Skip to content
aipowered

Find your next machine.

Try “5090”, “Jetson”, or “DGX”.

NVIDIA GH200 (96GB HBM3)

96GB HBM3
Memory
4,900GB/s
Memory bandwidth
Q2 2024
Released

Price history

On-demandSpot
USD / GPU / HR
On-demand · latest$4.252026-10-05

What it costs to rent an NVIDIA GH200

The NVIDIA GH200 is a Hopper-architecture fine-tuning NVIDIA GPU with 96GB of HBM3 memory and 4,900 GB/s of memory bandwidth, released Q2 2024. It is built for large-model training and high-throughput inference.

Renting one currently costs between $1.99 and $6.50 per GPU-hour across the 2 cloud providers we track, with a median of $4.25. The cheapest live rate is $1.99/hr from Vultr (On-demand). All rates below are normalised to a single GPU per hour so providers with multi-GPU nodes compare like-for-like.

Where to rent the GH200

How we rank

The median is one provider, one vote: each provider contributes the median of its own on-demand listings, then the headline is the median across providers — so no single marketplace or hyperscaler sets the price alone. Below two providers we show the range instead.

Default order is relevance, weighed the way a buyer would: price per GPU first, lightly blended with total node cost; on-demand preferred over spot and reserved when prices are close; established clouds edged above marketplaces; in-stock, recently-checked offers above stale ones; and no single provider is allowed to crowd the top. Click a column to sort by the raw numbers. No paid placements — ranking is never sold.

Scroll horizontally to compare all columns →

ProviderTierConfigTotal VRAMBillingAvailabilityGo to provider
VultrBestEnterprise1× GH20096GBOn-demand$1.99$1.99Rent
CoreWeaveEnterprise1× GH20096GBOn-demand$6.50$6.50Rent

Technical specifications

Throughput
FP6434 TFLOPS
FP3267 TFLOPS
TF32495 TFLOPS
FP16/BF16990 TFLOPS
Silicon & memory
Memory
96 GB
Memory bandwidth
4,900 GB/s
TDP
1,000 W
INT8
1,979 TOPS
Process
5 nm
Foundry
TSMC
Released
8 Aug 2023
TrainingInferenceHPC
Datasheet
SuperchipGrace CPU + Hopper GPU
GPU memory96 GB HBM3 or 144 GB HBM3e
CPU memoryup to 480 GB LPDDR5X
Link900 GB/s NVLink-C2C
UseAI training, inference and HPC
Powerplatform dependent
Class
Fine-tuning
Cloud providers
2 tracked
Price range
$1.99 – $6.50 / GPU / hr

SNAPSHOT 2026-10-12T05:40:41.151564+00:00 · BACK TO THE FULL INDEX