Hyperscaler
The global hyperscale clouds — AWS, Azure, Google Cloud, Oracle. Deep managed capacity, enterprise contracts, and the widest service surface. On-demand GPU rates sit at the top of the market.
Reference Rates · US · 2026-07-18
Reference Rates
Guaranteed on-demand reference rate per chip across the three operator segments. Every cell is a published, citable series — panel depth, stat, and history disclosed. Committed term on /term; interruptible, regional, and every other grain on /explorer.
| Silicon | Hyperscaler $/hr | Neocloud $/hr | Marketplace $/hr |
|---|---|---|---|
B300 | — | — |
$7.39 CRI-T3-B300-ALL-GTD-OD-US |
B200 |
$14.12 CRI-T1-B200-ALL-GTD-OD-US |
$7.64 CRI-T2-B200-ALL-GTD-OD-US |
$7.07 CRI-T3-B200-ALL-GTD-OD-US |
H200 |
$8.96 CRI-T1-H200-ALL-GTD-OD-US |
$4.29 CRI-T2-H200-ALL-GTD-OD-US |
$4.21 CRI-T3-H200-ALL-GTD-OD-US |
H100 |
$9.15 CRI-T1-H100-ALL-GTD-OD-US |
$3.63 CRI-T2-H100-ALL-GTD-OD-US |
$2.92 CRI-T3-H100-ALL-GTD-OD-US |
A100 |
$2.74 CRI-T1-A100-40GB-ALL-GTD-OD-US |
$2.15 CRI-T2-A100-ALL-GTD-OD-US |
$1.36 CRI-T3-A100-ALL-GTD-OD-US |
Neocloud
The dedicated GPU specialists — CoreWeave, Lambda, Nebius and peers. Single-tenant clusters built for training and inference, priced below the hyperscalers.
Marketplace
Marketplace and shared-pool venues — Vast, RunPod and aggregators. Self-serve capacity that reprices continuously, where posted neocloud cards move rarely. Interruptible product is common on these venues; the reference ladder compares guaranteed rates only.
| Attribute | Hyperscaler | Neocloud | Marketplace |
|---|---|---|---|
| Footprint | Managed multi-service hyperscale cloud | Dedicated single-tenant GPU clusters | Shared-pool or marketplace supply |
| Networking | Hyperscale backbone, managed networking | Dedicated cluster fabric, often InfiniBand | Variable — Ethernet or undisclosed |
| Services | Full enterprise services, support, compliance | Focused GPU stack, lighter wrap | Self-serve, minimal support |
| Examples in this segment | e.g. AWS · Azure · Google Cloud · Oracle | e.g. CoreWeave · Lambda · Nebius | e.g. Vast · RunPod · marketplaces |
| Typical use | Enterprise & regulated workloads | Dedicated training & inference | Cost-sensitive batch, dev/test, spot |
Each segment describes a distinct class of operator — hyperscaler, neocloud, or marketplace. The segment is a property of the operator, not the chip: the same silicon is priced across all three, so the ladder compares one chip's hyperscale, neocloud, and marketplace rate side by side.