Buying a rack server with the NVIDIA RTX PRO 6000 Blackwell Max-Q: which chassis list it

Who buys a rack server with the Max-Q

You can buy a rack system with NVIDIA RTX PRO 6000 Blackwell Max-Q cards from us when the chassis maker lists the card for that model. As of October 2026 few rack models qualify: Lenovo lists the Max-Q for its 2U ThinkSystem SR655 V3 and SR665 V3 with the entry 3, Lenovo and HP list up to four for rack-mountable workstations, and Supermicro names the card in the GPU lists of two GPU servers without saying how many Max-Q they take. Our team sizes the configuration to your workload and budget and replies within one business day.

The Max-Q makes sense in a rack for a fairly narrow set of buyers. A development team that already works on Max-Q workstations may want the same card in the server room, so that containers and drivers move from the desk to the rack unchanged. A lab with a fixed power budget may want 96 GB per card at 300 W and the full 1,792 GB/s of memory bandwidth. Both cases run on bare-metal Linux, usually for one team, and share the cards through MIG or a scheduler rather than through virtual machines.

For everything else in a rack we recommend the RTX PRO 6000 Server Edition: a service that runs continuously for the whole company, virtual machines with vGPU, NVIDIA AI Enterprise support, eight cards in one node, or a choice among many server models. The reasons are set out below, and the decision for a four-card node is worked through in our comparison of the Max-Q and the Server Edition in a 4-GPU server.

Which chassis makers list the Max-Q

The table shows the rack-mountable systems whose makers' own documents name the Max-Q, as we found them on 24 October 2026, and one widely used GPU server that does not list it.

Lenovo Press LP2364 (Max-Q, updated 8 May 2026) and LP2263 (Server Edition, updated 28 July 2026); Lenovo PSREF ThinkStation PX (9 October 2026); HP Z8 Fury G6i specifications; Supermicro SYS-551A-T, SYS-422GA-NRT and SYS-322GA-NR pages; HPE ProLiant Compute DL380a Gen12 QuickSpecs; all read on 24 October 2026.

The Lenovo pair is unusual in two ways. Lenovo's option guide LP2364 offers the card as "ThinkSystem NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 96GB PCIe Gen5 Active GPU" and gives the entry 3 for both servers, with N for every other current and withdrawn ThinkSystem model, including the SR650a V4 and SR675 V3 that take the Server Edition. The Server Edition's guide LP2263 shows the reverse: N for the SR655 V3 and SR665 V3. Across these two guides, no Lenovo server lists both editions. LP2364 also states "vGPU software support: No support", ships the card "with no auxiliary power cables", because the cables are server-specific, and lists a thermal kit per server: the SR655 V3 GPU Enablement Kit and the SR665 V3 GPU Full Length Thermal Option Kit. The guide has no legend for the 3; we read it as the supported number of cards and confirm the count in Lenovo's configuration before we quote.

The rack-mountable workstations are towers that lie on rails. Lenovo's PSREF sheet for the ThinkStation PX gives "Tower (54L) or 5U rack, optional rack rail kit needed" and "up to four NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (300W)", with an optional IPMI 2.0 BMC. HP lists "Up to 4 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition GPUs" for the Z8 Fury G6i, which fits a 5U rack with an optional rail kit. Supermicro's SYS-551A-T is a 5U full-tower rackmount for one Xeon W-3500 or W-3400 with up to four double-width GPUs, and its GPU list names the Max-Q and the 600 W Workstation Edition, not the Server Edition; the list gives no count per card. None of these lists the passive card.

