HopperLive Marketplace Asks Available

NVIDIA H200

H200 keeps H100's Hopper compute and replaces the memory system: 141 GB of HBM3e at 4.8 TB/s. The menu shape you see in the catalog is a larger host, about 44 vCPUs and 182 GB of RAM, next to 141 GB of VRAM. Pick it when H100 jobs fail on memory, not on FLOPS.

Deploy NVIDIA H200 on Marketplace →All Hardware Fleet

Flexible Compute Modes

Deploy NVIDIA H200 Your Way: Docker Pods or Root KVM VPS

NovaPod lets you decide how you want to harness GPU power. Whether you need an instant pre-cached container stack, a full root virtual machine with dedicated public IP, or a multi-node NVLink cluster, pay only for what you run with second-by-second crypto billing.

Docker GPU PodCold Boot < 15s

Fast, Lightweight AI Containers

Pure GPU compute without hypervisor overhead. Boot straight into PyTorch, vLLM, ComfyUI, or JupyterLab.

1-second precision metering
37+ Pre-cached ML framework templates
Zero idle fee when instance is stopped

Deploy as Docker Pod

Root KVM VPSPersistent NVMe

Full Root Virtual Machine

Dedicated operating system (Ubuntu 24.04 LTS / Windows GPU Server) with complete sudo root SSH access from minute one.

Dedicated public IPv4 (all ports open)
Pause anytime; pay only $0.00014/GB/hr for disk
Run Docker-in-Docker, k3s, custom drivers

Deploy as Root VPS

Multi-GPU Cluster900 GB/s NVLink

Supercomputing & Distributed Training

Pool up to 8× NVIDIA H200 on a unified node fabric with Quantum-2 InfiniBand RDMA for multi-node scale.

Zero-loss all-reduce gradient exchange
PyTorch FSDP, DeepSpeed Zero-3 ready
Native GPUDirect Storage & RDMA fabric

Configure Cluster

Hardware Specifications

Architecture: NVIDIA Hopper
VRAM: 141 GB HBM3e
Bandwidth: 4.8 TB/s
Tensor Cores: 528, 4th generation
CUDA cores: 16,896
FP64: 34 TFLOPS
FP32: 67 TFLOPS
TF32: 989 TFLOPS with sparsity
FP16 Tensor: 1,979 TFLOPS with sparsity
FP8 Tensor: 3,958 TFLOPS with sparsity
INT8: 3,958 TOPS with sparsity
NVLink: 900 GB/s
Reference host: 44 vCPUs, 182 GB RAM
Form factor: SXM5
TDP: 700 W
Versus H100 memory: 1.76× capacity, about 1.4× bandwidth

Same Hopper compute, different memory

H200 does not replace the H100's Tensor Cores. It replaces the memory: 141 GB of HBM3e at 4.8 TB/s instead of 80 GB of HBM3 at 3.35 TB/s. That is about 1.76× the capacity and about 1.4× the bandwidth. FP8, the Transformer Engine, and the Hopper software stack stay.

The menu host is larger than the H100 row: about 44 vCPUs and 182 GB of RAM with 141 GB of VRAM. Use the host RAM for the dataloader and the CPU side of a RAG stack. Do not expect it to act like a second GPU.

When 80 GB runs out

FP16 weights for a 70B model are about 140 GB before any cache. They are tight on an H200 and impossible on an H100. FP8 is the practical way to serve 70B with a 32k context and a real batch on one card. Multi-model serving, a resident embedding index, or a long agent trace are the other reasons the extra 61 GB pays for itself.

If the profile says you are compute-bound on a model that already fits in 80 GB, H200 will not feel faster in proportion to its price. If the profile says you are bandwidth-bound or out of memory, it is the Hopper card to take.

Neighbors

B200 is the next architecture: 192 GB HBM3e, 8 TB/s, and FP4. Move there for training that needs Blackwell numerics or for models that do not fit in 141 GB. A100 is the cheaper card when FP8 does not matter. MI300X has 192 GB on one GPU, but only if the image is ROCm.

Launch is the same as H100. Pick the live H200 offer, start a serving template, and stop the pod when the experiment ends. The ask on the card is the rate. The median next to it tells you whether that ask is typical.

Architectural Generation Comparison

H200 does not replace H100's compute. It replaces the memory system.

	H200	H100	B200
VRAM	141 GB HBM3e	80 GB HBM3	192 GB HBM3e
Bandwidth	4.8 TB/s	3.35 TB/s	8 TB/s
FP4	No	No	Yes
70B FP16 weights	Tight	Does not fit	Fits
70B FP8 + cache	Comfortable	Modest cache	Comfortable

Storage & Tooling Infrastructure

High-Speed Storage, Native Ports & Crypto Rails

Every NVIDIA H200 instance on NovaPod connects directly to high-throughput infrastructure designed to eliminate I/O bottlenecks and onboarding delays.

