The NVIDIA RTX PRO 6000 Blackwell is NVIDIA's 96GB GDDR7 professional GPU, launched in 2025 in three editions: Workstation, Max-Q Workstation and Server. Newegg listed the PNY Workstation Edition card at a retail asking price of $19,999.99 on 28 September 2026, about 2.3 times the $8,565 single-unit launch price in April 2025 reported by Tom's Hardware on 12 August 2026, a derived comparison. The live tables below show tracked rental listings, and the rule is simple: buy one only if you will keep it busy long enough for ownership to beat those rental prices; otherwise rent.
The three editions, and what each one costs
NVIDIA's family comparison page (retrieved 28 September 2026) describes Workstation, Max-Q Workstation, and Server editions with the same 96GB GDDR7 memory capacity. NVIDIA's product pages distinguish their power limits, cooling, and intended use. The prices are retailer asking prices for the card or card kit alone, on the dates shown.
| Edition | Power | Cooling | Form factor | Intended use | Purchase price |
|---|---|---|---|---|---|
| Workstation Edition | 600W | Double flow-through active cooling, air pushed across the heatsink and vented into the chassis | PCIe card | Single-GPU workstations | Newegg card listing: $19,999.99, 28 September 2026 |
| Max-Q Workstation Edition | Lower power limit | Self-contained blower that exhausts out the chassis rear | PCIe card | Dense multi-GPU workstations | CDW card listing: $16,233.99, 21 September 2026 |
| Server Edition | Configurable | Passive air cooling or partner liquid cooling | PCIe card | Rack servers | CDW card-kit listing: $13,824.99, 21 September 2026 |
That Workstation Edition price is above CDW's $16,268.99 retail card listing for the same edition on 21 September 2026. These are different retailers and dates, so treat the figures as snapshots, not a measured week-on-week market change. NVIDIA's Server Edition page describes passive cooling, with the air-cooled card relying on chassis airflow. Equal memory does not mean equal speed: NVIDIA's product pages report different clocks by edition, so check the Server Edition's figures for the clock mode you will run.
Reading the RTX PRO 6000 specs
The site's spec table below lists the RTX PRO 6000 Blackwell alongside the GeForce RTX 5090, RTX 6000 Ada Generation, and H100. Lenovo's RTX 6000 Ada product guide identifies Blackwell as its successor. These are the family-level numbers behind the site's full GPU spec reference, not a separate specification for every edition or rental configuration.
| Spec | RTX PRO 6000 | RTX 5090 | RTX 6000 Ada | H100 |
|---|---|---|---|---|
| VRAM | 96 GB | 32 GB | 48 GB | 80 to 94 GB |
| Memory type | GDDR7 | GDDR7 | GDDR6 | HBM3 |
| Memory bandwidth | 1,792 GB/s | 1,792 GB/s | 960 GB/s | 3,350 GB/s |
| FP16 (dense) | 503.8 TFLOPS | 209.5 TFLOPS | 364 TFLOPS | 989 TFLOPS |
| FP16 (with sparsity) | 1,007.6 TFLOPS | 419 TFLOPS | 728 TFLOPS | 1,979 TFLOPS |
| FP8 (dense) | 1,007.6 TFLOPS | 419 TFLOPS | 728.5 TFLOPS | 1,979 TFLOPS |
| FP8 (with sparsity) | 2,015.2 TFLOPS | 838 TFLOPS | 1,457 TFLOPS | 3,958 TFLOPS |
| FP4 (dense) | 2,015.2 TFLOPS | 1,676 TFLOPS | Not published | Not published |
| FP4 (with sparsity) | 4,030.4 TFLOPS | 3,352 TFLOPS | Not published | Not published |
| TDP | 600 W | 575 W | 300 W | 700 W |
| Interconnect | PCIe 5.0 | PCIe 5.0 | PCIe 4.0 | NVLink, PCIe 5.0, InfiniBand |
| Launch year | 2025 | 2025 | 2022 | 2022 |
Two things stand out. First, memory type. The family table lists GDDR7 at 1,792 GB/s for the RTX PRO 6000 Blackwell and RTX 5090. The RTX 6000 Ada uses older GDDR6 at 960 GB/s. The H100 uses HBM3 at 3,350 GB/s, about 1.9 times the RTX PRO 6000's bandwidth, a figure derived from those two table entries. Server Edition figures can differ from the family figure by clock mode.
