# Nistro AI Compute

The latest NVIDIA and AMD GPUs with cost-effective pricing. Nistro AI finds the best-priced capacity across clouds and regions, so you get the GPUs your training and inference workloads need, reserved for two or three years.

## Terms

Choose the term that fits your workload:

- **2 years** and **3 years**: reserve capacity for the long run.

## GPUs

Each GPU has its own page on [nistro.ai/compute](https://nistro.ai/compute) with its specifications and the workloads it suits best.

| GPU | Architecture | Memory |
| --- | --- | --- |
| [NVIDIA B300](https://nistro.ai/compute/b300) | Blackwell Ultra | 288 GB HBM3e |
| [NVIDIA B200](https://nistro.ai/compute/b200) | Blackwell | 180 GB HBM3e |
| [NVIDIA HGX H200](https://nistro.ai/compute/hgx-h200) | Hopper | 141 GB HBM3e |
| [NVIDIA H100 SXM](https://nistro.ai/compute/h100-sxm) | Hopper | 80 GB HBM3 |
| [NVIDIA A100 80GB](https://nistro.ai/compute/a100-80gb) | Ampere | 80 GB HBM2e |
| [AMD Instinct MI355X](https://nistro.ai/compute/mi355x) | CDNA 4 | 288 GB HBM3E |
| [AMD Instinct MI325X](https://nistro.ai/compute/mi325x) | CDNA 3 | 256 GB HBM3E |
| [AMD Instinct MI300X](https://nistro.ai/compute/mi300x) | CDNA 3 | 192 GB HBM3 |
| [NVIDIA RTX PRO 6000](https://nistro.ai/compute/rtx-pro-6000) | Blackwell | 96 GB GDDR7 |
| [NVIDIA RTX 5090](https://nistro.ai/compute/rtx-5090) | Blackwell | 32 GB GDDR7 |
| [NVIDIA RTX 4090](https://nistro.ai/compute/rtx-4090) | Ada Lovelace | 24 GB GDDR6X |
| [NVIDIA RTX 3090](https://nistro.ai/compute/rtx-3090) | Ampere | 24 GB GDDR6X |

## Get started

Get started on [nistro.ai/compute](https://nistro.ai/compute).

Want models served for you instead of raw GPUs? [Nistro AI Inference](https://nistro.ai/docs/inference) deploys any Hugging Face model behind an OpenAI-compatible endpoint.