Early-access requests
GPU rental: catalogue and specs
Specs from NVIDIA’s official materials. Availability and form factor of each model are being confirmed.
Open the sizing tool11 of 11 models
Memory is specified per GPU. Multiple GPUs have separate memory: using a larger model across them requires compatible software and model partitioning. NVLink does not make them one GPU.
We don’t publish prices yet: terms and availability come with our reply.
| GPU | Memory per GPU | Memory bandwidth | FP16/FP8, TFLOPSwithout sparsity | FP64, TFLOPS | GPU interconnect | Good for (sizing-tool estimate) | Actions |
|---|---|---|---|---|---|---|---|
| 96 GB and more | |||||||
| B300 Blackwell, SXM (being confirmed)270 GB HBM3eLLMs up to 123B in FP8 on one GPU, multi-GPU training (NVLink) | 270 GB HBM3e | 7.7 TB/s | 2,250/4,500 | 1.2 | NVLink 1.8 TB/s | LLMs up to 123B in FP8 on one GPU, multi-GPU training (NVLink) | |
| B200 Blackwell, SXM (being confirmed)180 GB HBM3eLLMs up to 123B in FP8 on one GPU, multi-GPU training (NVLink), FP64 computing | 180 GB HBM3e | 7.7 TB/s | 2,250/4,500 | 37 | NVLink 1.8 TB/s | LLMs up to 123B in FP8 on one GPU, multi-GPU training (NVLink), FP64 computing | |
| H200 Hopper, SXM (being confirmed)141 GB HBM3eLLMs up to 70B in FP8 on one GPU, multi-GPU training (NVLink), FP64 computing | 141 GB HBM3e | 4.8 TB/s | 989.4/1,978.9 | 34; Tensor Core 67 | NVLink 900 GB/s | LLMs up to 70B in FP8 on one GPU, multi-GPU training (NVLink), FP64 computing | |
| RTX PRO 6000 Blackwell, Server Edition (being confirmed)96 GB GDDR7LLMs up to 70B in FP8 on one GPU, 3D rendering, video encoding | 96 GB GDDR7 | 1.6 TB/s | 467.8/935.6 | not listed | PCIe 5, no NVLink | LLMs up to 70B in FP8 on one GPU, 3D rendering, video encoding | |
| 80 GB | |||||||
| H100 Hopper, SXM (being confirmed)80 GB HBM3LLMs up to 32B in BF16 on one GPU, multi-GPU training (NVLink), FP64 computing | 80 GB HBM3 | 3.35 TB/s | 989.4/1,978.9 | 34; Tensor Core 67 | NVLink 900 GB/s | LLMs up to 32B in BF16 on one GPU, multi-GPU training (NVLink), FP64 computing | |
| A100 Ampere, SXM (being confirmed)80 GB HBM2eLLMs up to 32B in BF16 on one GPU, multi-GPU training (NVLink), FP64 computing | 80 GB HBM2e | 2.04 TB/s | 312/none | 9.7; Tensor Core 19.5 | NVLink 600 GB/s | LLMs up to 32B in BF16 on one GPU, multi-GPU training (NVLink), FP64 computing | |
| 48 GB | |||||||
| L40S Ada Lovelace, PCIe (being confirmed)48 GB GDDR6LLMs up to 32B in FP8 on one GPU, 3D rendering, video encoding | 48 GB GDDR6 | 0.86 TB/s | 362.05/733 | not listed | PCIe 4, no NVLink | LLMs up to 32B in FP8 on one GPU, 3D rendering, video encoding | |
| RTX 6000 Ada Ada Lovelace, Workstation (being confirmed)48 GB GDDR6LLMs up to 32B in FP8 on one GPU, 3D rendering, video encoding | 48 GB GDDR6 | 0.96 TB/s | 364/728.5 | not listed | PCIe 4, no NVLink | LLMs up to 32B in FP8 on one GPU, 3D rendering, video encoding | |
| RTX A6000 Ampere, Workstation (being confirmed)48 GB GDDR6LLMs up to 14B in BF16 on one GPU, 3D rendering, video encoding | 48 GB GDDR6 | 0.77 TB/s | 154.8/none | not listed | supports an NVLink bridge for 2 GPUs | LLMs up to 14B in BF16 on one GPU, 3D rendering, video encoding | |
| 24 GB | |||||||
| L4 Ada Lovelace, PCIe (being confirmed)24 GB GDDR6LLMs up to 14B in FP8 on one GPU, 3D rendering, video encoding | 24 GB GDDR6 | 0.3 TB/s | 121/242 | not listed | PCIe 4, no NVLink | LLMs up to 14B in FP8 on one GPU, 3D rendering, video encoding | |
| RTX A5000 Ampere, Workstation (being confirmed)24 GB GDDR6LLMs up to 8B in BF16 on one GPU, 3D rendering, video encoding | 24 GB GDDR6 | 0.77 TB/s | 111.1/none | not listed | supports an NVLink bridge for 2 GPUs | LLMs up to 8B in BF16 on one GPU, 3D rendering, video encoding | |
† Usable memory per NVIDIA. The chip’s maximum memory is 288 GB.
‡ Usable memory per NVIDIA. The chip’s maximum memory is 192 GB.
* Dense FP16/BF16 and FP8 from the H100 SXM5 column of NVIDIA’s whitepaper: for the H200, NVIDIA lists only with-sparsity figures, the same as the H100 SXM’s.
** Derived: NVIDIA’s whitepaper per-SM Tensor Core rate × the 2,430 MHz boost clock in the product brief. For the Server Edition, NVIDIA gives FP16/FP8/FP4 only as rounded 1/2/4 PFLOPS peaks, the level of sparse figures; it publishes no dense ones.
*** Derived: half the with-sparsity FP16 figure NVIDIA gives for the RTX A5000 (222.2 TFLOPS); NVIDIA gives no dense one.
FP64 “not listed”: NVIDIA’s specs for these cards don’t give it; per NVIDIA’s whitepapers, FP64 on the GA102, AD102 and GB202 chips runs at 1/64 of the FP32 rate.
Specs from NVIDIA’s official materials, checked 26 September 2026.