NVIDIA RTX PRO 6000 cloud prices

The cheapest RTX PRO 6000 hour on 16 October 2026 was an interruptible Vast.ai bid at $0.33 on a single 300 W Max-Q card, which can be taken back whenever a higher bid arrives and carries its own disk charge 15. The cheapest non-interruptible marketplace listing was a Max-Q on Vast.ai at $0.95, with the cheapest full-power Workstation card at $1.00. packet.ai sells a shared card at $0.66 and HyperAI lists $0.80 without stating the machine 21 20. The cheapest dedicated card with a full machine on a provider's own page was Beam at $1.09, followed by GPUBank at $1.15 with traffic and IPv4 included, or $0.72 effective after the bonus on a $1,000 top-up 18 1. Which is cheapest for you depends on whether you accept interruption, a shared card or a 300 W edition, and whether you prepay.

A dedicated RTX PRO 6000 costs from about $0.95 to $2.50 per hour on demand at marketplaces and specialised clouds, and $3.36 to $5.50 per GPU at the hyperscalers. GetDeploying's crawl of 15 October 2026 put the median on-demand price at $2.21 across 165 configurations, with spot at $1.35 and reserved at $1.90 30. GPUBank's $1.15 all-in hour is about half that median 1. Spot capacity runs from Vast.ai's $0.33 bids to about $2.56 for AWS spot, and reservations from aggregator listings near $0.48 to $2.42 for Azure on three years. Always check what the hour includes: a GPU-only marketplace rate and an all-in machine price can differ by more than $0.30 an hour once disk and traffic are added, and a per-GPU figure taken from an 8-GPU node is only available if you rent all eight.

RTX PRO 6000 spot prices ran from $0.33 to about $2.56 per GPU-hour on 16 October 2026, depending on the provider and the unit rented. Vast.ai's interruptible minimum bids started at $0.33 on a Max-Q card, and an outbid instance is paused 15. GPUBank's interruptible tier costs $0.69 per GPU with five minutes of notice through the metadata endpoint, a webhook and the console, and a stop rather than a deletion 1. Nebius moved to dynamic spot pricing on 8 October with a floor of $0.79 that can change with capacity 22. Verda's spot instances cost $1.15, CoreWeave's 8-GPU Server Edition spot nodes work out to $1.195 or $1.386 per GPU, Google's g4-standard-48 about $1.71 and AWS's g7e.2xlarge about $2.56 25 27 34 32. Spot suits batch work such as video generation or checkpointed fine-tuning, not a production API.

Most clouds rent the Server Edition, while marketplaces also carry the Workstation and Max-Q cards. AWS describes its G7e instances as using the RTX PRO 6000 Blackwell Server Edition, and CoreWeave, Verda, Hyperstack and Lium name the same edition on their pages 31 27 26 23 28. The Server Edition is passively cooled, configurable from 400 to 600 W and rated at 1,597 GB/s, against 1,792 GB/s for the two workstation cards 5 6. Vast.ai lists all three editions separately, and on 16 October its offer search held 24 Server, 32 Workstation and 25 Max-Q listings 15. GPUBank lists 1.8 TB/s, which matches the workstation datasheets, but does not name the edition 1. If your job depends on sustained power or bandwidth, ask before reserving.

The Max-Q is a little slower on compute-bound work and about the same on memory-bound work. It has the same 96 GB and 1,792 GB/s as the 600 W Workstation Edition but half the power 4 3, so single-user token generation, which is limited by bandwidth, runs at much the same speed. Puget Systems measured the Max-Q 5 to 14 percent slower than the 600 W card in most content-creation tests, with the largest gaps of 13 to 14 percent on Topaz Video AI, Unreal Engine and DaVinci Resolve GPU effects 8. Long diffusion jobs, rendering and batched prefill sit at the compute-bound end. A Max-Q listing at $0.95 against a Workstation card at $1.00 is therefore a fair trade for LLM serving and a slightly worse one for video, and the host's power limit can matter as much as the edition 9.

