NVIDIA RTX PRO 6000 Price in 2026: Hourly Rates, Media Workloads and H100 Comparison

Most of the AI GPU conversation is about HBM: the H100, H200 and B200 all win on memory bandwidth, because LLM token generation is limited by how fast weights stream out of memory. Generative media works differently. A diffusion model spends its time on dense math over a whole image or video latent, so compute, memory capacity and video encoding count for more than raw bandwidth. The RTX PRO 6000 Blackwell is NVIDIA's answer for that workload: a PCIe card with 96 GB of GDDR7 at a fraction of the cost of an HBM part.

RTX PRO 6000 specs next to the H100

The RTX PRO 6000 Blackwell ships in three versions. The Server Edition is a passively cooled card for rack servers, at up to 600 W, configurable.¹ The Workstation Edition is the 600 W desktop card, and a 300 W Max-Q version uses a blower cooler.²³ All three carry 96 GB of GDDR7 with ECC.

The Workstation Edition lists 1,792 GB/s of bandwidth.²⁴

The H100 keeps the edge in FP8 and FP16 Tensor throughput and in bandwidth. The RTX PRO 6000 answers with FP4, 16 GB more memory and hardware encoders.

Why memory bandwidth matters less for image and video

When an LLM generates text one token at a time, every token reads every weight. At batch size 1, a model with 30 GB of weights can produce at most about 53 tokens per second on 1.6 TB/s, and about 112 on 3.35 TB/s. Here, bandwidth sets the speed limit, and HBM earns its price.

A diffusion model does the opposite. Each denoising step runs the full network over thousands of latent tokens for an image, and far more for a video. Each weight is reused across all of them, so the GPU waits on math, not memory. What matters instead:

• Does the model fit? 96 GB holds a large video model plus its text encoder and VAE on one card, where an 80 GB card may need offloading.
• Can it run at low precision? Blackwell's FP4 Tensor Cores give quantized diffusion models a path to more throughput. Hopper has no FP4.
• Where does encoding run? A GPU without NVENC hands finished frames to the CPU for software encoding. The RTX PRO 6000's four NVENC engines encode H.264, HEVC and AV1 in hardware.¹⁵

What an RTX PRO 6000 costs to buy

The price has nearly doubled since launch.³⁶ Server Edition cards are sold through system builders and resellers, who price them per order, and each one needs a PCIe Gen 5 server with airflow for 600 W, plus power and operations.

What an RTX PRO 6000 costs to rent

On fal, from fal.ai/pricing:⁷

At list price an RTX PRO 6000 costs about $0.00111 per second. Billing runs per second from model setup through idle time until the runner shuts down; image pulls and queueing are free.⁸

Worked costs for image and video generation

The rate per hour matters less than the cost per output. Measure GPU-seconds per image or per minute of video on your own model, then read across.

Images. Cost per 1,000 images at different GPU times per image:

Video. Cost per minute of finished video, by GPU minutes spent per output minute:

Replace the rows with your measured throughput. Idle time between bursts is billed, so low utilization raises the real cost per output; the cold-start guide covers the keep-alive trade-off.

RTX PRO 6000 vs H100 vs L40 for media work

Encoder counts and bandwidth are from fal's machine-type table.⁵

RTX PRO 6000 vs H100. The H100 lists at $4.50/h, 12.5% above the RTX PRO 6000. At "as low as" rates the gap is $2.49 against $1.99, or 25%. So an H100 is cheaper per image only if it finishes the same job at least 1.13 times faster at list prices, or 1.25 times faster at "as low as" rates. For a compute-bound diffusion model at FP8 or BF16, the H100's higher Tensor throughput often gets it there. For an FP4-quantized model, a model between 80 and 96 GB, or any pipeline that encodes video, the RTX PRO 6000 usually wins. Benchmark both on your model before choosing. The H100 price guide covers that GPU in depth.

RTX PRO 6000 vs L40. The Ada-generation L40 has half the memory and bandwidth, one fewer encoder and no FP4. It still suits image models that fit in 48 GB.

When not to use an RTX PRO 6000

• LLM serving at scale. Token generation is bandwidth-bound, and an H200 or B200 has three to five times the bandwidth of the Server Edition. The B200 price guide covers the high end.
• Multi-GPU training or tensor parallelism. The card links to its neighbours over PCIe Gen 5 rather than NVLink. An H100 SXM moves 900 GB/s over NVLink against 128 GB/s over PCIe Gen 5.⁴ Models split across many GPUs spend that difference waiting.

Buy or rent?

At $15,000 to $16,000 per card, the purchase price equals:

• At $4.00/h list: 3,750 to 4,000 rented hours, or about 156 to 167 days of continuous use.
• At $1.99/h (as low as): about 7,540 to 8,040 hours, or roughly 314 to 335 days.

That compares the card alone with a fully loaded rental, so the real break-even comes later. Media traffic is also bursty, peaking around launches and campaigns, and owned cards sit idle between peaks. Buying makes sense when a fleet stays busy around the clock for more than a year and you already run servers. The dedicated vs serverless break-even guide works through the utilization line.

Running on RTX PRO 6000s with fal

The machine type is GPU-RTXPRO6000: 96 GB of VRAM, 100 GB of system RAM, 6 CPU cores, and 4 NVENC plus 4 NVDEC engines with AV1.⁵ Set machine_type = "GPU-RTXPRO6000" in your app class. Serverless access is approved per account; request it on the Serverless get-started page, then follow the Quick Start.⁹

FAQ

How much does an NVIDIA RTX PRO 6000 cost?

About $15,000 to $16,000 to buy in the US in mid-2026: $16,000 for the Workstation Edition on NVIDIA's marketplace, and $14,999 for the Server Edition at Newegg.³⁶ To rent, fal charges $4.00 per GPU-hour at list price, or as low as $1.99.

What is the RTX PRO 6000 price per hour?

$4.00 per GPU-hour at list price on fal, billed per second, which is about $0.00111 a second. It goes as low as $1.99 per GPU-hour.

RTX PRO 6000 vs H100: which is better?

For LLM serving and multi-GPU training, the H100, with twice the bandwidth and NVLink. For image and video generation, the RTX PRO 6000 often costs less per output: 96 GB, FP4 and hardware encoders at $4.00/h against $4.50/h list.

References