25 Gb/s Shared Network Storage

Mount large model checkpoints and training sets across one or multiple instances simultaneously. Data outlives single pods and stays accessible when instances stop.

JupyterLab, VS Code & SSH

Connect from your local terminal with standard SSH keys or open web-based IDEs directly in browser tabs without managing complex reverse proxy configs.

Non-Custodial Multi-Chain Rails

Deposit USDT, USDC, SOL, ETH, or BNB to your dedicated personal deposit address. Confirmed on-chain in seconds with zero mandatory identity verification.

Target Workloads for NVIDIA H200

Long-context serving

141 GB holds a 70B model in FP8 and a much larger KV cache than 80 GB. FP16 weights alone are about 140 GB, so they are tight.

32k context and beyondHigh-batch 70B serving100B-class with a smaller batchLlama-class 70B in FP8

Several models on one GPU

Colocate a chat model, a smaller specialist, and an embedding model instead of paying for three cards.

Two or three 30B-class modelsSeveral 7B–13B modelsChat plus embeddingsA/B of two checkpoints

Memory-bound training

Use it when activations, optimizer state, or the KV cache spill off an H100. If the step is compute-bound inside 80 GB, H100 is the better rate.

Larger micro-batchesFewer tensor-parallel ranksRLHF that is cache-heavyFine-tunes that OOM on H100

Throughput at scale

Eight H200s on NVLink, then 400 Gb/s class InfiniBand with GPUDirect if you leave the node. Tested paths are vLLM, TensorRT-LLM, SGLang, Megatron, and DeepSpeed.

8-GPU NVLink nodeMulti-node by quoteSame Hopper software as H100Stop the pod when the eval ends

When to Choose NVIDIA H200

Pick H200 when H100 runs out of memory

Same Tensor Cores. The upgrade is 141 GB of HBM3e at 4.8 TB/s.

Pick H100 when 80 GB is comfortable

You would be paying for memory you never touch. Train on H100 and serve on H200 is a common split.

Pick B200 for FP4 or 192 GB

Blackwell is the next compute step: more memory, more bandwidth, and a format Hopper does not have.

Pick A100 when FP8 does not matter

Classic training and smaller inference are cheaper on Ampere.

The 141 GB rule

FP16 weights for a 70B model are about 140 GB before cache. H200 can hold that and little else. FP8 is the practical way to serve 70B with a long context on one card.

Menu host

The catalog row is 44 vCPUs and 182 GB of RAM with 141 GB of VRAM. That host is for the KV cache and the data loader, not a second GPU.

Frequently Asked Questions

How much does an H200 cost here?

The live ask on the marketplace, billed per second. Specialist clouds and hyperscalers publish their own H200 rates, often in a wide range. Use those only as context. The number you pay is the offer you deploy.

What is the cheapest way?

A reclaimable offer, if one is listed and your job checkpoints. Production serving stays on a card the host will not take back. Tear the machine down when the run ends.

Can I rent it by the hour?

Yes, and the meter is seconds. A one-hour benchmark costs about one hour. There is no commit fee on a normal deploy.

What is the difference from H100?

Compute stays in the same Hopper class: fourth-generation Tensor Cores and FP8. Memory changes from 80 GB HBM3 at 3.35 TB/s to 141 GB HBM3e at 4.8 TB/s.

Inference or training?

Both, and it is especially strong for inference, because batch, context, and colocated models all consume the extra memory. Training helps when memory is the stall. If it is not, H100 often has the better price.

What fits on one GPU?

70B in FP8 with room for cache. 70B in FP16 is about 140 GB before cache, so it is tight. 100B-class needs a smaller batch. 200B and up wants tensor parallel across several H200s.

Can I serve more than one model?

Yes. 141 GB can hold two or three models around 30B, or several 7B–13B models, which is how a chat, code, and embedding stack shares one card.

Clusters?

Eight H200s per node on 900 GB/s NVLink is the unit. Multi-node wants 400 Gb/s class networking and GPUDirect. Larger shapes are quoted.

Is it worth it over H100?

When you are out of memory, when the KV cache is the limiter, or when you want several models resident. If you are comfortably under 80 GB, H100 is the cheaper Hopper card.

What is the host in the menu?

44 vCPUs and 182 GB of RAM next to 141 GB of VRAM. That RAM is for the dataloader and the CPU side of a RAG stack. It is not a second GPU. A specific offer can differ.

When do I skip it for MI300X?

When you need more than 141 GB on one GPU and the image is already ROCm. CUDA serving stays here or on B200.

Alternative Hardware Options

Ready to Deploy NVIDIA H200?

Launch an instance on the live marketplace in under 15 seconds. Fund your balance with cryptocurrency and pay per second with zero commitments.

Launch on Marketplace →