NVIDIA GPU clusters: H100, H200, B200 and GB200
Rent multi-GPU nodes at a live, per-GPU hourly rate across a marketplace of independent providers, no long-term reservation required. Below: what each generation of hardware actually offers, real pricing today, and where Aquanode's marketplace model stops and a dedicated multi-node cluster provider starts.
What counts as a GPU cluster
A GPU cluster is more than one GPU node working on the same job. That covers two very different shapes of hardware. A multi-GPU node is a single server carrying several GPUs (commonly 4 or 8) that talk to each other over a fast on-board interconnect like NVLink or Infinity Fabric. A multi-node cluster connects several of those servers together over a network fabric like InfiniBand, so a training job can span hundreds or thousands of GPUs at once. Rack- scale systems like NVIDIA's GB200 NVL72 push that further still, wiring dozens of GPUs into what behaves like one giant accelerator.
Which one you need depends on the job. Fine-tuning and most inference workloads fit comfortably on a single multi-GPU node. Pretraining a frontier-scale model from scratch is usually the only workload that actually needs a multi-node, InfiniBand- connected cluster. See our full breakdown in What Are AI Compute Clusters, and How Do You Choose One.
Which GPU generation to choose
H100 (Hopper)
80GB of HBM3 at 3.35 TB/s, connected over NVLink 4 at 900 GB/s per GPU. Still the most widely stocked card across the marketplace, and the default choice for both training and inference unless a job specifically needs more VRAM or the newest compute (NVIDIA H100 product page).
H200 (Hopper refresh)
Identical compute to H100 on the same die, but 141GB of HBM3e at 4.8 TB/s, 76% more memory and 43% more bandwidth. That extra headroom is wasted on a compute-bound job, but it is the better single-card fit for a large model in FP16 or a long-context inference workload where the key-value cache, not raw compute, is what runs out of room (NVIDIA H200 product page).
B200 (Blackwell)
180GB of HBM3e, NVLink 5 at 1.8 TB/s per GPU, and more than double H100's FP8 throughput with native FP4 tensor-core support on top. It is the highest-throughput card on the marketplace and priced accordingly, with the thinnest supply of the three, so it earns its rate on frontier-scale training or the highest-throughput inference deployments rather than workloads an H100 already handles comfortably (NVIDIA HGX B200 datasheet).
AMD MI300X
192GB of HBM3 at 5.3 TB/s, the largest single-card VRAM pool on the marketplace, more than double an H100's 80GB. Cards talk to each other over AMD's Infinity Fabric rather than NVLink. That VRAM headroom is what lets the biggest models fit on fewer cards instead of splitting across more of them (AMD Instinct MI300X data sheet).
GB200 NVL72 (rack-scale, not a rentable card)
A full GB200 NVL72 rack wires 72 Blackwell GPUs and 36 Grace CPUs into a single NVLink domain with roughly 13.4TB of aggregate GPU memory, built for frontier-scale training and the largest inference deployments. It is sold and deployed as one liquid-cooled system, not a card anyone rents by the GPU-hour, so Aquanode does not list it at per-GPU granularity; the live H100/H200/B200 pricing below is the closest per-GPU-hour path to the same Hopper/Blackwell compute today (NVIDIA GB200 NVL72 product page).
Interconnect: what actually moves data between GPUs
Raw TFLOPS only tells part of the story on a multi-GPU job. Training splits a model or a batch across cards, which means the cards constantly exchange gradients and activations with each other, and how fast that exchange happens can matter more than how fast any single card computes.
On-board interconnect is what carries that traffic inside one node. NVLink 4 moves 900 GB/s per GPU on H100 and H200, and NVLink 5 roughly doubles that to 1.8 TB/s per GPU on B200, both figures from NVIDIA's own datasheets. AMD's MI300X uses Infinity Fabric instead of NVLink for the same job. A card with no board-level interconnect at all (older PCIe-only parts) falls back to the PCIe bus for cross-card traffic, which is fine for independent inference jobs that do not need cards to talk to each other, but a real bottleneck for tightly coupled training.
InfiniBand is a different layer entirely: it connects separate servers to each other, not GPUs within one server, which is what lets a training run span hundreds or thousands of GPUs across a multi-node cluster instead of the handful that fit in one box. Every multi-GPU node listed on Aquanode's marketplace keeps its vendor's real on-board interconnect intact rather than splitting a card's bandwidth across tenants, so a training job that needs a fast card-to-card path actually gets one.
Why rent GPU compute through Aquanode
Experimentation
Spin up a multi-GPU node in minutes, pay by the hour, and close it the moment you're done. Nothing to reserve and nothing idle burning a monthly commitment.
Fine-tuning and training
8-GPU H100, H200 and B200 nodes with NVLink between cards, priced per GPU and compared live across every provider we list, so you're never stuck paying one vendor's rate because that's the account you already have.
Inference and serving
Deploy on the cheapest node that fits your model today, and carry your environment with you if a better rate shows up somewhere else next week.
What you get with an Aquanode marketplace node
Real per-GPU pricing, not a node price split unevenly
Every multi-GPU offer is normalized to a real per-GPU hourly rate before it's published, refreshed hourly, and billed per second once your node is running.