One RTX PRO 6000 runs a 70B model comfortably in 4-bit and tightly in 8-bit. In 4-bit Llama 3.3 70B's weights take about 41 GB, leaving around 50 GB for KV cache and long contexts; in 8-bit they take about 75 GB, which leaves only a modest cache. CloudRift measured 1,031 output tokens per second in aggregate on Llama 3.3 70B AWQ-INT4 with vLLM at a concurrency of 400 10, about $0.31 per million tokens at GPUBank's $1.15 hour. For a single user, Hardware Corner measured 32.0 tokens per second with llama.cpp at 4K of context, falling to 16.6 at 128K 11, which costs about $11 per million tokens. The card is therefore a good 70B server for many concurrent users and a modest one for a single chat session, where an H100 SXM's higher bandwidth would answer faster.

Yes, gpt-oss-120b runs on one RTX PRO 6000 in its native format. OpenAI released the model with 117 billion parameters, 5.1 billion of them active per token, and MoE weights post-trained in MXFP4, and says it runs on a single 80 GB GPU 14. That leaves the RTX PRO 6000's 96 GB with roughly 30 GB beyond the weights for cache and batching. Because only 5.1 billion parameters are active, generation is much faster than on a dense 70B model: Hardware Corner measured 221.3 tokens per second with llama.cpp on a Q8_0 build at 4K of context and 123.3 at 128K 11. Running the model in 16-bit or 8-bit would need more than one card. For a team that wants a capable open model on one rented GPU, this is one of the strongest reasons to choose the card over two smaller ones.

Wan 2.2 takes from under three minutes to about 40 minutes per clip on an RTX PRO 6000, depending on the model, the resolution and the power limit. HOSTKEY measured the TI2V-5B model at 160 seconds for a five-second clip at 720p on a Server Edition, about $0.05 per clip at GPUBank's rate 9. The 14B text-to-video model at 720p and 24 fps took about 40 minutes per clip in the card's default mode and about 30 minutes after switching to 600 W, or $0.58 to $0.77 per clip. HOSTKEY did not publish the step count or precision, so your times will move with the workflow, the sampler and speed-up LoRAs. The 14B model's two experts total 27 billion parameters, and Wan's own single-GPU command expects a card of at least 80 GB, which the RTX PRO 6000 satisfies 13.

The RTX PRO 6000 is better only when your model needs more than 32 GB, or more than 48 GB in the L40S's case. On CloudRift's Qwen3-Coder-30B AWQ test it produced 8,425 tokens per second against 4,570 on an RTX 5090 10; priced at GPUBank's $1.15 and $0.48, that is $0.038 per million tokens against $0.029, so the cheaper card wins per job 1. On Llama 3.3 70B in 4-bit, which does not fit on one 5090, one RTX PRO 6000 gave 1,031 tokens per second against 1,230 for two 5090s, close to a tie at $1.15 against $0.96 an hour, with the single card keeping more room for context. The L40S at $0.55 has 48 GB and 864 GB/s, enough for 30B models and image work but too tight for a 70B model with any useful context.

The RTX PRO 6000 has more memory and costs less per hour than an H100, but the H100 is faster per user on large models and better for training. It has 96 GB against 80 GB, which lets a 4-bit or tight 8-bit 70B model fit on one card with cache room. The H100 SXM's 3.35 TB/s of HBM3 is about 1.9 times the bandwidth, and it has NVLink for multi-GPU training, which the RTX PRO 6000 lacks. On GPUBank the RTX PRO 6000 costs $1.15 against $1.35 for the H100 PCIe and $1.85 for the H100 SXM 1. The measurements split by load: CloudRift's batched GLM-4.5-Air test gave 3,140 tokens a second on one RTX PRO 6000 against 2,987 on one H100, while Runpod's single-user vLLM test gave 89 against 109 on the H100 PCIe and 165 on the H100 SXM 46 45. Choose the RTX PRO 6000 for single-card inference, video generation and graphics, and the H100 SXM for training across GPUs and latency-sensitive serving.

