Buying an NVIDIA RTX PRO 5500 Blackwell workstation: what NVIDIA has published so far

What a company can buy today

The NVIDIA RTX PRO 5500 Blackwell is an 84 GB workstation card (1,398 GB/s, up to 600 W) whose specifications NVIDIA still marks as preliminary. You can buy a workstation with one or two of them from us now: our team sizes it on NVIDIA's published figures and replies within one business day, and the final specifications, quantities and lead time are confirmed in writing with the quote.

On 24 October 2026 NVIDIA's US and EU product pages for the card show a "Coming Soon" heading with a sign-up "to be notified when the NVIDIA RTX PRO 5500 is available", and the specification section ends with the note "Preliminary product specifications, subject to change". NVIDIA describes the card as "designed for rack-mounted workstation deployment with a choice of air- or liquid-cooled thermal solutions". Our article on the RTX PRO 5500 Blackwell 84 GB follows its launch status.

This page is for companies that already know they want 70B-class models in 8-bit on one card, or large scenes in 84 GB, and want to plan the workstation around the card before the datasheet appears. If you need a system with fully published specifications, the last sections compare the cards that have them.

What NVIDIA has published, and what it has not

We read NVIDIA's product page, the MIG user guide and the vGPU supported-GPU list on 24 October 2026. We plan only on the figures in the table; everything else below is our arithmetic or is confirmed with the quote.

NVIDIA RTX PRO 5500 product page (US and EU), read 24 October 2026. All card figures are preliminary per NVIDIA.

What is missing matters as much. NVIDIA's page gives no CUDA core count, no FP32 rate, no AI TOPS figure, no clocks, no memory bus width and no power connector, and NVIDIA's product literature lists no datasheet for the card. We size only on the figures in the table. Board dimensions, cooler design, video engines and outputs of the version you order are confirmed with the quote; the page mentions a liquid-cooled version in an "RXM form factor" without defining it, and we treat it as a separate form factor whose details also come with the quote.

NVIDIA's software documents do not cover the card yet: the MIG user guide, last updated on 13 October 2026, does not list it, and the vGPU supported-GPU list of 2 October 2026 does not include it.

Where the 5500 sits among the workstation cards

NVIDIA product pages, datasheets and MIG user guide (13 October 2026) as cited in our card guides; RTX PRO 5500 from its product page, read 24 October 2026.

On these figures the 5500 has 17 per cent more memory than the 72 GB RTX PRO 5000 and 4 per cent more bandwidth, by our arithmetic. The RTX PRO 6000 Max-Q has 12 GB more memory and 28 per cent more bandwidth at half the 5500's stated maximum power. Token generation on a model that fits a card follows bandwidth, so on such a model the 5500 should generate at close to the rate of the 5000 and below the rate of either RTX PRO 6000. Whether the 5500's higher power goes into faster prompt processing, rendering or training will only show once NVIDIA publishes its compute figures; we found no independent test of the card.

The 5500 earns its place by capacity. 84 GB is the step at which a 70B model in FP8 fits on one card, which the 72 GB card cannot do. Our comparison of the RTX PRO 5500 and the RTX PRO 5000 72 GB and the earlier one against the RTX PRO 6000 Max-Q go through the trade-offs.

Workstation configurations we can plan now

We plan the 5500 in one-card and two-card workstations. NVIDIA states no maximum number of cards per system. On 24 October 2026 none of the three large-tower specifications we checked listed the RTX PRO 5500: Lenovo's PSREF sheet for the ThinkStation PX of 9 October 2026, HP's page for the Z8 Fury G6i and Dell's configurator for the Precision 7960. A tower for it therefore means a platform we build on, sized for 600 W cards, or a maker's tower once its maker lists the card.

Our AI servers page (workstation platforms); memory and cores by NVIDIA's configuration guide for NVIDIA-Certified Systems (30 September 2026), written for servers and applied here by the same arithmetic.

NVIDIA's guide asks for system memory of at least twice the GPU memory and at least six physical cores per GPU. Two 84 GB cards make 336 GB, which we round up to 384 GB as eight 48 GB modules on eight channels; one card makes 168 GB, rounded to 192 GB. AMD lists the Threadripper PRO 9975WX with 32 cores, 350 W, 128 PCIe 5.0 lanes and eight memory channels, so two cards at x16 leave lanes for NVMe drives and a fast network card.

