetic. Before you plan A and B feeds for a four-supply server, ask the maker whether the configuration supports grid redundancy at all.

Why 230 V gets the full supply rating

Dell rates its 3,200 W Titanium supply at 3,200 W between 220.1 and 240 V and at 2,900 W between 200 and 220 V; its 2,400 W supply gives 2,400 W at 200 to 240 V and 1,400 W at 100 to 120 V. A European 230 V circuit therefore gets the full rating, and sizing advice written for 208 V circuits, where the same supply is rated at 2,900 W, does not carry over one to one. Lenovo states for its ThinkSystem SR650a V4 that "All AC power supplies support 230V power". At 230 V, a 3,200 W supply delivering full output draws more than 13.9 A: 3,200 W divided by 230 V gives 13.9 A, and the conversion losses come on top.

C13, C19 and CEE: cords, outlets and rack PDUs

Under IEC 60320, a C13 connector is rated at 10 A and a C19 at 16 A, both at 250 V. At 230 V that is about 2.3 kW per C13 outlet and 3.68 kW per C19 outlet. Supplies of 2,400 W and more need C19: Dell lists C19 cords for its 3,200 and 2,400 W supplies, and HPE states that its "2400W and 3200W supplies support C19/ C19-C20 Power Cords only". A rack PDU with only C13 outlets cannot feed these servers.

The feed into the rack PDU is a separate connector. A three-phase PDU such as APC's AP7526 takes 400 V three-phase at 32 A through an IEC 60309 32A 3P+N+PE plug, the connector known as CEE in much of Europe, is rated at 22,000 VA and has six C19 outlets at 230 V. An eight-card Dell server on A and B grids puts four cords on each PDU, so one such PDU per side feeds one eight-card server and leaves two outlets. Count outlets per side as carefully as kilowatts.

IEC 60320 ratings as listed in NetApp's knowledge base article on AC power cables (modified 9 June 2026); APC AP7526 product page, read on 24 October 2026.

A and B feeds: each side carries the whole server

With A and B feeds, half of the supplies hang on each feed and share the load in normal operation. When one feed is lost, the other side carries the whole server, so each feed, PDU and UPS has to be sized for the full load, not for half of it. For an eight-card server at 6.3 to 8 kW, that means a 400 V three-phase feed on each side: the room provides twice the server's load in feed capacity, and half of it stays in reserve.

The same rule applies to the rack total. Four four-card servers at 600 W, about 13.6 kW by our estimate, need a feed of at least that size on side A and the same on side B: one 400 V 32 A three-phase feed per side, planned at 17.74 kW, covers them. Without A/B, the same rack can run on one such feed, but a feed failure then stops every server in it.

Running the cards at 450 W: what the cap saves

The RTX PRO 6000 Server Edition has two power modes, set by how the 16-pin cable is strapped: 600 W or 450 W maximum by default, each with a 300 W minimum, according to NVIDIA's product brief SP-12355-001_v02. The H200 NVL's brief gives 600 W as maximum and default and 200 W as minimum. Both cards refuse to boot if the cable signals less power than their default cap, so an H200 NVL on a cable strapped for 450 W does not start. A cap set with nvidia-smi lasts until the next driver load, as our guide to the RTX PRO 6000 power limit explains.

For four cards the 450 W cap saves 600 W at the cards and about 700 W at the wall by our estimate, which is the difference between 32 A and 16 A. StorageReview measured what it costs on an RTX PRO 6000 Workstation Edition, which has the same GPU and memory: the card "gave up 2 to 5 percent at short context" and "4 to 9 percent at long context" at 450 W, and delivered 4.8 output tokens per second per rated watt against 3.8 at 600 W on a 14B model. Our page on RTX PRO 6000 Server Edition power at 600, 450 and 300 W covers the modes per card, and H200 NVL at 450 W the servers that cap that card.

Heat, UPS and the rest of the rack

Practically all the power a server draws becomes heat, 3.412 BTU/h per watt. An eight-card server at 6.3 kW therefore puts about 21,500 BTU/h into the room and needs about 1,000 CFM of air at an 11 °C rise, by the 160 CFM per kW in our rack guide. A UPS has to carry the same kilowatts for the ride-through time, with each side's UPS sized for the whole load under A/B; our guide to a UPS for a GPU server works through the sizes. Network switches, storage and management equipment in the same rack come on top of the server figures.

The server room itself is covered in GPU rack power and cooling, air-cooled AI servers in a 10 kW rack and our checklist for preparing a server room for a GPU server. For a machine outside a data centre, GPU server in an office also covers noise. We check feed, outlets and UPS against the configuration before we quote; describe your rack position and its feeds, and we return the power supplies, cords and card power mode per server with the configuration.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two, four or eight RTX PRO 6000 Server Edition or H200 NVL cards, or eight RTX PRO 4500 Server Edition, with the power supplies and cords chosen for your feeds, assembled and tested before delivery. Our entry server has two RTX PRO 6000 Server Edition, the inference node four, and the training node eight H200 NVL with NVLink bridges. Every component carries the manufacturer warranty, on one EU contract and invoice.

FAQ