Multi-GPU and Cluster Pricing: How 8x Node Rates Actually Work

Some of the best per-GPU prices on our index are only available if you rent eight GPUs at once. A provider advertises an H100 rate that undercuts everyone, and the fine print says the smallest thing you can actually rent is a full 8-GPU node, so the real hourly hold on your card is eight times the headline. Most comparison sites quote that per-GPU number without the disclosure. We normalize every price to a per-GPU rate too, because that is the only way to compare providers, but we badge every rate that carries a minimum node size. This guide explains how those numbers work and when they are a good deal.

Why providers sell GPUs in blocks of eight

The 8x block is not a pricing gimmick; it is how the hardware is built. Datacenter GPUs like the H100, H200, and B200 ship to providers on baseboards that carry eight GPUs wired together with NVLink, a direct GPU-to-GPU interconnect that is far faster than routing traffic through the host machine. The eight cards share a chassis, power delivery, and network fabric. Carving one GPU out of that node and selling it separately is possible (some providers do, through virtualization), but many simply rent the node as the unit it physically is. That is also why the 8x configuration is what serious training work wants: the whole point of the baseboard is that the GPUs can talk to each other at full speed.

Per-GPU price vs the total node cost

Every price on this site is normalized to dollars per GPU per hour, whatever the node size behind it. That makes an 8x node rate directly comparable to a single-card rate, but it does not make them the same purchase.

Take the cheapest on-demand H100 on our index today: $1.99/hr per GPU. If a rate like that is sold only as an 8-GPU node, your actual meter runs at eight times the per-GPU figure for every hour the node is up, whether your job uses one of the GPUs or all eight. The per-GPU rate tells you whether the price is competitive. The node minimum tells you the size of the check. Both matter, which is why we publish both.

The corollary: a node-only rate is only cheap if you can keep all eight GPUs busy. At 50 percent utilization of the node, your effective per-GPU cost doubles, and a pricier single-card rental would have beaten it.

What the min-GPU badges mean on our tables

On GPU pages, any row where the price requires a multi-GPU rental carries a badge like "8x min": that rate is per GPU, and the smallest rentable unit at that price is an 8-GPU node. On provider pages, the same disclosure appears next to the pricing kind as "8-GPU node". No badge means the default: you can rent a single GPU at the listed rate. The multi-GPU node glossary entry covers the term. When two providers show similar per-GPU rates and one carries the badge, they are different products: one is a card, the other is a cluster with a much larger minimum spend.

NVLink vs PCIe: what you are paying for

Inside a multi-GPU node, the interconnect is the difference between eight GPUs that act like one big accelerator and eight GPUs that happen to share a chassis. Our spec tables carry the interconnect for every card, from NVIDIA's published datasheets: the H100 SXM variant links GPUs at 900GB/s over NVLink, while the H100 PCIe variant communicates over PCIe Gen5, an order of magnitude less GPU-to-GPU bandwidth. For workloads that shuttle gradients or model shards between GPUs constantly, that gap shows up directly in training throughput. For single-GPU work it is irrelevant, which is why the PCIe cards price lower and are often the better single-card rental. The SXM vs PCIe glossary entry has the details.

Training vs inference: who actually needs the node

Multi-GPU training is the clear case for the 8x node. Distributed training synchronizes gradients across GPUs every step, so interconnect bandwidth directly gates throughput, and NVLink-connected nodes are built for exactly this. If you are training or fine-tuning a model too large for one card, the node is the product you want.

Inference usually is not. Most inference workloads scale by replication: many independent single-GPU workers behind a load balancer, not one tightly coupled 8-GPU system. Unless a single model's weights exceed one card's VRAM even after quantization, single-GPU rentals scale inference more cheaply and with less idle risk than a node.

A single big GPU often beats a slice of a node. If your model fits on one 80GB card, one H100 with no minimum commitment is simpler and usually cheaper than buying into node-sized capacity you cannot fill. Compare what the same budget buys on the H100 vs A100 page before assuming you need more than one card.

The cluster fine print

Node rentals amplify every hidden cost in the bill. A minimum commitment on an 8x node is eight times the commitment of the same term on one card, and cluster products in particular like to pair an attractive per-hour headline with weekly or monthly minimums. Storage and egress scale with the job size too: training runs that produce terabytes of checkpoints pay real money to move them out on providers that bill egress. The fees page covers what each provider charges beyond compute, and the rent-vs-buy calculator will tell you when a sustained multi-GPU bill crosses the line where owning hardware wins.

FAQ

Why is the per-GPU price sometimes lower on an 8x node than for a single GPU?

Selling a whole node is operationally simpler for the provider: one tenant, full utilization of the baseboard, no virtualization overhead. Some providers pass that back as a lower per-GPU rate. You are being paid, in effect, for committing to the full node.

Can I rent just one H100 instead of a node?

Usually yes. Many providers on our H100 page sell single GPUs with no minimum; the rows without a min-GPU badge are exactly those. The badged rates are node-only.

Does the min-GPU badge change how the table sorts?

No. Tables sort by per-GPU price, badged or not, and referral status never affects ordering. The badge is a disclosure, not a ranking penalty: you decide whether a node-only rate fits your workload.

Do I need NVLink for fine-tuning?

If the fine-tune fits on one GPU (as many LoRA-style jobs do), no: interconnect only matters when GPUs must exchange data mid-job. If you are sharding a model or its optimizer state across GPUs, NVLink-connected nodes will train meaningfully faster than PCIe-connected cards.