RTX PRO 6000 Blackwell: what 96 GB at $1–2.50 an hour buys you for LLMs and video

The RTX PRO 6000 Blackwell is showing up everywhere in GPU rental listings: 96 GB of memory, a 600 W board, and prices from about $1 an hour on marketplaces to $2.50 on managed clouds. It isn't a datacenter flagship, so the question is simple: per token and per video, how does it compare with the cards we've already measured?

We rented one on RunPod Secure Cloud ($2.49/h) and ran the same tests we use everywhere else: a health check, the same 8B LLM under the same latency target, and the same image-to-video model as our H200 video test.

The card itself

tokenwatt-check passed on both editions we could rent:

* 378 with PyTorch 2.8 (CUDA 12.8), 408 with PyTorch 2.13 (CUDA 13.0) on a second node.

For reference, our healthy H100 SXM measures about 680 TFLOPS and 3.0 TB/s, an A100 260 TFLOPS and 1.75 TB/s, an L40S 170 TFLOPS and 0.65 TB/s. So the RTX PRO 6000 has more than half an H100's compute and about half its memory bandwidth, with more memory than either.

A third edition, the Max-Q, runs the same chip at 300 W. It was out of stock when we tried, so we have no numbers for it. Check which edition a listing means before comparing prices.

LLM serving: close to an H100 NVL, half the cost of an A100

Same setup as our other Qwen3-8B measurements: BF16 weights, vLLM 0.31.0 with default settings, a chat-like workload, and a latency target of p99 time to first token ≤ 2 s and p99 time per output token ≤ 50 ms.

At $2.49 an hour it lands next to an H100 NVL per token and costs about half as much as an A100. At the roughly $1 an hour some marketplaces advertise for this card, the same throughput works out to about $0.23 per million tokens, below even the H100 SXM at its price. That's arithmetic on the measured throughput, not a measurement on a $1 node. Accuracy on our GSM8K check was 94%, in line with every other card.

It also served 2.7× as many requests as the A100 with less memory bandwidth (1.46 vs 1.75 TB/s). As on the A100, the limit is the slowest tokens during long prompts, and that depends on compute, where the Blackwell card has 1.5× the A100's.

Video: a 15-second 720p clip on one card

The 15-second 720p Wan2.2 clip from our video tests peaks at 68.5 GB of GPU memory with FP8 weights. It fits on one 96 GB RTX PRO 6000 with room to spare. (With the official BF16 weights the model is larger: in a separate test, one and four 80 GB H100s ran out of memory on the same clip and only eight worked.)

Same model (a Wan2.2 14B image-to-video FP8 fine-tune, 8 distilled steps), same ComfyUI version, input, prompts and seeds as on the H200:

The Blackwell card is 31% slower on the 15-second clip and 39% cheaper per clip. At ~$1 an hour, a clip would cost about $0.21, or $0.014 per second of video. For comparison, hosted APIs for Wan2.2 14B image-to-video list $0.036–0.08 per second of 720p video, mostly for clips of 5–8 seconds.

One difference: SageAttention 2 sped the 5-second clip up 1.3× here, against 1.6× on the H200. It built and ran on Blackwell, but appears less tuned for it.

Should you rent one?

• For video generation and models that need 80–96 GB on one card, yes. One RTX PRO 6000 replaces a multi-GPU setup and costs less per clip than an H200.
• For LLM serving at a latency target, it's competitive on price per token, not on speed. An H100 SXM still serves 2.8× as much per card and is cheaper per token at the prices we paid. The RTX PRO 6000 wins when it's rented near $1 an hour.
• Check the edition and the node. "RTX PRO 6000" covers 600 W and 300 W boards. nvidia-smi -q -d POWER or pip install tokenwatt-check shows which one you got.

Limits

One node per test on one platform (RunPod Secure Cloud, EUR-IS-2), quick-mode LLM measurement (2-minute windows, 2 repeats), one LLM, one video model and one reference image. Prices are RunPod's on the day. The $1/hour figures are arithmetic on the advertised price, not measurements. The Max-Q edition was not tested. Power is GPU board power.

Raw data: Qwen3-8B report · Method: how we measure · Prices across providers: RTX PRO 6000 price per hour · all GPU prices.