RTX PRO 6000 Blackwell: Workstation, Max-Q and Server Edition compared

The three RTX PRO 6000 Blackwell editions are the same 96 GB GPU in three coolers: the Workstation Edition is a 600 W card for one desktop, the Max-Q Workstation Edition is a 300 W card you can fit four of in one tower, and the Server Edition is a passive 600 W card for rack servers. They differ in power, clock speed, size and, for the Server Edition, memory bandwidth: 1,597 GB/s against 1,792 on the two workstation cards. On 2026-11-12 one US dealer listed each edition at $14,999.99, in stock, and another at $17,135.99 to $17,158.99 on backorder, against a launch price of $8,565 in April 2025.

The three editions side by side#

The numbers below are from NVIDIA's three datasheets, its Server Edition product brief and its RTX Blackwell PRO architecture whitepaper, all read on 2026-11-12.

All three use the same chip with 24,064 CUDA cores, 752 Tensor Cores and compute capability 12.0, and all three support FP8 and FP4. NVIDIA publishes dense Tensor Core rates only for the two workstation cards. The Server Edition's dense figures are our calculation from its 2,430 MHz boost clock (1,007.6 x 2,430 / 2,617 = 935.6, and 2,015.2 x 2,430 / 2,617 = 1,871.2), which is how the RTX PRO 6000 page works them out too.

Workstation Edition: one 600 W card on a desk#

The Workstation Edition is the fastest of the three and the largest. NVIDIA's datasheet calls it a single-GPU powerhouse for AI developers, researchers, engineers and creative professionals. It draws 600 W through one 16-pin connector, and its double flow-through cooler makes the card 5.4 inches tall and 12 inches long, taller than a standard full-height card. Check the case before you order: NVIDIA lists it as extended height.

It is the edition to buy for one card in a tower. It has the highest clock, 2,617 MHz, and the full 1,792 GB/s of memory bandwidth, and its four DisplayPort 2.1b outputs drive monitors directly.

Max-Q Workstation Edition: 300 W, and four in one tower#

The Max-Q gives up 13% of the compute for half the power. Its dense FP8 rate is 877.9 TFLOPS against 1,007.6, the same ratio as its 2,280 MHz clock against 2,617 (our calculation: 877.9 / 1,007.6 = 0.87). Memory and bandwidth are unchanged at 96 GB and 1,792 GB/s, so a job bound by memory bandwidth, such as one stream of generated tokens, should lose less than that 13%.

What the lower power buys is density. The Max-Q is a standard 4.4 by 10.5 inch dual-slot card with a blower, and NVIDIA's datasheet says to scale up to four of them in one workstation. Four cards draw 1,200 W between them and hold 384 GB (our calculation: 4 x 300 W and 4 x 96 GB). There is no NVLink on this family, so four cards pool memory through software over PCIe, and vLLM's multi-GPU guide covers what that means for serving one model across them.

Server Edition: passive, for chassis airflow#

The Server Edition has no fan and depends on the server to cool it. NVIDIA's product brief gives the minimum airflow through its heat sink: 43 CFM at a 25 C inlet, rising to 125 CFM at 45 C, fully ducted. That rules out an ordinary desktop case. The same brief gives two power modes, 600 W and 450 W, with a 300 W minimum, and lists NVLink as not supported.

Two details set it apart from the workstation cards beyond cooling. Its memory bandwidth is 1,597 GB/s, 11% below the 1,792 GB/s of the other two (our calculation: 1,597 / 1,792 = 0.89), and its boost clock is 2,430 MHz, 7% below the Workstation Edition. Its four DisplayPort outputs are off by default, because the default mode is the one NVIDIA's virtual GPU software needs. The datasheet also lists confidential computing and secure boot, which the workstation datasheets do not. NVIDIA describes the card as built for high-density server platforms, and a liquid-cooled single-slot version exists for denser builds.

What each edition costs#

All three editions cost about the same as each other, and about twice what the card launched at. NVIDIA's pages carry no price, so these are dated figures from named sources:

The two dealers disagree by more than $2,100 on the same PNY part numbers on the same day, and the cheaper one had stock. So the honest range for one card on 2026-11-12 is $15,000 to $17,200, whichever edition you pick. That is 1.75 to 2 times the $8,565 launch price (our calculation: 14,999.99 / 8,565 and 17,135.99 / 8,565). Tom's Hardware ties the increases to the memory shortage and reports that NVIDIA changed its own listing twice in 2026 without an announcement. Central Computer also asks for a compliance approval before it ships these cards, so allow time for paperwork.

