ComputeAssay

Head to head · updated 2026-09-10

H100-SXM vs H100-PCIe: price and deliverability

H100-PCIe currently rents at 0.7× the median price of H100-SXM ($2.89 vs $3.99 per GPU-hour on demand). Whether that premium is worth it depends on memory, fabric, and what you are actually running — not the sticker.

$/GPU-hour, on demand · ● A/B fabric disclosed · C · D · U undisclosed H100-PCIe · TensorDock $1.99 H100-PCIe · Vast.ai (l $2.40 H100-PCIe · Hyperstack $2.50 H100-PCIe · RunPod $2.89 H100-PCIe · OVHcloud $2.99 H100-SXM · Hyperstack $3.20 H100-SXM · DataCrunch $3.25 H100-PCIe · Lambda $3.29 H100-PCIe · Scaleway $3.34 H100-SXM · RunPod $3.49 H100-SXM · Scaleway $3.61 H100-SXM · Nebius $3.85 H100-SXM · Crusoe $3.90 H100-SXM · Lambda $3.99 H100-SXM · CoreWeave $6.16 H100-SXM · AWS $6.88 H100-SXM · Oracle Clou $10.00 H100-SXM · Google Clou $11.06 H100-SXM · Google Clou $11.62 H100-SXM · Microsoft A $12.29
H100-SXM vs H100-PCIe, on-demand, 2026-09-10 · computeassay.com
H100-SXMH100-PCIe
Lowest on-demand $/GPU-hr$3.20$1.99
Median on-demand$3.99$2.89
Sellers tracked127
Grade A/B offerings160
Max VRAM per GPU (tracked SKUs)80 GB80 GB

Same-name capacity is not one good: within each column, prices spread by multiples depending on fabric. Check the per-GPU pages for every listing with its published interconnect, and the substitutability rules for when one can genuinely replace the other.

These ratios move weekly

The Delivered Compute Report tracks the spread — free, every row sourced.

Subscribe free