NVIDIA H200 rental: early access
Hopper · SXM (being confirmed)
We’re collecting early-access requests. We’re placing our capacity in Uzbekistan. We don’t publish prices yet. We’ll confirm availability in our reply.
H200
Request H200We’ll reply with terms for your workload. Nothing to pay before the contract.
Estimate usage- Memory per GPU
- 141 GB HBM3e
- Memory bandwidth
- 4.8 TB/s
- GPU interconnect
- NVLink 900 GB/s
NVIDIA H200 specifications
- Architecture
- Hopper
- Announced
- 2023
- Memory
- 141 GB
- Memory type
- HBM3e
- Memory bandwidth
- 4.8 TB/s
- FP32
- 67 TFLOPS
- FP16/BF16, dense
- 989.4 TFLOPS
- FP8, dense
- 1,978.9 TFLOPSDense 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.
- FP4, dense
- none
- FP64
- 34 TFLOPS
- FP64 Tensor Core
- 67 TFLOPS
- RT cores
- no
- NVENC/NVDEC
- NVENC 0, NVDEC 7NVENC from the H100 row of NVIDIA’s encode and decode matrix, which has no H200 row; the H200 page lists decoders only.
- MIG
- up to 7 instances
- GPU interconnect
- NVLink 900 GB/s
- Power (TDP/TGP)
- up to 700 W
- Form factor
- SXM (being confirmed)
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.
Other variants: H200 NVL (opens in a new tab)
The H200 form factor is still being confirmed. Specs from NVIDIA’s official materials, checked 26 September 2026.
The H200 by workload
- Inference: LLMs up to 70B in FP8 on one GPU
- Fine-tuning: QLoRA up to 123B, LoRA up to 32B on one GPU
- Training from scratch: models up to 3B on one GPU
- Rendering: no RT cores
- FP64 computing: 34 TFLOPS, Tensor Core 67 TFLOPS
An estimate from open formulas, not a benchmark or an invoice.
Servers with the NVIDIA H200
Server configuration (CPU, RAM, disk, network) is being confirmed.
Sizing assumption: up to 8 GPUs per server. We’ll confirm the actual configuration in our reply.
More than 8 GPUs? Discuss a clusterSources
- NVIDIA H200 (opens in a new tab)
- NVIDIA H100 Tensor Core GPU Architecture (whitepaper V1.04) (opens in a new tab)
- NVIDIA Video Encode and Decode Support Matrix (opens in a new tab)
- NVIDIA Omniverse Developer Guide: Technical Requirements (opens in a new tab)
- NVIDIA Newsroom, November 13, 2023 (opens in a new tab)
Similar GPUs
- NVIDIA H100
80 GB HBM3 · 3.35 TB/s · NVLink
H200 leads on 2 of 7 metrics. H200 has more memory (141 vs 80 GB) and memory bandwidth (4.8 vs 3.35 TB/s). Equal: FP32, FP16/BF16, FP8, FP64 Tensor Core, NVLink.
- NVIDIA B200
180 GB HBM3e · 7.7 TB/s · NVLink
B200 leads on 6 of 7 metrics, H200 on 1. B200 has more FP16/BF16 (2,250 vs 989.4 TFLOPS) and FP8 (4,500 vs 1,978.9 TFLOPS). But H200 has more FP64 Tensor Core throughput (67 vs 37 TFLOPS). Without Tensor Cores, B200 has more FP64: 37 vs 34 TFLOPS.
- NVIDIA RTX PRO 6000 Blackwell
96 GB GDDR7 · 1.6 TB/s · PCIe 5
H200 leads on 4 of 5 metrics, RTX PRO 6000 on 1. H200 has more memory bandwidth (4.8 vs 1.6 TB/s) and FP8 (1,978.9 vs 935.6 TFLOPS). But RTX PRO 6000 has more FP32 (120 vs 67 TFLOPS). FP64: 34 TFLOPS on H200; not in NVIDIA’s card specs for RTX PRO 6000 (NVIDIA’s whitepaper gives 1/64 of FP32 for the GB202 chip).
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.
The RTX PRO 6000’s FP16, FP8 and FP4 are derived figures. 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.
We’ll reply with terms for your workload. Nothing to pay before the contract.