Which edition QuantaCloud rents#

QuantaCloud's catalog lists the card only as "RTX PRO 6000 Blackwell", and nvidia-smi -q on any VM you launch shows which edition it is. That is what the RTX PRO 6000 page says, and I will not guess past it here. The difference that matters on a rented VM is bandwidth: 1,597 GB/s on the Server Edition and 1,792 GB/s on the Workstation and Max-Q.

What fits does not depend on the edition. Public nvidia-smi output shows 97,887 MiB on Server and Workstation Edition cards alike, which gives a working budget of about 88 GiB at vLLM's default of 92% (our calculation: 97,887 / 1,024 x 0.92 = 87.9). The live configurations and price:

Prices checked 14 Nov 2026, 08:59 UTC

Buy one or rent one#

The break-even is 6,276 to 7,170 hours of rental, depending on which listing you buy from. A 1x RTX PRO 6000 VM cost $2.39 an hour on 2026-11-12 (today $2.16/GPU-hr), and that hour included 16 vCPUs, 144 GB of RAM and 725 GB of disk. Dividing each card price by $2.39 (our calculation):

The arithmetic for the first row: 14,999.99 / 2.39 = 6,276 hours. At 10 hours a week that is 627.6 weeks, 12.1 years, and at 40 hours a week it is 156.9 weeks, 3.0 years. Running around the clock uses 168 hours a week, so 6,276 hours pass in 37.4 weeks, 8.6 months. If the live price differs from $2.39, divide by the live number.

The card price is not the whole cost of owning one. It needs a host with a PCIe 5.0 x16 slot, a 16-pin power feed and room for the cooler, and the Server Edition needs a server chassis. Power is yours too: 6,276 hours at the 600 W limit is 3,766 kWh, about $684 at the average US residential price of 18.16 cents per kWh in August 2026 (our calculation from EIA's figure). The rental price leaves out one thing as well: stopping a QuantaCloud instance deletes its disk, so a rented card is for work you can pick up again from files you copied off.

My rule is the hours. Under about 20 hours a week, a card at these prices takes more than six years to pay for itself against the rent, and I would rent. Past 40 hours a week of real GPU load, year after year, owning starts to make sense, and at that point the choice between editions is the choice of machine: Workstation Edition for one card on a desk, Max-Q for two to four in a tower, Server Edition for a rack. The GPU cost calculator runs the same sum for your own hours at live prices, and AI server cost covers the machine around the card.

FAQ#

Is the Server Edition slower than the Workstation Edition?

A little, on paper. Its memory bandwidth is 1,597 GB/s against 1,792, and its boost clock is 2,430 MHz against 2,617, which puts its dense FP8 rate at about 936 TFLOPS against 1,007.6. The memory is the same 96 GB.

Can I put a Server Edition card in a desktop PC?

Not without ducted airflow. The card is passive, and NVIDIA's product brief asks for at least 43 CFM through the heat sink at a 25 C inlet and 125 CFM at 45 C. Its display outputs are also off by default. For a desktop, the Workstation Edition and the Max-Q are the editions with their own fans.

How many RTX PRO 6000 cards can work together?

NVIDIA says up to four Max-Q cards in one workstation, and the Server Edition goes into rack servers that take eight, such as the 4U system priced in AI server cost. The Server Edition brief lists NVLink as not supported, so the cards talk over PCIe, and software has to split the model across them. InfiniBand vs NVLink explains what each link is for.

Does the edition change what fits in 96 GB?

No. All three carry 96 GB of GDDR7 with ECC, and the cards report the same 97,887 MiB. The edition changes how fast the memory is read and how much power the card may draw, not how much model it holds. The RTX PRO 6000 page lists what fits.

Why does a 96 GB card cost more than it did at launch?

Because the listing moved, twice. Tom's Hardware reported NVIDIA's own price for the Workstation Edition at $8,565 in April 2025, $13,250 in June 2026 and $16,000 in August 2026, and ties the increases to the memory shortage. Dealer prices on 2026-11-12 sat on either side of that last figure.

Can I get dedicated RTX PRO 6000 servers instead of buying cards?

Yes, as reserved capacity built to order. We order and build the servers to your spec, and the configuration, lead time and terms are quoted in writing. Dedicated GPU servers covers the options, or you can send a capacity brief.

The decision rule is short: pick the edition by the machine it goes in, and pick buying or renting by the hours. If you are not sure the card will be busy 40 hours a week, rent one first, run your own workload, and count the hours before you spend $15,000. RTX PRO 6000 vs RTX 5090 covers the consumer card people weigh against it, and the deploy guide covers launching.