Your setup outlives any single provider
Snapshot your environment (custom software, weights, the config you spent an evening tuning) and restore it onto a fresh node on a different provider, instead of rebuilding from scratch every time supply or price shifts. Set an automatic snapshot interval and you don't have to remember to do it by hand.
One account across every provider
One login, one bill and one set of API keys instead of a separate account per vendor every time you go chasing capacity or a better rate.
Jobs that survive a provider pulling the box back
Run a saved environment as a job instead of a live session. If a provider reclaims the node mid-run, Aquanode detects the loss and retries on a different provider, picking up from the last checkpoint if the job checkpoints, and every job is bounded by a time limit times attempts times machines, so a retry loop can never run away on you.
Storage for multi-GPU jobs
A dedicated cluster provider typically answers storage with one managed parallel filesystem shared across a whole rack. Aquanode is a marketplace of independent providers rather than one filesystem vendor, so the answer here is per-workspace portability instead: the data follows the job, not the box.
Network volumes
Block storage that belongs to your workspace rather than to one instance. Create it once, attach it to whichever box needs it, and keep it when that box goes away. Available today with Hyperstack and Verda.
Portable workspaces
Name a directory once and it syncs on a tick. Deploy on a different provider, attach the same workspace there, and your training data and configs are already waiting instead of getting rebuilt from scratch.
Workspace backups
Snapshot a directory on a running box, then restore it onto a different GPU or provider and pick up where you left off. Snapshots after the first are incremental, so a frequent schedule stays cheap.
Software stack: CLI, SSH and live metrics
The aq CLI is how most jobs get driven end to end. Run aq gpus with no account and no login to see a live per-GPU market summary across every provider Aquanode integrates with; log in and the same CLI creates, SSHes into and tears down a node.
Once a node is running, live GPU metrics in the console show per-GPU utilization, VRAM used, temperature, power draw, and SM and memory clock speed, refreshed every few seconds, alongside host CPU and system memory, so you can tell whether the GPU you are paying for is actually doing the work.
Being direct about the boundary: there is no managed Kubernetes or Slurm layer sitting on top of a node. You get the box, the CLI, the API and the metrics, and you bring your own orchestrator the moment a job needs to span more than one node.
Multi-GPU nodes available now
| GPU | VRAM | Interconnect | Aquanode cheapest $/hr per GPU | |
|---|---|---|---|---|
| B200 | 180GB HBM3e | NVLink 5, 1.8 TB/s per GPU | $6.75/hr | View B200 offers |
| H200 | 141GB HBM3e | NVLink 4, 900 GB/s bidirectional | $3.59/hr (sold out) | View H200 offers |
| H100 | 80GB HBM3 | NVLink 4, 900 GB/s bidirectional | $1.99/hr (sold out) | View H100 offers |
| AMD MI300X | 192GB HBM3 | Infinity Fabric | $2.39/hr (sold out) | View AMD MI300X offers |
Need a multi-node InfiniBand cluster instead?
Being direct about it: Aquanode is a marketplace for single-node rentals, including multi-GPU nodes with up to 8 GPUs. We do not operate our own InfiniBand-connected multi-node training clusters or rack-scale systems like GB200 NVL72. If a multi-week, multi-hundred- GPU pretraining run is what you actually need, that is a genuinely different product, and providers who run it full-time (Together AI's GPU Clusters, Lambda's 1-Click Clusters) are the better fit for that specific job. See how Aquanode compares to Together AI for the fuller picture.
Frequently asked questions
What's the difference between a multi-GPU node and a multi-node cluster?
A multi-GPU node is one server with several GPUs (commonly 4 or 8) wired together over an on-board interconnect like NVLink or Infinity Fabric. A multi-node cluster connects several of those servers over InfiniBand so a job can span hundreds or thousands of GPUs. Aquanode's marketplace covers multi-GPU nodes up to 8 GPUs; a multi-node, InfiniBand-connected cluster is a different product, covered above.
Which interconnect does each GPU use?
H100 and H200 use NVLink 4 at 900 GB/s per GPU. B200 uses NVLink 5 at 1.8 TB/s per GPU, roughly double. AMD's MI300X uses Infinity Fabric instead of NVLink. All of these are on-board, GPU-to-GPU links inside one server; none of them is the InfiniBand fabric that connects separate servers together.
Does Aquanode manage Kubernetes or Slurm for me?
No. You get the node, the aq CLI, the API and live GPU metrics, and you bring your own orchestrator. That's a real difference from a dedicated cluster provider that layers managed Kubernetes or Slurm on top; if you need that layer managed for you, a dedicated cluster product is the better fit.
What happens to my data if I move to a different provider?
It depends on what you set up beforehand. A network volume stays with the provider and region where you created it; a portable workspace or a workspace backup follows you, syncing or restoring onto the new box on a different provider so you're not rebuilding a training environment from scratch every time supply or price shifts.
How is per-GPU pricing calculated on a shared node?
Every multi-GPU offer is normalized to a real per-GPU hourly rate before it's published on the marketplace, refreshed hourly, and billed per second once the node is running, rather than a node price you have to split unevenly yourself.
Ready to rent?
Compare live multi-GPU pricing across every provider Aquanode lists, or talk to us about a larger deployment.
