Air-cooled AI server: 2 to 8 PCIe GPUs in an existing 10 kW rack

PCIe GPU servers run on air in a 5 to 15 kW rack

An AI server with two to eight PCIe GPUs runs air-cooled in an existing rack of 5 to 15 kW, as long as the rack position supplies the power and air the server needs: about 1.8 kW for two 600 W cards, about 3.4 kW for four and 6 to 8 kW for eight, by our estimate and Dell's published figure for its eight-card nodes. Liquid cooling belongs to rack-scale systems such as the GB200 NVL72 at about 120 kW per rack and becomes the practical route above about 50 kW per rack. One or two servers with NVIDIA RTX PRO 6000 Blackwell Server Edition, NVIDIA H200 NVL or RTX PRO 4500 Blackwell Server Edition cards need no liquid loop.

Most existing rooms are in this range. Uptime Institute's survey of July 2026, as cited in our guide to liquid cooling for GPU racks, shows the average typical rack just past 11 kW. NVIDIA wrote when it launched the H200 NVL in November 2024, citing a survey, that "roughly 70% of enterprise racks are 20kW and below and use air cooling", and positioned the card for them. The question for such a room is how many servers of which size each rack takes, and whether the inlet air stays inside the limits the server makers set.

Which AI hardware needs liquid cooling, and which does not

The power per rack draws the dividing line. The table puts the systems people ask about side by side.

NVIDIA's rack rating for the GB200 NVL72 and DGX figures as cited in our guides to liquid cooling for GPU racks and to the DGX B300; PCIe rows: our estimates as in our GPU server power guide, with Dell's estimated maximum of about 8 kW at full load for its eight-card XE7740 and XE7745 (AI Factory with RTX PRO 6000 datasheet).

The NVL72 racks are built around liquid: the compute and switch trays are liquid-cooled, and there is no air-cooled version. Eight-GPU HGX and DGX systems with SXM modules exist in air-cooled form, but each one draws 14 to 15 kW in 10 rack units, so it needs a high-density rack on its own; our liquid cooling guide cites Supermicro at 32 air-cooled B200 GPUs per rack against 96 with cold plates. We do not supply NVL72 racks or B200 systems; for the DGX B300 and HGX B300 we check price and availability per project.

PCIe servers are the other branch. The cards are passive, dual-slot or single-slot, and take their air from the server's front-to-back fans. Eight of them in a 4U server stay within one 10 kW rack position, and two to four cards fit a 2U or 4U server at a fraction of that. The cards are the ones we supply as standard: RTX PRO 6000 Server Edition, H200 NVL with NVLink bridges and RTX PRO 4500 Server Edition.

Power per server and servers per rack

For the server load we use the same method as our GPU server power guide: board power of the cards, two AMD EPYC processors sized to the card count, and 15 per cent for memory, drives, network cards, fans and conversion losses. A measurement on the delivered system replaces the 15 per cent with a real figure.

Our estimate: 600 W per RTX PRO 6000 Server Edition or H200 NVL, 165 W per RTX PRO 4500 Server Edition, 2 × 200 W processors for two cards, 2 × 280 W for four cards and for eight RTX PRO 4500 SE, 2 × 360 W for eight 600 W cards, plus 15 per cent. Servers counted to the rack's design load, before switches and other equipment.

One published figure moves a count down. Dell estimates its eight-card nodes at about 8 kW at full load, which still leaves one in a 10 kW rack but only one in 15 kW. For servers with two or four cards we found no published wall-power figure; where the rating of the power supplies the maker specifies for the configuration lies above our estimate, count the rack on that rating until a measurement on the delivered system gives a real value. Network switches, storage and a management switch in the same rack take their share of the budget too.

The rack budget is also an electrical one. A 10 kW rack is more than the 8.87 kW we plan on a 400 V 16 A three-phase feed, so it needs a 32 A three-phase feed, planned at 17.74 kW, which also covers 15 kW. With A and B feeds, each side has to carry the whole rack. Our power guide has the feed per configuration and the power supply rules of Dell, HPE and Cisco.

We build these servers with two, four or eight cards; send us the kW per rack, the feed type and the inlet temperature, and our team replies with the card count and chassis that fit each position.

Inlet temperature: 30 °C with RTX PRO 6000 Server Edition and H200 NVL

The server makers set the inlet limits per configuration, and the 600 W cards have the lowest ones. In Dell's eight-card servers, the PowerEdge XE7740 with Intel Xeon 6 and the XE7745 with AMD EPYC, the maximum ambient temperature is 30 °C for the RTX PRO 6000 Server Edition and the H200 NVL, against 35 °C for the H100 NVL and the L40S.