Second, the FP16, FP8 and FP4 rows each split into dense and sparse figures. NVIDIA's sparsity guide of 20 July 2021 describes 2:4 structured sparsity: zeroing half the weights in a fixed pattern, then retraining to recover accuracy. Treat the dense row as the relevant theoretical figure for an ordinary, unpruned model, not a measured application result; sparse peak figures are not guaranteed workload speedups. The table also shows why the H100 and RTX 6000 Ada carry no FP4 figures at all: NVIDIA's TensorRT support matrix, updated 8 September 2026, marks FP4 on the Hopper and Ada architectures as hardware emulation only, not accelerated, while NVIDIA's Blackwell product material describes native FP4 support for the RTX PRO 6000 and RTX 5090.
RTX PRO 6000 vs RTX 5090: what your money buys
Both the RTX PRO 6000 Blackwell and GeForce RTX 5090 have a 2025 launch year in the site's spec table. WCCFTech reports that they share the Blackwell GB202 die (retrieved 28 September 2026). The site's datasheet-verified figures give 21,760 CUDA cores on the RTX 5090 against 24,064 on the RTX PRO 6000, about 10.6 percent more, a figure derived from those two counts.
Memory capacity is the bigger proportional gap: 96GB against 32GB, a derived 3.0 times as much from the table above. AI Infrastructure's technical comparison, retrieved 28 September 2026, reports ECC memory on the RTX PRO 6000 and no ECC on the RTX 5090. It states that a single-bit error in a model's weights goes undetected and uncorrected on the RTX 5090 but is corrected automatically on the RTX PRO 6000. The same comparison notes that the RTX PRO 6000 supports MIG partitioning and NVIDIA vGPU software, while the RTX 5090, as a consumer GeForce card, supports neither. The site's family table lists both cards at 1,792 GB/s of memory bandwidth; the edition caveat above still applies.
The dated price gap is real money. Newegg's $19,999.99 Workstation Edition retail card listing from 28 September 2026 is roughly 3.1 times Tom's Hardware's best tracked US RTX 5090 card price of $6,389 on 14 September 2026, another derived ratio. Both are dated asking prices, not quotes, and the RTX 5090 has sold for less at other times. See the two cards compared in full at RTX 5090 vs RTX PRO 6000 Blackwell.
Renting: variants, MIG slices, and providers
Renting avoids the upfront card purchase. The table below shows the cheapest live price for each RTX PRO 6000 Blackwell variant the site tracks.
| Variant | VRAM | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|---|
| RTX PRO 6000 Blackwell MIG 24GB | 24 GB | $0.59 | RunPod | 1 |
| RTX PRO 6000 Blackwell MIG 48GB | 48 GB | $1.09 | RunPod | 1 |
| RTX PRO 6000 Blackwell | 96 GB | $1.99 | Vast.ai | 5 |
NVIDIA's MIG user guide, retrieved 28 September 2026, lists partitioning profiles for the RTX PRO 6000 Blackwell. Check the live variant table to distinguish a full card from a slice. NVIDIA's MIG introduction describes hardware isolation and separate paths through the memory system, rather than software time-sharing; see what MIG fractional GPUs are for how that isolation works and when a slice is enough for your workload.
| Provider | RTX PRO 6000 Blackwell $/GPU-hr | RTX 5090 $/GPU-hr | RTX 6000 Ada Generation $/GPU-hr |
|---|---|---|---|
| RunPod | $0.59 | none in stock | none in stock |
| Vast.ai | $1.99 | $0.53 | none in stock |
| LeaderGPU | $2.04 | $2.29 | $1.20 |
| Massed Compute | $2.19 | none in stock | $0.79 |
| QuantaCloud | $2.39 | none in stock | $0.79 |
The provider table above compares tracked RTX PRO 6000 Blackwell, RTX 5090 and RTX 6000 Ada listings, provider by provider.
