Rent RTX PRO 6000 Blackwell GPUs and 8-GPU servers

To rent RTX PRO 6000 Blackwell GPUs, you can take one GPU in a dedicated virtual machine for $1,390 per month, or a full RTX PRO 6000 server with eight GPUs for $10,490 per month. Each card is NVIDIA's Blackwell Server Edition with 96 GB of GDDR7 ECC memory, assigned only to you, and both configurations are in stock.

What the RTX PRO 6000 Blackwell is

The RTX PRO 6000 Blackwell Server Edition is a dual-slot PCIe Gen5 card for data-center servers, drawing up to 600 W per NVIDIA's specifications. It pairs a large memory pool, 96 GB of GDDR7 with ECC at about 1.6 TB/s, with RT cores for ray tracing and Tensor Cores that handle FP8 and FP4. NVIDIA quotes 2 PFLOPS of FP8 and 4 PFLOPS of FP4, both with sparsity.

Compared with the Ada Lovelace L40S that we also rent, it doubles memory from 48 GB to 96 GB, raises bandwidth from 864 GB/s to about 1.6 TB/s, and adds FP4. Compared with HBM GPUs such as the H100, it trades bandwidth for capacity: 96 GB at 1.6 TB/s, against 80 GB at 3.35 TB/s. There is no NVLink, so GPUs exchange data over PCIe. Our comparisons with the L40S and with the H100 go through each trade-off.

Rendering, generative media and inference on 96 GB

Rendering and VFX

Blender Cycles with OptiX, Redshift, Octane and Unreal Engine all use RT cores for ray tracing. Without NVLink, each GPU keeps its own copy of the scene, so 96 GB per GPU is your scene budget, and eight GPUs render eight frames or tiles in parallel rather than pooling memory. For production scenes with large textures and geometry caches, 96 GB keeps in GPU memory what would otherwise need slower out-of-core rendering. See rendering and VFX for typical setups.

Generative media

Video diffusion models are where 96 GB pays off. A 14B-parameter video model needs about 28 GB for its BF16 weights alone; the text encoder, the VAE and the activations for long, high-resolution clips come on top, which leaves little room on a 32 GB card and can be tight on 48 GB. On 96 GB, you can keep the whole pipeline in memory in BF16. A 12B-parameter image model such as FLUX.1, about 24 GB in BF16, leaves room for large batches or several models loaded at once. More on generative media workloads.

Cost-efficient inference

For LLMs, a 70B model in FP8 takes about 70 GB, leaving roughly 15–20 GB for KV cache after headroom; in 4-bit it takes about 35 GB, leaving most of the card for concurrent requests. A 32B model in BF16, about 64 GB, also fits. The limit is bandwidth: at about 1.6 TB/s, the card has roughly half the H100's and a third of the H200's, so for latency-sensitive serving of large models, HBM GPUs are the better tool. For batch jobs, internal tools and moderate-traffic endpoints, the RTX PRO 6000 gives you more GPU memory per dollar.

One GPU or the 8-GPU RTX PRO 6000 server?

The single-GPU plan includes 16 vCPUs, 188 GB of RAM, 1.92 TB of NVMe and a 10 Gbps port: enough for look-dev, one generation pipeline or one inference endpoint.

The RTX PRO 6000 server is a bare-metal machine with eight GPUs (768 GB of GDDR7 in total), two AMD EPYC 9555 processors with 128 cores in total, 1.5 TB of DDR5, four 3.84 TB NVMe drives and two 100 GbE ports. The GPUs connect over PCIe Gen5 with no NVLink; a Gen5 x16 link carries roughly 64 GB/s in each direction, far below the NVLink bandwidth of HGX servers. That makes this server a strong fit for work that splits into independent pieces, such as render frames, batch generation and one inference replica per GPU, and a weaker fit for large-model training, where gradients or activations cross GPUs at every step. Our explainer on NVLink vs PCIe shows where the line falls. Eight single plans would cost $11,120 per month, against $10,490 for the server.

Software and remote workflow

You get root access over SSH on Ubuntu 22.04 or 24.04 with NVIDIA drivers, CUDA, Docker and the NVIDIA Container Toolkit preinstalled; Debian or Rocky Linux are available on request. Render nodes and generation tools run headless, and you reach web interfaces through an SSH tunnel instead of exposing them publicly:

ssh -L 8188:localhost:8188 root@<server-ip>     # ComfyUI on your laptop at localhost:8188
blender -b shot.blend -E CYCLES -a -- --cycles-device OPTIX   # render the animation on the GPU

Commercial renderers need their own licenses, which you bring. Our remote development guide goes further.

What it costs to rent RTX PRO 6000 Blackwell

• Single RTX PRO 6000 $1,390 per month, about $1.90 per GPU-hour
• 8-GPU server $10,490 per month, about $1.80 per GPU-hour

Per gigabyte of GPU memory, the single plan works out to about $14.50 per month, against about $27.40 for our single H100 and $18.50 for the L40S. As with any monthly plan, the per-hour figure assumes the GPU runs all 730 hours. A render crunch that keeps a GPU busy around the clock for three weeks uses 504 hours, so monthly wins whenever hourly capacity would cost you more than about $2.76 per GPU-hour; a single weekend of rendering is usually better billed by the hour. In general, divide the monthly price by the hourly rate you would pay to find your break-even point.

Outbound traffic is included with no egress fees: 20 TB per month on the single GPU and 50 TB on the 8-GPU server, enough to move finished frames and model outputs off the server. Every plan also includes a public IPv4 address and an IPv6 /64, DDoS mitigation, 24/7 support and the 99.9% availability SLA, with no setup fee.

How ordering works

Request a configuration on the site, and we confirm in writing the configuration, price, location and delivery date before you pay. You then pay by card or in crypto and receive root access details by email; if we miss the confirmed delivery date, you can cancel for a full refund. Plans renew monthly with an email reminder and can be cancelled any month without a fee, which suits project-based work. For a fleet of render nodes or a 6–12 month term, contact us for volume pricing.