3 providers · per GPU-hour
AMD MI300X cloud pricing
3 offers from 3 providers, 2 with a published on-demand price. Per GPU-hour, US$. Prices checked 16 Sept 2026.
- Lowest on-demand
- $2.89/GPU-hr
- Lowest spot
- $1.85/GPU-hr
- Memory
- 192 GB HBM3
Spec sheet: MI300X ↗
| # | Provider | Offer | On-demand | Spot | Committed | As published (original currency) · source | Actions |
|---|---|---|---|---|---|---|---|
| 1 | Cyfuture | MI300X x1 1× MI300X · 16 vCPU · 256 GB RAM | $2.89/GPU-hr | — | — | ||
| 2 | Crusoe | MI300X x1 1× MI300X | $3.45/GPU-hr | — | — | ||
| 3 | Vultr | MI300X x8 8× MI300X | Contact for pricing | $1.85/GPU-hr | — |
What is the MI300X used for?
The MI300X is a data-centre GPU positioned by its maker for both training and inference, built on the CDNA3 architecture and released in 2023. Multi-GPU servers link cards over Infinity Fabric: 8 links, 128 GB/s peak per link, which matters when a job is split across several GPUs.
How large a model fits on one MI300X?
With 192 GB of memory, one card holds the weights of a model up to about 96 billion parameters at 16-bit precision, or about 192 billion at 8-bit. That is weights only: inference also needs room for the KV cache, and training needs several times more for gradients and optimiser state, so larger jobs are split across multiple GPUs.
| Spec (as published) | MI300X |
|---|---|
| Architecture | CDNA3 |
| Form factor | OAM |
| Memory | 192 GB HBM3 |
| Memory bandwidth | 5,300 GB/s |
| Interconnect | Infinity Fabric: 8 links, 128 GB/s peak per link |
| Max power (TDP) | 750 W |
| Precisions listed | FP64, FP32, TF32, FP16, BF16, FP8, INT8 |
| Partitioning | SR-IOV |
| Released | 2023 |
| Source | maker spec sheet ↗ |
Need MI300X capacity for a longer run?
Committed and cluster pricing is often negotiated. Tell us the GPU count, duration and region.