Buy or rent: run the numbers
Newegg listed the PNY Workstation Edition card at a retail asking price of $19,999.99 on 28 September 2026. The table below turns that figure into rented GPU-hours at today's cheapest live prices for the RTX PRO 6000 Blackwell, the RTX 5090, and the H100.
| GPU | $/GPU-hr | Provider | GPU-hours for $20,000 | Days at 24 h a day |
|---|---|---|---|---|
| RTX PRO 6000 Blackwell | $0.59 | RunPod | 33,898 | 1,412.4 |
| RTX 5090 | $0.53 | Vast.ai | 37,735 | 1,572.3 |
| H100 | $2.50 | Hyperstack | 7,999 | 333.3 |
Use four factors in your buy-or-rent calculation. Utilisation: enter the hours you expect to use, and check whether a rental contract bills running time or a reserved duration. Google's Dynamic Workload Scheduler documentation distinguishes usage billing from payment for an entire reservation (accessed 21 September 2026). Data residency: check your own contract's location requirements before comparing costs. Power and cooling: a Workstation Edition card assumed to run continuously at its full 600W for a non-leap year draws 5,256 kWh, derived from 600W times 8,760 hours, divided by 1,000. At the US Energy Information Administration's 2025 industrial average of 8.62 cents per kWh, retrieved 21 September 2026, that GPU-only scenario is about $453 a year in electricity, a derived cost before host-system power, cooling, or colocation. Use your own electricity tariff and expected draw, not this maximum-power scenario, for a purchase decision. Resale: GPUDojo recorded used-card asking prices from $4,400 for the RTX A6000 and $750 for the Quadro RTX 6000 on 27 September 2026. Their launch years in the site's spec table are 2020 and 2018. These are listing snapshots, not evidence of completed sales or a guaranteed resale value for the RTX PRO 6000 Blackwell, whose table launch year is 2025.
Run your own scenario, your own hours and your own power rate, with the rent versus buy GPU calculator before committing to either path.
Where it fits for AI work
Treat the RTX PRO 6000 Blackwell's 96GB as a memory budget, not a guarantee that an inference or fine-tuning job fits. Runpod's sizing guidance, accessed 21 September 2026, suggests roughly 2GB per billion parameters at 16-bit precision. Derived weight-only estimates for a 70 billion parameter model are 140GB at 16-bit, more than one card holds, and 35GB at 4-bit. NVIDIA's inference guide describes additional KV-cache memory; its NVFP4 guide also accounts for quantization scales. Those extras mean the 35GB estimate is not the complete runtime footprint. Use the LLM VRAM calculator for your own model size, and the LLM GPU requirements reference for which precision and GPU pairing is usual at each parameter count.
The site's interconnect entry lists PCIe 5.0 and no NVLink for this family. CoreWeave's RTX PRO 6000 Server Edition documentation states plainly that it lacks NVLink for memory pooling across GPUs, and Nebius describes the interface as standard PCIe, not NVLink, both retrieved 28 September 2026. There is no NVLink pooling path here; do not budget multiple cards as one automatically combined memory pool.
