RTX PRO 6000 Blackwell for inference and rendering: a buyer’s review

In June 2026 NVIDIA’s own marketplace listed the RTX PRO 6000 Blackwell at $13,250, a figure Wccftech reported on 13 June against a launch price of about $8,000. By August the German integrator INGATE recorded a list price of $16,000 for the Workstation Edition, about 87% above launch with no change to the hardware. NVIDIA has not given a public reason; both reports point to the card’s 96 GB of GDDR7 during a memory shortage. INGATE notes that the card carries 32 GDDR7 modules of 3 GB each.

Three cards, one memory size

All three versions carry 96 GB of GDDR7 with ECC and four NVENC and four NVDEC engines. The Workstation Edition draws up to 600 W, uses a double flow-through cooler, offers four DisplayPort 2.1 outputs and lists 1,792 GB/s of memory bandwidth. The Max-Q Workstation Edition runs at 300 W, according to INGATE, which makes it the version for dense multi-GPU towers. The Server Edition has a configurable TDP of up to 600 W, comes as a dual-slot air-cooled card or a single-slot liquid-cooled card, lists 1,597 GB/s and supports up to four MIG instances. NVIDIA does not publish separate list prices for the Server or Max-Q editions.

For inference, the most detailed public test is Akamai’s, published on 30 October 2025, which ran a 49-billion-parameter Nemotron model through NVIDIA NIM. On Akamai’s own hardware at 100 concurrent requests, FP4 delivered 3,030 tokens per second against 2,256 at FP8. The headline claim of 1.63 times H100 throughput, however, compares the Blackwell card at FP4 in Akamai’s data centre with an H100 NVL at FP8 in a different environment. In the same environment, the RTX PRO 6000 at FP8 reached about 0.56 times the H100 NVL, and at FP4 it roughly matched the H100 at FP8. The card is competitive when the model can run at FP4; it is not a like-for-like replacement for an HBM-based GPU at higher precision.

For rendering and visualisation the case is simpler. The Server Edition has 188 fourth-generation RT Cores, and 96 GB of memory allows large scenes, simulations and video pipelines to stay on one card. NVIDIA positions it as the successor to the L40S but does not publish comparative figures.

What it means for buyers

Specify the edition, not just the name: Workstation, Max-Q and Server Edition differ in power, cooling and airflow, and a Server Edition card belongs only in a server the OEM has validated for it. Confirm whether your inference stack runs at FP4 before relying on vendor throughput figures, and ask for test results at the precision you will use. Quotes for this card have changed within weeks during 2026, so ask suppliers to state how long a price is valid and whether stock is allocated. Check whether vGPU or NVIDIA AI Enterprise licences are needed for your use and quote them separately. InTechnologies can supply both workstation and server editions on request.