A100 vs RTX-5090-32GB
RTX-5090-32GB is the cheaper of the two: $0.25/hr at Salad, against $1.09/hr for A100 at Civo — 4.4× the price.
Cheapest on each
Aligned facts
Same sheet, same order, both cards — including the gaps.
| Attribute | A100 | RTX-5090-32GB |
|---|---|---|
| VRAM | ○ no answer | 32 GB |
| Bandwidth | ○ no answer | 1.79 TB/s |
| FP16 | ○ no answer | ○ no answer |
| FP8 | ○ no answer | ○ no answer |
| TDP | ○ no answer | 575 W |
| NVLink | ○ no answer | PCIe 5 (no NVLink) |
| Architecture | Ampere | Blackwell (GB202) |
| Released | ○ no answer | 2025 |
A blank is not a zero. Where one card states a figure and the other states none, neither wins the row — the second number is unknown, and ranking an unknown as nothing is how a comparison starts to lie.
Every seller, both cards
One row per seller, cheapest tier, on-demand list prices only. Sellers of both cards lead, cheapest combined rate first.
| Seller | A100 | RTX-5090-32GB |
|---|---|---|
| Spheron | $1.432026-10-26 | $0.842026-10-26 |
| Runcrate | $1.762026-10-19 | $1.152026-10-26 |
| Civo | $1.092026-10-27 | ○ no answer |
| Hyperstack | $1.352026-10-26 | ○ no answer |
| Massed Compute | $1.352026-10-26 | ○ no answer |
| Sesterce | $1.492026-10-07 | ○ no answer |
| Koyeb | $1.602026-10-26 | ○ no answer |
| Cerebrium | $2.002026-10-27 | ○ no answer |
| Modal | $2.102026-10-26 | ○ no answer |
| Lyceum | $2.292026-10-26 | ○ no answer |
| CoreWeave | $2.702026-10-26 | ○ no answer |
| Replicate | $5.042026-10-26 | ○ no answer |
| Salad | ○ no answer | $0.252026-10-26 |
| HyperAI | ○ no answer | $0.352026-10-26 |
| Theta EdgeCloud | ○ no answer | $0.492026-10-26 |
| GPUhub | ○ no answer | $0.522026-10-26 |
| Shadeform | ○ no answer | $0.652026-10-26 |
| Nova Cloud | ○ no answer | $0.662026-10-26 |
| Beam | ○ no answer | $0.722026-10-27 |
| RunPod | ○ no answer | $1.192026-10-27 |
What this page will not tell you
On-demand list rates only · spot and committed terms priced separately and not mixed in here · no performance-per-dollar figure appears on this page, because it needs a throughput benchmark we do not hold and dividing a price by a vendor’s own TFLOPS claim produces a number that looks authoritative and means nothing