On-demand, the H200 rents for $4.17/hr and the MI325X for $3.26/hr — the MI325X is ~28% cheaper per hour, averaged across cloud providers.
| H200 | MI325X | |
|---|---|---|
| On-demand /hr | $4.17 | $3.26 |
| Spot /hr | $2.81 | $— |
| Reserved /hr | $3.42 | $1.98 |
| Per month (on-demand) | $3,044 | $2,380 |
| H200 | MI325X | |
|---|---|---|
| Vendor | NVIDIA | AMD |
| Architecture | Hopper | CDNA 3 |
| Memory | 141 GB HBM3e | 256 GB HBM3e |
| Memory bandwidth | 4.8 TB/s | 6 TB/s |
| Launched | 2024 | 2025 |
On-demand, the H200 averages $4.17/hr vs the MI325X at $3.26/hr — the MI325X is about 28% cheaper per hour. Spot rates are $2.81/hr vs $—/hr.
H200 costs more per hour but offers 141 GB HBM3e memory and newer architecture — for large-model training the higher hourly rate often yields a lower cost-per-job. For memory-bound inference or budget runs, the MI325X is usually the better value.
At on-demand rates that's about $3,044/mo for the H200 and $2,380/mo for the MI325X (730 hours).
See full pricing & history: H200 price · MI325X price · all GPUs.
Cross-provider averages from public cloud GPU listings. Informational only.