Bandwidth is the other limit. The family table's RTX PRO 6000 GDDR7 figure is 1,792 GB/s against the H100's HBM3 at 3,350 GB/s, about 1.9 times as much bandwidth on the H100, the same derived ratio as above, subject to the edition caveat. NVIDIA's own engineering blog describes token-by-token decoding as memory-bound work, where transfer speed from memory dominates latency more than raw compute speed, accessed 21 September 2026. That makes bandwidth relevant to token latency; this ratio alone does not predict application speed. Compare the two cards directly at H100 vs RTX PRO 6000 Blackwell, and see how the RTX PRO 6000 stacks up against NVIDIA's Ada-generation L40S at L40S vs RTX PRO 6000 Blackwell.
If you meant an older "6000" card
The "6000" naming sequence starts with NVIDIA's Quadro 6000 announcement in 2010 and continues through Quadro M6000, Quadro P6000, Quadro RTX 6000, RTX A6000, and RTX 6000 Ada Generation, before the Blackwell-generation RTX PRO 6000 covered on this page. The prefix and suffix matter; the early cards were not called RTX 6000. If a used listing, an old build guide, or a coworker's recommendation sent you looking for a plain "RTX 6000" or "RTX A6000", see RTX 6000 vs RTX A6000 vs RTX 6000 Ada for what actually separates them.
Whichever edition or variant you need, check current listings before you buy or rent. Buy only if the chosen configuration meets your workload and location requirements and its total ownership cost beats renting over your expected use; otherwise rent.
Sources
- Lenovo: RTX 6000 Ada product guide, retrieved 28 September 2026
- CG Channel: NVIDIA unveils the Quadro M6000, March 2015
- NVIDIA: Quadro P6000 datasheet, retrieved 28 September 2026 via search
- NVIDIA: Quadro 6000 announcement, 27 July 2010
- NVIDIA MIG User Guide: introduction, updated 11 September 2026
- NVIDIA: introducing NVFP4, 24 June 2025
- NVIDIA GeForce RTX 5090 product page, accessed 21 September 2026
- Google Cloud: Dynamic Workload Scheduler billing, accessed 21 September 2026
- NVIDIA RTX PRO 6000 family comparison page, retrieved 28 September 2026
- NVIDIA RTX PRO 6000 Blackwell Workstation Edition product page, retrieved 28 September 2026
- NVIDIA RTX PRO 6000 Blackwell Server Edition product page, retrieved 28 September 2026
- Newegg: PNY RTX PRO 6000 Blackwell Workstation Edition listing, retrieved 28 September 2026
- CDW: PNY RTX PRO 6000 Blackwell Workstation Edition listing, retrieved 21 September 2026
- CDW: PNY RTX PRO 6000 Blackwell Max-Q Workstation Edition listing, retrieved 21 September 2026
- CDW: PNY RTX PRO 6000 Blackwell Server Edition listing, retrieved 21 September 2026
- Tom's Hardware: NVIDIA doubles RTX PRO 6000 Blackwell's MSRP to $16,000, 12 August 2026
- Tom's Hardware: lowest GPU prices tracker, 14 September 2026
- WCCFTech: RTX PRO 6000 beats GeForce RTX 5090, retrieved 28 September 2026
- AI Infrastructure: RTX workstation GPU comparison, retrieved 28 September 2026
- NVIDIA MIG User Guide: supported MIG profiles, retrieved 28 September 2026
- CoreWeave docs: RTX PRO 6000 Blackwell Server Edition instance, retrieved 28 September 2026
- Nebius: RTX PRO 6000 Blackwell Server Edition compute page, retrieved 28 September 2026
- NVIDIA developer blog: accelerating inference with sparsity, 20 July 2021
- NVIDIA TensorRT support matrix, updated 8 September 2026
- Runpod docs: choosing a pod, accessed 21 September 2026
- NVIDIA developer blog: mastering LLM inference optimization, 17 November 2023
- GPUDojo: RTX A6000 price tracker, as of 27 September 2026
- GPUDojo: Quadro RTX 6000 price tracker, as of 27 September 2026
- US Energy Information Administration: Electric Power Monthly, Table 5.3, retrieved 21 September 2026