The cooler design is not in a datasheet yet. A two-card tower for 600 W cards needs the spacing, airflow and power cables that the final datasheet and the card's board design call for, and we confirm them before the order. Send us the tower your company runs or the platform you prefer, and our team checks slots, cables and power against the card as specified at the time of the quote.

What 84 GB holds by our sizing rule

Our users-per-card rule takes 90 per cent of the memory the driver reports, less 3 GiB per card for the runtime and activations, and the rest holds weights and KV cache. The driver reports 95.6 GiB on a 96 GB card; for 84 GB we scale that by capacity, an assumption on preliminary specifications, which gives 72.3 GiB per card and 144.6 GiB for two. Conversations are 8,192 tokens with an FP8 KV cache: 0.625 GiB each for Qwen3-14B, 1.0 GiB for Qwen3-32B and 1.25 GiB for Llama 3.3 70B. gpt-oss-120b is counted at 32K tokens with a 16-bit cache of 1.125 GiB.

Our estimate by the users-per-card rule with 72.3 GiB usable per RTX PRO 5500 and 61.5 GiB per RTX PRO 5000 72 GB, ECC off; weights Qwen3-14B FP8 15.2 GiB, Qwen3-32B FP8 32.0 GiB, Llama 3.3 70B NVFP4 39.8 GiB and FP8 67.7 GiB, gpt-oss-120b 60.8 GiB, as in our model guides. Memory arithmetic, not measured throughput.

The FP8 row is the main difference from the 72 GB card. One 5500 loads Llama 3.3 70B in FP8 with 4.6 GiB left for cache, enough for three conversations of 8K, which suits one developer or a small test team. Two cards running the model as one TP=2 copy hold 61. With the example values of our guide to a private ChatGPT server by company size, about 20 requests are in flight at peak for 500 staff, so the two-card build covers that company size by memory. If ECC reserved the 6.25 per cent that NVIDIA's CUDA Best Practices Guide gives for GDDR memory, usable memory would be about 67.6 GiB per card and the one-card FP8 case would no longer fit; NVIDIA publishes no figure for GDDR7.

Speed per conversation follows bandwidth. By our arithmetic, reading 70.6 GB per token for Llama 3.3 70B in FP8 puts the single-user ceiling at about 20 tokens per second on 1,398 GB/s, and reading 40.6 GB in NVFP4 at about 34; the RTX PRO 6000 reaches about 44 in NVFP4 on the same rule. These are ceilings from the specifications, not measurements. The cards have no NVLink, so a TP=2 copy exchanges its partial results over PCIe; our article on one model on several GPUs over PCIe or NVLink explains the cost.

MIG, vGPU and software

NVIDIA's product page FAQ gives MIG as "one instance with up to 84 GB, or two instances with up to 42 GB each". By our rule a 42 GB half leaves about 34.6 GiB usable, enough for Qwen3-14B in FP8 with about 31 conversations of 8K, or Qwen3-32B in FP8 with about two. The profile names, the minimum driver and any vBIOS or display-mode requirement are not documented yet, because the MIG guide does not list the card. On the RTX PRO 6000 and 5000, MIG needs Linux, driver 575.51.03 or later and the card switched to compute display mode, which turns its display outputs off; expect the 5500 to follow a similar procedure, and treat MIG as a point the quote confirms.

The card is not on NVIDIA's vGPU list. A workstation with the 5500 runs its workloads on the host or in containers, or passes whole cards to virtual machines. For virtual desktops and compute VMs with vGPU, NVIDIA lists the RTX PRO 6000 Server Edition from release 19.0, the RTX PRO 4500 Server Edition from 20.0, and the RTX PRO 5000 72 GB from 20.2 on Red Hat Enterprise Linux with KVM only; our guide to MIG and vGPU: VMs per GPU counts the seats.

On request the workstation arrives with a Linux release or your company's image, the NVIDIA driver, CUDA and a container runtime with the NVIDIA Container Toolkit. Which driver branch first supports the card is part of the quote. As of October 2026 the R580 branch has long-term support to June 2028, and NVIDIA's security bulletin of 30 September 2026 names 580.178.04 and 595.91.07 as fixed versions; our guide to NVIDIA driver branches and CUDA versions explains how to pin a branch.

