Rent NVIDIA RTX 6000 Ada GPUs by the hour

The RTX 6000 Ada is a 48 GB NVIDIA workstation GPU built on the Ada Lovelace architecture, and DeployGPU rents it as a VM by the hour. NVIDIA's datasheet lists 960 GB/s of memory bandwidth, the highest of the four 48 GB cards DeployGPU offers, and 300 W of board power. Every offer on this page is a VM you can launch now with SSH access, billed by the hour and settled to the second when you stop.

Offers now

5

From

$0.78/GPU-hr

GPUs per VM

1x, 2x

Providers

3

Massed Compute, QuantaCloud, Verda · checked 2026-11-13 06:40 UTC

RTX 6000 Ada offers you can launch now

The table lists every RTX 6000 Ada VM in stock, cheapest first, and refreshes every 15 minutes. The Launch link opens that exact offer on the deploy page, where the full hourly breakdown is shown before anything is charged.

DeployGPU only lists providers it can launch machines on. For RTX 6000 Ada prices across the whole market, compare all providers on GpuPerHour.

RTX 6000 Ada specs next to the RTX A6000, L40S and RTX PRO 6000

The RTX PRO 6000 vs RTX 6000 Ada confusion starts with the names. The RTX 6000 Ada is an Ada Lovelace card with 48 GB, and the RTX PRO 6000 Blackwell is a different card with 96 GB that NVIDIA sells as a Server Edition and a Workstation Edition. Check the generation and the memory in any listing before you compare prices.

Specifications from NVIDIA's RTX 6000 Ada datasheet, RTX A6000 datasheet, L40S page and RTX PRO 6000 Blackwell Server Edition page, with compute capability from NVIDIA's CUDA GPU list, checked 12 November 2026. The RTX PRO 6000 column uses the Server Edition's 1,597 GB/s, as the RTX PRO 6000 page does, and the Workstation Edition datasheet lists 1,792 GB/s.

In an RTX 6000 Ada vs L40S comparison, both list the same 18,176 CUDA cores and 568 Tensor Cores, and they differ in memory bandwidth, 960 GB/s against 864 GB/s, and in board power, 300 W against 350 W. In an RTX 6000 Ada vs RTX A6000 comparison, the newer card adds FP8 Tensor Cores and 960 GB/s of bandwidth against 768 GB/s, and it gives up the optional NVLink bridge.

What fits on a 48 GB GPU

These are weights-only sizes of the published checkpoints, the summed safetensors files on each model's Hugging Face card, checked 12 November 2026, and the fit is calculated rather than measured. One 48 GB card gives 44.16 GB at the site's 92% rule, and what is left after the weights is what vLLM has for the KV cache.

MXFP4, the format gpt-oss ships in, runs natively only on Blackwell, whose Tensor Cores NVIDIA's Blackwell architecture page says add community-defined microscaling formats. On an RTX 6000 Ada the serving engine handles it in software. Image work fits too: Hugging Face's diffusers quantization post measures the full FLUX.1-dev pipeline in BF16 at about 31.4 GB after loading and 36.2 GB at peak on an H100, under the 44.16 GB a 48 GB card gives. The ComfyUI template launches on it from the dashboard.

More than one RTX 6000 Ada in a VM

NVIDIA's datasheet lists NVLink as No for the RTX 6000 Ada, so a VM with 2, 4 or 8 of them connects its cards over PCIe 4.0 x16. Data parallelism tolerates that link, and tensor parallelism depends on a fast one. vLLM's parallelism guide says that on GPUs without NVLink, such as the L40S, pipeline parallelism gives higher throughput and lower communication overhead than tensor parallelism. The multi-GPU page explains the Interconnect column. One VM is the largest unit DeployGPU launches, and for more than 8 GPUs you can request a cluster quote.

RTX 6000 Ada, RTX A6000, L40S or a 96 GB card

Pick the RTX 6000 Ada when the job leans on memory bandwidth, as token generation does, or needs FP8, and let the live table settle its price against the L40S. Pick the RTX A6000 when BF16 or FP16 is enough or a 2-GPU job wants the NVLink bridge. Move to 96 GB when the weights plus cache pass about 44 GB on one card. Qwen3.8-27B in BF16 is 55.6 GB, which needs 2 RTX 6000 Ada cards but fits one RTX PRO 6000 with about 33 GB left (calculated). The 48 GB tier page lines up all four 48 GB cards and the 96 GB and larger tier covers what sits above them.

How long until it is ready

A VM usually reaches Running in 1 to 3 minutes. App templates take about 6 to 8 minutes, because the platform waits until the app itself answers before it marks the deployment ready. On some providers SSH accepts connections a little after the machine reports Running, so a Connection refused in the first minute usually clears on its own, and the Connection refused fix covers what to check if it does not. In the dashboard, every RTX 6000 Ada VM boots an Ubuntu image with NVIDIA drivers and Docker installed by the provider, and three app templates set up a ready app instead: JupyterLab with PyTorch, ComfyUI and Open WebUI with Ollama. For an OpenAI-compatible server on your own card, follow the vLLM guide.

What you pay

The price on each row is the hourly rate you are billed, including any platform fee. Launching reserves the first hour from your credit balance, and when you stop, the unused seconds of that hour go back to the balance. Credits are bought by card with a 5% + $0.30 processing fee on each top-up. The pricing page has the full breakdown. Paid time starts when provisioning starts, but only a launch that reaches Running is charged, and a failed launch refunds the reserved hour automatically.

Stopping an RTX 6000 Ada VM deletes its disk, so copy results off first. The guide to stopping a GPU without losing work shows how.