Supermicro's SYS-422GA-NRT and SYS-322GA-NR are the GPU servers in the table that name both editions. The 4U SYS-422GA-NRT's datasheet states "Support up to 8 double-width 600W PCIe GPU accelerator cards at 35ºC" and, separately, eight RTX PRO 6000 Server Edition at 35 °C, with a dual-root PCIe 5.0 switch between processors and GPUs, two Xeon 6900 processors and four 3,200 W supplies in a 3+1 arrangement; its GPU list names the Max-Q, the Server Edition, the RTX PRO 4500 Server Edition and the H200 NVL. The 3U SYS-322GA-NR takes up to eight double-width GPUs and lists the Max-Q next to the Server Edition; Supermicro's RTX PRO systems page gives it up to four RTX PRO 6000 Blackwell, without naming an edition. Neither datasheet gives a card count for the Max-Q, so we ask Supermicro to confirm the number of Max-Q for the exact configuration before we quote. Eight cards at 300 W are the subject of our page on eight GPUs at 300 W.

Outside the makers' own lists, system builders also sell rack servers with four Max-Q. Puget Systems, a US builder, offers a 2U "Quad GPU Large Language Model Server" with up to four Max-Q as its default card, two redundant 2,600 W supplies and the note that 240 V is required for power-supply redundancy. A builder's offer is a build on someone else's chassis, so the question to ask is whether the chassis maker lists the card for those slots. We build Max-Q systems only on chassis whose maker lists the card; tell us which model you have in mind, and we check its GPU list, slots and cables before we quote.

Why most GPU servers do not list it

Server makers do not publish their reasons, but the two cards are built for different airflow. A GPU server pushes air from the front to the back through passive heat sinks. NVIDIA's product brief for the Server Edition asks for a minimum airflow through each card's heat sink of 43 CFM at a 25 °C inlet, 65 CFM at 35 °C and 125 CFM at 45 °C, from either side, and the server's fans and ducts are designed to deliver it.

The Max-Q brings its own fan. NVIDIA's datasheet lists the thermal solution as "Active", and Puget Systems describes it as a blower that takes air from inside the system and pushes it out through the bracket. In a front-to-back chassis that works only if the bracket faces the hot aisle and the blower can draw enough air from the slot beside it. In a dense riser with double-width slots side by side, the intake of each blower sits against the next card. Puget limits its desktops to three Max-Q with a slot gap and writes that four 300 W cards are hard to cool; a four-card build log on the Level1Techs forum (July 2026) reports 91 °C on the two inner cards at 300 W, close to the 88 to 92 °C at which its cards throttled. A server maker that lists the Max-Q has qualified a slot layout, a fan profile and a cable set for it; one that does not has not.

There is also a support side. NVIDIA's list of GPUs supported by vGPU, updated on 2 October 2026, names the RTX PRO 6000 Server Edition from release 19.0 and its liquid-cooled version from 20.0, and no Max-Q. NVIDIA's AI Enterprise 8.2 support matrix (2 September 2026) and its list of NVIDIA-Certified Systems (4 October 2026) name the Server Edition and not the Max-Q, as our guide to the Max-Q in a rack server sets out with its checklist for slots, cables and BMC fan control.

What a Max-Q rack server contains

The configuration follows the chassis. The table shows the three shapes a Max-Q rack system takes and what we size in each.

Makers' documents cited above; system memory by NVIDIA's configuration guide for NVIDIA-Certified Systems (30 September 2026): at least twice the total GPU memory.

We apply NVIDIA's host rule from its configuration guide for NVIDIA-Certified Systems, even though no Max-Q system is certified: at least six physical cores per GPU and system memory of at least twice the GPU memory, so 576 GB for three cards and 768 GB for four. For storage we plan mirrored NVMe drives for the system and the model store, and for the network 25 GbE for users and 100 GbE or more where models load from shared storage. Every card needs its own 16-pin cable from the chassis maker's kit, since server options ship without them.

What three, four or eight Max-Q run

Each Max-Q has 96 GB, of which the driver reports 95.6 GiB. By our users-per-card rule, 0.9 × 95.6 GiB less 3 GiB per card, three cards hold 249.1 GiB of weights and KV cache, four hold 332.2 GiB and eight 664.3 GiB.

Our estimates by the users-per-card rule: concurrent conversations at full context length, FP8 KV cache for Llama 3.3 70B, 16-bit for the others; same arithmetic as our four-card Server Edition guide.

