Dedicated NVIDIA GPU servers

We rent dedicated GPU servers whole, on long terms, and every price comes in a written proposal before you sign.

Bare metal vs shared GPU infrastructure explains what dedicated hardware changes in production.

The catalog falls into four groups by what the hardware is built for: training, rack-scale systems, inference, and graphics and media.

A100 80GB SXM servers, eight GPUs on an HGX board, are also available on request.

Training: HGX B300 and HGX B200

HGX B300 is our main line: eight Blackwell Ultra GPUs with 270 GB of HBM3e each, 2.1 TB per server, linked by fifth-generation NVLink at 1.8 TB/s per GPU. Our reference GPU clusters are built from it, on a fabric of 800 Gb/s per GPU. HGX B200 carries 180 GB of HBM3e per GPU, 1,440 GB per server, and keeps far more double precision: NVIDIA lists 296 TFLOPS of FP64 per 8-GPU board, against 10 TFLOPS for HGX B300. Choose B200 when simulation or other FP64 work shares the cluster with AI. 8 x 270GB: what fits on a B300 node works through 405B and 70B models.

By reservation: GB300 NVL72 and Vera Rubin

GB300 NVL72 is a liquid-cooled rack of 72 Blackwell Ultra GPUs and 36 Grace CPUs in one NVLink domain, with 20 TB of HBM3e and 130 TB/s of total NVLink bandwidth, by NVIDIA's figures. Vera Rubin is the next NVIDIA platform, with 288 GB of HBM4 per Rubin GPU, and we have Rubin servers on order. Both are by reservation. Our facility requirements include a path to direct-to-chip liquid cooling for GB300 NVL72 and later systems.

Inference: RTX PRO 6000 Blackwell and L40S

RTX PRO 6000 Blackwell servers carry eight GPUs with 96 GB of GDDR7 each, 768 GB per server, served from Montreal. L40S servers carry eight GPUs with 48 GB of GDDR6 each. Neither card has NVLink, so GPUs exchange data over PCIe: these servers suit models where each replica fits on one GPU or a few, such as a 70B model in FP8 on a single 96 GB card. 8 x 96GB: what fits on an RTX PRO 6000 Blackwell node works through the memory arithmetic.

Graphics and media: GeForce and RTX workstation cards

RTX 5090 (32 GB), RTX 4090 (24 GB) and RTX A6000 (48 GB) servers carry eight GPUs; RTX 4000 Ada (20 GB) and RTX A4000 (16 GB) servers carry four.

Each server carries one of three labels, which say where its hardware stands today.

Four questions narrow the choice: how much memory each GPU needs, how the GPUs talk to each other, how many servers the job spans, and what the workload is.

Memory per GPU

Memory decides what fits. Weights, plus the KV cache for inference or optimizer state and activations for training, must sit in GPU memory, and a 70B-parameter model in FP8 needs about 70 GB for its weights alone. 70GB of weights, four GPUs shows what each memory size leaves for the KV cache.

Interconnect

When one model is split across GPUs, GPU-to-GPU bandwidth sets the pace. HGX B300 and B200 servers give each GPU 1.8 TB/s of NVLink and GB300 NVL72 joins 72 GPUs in one NVLink domain, while PCIe servers rely on the PCIe bus: 64 GB/s on the L40S. Between servers, HGX B300 clusters use a rail-optimized, fully non-blocking fabric at 800 Gb/s per GPU.

Cluster size

Past one server, our reference clusters run 64, 128 or 288 HGX B300 nodes, up to 2,304 GPUs on one dedicated fabric, and the design extends beyond that. See GPU clusters and the delivery specification.

Workload