While the AMD Instinct MI300X was promoted as the "H100 killer" for its massive 192 GB of VRAM, its actual value depends on your project.
The following breakdown compares the current landscape of MI300X on-demand pricing. Niche clouds are aggressively undercutting hyperscale giants to win over developers running large-scale LLM inference.
Takeaways
- Azure and Vultr offer the lowest MI300X price below $2/hr.
- DigitalOcean also lists MI300X below $2/hr.
- Oracle and CoreWeave charge materially more than the lowest listed MI300X clouds.
- At $2.19/hr, Thunder Compute H100s remain a strong CUDA-first alternative.
Current on-demand MI300X prices (per GPU)
| Provider | SKU / Instance | On-Demand $/GPU-hr* | Notes |
|---|---|---|---|
| TensorWave | N/A | $1.71 | |
| Vultr | 8 x MI300X cluster | $1.85 | $14.80/hr node / 8 GPUs |
| Digital Ocean | N/A | $1.99 | |
| RunPod (Secure Cloud) | N/A | $2.19 | |
| Crusoe Cloud | N/A | $3.45 | |
| Cirrascale | N/A | $3.85 | $22,499/month (shortest commitment) |
| Oracle Cloud | BM.GPU.MI300X.8 | $6.00 | 8-GPU bare-metal ($48/hr) |
| Azure | ND MI300X v5 (8 x MI300X) | $6.00 | 8-GPU VM pricing, normalized per GPU |
| CoreWeave | 8 x MI300X | $6.31 | 8-GPU node pricing, normalized per GPU |
| * Normalized to a single MI300X, even when only multi-GPU nodes are offered. Last update: August 1, 2026 |
|||
Methodology
Why you can trust these numbers:
- On-demand only - no reservations or prepaid contracts.
- Same silicon - every figure refers to AMD MI300X (192 GB) SKUs.
- Public price lists - pulled in June 2026 from each provider's published pricing pages.
- U.S. regions, USD - prices elsewhere vary by 5-20 %.
Azure ND MI300X Instances
Through the ND_MI300X_v5-series, Azure offers hyper-scale virtual machines designed for high-end Deep Learning training and tightly coupled, scale-up and scale-out Generative AI or HPC workloads.
Powered by AMD Instinct MI300X accelerators and 4th Gen Intel Xeon Scalable (Sapphire Rapids) processors, these instances provide immense high-bandwidth memory (HBM) capacity, making it possible to fit larger LLMs into a single node with fewer GPUs.
| Spec | Standard_ND96isr_MI300X_v5 |
|---|---|
| Provider | Azure |
| MI300X GPUs | 8 |
| vCPUs | 96 |
| RAM | 1,850 GiB |
| Hourly price | $48.00 |
| Price per-GPU | $6.00 |
Why this matters for developers
- More memory, new math - MI300X packs 192 GB HBM3 and FP8 support, ideal for 70 B+ parameter models.
- Price gap vs. H100 - Vultr rents clusters of 8x MI300X at $14.80/hr, which means it's about 16% cheaper per GPU than Thunder Compute's H100 at $2.19/hr. But the best choice depends on software stack and workload fit.
- H100 still has strong ROI - Unless your model truly needs 192 GB or AMD's MI300X architecture, an H100 80 GB at $2.19/hr can still deliver better value for CUDA-first workflows.
- Bookmark this page - we refresh the numbers monthly so you don't have to.
Get the full picture in our AI GPU rental market trends overview.