The counts are the same as for Server Edition cards, because both editions have 96 GB; our guide to what a 4× RTX PRO 6000 Server Edition server runs has the full model list and layouts. Three cards are awkward for one large model, since serving engines usually split a model over two, four or eight GPUs, so on the Lenovo pair we plan copies or a pair plus a single card. Speed per conversation follows memory bandwidth, where the Max-Q's 1,792 GB/s is about 12 per cent above the Server Edition's 1,597 GB/s.

We supply the Max-Q and build these systems to order; send us the models you plan to run and the peak number of users, and we reply with a configuration and the conversations per layout.

When we recommend the Server Edition instead

For most rack projects the Server Edition is our recommendation. Speed is rarely the reason; cooling, software support and the choice of chassis are.

• Continuous operation. A passive card in a server designed for it is cooled by the server's own fan control; a blower card in a dense chassis depends on intake air the server was not built to deliver.
• Virtual machines. vGPU runs only on the Server Edition. On the Max-Q, virtual machines get whole cards through passthrough, as our page on the Max-Q in virtual machines explains.
• Support. NVIDIA AI Enterprise support and the NVIDIA-Certified Systems list cover the Server Edition. Lenovo's own Max-Q option carries a one-year limited warranty, or the server's warranty once installed in a Lenovo server.
• Choice of chassis. HPE, Lenovo, Cisco, Dell and Supermicro list the Server Edition in two-, four- and eight-card servers, as our list of Server Edition compatible servers shows. For the Max-Q we found the handful above.
• Power. The Server Edition runs at 600 W or capped at 450 W, and its product brief gives a 300 W minimum, which brings it to the Max-Q's power while keeping passive cooling. We found no published measurement of its speed at 300 W.

The Max-Q remains the right card where a maker lists it, the work runs on bare metal and the 300 W budget or the higher bandwidth matters. A tower under a desk is a different purchase, covered in our guide to buying a 4× Max-Q workstation.

Power, room and software

Three Max-Q draw 900 W, four 1,200 W and eight 2,400 W for the cards. By our estimate (cards plus processors plus 15 per cent for the rest), a four-card system with two 280 W AMD EPYC 9355 comes to about 2.0 kW at the wall, which one 230 V 16 A feed carries, also on each side of a redundant pair. Eight cards come to about 3.6 kW with two 360 W EPYC 9555 and about 3.9 kW with two 500 W Xeon 6900, beyond one 16 A circuit. The arithmetic for each count, the cords and the UPS are in our page on Max-Q power on 230 V.

On request the system arrives with a Linux release the chassis maker supports, the NVIDIA driver, CUDA and a container runtime with the NVIDIA Container Toolkit. MIG splits each Max-Q into up to four 24 GB instances; it needs driver 575.51.03 or later, a current vBIOS and NVIDIA's DisplayModeSelector to switch the card to compute mode, which turns its display outputs off. We set the cards to compute mode before delivery if you plan MIG, as described in our guide to RTX PRO 6000 MIG profiles.

How ordering works and what to send us

The order runs as for any of our AI servers:

1. You describe the workload and the chassis you prefer, or ask us to choose one. Sizing is free of charge.
2. We check the maker's GPU list for the exact model, the slot layout, the cables and the power supplies, and say in advance if the Max-Q does not fit.
3. You receive a written quote within one business day with price, lead time and warranty terms per component; quantities and lead time are confirmed in writing with the quote.
4. You order on one EU contract with an EU invoice from Eurokommerz in Vienna.
5. We assemble the system and test it before delivery (power-on and basic load test).
6. We deliver anywhere in the EU, with commissioning on site on request. Every component carries the manufacturer warranty, handled through us, and the assembly is covered by us.

For the first reply, these points help: the models and context lengths, the peak number of requests in flight, whether virtual machines are planned, the chassis model if you already own one, the rack position with feed type and inlet temperature, and the delivery site. Write these points in the form below, and our team returns a configuration with the edition we recommend for it.

FAQ