Compare prices

Good for: The largest training and serving jobs — frontier-class memory and bandwidth.

Not good for: Very costly for small jobs — best value only on big work.

Good for: Big training jobs and heavy serving — the strongest all-round card on this board.

Not good for: Frontier training money. If your job fits on anything cheaper, a B200 is budget burned for a spec sheet.

Good for: 70B+ serving and serious fine-tunes — H100 speed with far more memory.

Not good for: More card than most jobs need; the main trap is paying for it on small work.

Good for: High-throughput serving and training — FP8 and the fastest memory in the Hopper family.

Not good for: Pays off only when the card stays busy; on bursty work you pay H100 rent for idle silicon.

Good for: The same high-throughput Hopper serving and training, from providers that sell the HGX card without splitting variants.

Not good for: Pays off only when kept busy — avoid it for bursty development work.

Good for: Single-card work where NVLink would sit unused — same silicon, lower price.

Not good for: No NVLink, and lower power than the SXM — slower when cards must work together.

Good for: Hyperstack's own label for its NVLink-linked PCIe-form H100 — priced between its PCIe and SXM cards.

Not good for: One provider's label, published alongside its plainer “H100 PCIe” — check which card you are actually booking.

Good for: 94GB of HBM — the H100 for models that nearly fit on 80GB.

Not good for: Sold as linked pairs, not as a lone card.

Good for: The honest workhorse — proven, plentiful, and the safe pair of hands for older stacks.

Not good for: An older generation — no FP8, so the newer speed tricks for serving don't apply.

Good for: Entry-price A100 — fits 32B-class work in INT4 with honest KV-cache budgeting.

Not good for: The 40GB fills up quickly on larger models.

Good for: 7B–32B serving and image work — sub-$1/hr with datacenter manners. Our all-rounder.

Not good for: 70B at FP16 does not fit, and no flag fixes that. Weak at double-precision maths; no NVLink.

Good for: The L40S's slower sibling — the same 48GB for lighter serving and image work.

Not good for: Weak at double-precision maths — not the card for scientific computing. No NVLink.

Good for: Prototyping and image generation for almost nothing — the cheapest useful card on the board.

Not good for: 24GB fills up fast on big models. No NVLink, so two cards can't pool memory — and it is a gaming card at heart.

Good for: Fast 32GB experiments on a budget — 32B in INT4 fits.

Not good for: No NVLink, so two cards can't pool memory. A consumer card — driver and ECC behaviour matter for production serving.

Good for: 70B on ONE card — the most VRAM per dollar on this board. Our sleeper pick.

Not good for: No NVLink scale-out; strong all-round, but the Max-Q version (listed separately) runs slower.

Good for: Nearly the full card's 96GB-class memory at a new-marketplace price.

Not good for: This is the Max-Q version: it runs at lower power and is slower than the full card.

Good for: CPU and GPU on one coherent fabric — a specialist pick for workloads built for it.

Not good for: Its processor is ARM, not the usual x86 — some software needs changes to run on it.

Good for: 48GB workstation card at the lowest hourly rate on this board.

Not good for: An older generation — no FP8, so the newer speed tricks for serving don't apply.

Good for: Rack-scale systems for the largest training programmes.

Not good for: A whole 72-GPU rack — far more than ordinary hourly work needs, and not a card for small jobs.

Good for: Rack-scale systems for the largest training programmes.

Not good for: Quote only, and a whole rack — far more than ordinary hourly work needs.