>>
NVIDIA H100 GPU PODS
Rent H100 pods by the minute from independent providers on the Bittensor Subnet 51 marketplace, with pods starting in seconds.
pip install lium.io && lium up --gpu H100# spin up a pod from your terminal59 H100 GPUs available · from $1.20/GPU/hr · ≈2.0¢/min · 10 locations · billed per minuteGPU memory
80 GB HBM3 / HBM2e
Bandwidth
up to 3.35 TB/s
FP16 tensor
up to 989 TFLOPS
FP8 tensor
up to 1,979 TFLOPS
Interconnect
NVLink up to 900 GB/s
Max power
350–700 W
Tensor figures are NVIDIA’s published dense numbers for the card, without sparsity; sparsity doubles them. What a pod actually delivers depends on the model, the kernels and the host it runs on.
Cheapest H100 pods right now
Read live from the marketplace. Prices are per GPU per hour, billed by the minute.
What the H100 is built for
The H100 is NVIDIA’s Hopper-generation data center GPU and still the default choice for serious training and high-throughput inference. Its transformer engine adds native FP8, which roughly doubles arithmetic throughput over FP16 for models that tolerate the lower precision, and 80 GB of high-bandwidth memory keeps weights, activations and KV cache on the card instead of spilling to host RAM. Teams reach for it to fine-tune language models, to serve production inference where latency is part of the product, and to run diffusion, speech and scientific workloads that are limited by memory bandwidth rather than by raw arithmetic.
Typical workloads
Fine-tuning 7B–70B language models
Production LLM inference at low latency
Diffusion, speech and vision training
Bandwidth-bound HPC and simulation
SXM or PCIe?
SXM
H100 SXM 80GB HBM3
from
$1.20
/GPU/HR
· 56 free on 20 machines
Best for multi-GPU training, tensor parallelism and anything bandwidth-bound.
GPU memory
80 GB HBM3
Bandwidth
3.35 TB/s
FP16 tensor
989 TFLOPS
FP8 tensor
1,979 TFLOPS
Interconnect
NVLink 900 GB/s
Max power
700 W
PCIe
H100 PCIe 80GB HBM2e
from
$1.50
/GPU/HR
· 3 free on 3 machines
Best for single-GPU inference, and usually the cheaper of the two per GPU-hour.
GPU memory
80 GB HBM2e
Bandwidth
2 TB/s
FP16 tensor
756 TFLOPS
FP8 tensor
1,513 TFLOPS
Interconnect
NVLink bridge 600 GB/s
Max power
350 W
Both carry 80 GB, but they are not the same part. The SXM card uses HBM3 at 3.35 TB/s and sits on an NVLink mesh at 900 GB/s, so tensors move between GPUs in one node without crossing PCIe. The PCIe card uses HBM2e at 2 TB/s, draws 350 W instead of 700 W, and links cards in pairs over a 600 GB/s bridge. For single-GPU inference the gap is modest and the PCIe card is usually the cheaper of the two per GPU-hour. For multi-GPU training, tensor parallelism, or anything bandwidth-bound, take SXM. The marketplace lists them as NVIDIA H100 80GB HBM3 and NVIDIA H100 PCIe.
Frequently asked questions
How much does an H100 pod cost?
Each provider sets its own price per GPU-hour, and Lium bills per minute with no minimum rental period, so a twenty-minute experiment costs twenty minutes. The prices on this page are read live from the marketplace and move as providers add and remove machines, so the cheapest H100 today may not be the cheapest one tomorrow.
Is there a minimum rental period?
No. There is no minimum and no reservation to make: start a pod when you need it, stop it when the job is done, and pay for the minutes it ran. That makes short benchmarks, one-off fine-tunes and interactive debugging practical in a way that a fixed block of time is not.
How do I rent an H100 on Lium?
Two ways, and they do the same thing. In the browser, open the H100 pods, pick a machine, choose a template and start it — a pod starts in seconds when the image is already cached on the host. From a terminal, run pip install lium.io and then lium up --gpu H100; the CLI is what most people wire into scripts and agents. Both need an account with a funded balance.
Where do the machines run, and what happens to my data?
Providers run their own hardware in their own facilities, so the fleet moves. The live line at the top of this page counts how many distinct locations have free H100 capacity right now, and every pod in the list names its city and country, so you can pick a region before you start. Your data lives on the machine you rented and is released when you delete the pod; attach a volume if you need storage that outlives a single pod.
Start an H100 pod in seconds
Billed per minute with no hourly minimum. 59 H100s are free right now across 10 locations.
Browse H100 pods
pip install lium.io && lium up --gpu H100# spin up a pod from your terminalNeed a different card? NVIDIA B200 pods are on the same marketplace.
For AI agents — CLI install, agent skill, headless signup: llms-full.txt