Power on 230 V at the stated maximum

NVIDIA gives the card's power as "up to 600 W". We plan with that maximum until the datasheet appears. By our estimate a workstation draws its cards' board power, the power of a 350 W processor and 15 per cent on top for memory, drives, fans and power-supply losses.

Our estimate from NVIDIA's preliminary maximum of 600 W per card and the 350 W TDP of the Threadripper PRO 9975WX, plus 15 per cent, at full load. A 230 V 16 A circuit delivers 3.68 kW, planned at 80 per cent as 2.94 kW. The same figures as for one and two RTX PRO 6000 Workstation Edition cards.

One and two cards stay on one ordinary circuit. Four would not, by this estimate, which is one more reason we plan the card in one-card and two-card builds until a maker or NVIDIA states more. A two-card tower turns about 1.8 kW into heat, roughly 6,100 BTU per hour, which belongs in a technical room or an air-conditioned office. Large workstation supplies give their full rating only at 200 V or more, which European sockets provide. An IEC C13 cord is rated at 10 A, about 2.3 kW at 230 V, so the largest supplies use C19 cords rated at 16 A. Our guide to a GPU server in an office covers noise and heat in shared rooms.

Rack-mounted workstations

NVIDIA positions the 5500 for "rack-mounted workstation deployment", in which the GPU sits in a rack and the user works at a desk elsewhere through a remote display protocol. Until a datasheet describes the cooler, a rack chassis for it has to be one whose maker lists the RTX PRO 5500, or a rack-mountable tower. For a GPU host that many people share in continuous operation in a server room, the passive RTX PRO 6000 Blackwell Server Edition, which is on NVIDIA's vGPU list, remains the card we recommend; our comparison of the RTX PRO 6000 Server, Workstation and Max-Q editions sets out the differences.

When a card with published specifications fits better

• Your model fits in 72 GB per copy: the RTX PRO 5000 72 GB at 300 W has a published datasheet, vGPU on Red Hat KVM and, by bandwidth, a token-generation ceiling within about 4 per cent. See our page to buy an RTX PRO 5000 72 GB workstation.
• You need 96 GB, more bandwidth or three to four cards: the RTX PRO 6000 Max-Q, up to four per workstation according to NVIDIA, at 300 W each; see our page to buy a 2× RTX PRO 6000 Max-Q workstation.
• You want the fastest single card in a tower: the 600 W RTX PRO 6000 Workstation Edition, as our page to buy an RTX PRO 6000 Workstation Edition PC explains.
• You need 84 GB on one card and can wait for the datasheet: the RTX PRO 5500, quoted with the final specifications.

We put the 5500 in the same written quote as the RTX PRO 5000 72 GB and the RTX PRO 6000 on request, so the three can be compared for your workload; our guide to the best workstation GPU for AI in 2026 places all of them.

How ordering works with us in the EU

1. You describe the workload: models, users, data, the room and the timeline. Sizing is free of charge.
2. We propose the configuration on NVIDIA's figures at that date and check that cards, tower, power supply and cooling fit the room and the circuit.
3. You receive a written quote within one business day, with price, lead time and warranty terms per component, and the card's specifications as confirmed at that point. 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. The workstation is assembled and tested before delivery, with a power-on and basic load test.
6. We deliver anywhere in the EU, with commissioning on site on request; other destinations through our logistics partner.
7. Every component carries the manufacturer warranty, handled through us; DOA units are replaced, and the assembly is covered by us.

Our GPU server acceptance testing guide lists checks you can repeat on delivery.

What to send us for a quote

• The models, with precision and context length, or the scenes and applications for rendering and simulation
• How many people use the machine at peak, and whether through one service, MIG instances or virtual machines
• One or two cards, and whether a published-spec alternative should be quoted beside the 5500
• The tower or platform your company standardises on, and the operating system
• The room: circuit rating, cooling and whether the machine stands at a desk or in a rack
• Delivery site, invoicing entity and timeline

Write these points in the form below, and we reply with a configuration and a written quote within one business day.

FAQ