Enter a GPU type, a count and monthly usage to estimate what the cluster costs on AWS, Google Cloud and Azure at published on-demand list prices — and to see what those hourly rates leave out.
OMC has not published network rates and is not selling compute today. Enter a hypothetical rate below to see the arithmetic — the number is your own assumption, not a quotation, forecast or commitment by OMC.
USD per GPU-hour. Hyperscaler figures are 8-GPU node prices divided by eight; Lambda and CoreWeave are quoted per GPU or per node as noted.
| GPU | AWS | Google Cloud | Azure | Lambda | CoreWeave |
|---|---|---|---|---|---|
| H100 80GBSXM5 · node spec in notes | $6.88p5.48xlarge · $55.04 ÷ 8 | $11.06a3-highgpu-8g · $88.49 ÷ 8 | $12.29ND96isr H100 v5 · $98.32 ÷ 8 | $3.99per GPU, on demand | $6.16HGX H100 · $49.24 ÷ 8 |
| H200 141GBSXM · node spec in notes | $7.91p5en.48xlarge · $63.30 ÷ 8 | $10.60a3-ultragpu-8g · $84.81 ÷ 8 | $13.78ND96isr H200 v5 · $110.24 ÷ 8 | — | $6.31HGX H200 · $50.44 ÷ 8 |
Three cost mechanisms most price tables ignore, and which decide the real bill far more than the headline rate.
A p5.48xlarge bills $55.04 per hour whether its eight GPUs are training or doing nothing — roughly $4,000 over a single long weekend. On-demand pricing charges for wall-clock time, not for work delivered.
Wall-clock billingOutbound data is priced separately: about $0.09/GB on AWS, $0.08/GB on Azure and $0.08–0.12/GB on Google Cloud. For a mid-sized H100 deployment moving checkpoints and outputs, that commonly adds four figures a month.
Data transferThe capacity AI teams actually buy is often excluded from the discounts designed to save money: AWS Spot is outside Savings Plans, Capacity Blocks are outside both Savings Plans and Reserved Instances, and Google excludes A3/A4 from flexible committed-use discounts.
Commitment structureA comparison of billing mechanics, not prices. OMC has published no network rates; the items below are design commitments from the whitepaper and may change before mainnet.
| Dimension | Hyperscaler on-demand | OMC (designed) |
|---|---|---|
| Billing unit | Per hour, whole instance | Per job, with per-second micropayments on the L2 rollup |
| Idle time | Running hours are billed | Intended to settle only for accepted, verified work |
| Capacity access | Regional quota limits and availability gates | Permissionless, subject to supply on the network |
| Commitment | None required; discounts need 1–3 year commitments | No commitment required by design |
| Payment | Fiat, via vendor billing | OMC or supported stablecoins (USDT / USDC on BNB Chain) |
| Result verification | Provider trust | Tiered verification policy per job type (spot-check re-execution, redundancy, TEE) |
Because there is no published OMC network rate to quote. The mainnet is targeted for Q1 2027, and until rates are set and published, any figure we printed would be an invention. Instead the calculator lets you type your own hypothetical rate and shows you the arithmetic — clearly labelled as your assumption, not our price.
Hyperscalers sell GPU capacity inside fixed machine shapes. AWS p5.48xlarge, Google's a3-highgpu-8g and Azure's ND96isr H100 v5 each bundle eight H100 GPUs, so the honest comparable number is the node price divided by eight. Comparing a per-node price against a per-GPU price is the single most common way GPU price tables mislead — a headline "$49.24" can be more expensive per GPU than a "$3.99" once the unit is aligned.
No. They are compute list prices only. Storage, egress, load balancers, snapshots and any support tier sit on top, and for sustained deployments those lines are material — egress alone commonly adds four figures a month for a mid-sized training cluster.
Spot and committed rates are real and often far lower, but they are not like-for-like: spot capacity can be reclaimed, and committed-use discounts require one- to three-year commitments and often exclude the newest GPU families. Mixing them into one table is how "cheapest" claims get made. This page sticks to on-demand list prices, which are the only rates that are both published and directly comparable.
The rates were observed in September 2026 from vendor pricing pages. GPU pricing changes frequently, so we date every figure rather than presenting it as permanent. Treat this as a well-sourced starting point for your own modelling, and confirm live pricing with the vendor before you commit budget.
No. OMC is operating a public testnet. Mainnet — when the network can accept paying jobs and providers can earn — is targeted for Q1 2027. You can join the testnet now and claim the community airdrop, but there is no compute to purchase and no price to pay.
The methodology behind OMC's delivered-cost comparisons, and the full billing design, are documented in the whitepaper.