The short answer: a per-GPU-hour rate on an Indian provider's pricing page tells you what a share of a GPU costs, not what a job spanning multiple GPUs actually gets for that price. Whether the GPUs sit on one physical board with a fast link between them, and what happens once a job needs more GPUs than one board holds, are separate questions a headline ₹/hour figure doesn't answer by itself.
What "HGX" in a plan name is signalling
HGX is Nvidia's reference design for an eight-GPU server board with NVLink and NVSwitch interconnect built in, so a provider advertising an "H100 HGX" plan is telling you the underlying hardware supports fast GPU-to-GPU communication inside that box — useful for model-parallel or tensor-parallel training that needs GPUs to exchange data constantly. What the plan name alone doesn't tell you is whether the quoted rate buys the whole eight-GPU board, a dedicated fraction of it, or a slice shared with other tenants whose jobs are also pulling on the same interconnect. That distinction changes the effective throughput a multi-GPU job gets for the same sticker price, and it isn't visible from the model name or the headline number.
A concrete example from this site's own tracked pricing
In the pricing this site tracks, AceCloud's H100 listing is sourced from a page named for the HGX board specifically, and it's billed monthly only — there's no metered hourly option. The table on this site shows that plan as roughly ₹247/hour purely as a monthly total divided by 730 hours for comparability, not a real per-hour quote, which is flagged in this site's own data notes. That's a useful example of the gap this post is about: the plan name signals HGX-class hardware, but the public price and the public spec don't say, in the same place, how many of the board's eight GPUs that monthly figure covers or what interconnect bandwidth is guaranteed rather than shared. That's not specific to one provider — check GPU count and interconnect scope directly with sales or the provisioning console before assuming an HGX-labelled rate matches what a specific multi-GPU job needs. See the live H100 pricing page and full comparison table for current listings across providers.
What changes once a job outgrows one node
NVLink and NVSwitch only cover GPUs on the same physical board. A job too large for eight GPUs needs multiple servers to exchange data over a separate networking layer — typically InfiniBand or high-speed RoCE Ethernet — and that layer is a large part of why a self-hosted multi-GPU cluster costs far more than the GPUs alone. Independent cost breakdowns of owned eight-GPU H100 clusters in India put the server hardware itself at roughly ₹2.5–5 crore, with InfiniBand-class networking adding a further ₹50 lakh or more on top, before power and cooling for that density are even factored in. None of the India-focused providers in this site's tracked dataset publish a per-node or per-cluster InfiniBand rate on their public pricing pages — where multi-node clustering is offered at all, it's quoted after a direct sales conversation, which means it doesn't appear in any ₹/GPU-hour comparison a public pricing table, including this one, can make today.
What to actually check before comparing quotes for a multi-GPU job
- Ask whether the quoted rate is per single GPU or covers an entire node, and how many GPUs that node holds.
- Ask what interconnect bandwidth is guaranteed for that price — shared with other tenants on the same board, or dedicated to your job.
- Ask explicitly whether multi-node clustering is offered at all, and get its cost in writing; it's usually outside the standard pricing page entirely.
- If the workload is single-GPU — true for most fine-tuning and inference jobs — none of this applies, and the standard per-GPU-hour checks in this site's five-things-the-sticker-price-hides post are sufficient on their own.
None of this makes multi-GPU cloud rental a bad option relative to owning hardware — avoiding a ₹2.5–5 crore-plus capital outlay and the networking build-out that goes with it is a real advantage of renting, not just a marketing line. It just means the comparison that matters for a distributed-training job isn't the same one that works for a single-GPU inference workload, and the difference doesn't show up in a single ₹/GPU-hour column.