AMD GPU Cloud:
Rent MI300X on ROCm from $3.59/hr.
Spheron rents the AMD Instinct MI300X on-demand at $3.59/hr per GPU as of 20 Oct 2026: 192GB of HBM3, a ROCm image on Ubuntu, and 1x or 2x instances you launch yourself with no quota request. It sits in the same console as the NVIDIA catalog, so one account covers both vendors.
- HBM3 per GPU
- 192GB
- Ubuntu image
- ROCm
- Self-serve
- 1x / 2x
- Billing
- Per-Min
TL;DR: AMD GPU Cloud on Spheron
Spheron rents the AMD Instinct MI300X at $3.59/hr per GPU on-demand as of 20 Oct 2026, in 1x and 2x instances.
Each MI300X has 192GB of HBM3 at 5.3 TB/s, enough for a 70B model in BF16 on one GPU, where an 80GB H100 needs two.
Instances boot Ubuntu with ROCm. PyTorch, vLLM, SGLang and Triton run; TensorRT-LLM and NVIDIA NIM do not.
AMD Developer Cloud billed MI300X at $1.99/hr on 17 Sep 2026, with free credits limited to testing. Azure and Oracle sell only 8-GPU nodes, from $48/hr.
AMD GPUs You Can Rent Today
One AMD model is self-serve today. Here is what you can launch, what goes through the team, and what is not on offer.
| GPU | Memory | How you get it | Spheron rate |
|---|---|---|---|
| Instinct MI300X, 1x and 2x | 192GB HBM3 per GPU | Self-serve VM, Michigan, US | $3.59/hr per GPU on-demand |
| Instinct MI300X, 8-GPU nodes | 1.5TB HBM3 per node | Reserved capacity on request | Quoted |
| Instinct MI325X, MI355X | 256GB, 288GB HBM3E | Not offered today | n/a |
The ROCm Software Stack
Every MI300X instance boots one fixed image: Ubuntu with ROCm. Check the release with cat /opt/rocm/.info/version, then run your framework in a container that matches it. These run on ROCm today:
PyTorch
vLLM
SGLang
Triton
What still needs CUDA
TensorRT-LLM and TensorRT engines
FlashAttention 3's CUDA kernels
NVIDIA NIM containers
Hand-written CUDA kernels, until ported with HIPIFY
Moving From CUDA to ROCm
Start with a stack audit, then run your real job on a 1x instance. In our own Llama 3.1 8B LoRA run, the port took one container swap and one batch-size change.
ROCm vs CUDA in 2026
Train an LLM on AMD GPUs
MI300X vs H200
H100 vs MI300X fine-tuning
Triton on ROCm vs CUDA
MI300X and MI355X pricing
Spheron vs AMD Developer Cloud vs Hyperscalers
Per-GPU on-demand MI300X rates. AMD Developer Cloud is cheaper per hour on paper; the difference is what you are allowed to run and how small an order can be.
| Provider | MI300X per GPU | Smallest order | Notes |
|---|---|---|---|
| Spheron | $3.59/hr | 1 GPU, self-serve | Per-minute after a 20-minute minimum (60 on 2x). As of 20 Oct 2026 |
| AMD Developer Cloud (run by DigitalOcean) | $1.99/hr | 1 GPU | Billed rate on 17 Sep 2026. One-time free credits for testing and development only, no production workloads |
| Crusoe | $3.45/hr | 8-GPU node | On-demand. Checked 16 Aug 2026 |
| Azure ND MI300X v5 | $6.00/hr | 8-GPU VM ($48/hr) | Cheapest regions, quota request first. Checked 16 Aug 2026 |
| Oracle BM.GPU.MI300X.8 | $6.00/hr | 8-GPU bare metal ($48/hr) | On-demand bare metal. Checked 16 Aug 2026 |
Pricing fluctuates based on GPU availability. The Spheron rate is as of 20 Oct 2026; other providers reflect their published rates on the dates shown and may have changed. Check current GPU pricing for live rates, and the Azure MI300X pricing breakdown for that rate card by region.
AMD GPU Cloud FAQ
What you can rent, what it costs, and what runs on ROCm.
The AMD Instinct MI300X, with 192GB of HBM3 per GPU. You launch it yourself in 1x or 2x instances from the same Spheron console as the NVIDIA catalog, and larger MI300X configurations, including 8-GPU nodes, are arranged as reserved capacity on request. MI325X and MI355X are not offered on Spheron today.
An AMD MI300X costs $3.59/hr per GPU on-demand on Spheron as of 20 Oct 2026, and a 2x instance bills both GPUs at that rate. Billing is per minute after a minimum runtime of 20 minutes on 1x and 60 minutes on 2x, with no contract. No MI300X spot tier is listed right now.
See live GPU pricingYes, for anything past evaluation. AMD Developer Cloud runs on DigitalOcean and billed the MI300X at $1.99/hr for one GPU in a 17 Sep 2026 deployment. Its free credits are one-time, per account, and restricted to testing and development, with production workloads excluded. Spheron has no workload restriction, bills per minute after the minimum runtime, and puts the MI300X in the same account as H100, H200 and B200, so you can run one job on both vendors and compare.
Most PyTorch code will, without changes, because ROCm builds of PyTorch expose AMD GPUs through the torch.cuda API. vLLM, SGLang and Triton also run on ROCm. TensorRT-LLM, FlashAttention 3's CUDA kernels, NVIDIA NIM and hand-written CUDA kernels do not, until you port the kernels with AMD's HIPIFY tools. Test your exact stack on a 1x instance before moving a workload.
Read ROCm vs CUDA in 2026$48 an hour at the lowest published rates. Azure ND MI300X v5 (in its cheapest regions) and Oracle BM.GPU.MI300X.8 both published $6.00 per GPU-hour on-demand on 16 Aug 2026, and both sell only full 8-GPU nodes. On Spheron the smallest order is one MI300X at $3.59/hr.
Rent MI300X from $3.59/hr
192GB of HBM3 on ROCm, in the same account as your NVIDIA GPUs. Launch a 1x instance and test your stack today.
