RTX 5090 vs A100 80GB
The RTX 5090 rents for 61% less per hour today — $0.69 against $1.79 — with 32 GB against 80 GB of memory.
Prices collected 2026-09-18 · 8 providers price both
Index median per GPU-hour, on-demand, per snapshot.
Side by side
| RTX 5090 | A100 80GB | |
|---|---|---|
| Median rental, per GPU-hour | $0.69 | $1.79 |
| Cheapest listing | $0.47 | $0.43 |
| Per month at 720 h | $497 | $1,289 |
| Providers pricing it | 9 | 37 |
| Memory | 32 GB GDDR7 | 80 GB HBM2e |
| Memory bandwidth | 1,792 GB/s | 2,039 GB/s |
| GB of memory per $/hr | 46.4 | 44.7 |
| Architecture | Blackwell | Ampere |
| Class | Consumer | Datacenter |
| Released | 2025-01 | 2020-11 |
Same provider, both GPUs
The cleanest comparison: 8 providers price both cards, so the difference is the GPU and nothing else. The RTX 5090 is cheaper at 8 of them.
| Provider | RTX 5090 | A100 80GB | A100 80GB premium |
|---|---|---|---|
 Vast.ai | $0.68 | $0.80 | +19% |
| $0.68 | $1.43 | +110% | |
| $0.68 | $1.71 | +151% | |
 RunPod Community Cloud | $0.69 | $1.29 | +87% |
 GPU.ai | $0.73 | $1.21 | +66% |
 io.net | $0.89 | $1.31 | +47% |
 Beam | $0.91 | $1.36 | +50% |
 RunPod Secure Cloud | $0.99 | $1.59 | +61% |
What each one runs, and what a million tokens costs
Weights + KV cache at 8K context + overhead against 90% of each card's memory. Where a model needs more than one card the cell says how many and prices the set at today's median. $/M tokens is the hourly median divided by estimated single-stream output, from memory bandwidth — an estimate for ranking, not a benchmark.
| Model | Bits | RTX 5090 | A100 80GB |
|---|---|---|---|
| Kimi K3 | 4-bit | 55× GPU$37.95/hr | 22× GPU$39.38/hr |
| Kimi K3 | 16-bit | 195× GPU$134.55/hr | 78× GPU$139.62/hr |
| DeepSeek V4.1-Flash | 4-bit | 11× GPU$7.59/hr | 5× GPU$8.95/hr |
| DeepSeek V4.1-Flash | 16-bit | 39× GPU$26.91/hr | 16× GPU$28.64/hr |
| Llama 4 Maverick | 4-bit | 8× GPU$5.52/hr | 4× GPU$7.16/hr |
| Llama 4 Maverick | 16-bit | 28× GPU$19.32/hr | 12× GPU$21.48/hr |
| gpt-oss-120b | 4-bit | 3× GPU$2.07/hr | ✓ 67.9 GB~569 tok/s · $0.87/M |
| gpt-oss-120b | 16-bit | 9× GPU$6.21/hr | 4× GPU$7.16/hr |
| Llama 3.1 70B | 4-bit | 2× GPU$1.38/hr | ✓ 44.5 GB~41 tok/s · $12.13/M |
| Llama 3.1 70B | 16-bit | 6× GPU$4.14/hr | 3× GPU$5.37/hr |
| Gemma 4 31B | 4-bit | ✓ 21.2 GB~82 tok/s · $2.34/M | ✓ 21.2 GB~94 tok/s · $5.29/M |
| Gemma 4 31B | 16-bit | 3× GPU$2.07/hr | ✓ 65.7 GB~26 tok/s · $19.12/M |
| Qwen3 Coder 30B-A3B | 4-bit | ✓ 19 GB~772 tok/s · $0.25/M | ✓ 19 GB~879 tok/s · $0.57/M |
| Qwen3 Coder 30B-A3B | 16-bit | 3× GPU$2.07/hr | ✓ 62.9 GB~247 tok/s · $2.01/M |
| Qwen3.8 27B | 4-bit | ✓ 18.7 GB~94 tok/s · $2.04/M | ✓ 18.7 GB~107 tok/s · $4.65/M |
| Qwen3.8 27B | 16-bit | 2× GPU$1.38/hr | ✓ 57.6 GB~30 tok/s · $16.57/M |
| gpt-oss-20b | 4-bit | ✓ 13.7 GB~708 tok/s · $0.27/M | ✓ 13.7 GB~806 tok/s · $0.62/M |
| gpt-oss-20b | 16-bit | 2× GPU$1.38/hr | ✓ 43.9 GB~227 tok/s · $2.19/M |
| Llama 3.1 8B | 4-bit | ✓ 6.9 GB~319 tok/s · $0.6/M | ✓ 6.9 GB~362 tok/s · $1.37/M |
| Llama 3.1 8B | 16-bit | ✓ 18.4 GB~90 tok/s · $2.13/M | ✓ 18.4 GB~102 tok/s · $4.87/M |
Buy or rent?
Hours of renting at today's median that add up to the card's approximate list price. List prices are typical quoted figures, not tracked, and do not include the server, power or the person to run it — so the break-even is a floor, not a verdict.
Questions
Is the RTX 5090 or the A100 80GB cheaper to rent?
On 2026-09-18 the RTX 5090 has a market median of $0.69 per GPU-hour across 9 providers and the A100 80GB $1.79 across 37, so the RTX 5090 is 61% cheaper per hour. Among the 8 providers that price both, the RTX 5090 is cheaper at 8 and the A100 80GB at 0. Per hour is only half the answer — the table below prices a whole model on each.
Which should I rent for a 70B model?
At 4-bit, Llama 3.1 70B needs about 44.5 GB. On the RTX 5090 that is 2 cards at $1.38/hr; on the A100 80GB one card at $1.79/hr.
What is the real difference between them?
32 GB GDDR7 at 1,792 GB/s (Blackwell) against 80 GB HBM2e at 2,039 GB/s (Ampere). Memory decides what fits; bandwidth decides how fast it generates, because each token streams the active weights once. Compute matters for prompt processing and training, not for how fast text appears.