Compare · Hyperscaler
Amazon Web Services rents GPUs through its P- and G-series EC2 instances, from single T4s to 8× B200-class training nodes. GPU.ai takes the other approach: one account that prices 12+ clouds on every launch and passes the raw rate through at zero markup.
A fair read of both sides — their genuine strengths, the trade-offs, and what an aggregated market does differently.
Strengths
Trade-offs
The aggregator model
Live per-GPU floor prices across our aggregated fleet, next to AWS's public on-demand list rates where they sell the same model.
| GPU | VRAM | GPU.ai / GPU / hr | AWS list | Stock |
|---|---|---|---|---|
| RTX A4000 | 16 GB | $0.075 | Not sold | 45 available |
| RTX 5080 | 16 GB | $0.083 | Not sold | 32 available |
| RTX 3080 | 10 GB | $0.083 | Not sold | 19 available |
| V100 | 16 GB | $0.105 | Not sold | 7 available |
| RTX 3090 | 24 GB | $0.120 | Not sold | 65 available |
| RTX 4080 | 16 GB | $0.130 | Not sold | 8 available |
| RTX A4500 | 20 GB | $0.190 | Not sold | 8 available |
| RTX 4000 Ada | 20 GB | $0.200 | Not sold | 8 available |
| RTX A5000 | 24 GB | $0.233 | Not sold | 20 available |
| RTX 2000 Ada | 16 GB | $0.240 | Not sold | 8 available |
| RTX 4090 | 24 GB | $0.270 | Not sold | 80 available |
| A30 | 24 GB | $0.290 | Not sold | 3 available |
| RTX 5090 | 32 GB | $0.295 | Not sold | 88 available |
| L4 | 24 GB | $0.310 | Not sold | 35 available |
| RTX PRO 4500 | 32 GB | $0.340 | Not sold | 49 available |
| A40 | 48 GB | $0.440 | Not sold | 52 available |
| RTX A6000 | 48 GB | $0.450 | Not sold | 34 available |
| L40 | 48 GB | $0.575 | Not sold | 132 available |
| A100 40GB | 40 GB | $0.610 | Not sold | 26 available |
| RTX 6000 Ada | 48 GB | $0.650 | Not sold | 61 available |
| L40S | 48 GB | $0.740 | $1.52 | 232 available |
| A100 80GB | 80 GB | $0.880 | $5.12 | 94 available |
| RTX PRO 6000 | 96 GB | $1.69 | $4.14 | 183 available |
| H100 SXM | 80 GB | $1.74 | $3.93 | 45 available |
| H100 PCIe | 80 GB | $1.87 | Not sold | 8 available |
| H100 NVL | 94 GB | $2.59 | Not sold | 26 available |
| H200 SXM | 141 GB | $3.59 | $4.97 | 41 available |
| H200 NVL | 141 GB | $3.61 | Not sold | 10 available |
| B200 | 192 GB | $5.88 | Not sold | 12 available |
GPU.ai prices are live per-GPU floor rates from our aggregated fleet, normalized per GPU per hour, billed per second, with zero markup on the provider's raw rate. AWS figures are their public on-demand list rates for the closest equivalent SKU — never spot, reserved, or committed-use discounts. We only show models with launchable stock. Last updated 2026-08-17; this page refreshes hourly.
GPU.ai takes a different approach: instead of running one fleet, it prices the whole GPU market on every launch and routes you to the cheapest provider with live stock, at the provider's raw rate with zero markup. If AWS's strengths — deepest ecosystem of managed services around the gpu (s3, sagemaker, eks) — are not the thing you're paying for, an aggregated market usually wins on price and availability.
GPU.ai publishes its live per-GPU floor price for every model it carries — currently from $0.075/GPU/hr at the low end. Prices are normalized per GPU per hour and billed per second, and the comparison table on this page is refreshed hourly from live fleet data.
Launches are standard SSH instances with your keys, so most workflows move without changes: pick a GPU, launch, rsync your data, run. There are no commitments to unwind and no minimum spend — you can trial a single instance for a few cents before moving anything.
Honestly: your stack is deeply built on aws managed services, or if you need specific compliance regimes only a hyperscaler offers. This page's job is a fair comparison, not a hard sell — the live price table above is the argument that matters.