Renting is cheaper than buying for most users at 2026 card prices. NVIDIA's US marketplace raised the Workstation Edition to $16,000 in August 2026, nearly double its $8,565 launch listing, and GPU Poet's October average of the cheapest eBay listings was $16,963 36 37 16. With a $4,000 host, resale after two years of $7,400 and power at $0.1819 per kWh, owning costs about $525 a month plus $0.146 per loaded hour. That breaks even with GPUBank's $1.15 on-demand price at about 523 hours a month. Against the $0.72 effective hour from $1,000 top-ups there is no break-even within a 730-hour month. Buying makes sense if you will keep the card busy most of the day for two years, need it on your own premises, or can find one near the launch price.

GPUBank charges only for an optional persistent disk on top of the GPU price. The $1.15 hourly price includes 16 vCPU, 188 GB of RAM, the 2 TB local NVMe scratch disk, one public IPv4 address and all inbound and outbound traffic 1. The persistent disk costs $0.03 per GB-month, prorated over a 730-hour month, so 200 GB costs $6.00 for a full month and 500 GB costs $15.00. A stopped machine costs only that disk, a launch or health check that fails is not billed, and there is no minimum billing period. By comparison, Runpod bills $0.20 per GB-month for an idle pod's volume disk, Hyperstack about $0.07 per GB-month for volumes, and Vast.ai hosts between $0.13 and $0.87 per GB-month plus a traffic charge on the RTX PRO 6000 listings we read 24 23 15.

On GPUBank a host failure during a job restarts the machine on a healthy host with the same disk and credits the interrupted session in full, and a failed launch is never billed because the meter starts only after the health check 1. Interruptible machines get five minutes of notice through the metadata endpoint, a webhook and the console, run your checkpoint hook and are stopped rather than deleted, and if the notice is missed the last 30 minutes are credited. Elsewhere the rules differ: an outbid Vast.ai instance is paused while its storage keeps billing 17, Nebius's spot price can move every 15 minutes up to one cent below its on-demand rate 22, and on a marketplace the hours lost to a host that drops offline are usually the renter's. For video batches, where each clip stands alone, a reclaim costs one clip; for a long fine-tune, checkpoint every few minutes whatever the provider.

Yes, the top-up bonus lowers the effective hourly price, but only on money you will spend. GPUBank takes payment in crypto from $25 into a prepaid US dollar balance that never expires, and cards are not accepted for now 1. The bonus is 20 percent from $25, 30 percent from $50, 35 percent from $100, 40 percent from $250, 50 percent from $500 and 60 percent from $1,000, which turns the $1.15 on-demand hour into about $0.96, $0.85 and $0.72 at the $25, $100 and $1,000 levels. At $0.72 the RTX PRO 6000 is cheaper than any dedicated all-in hour we could read on a provider's own page, though still above Vast.ai's bids and packet.ai's shared card. Monthly spend caps can be set per account and per team member, which keeps a large prepaid balance under control.

Aggregator prices disagree with provider pages because they crawl at different times, pick different configurations and sometimes mix fractional or 8-GPU listings with single cards. On 16 October 2026 Vast.ai's own pricing pages for all three editions read "No current offers" while its offer search returned 81 rentable listings from $0.95 16 15. Runpod's page showed $2.49 for a card that a survey two weeks earlier had at $2.09 24, and Clore.ai's latest own figure is a snapshot from 24 August 29. Some rows, such as Hostinger at $0.65 or HyperAI's $0.48 reservation, exist only on aggregators 30. Treat every aggregator figure as a lead: read the provider's page or console on the day, and check whether the price is for one GPU, a MIG slice, a shared card or one-eighth of a node.