Dell PowerEdge XE7740 Technical Guide Rev. A04 (April 2026) and XE7745 Technical Guide Rev. A09 (September 2026), read on 24 October 2026; Dell writes "RTX Pro 6000" for the 600 W passive card. Supermicro SYS-422GA-NRT datasheet: "up to 8 double-width 600W PCIe GPU ... at 35ºC", read in October 2026.

Dell's general operating range for the XE7740 is 10 to 35 °C up to 900 m altitude, with the maximum falling by 1 °C per 300 m above that. At the card, NVIDIA's product brief for the RTX PRO 6000 Server Edition asks for 43 CFM through the heat sink at a 25 °C inlet and 125 CFM at 45 °C: the warmer the air, the more of it the server fans have to move. ASHRAE recommends 18 to 27 °C at the server inlet, so a room that already runs at the top of that band leaves 3 °C to Dell's 30 °C limit. Measure the inlet at the rack position under today's load, at the height where the GPU server will sit, and compare it with the limit for the exact configuration. Our comparison of GPU server inlet temperature at 25, 30 or 35 °C lists more servers.

Airflow, blanking panels and fillers

Every kilowatt the server draws leaves it as heat in the exhaust air. By the 160 CFM per kW at an 11 °C rise that our guide to GPU rack power and cooling uses, a 10 kW rack needs about 1,600 CFM of cooled air at its front, and an eight-card server at 6.3 kW about 1,000 CFM on its own. Check what the floor tiles or the in-row unit deliver at that rack position, not the average for the room.

Air that is delivered has to reach the intakes cold. Schneider Electric's white paper on blanking panels states that "unused vertical space in open frame racks and rack enclosures creates an unrestricted recycling of hot air". The checks for a GPU server in an existing rack are short:

• Blanking panels. Close every empty rack unit above, below and next to the GPU server, so hot exhaust cannot flow forward through the rack into the intakes.
• GPU blanks and fillers. Dell's XE7740 guide lists GPU blanks, shroud fillers and rear GPU shrouds for each GPU configuration, and the XE7745 guide lists air gap sealing fillers. Keep them installed as delivered, also when a server starts with fewer cards than it has slots.
• Front to back. Mount the server in the same airflow direction as the rest of the row, with nothing blocking the front bezel or the rear exhaust.
• Containment. Our rack guide puts the limit of ordinary room air at about 20 kW per rack; above that, hot or cold aisle containment is needed. A 5 to 15 kW rack stays below it.
• Noise. Dell gives 87 dB(A) for the XE7740 and XE7745 at maximum load and 35 °C, recommends ear protection during operation and does not recommend the XE7740 for acoustically sensitive environments. Plan service work in the rack with that in mind.

Splitting the load: two smaller servers or a power cap

When one rack position cannot take an eight-card server, there are three ways to keep the same number of GPUs without liquid cooling.

1. Two four-card servers instead of one eight-card server. Two servers at about 3.4 kW each can go into two racks of 5 kW, or into one rack of 10 kW. They also keep half the capacity when one server fails, as our comparison of one 8-GPU server or two 4-GPU servers explains.
2. The 450 W cap. Four RTX PRO 6000 Server Edition cards at 450 W bring the server from about 3.4 to 2.7 kW by our estimate. StorageReview measured the cost on the Workstation Edition, which has the same GPU and memory, at 2 to 5 per cent of throughput at short context and 4 to 9 per cent at long context. The H200 NVL does not start on a cable strapped for less than its 600 W default.
3. Smaller cards for many small workloads. Eight RTX PRO 4500 Server Edition cards at 165 W each draw 1.32 kW and give 256 GB of GPU memory in eight 32 GB pools, which suits virtual desktops, smaller models and many parallel services; our page on the 8× RTX PRO 4500 Server Edition shows what it runs.

When an existing room is not enough

An existing room reaches its limit in one of three places: the electrical feed, the cooled air per rack position, or the inlet temperature at the server. The first two can often be solved by spreading servers over more racks. The third cannot, because a 600 W card in an eight-card chassis needs air below the maker's limit, 30 °C in Dell's servers. A room whose inlet temperature rises above that on warm days needs more cooling capacity before the server arrives, whatever the rack density.

Above about 20 to 25 kW per rack, Uptime Institute's figures in our rack guide point to containment and in-row cooling, and above about 50 kW to rear-door heat exchangers or direct-to-chip liquid. A rack of four eight-card PCIe servers at 25 to 32 kW is therefore a room project, while one or two GPU servers in a 10 kW rack are not. Our checklist for preparing a server room for a GPU server covers floor loading, power and cooling in order, and GPU server power the feeds. We check rack, feed and inlet temperature before we quote; describe the room and the rack positions you have, and we tell you which configuration runs there on air.

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, all air-cooled, with the chassis, fans and power mode chosen for your rack positions. Each server is assembled and tested before delivery, and every component carries the manufacturer warranty on one EU contract and invoice. Commissioning on site is available on request.

FAQ