---
title: "GPU: A40 (GPU3) price reduction - 30% off from June 2026 "
date: 2026-05-26
tags: ["GPU"]
url: https://changelog.exoscale.com/gpu-a40-gpu3-price-reduction-30-off-from-june-2026/
---
# GPU: A40 (GPU3) price reduction - 30% off from June 2026 

 All [A40 (GPU3)](https://www.exoscale.com/gpu/gpu3/) instances in Frankfurt will be **30% cheaper starting June 1st, 2026**. Existing instances continue to run unchanged and automatically benefit from the new pricing.
 
| Size   | GPUs | Current EUR/CHF/USD/hr | New EUR/CHF/USD/hr |
|--------|------|----------------|------------|
| Small  | 1    | 1.4933         | 1.0453     |
| Medium | 2    | 2.9867         | 2.0907     |
| Large  | 4    | 5.9733         | 4.1813     |
| Huge   | 8    | 11.9467        | 8.3627     |
 
The A40 has 48 GB of GDDR6 memory with Ampere RT Cores, 336 Tensor Cores, and 10572 CUDA Cores into a single card, making it a capable choice across a wide range of GPU workloads: AI model fine-tuning, batch inference, large-context LLM serving, 3D rendering, visual effects, AR/VR simulations, and scientific computing that needs GPU memory beyond the 24 GB range.

It also runs as the GPU backend for [Dedicated Inference](https://www.exoscale.com/ai-cloud-infrastructure/dedicated-inference/), which lets you deploy any Hugging Face model as a production-ready, OpenAI-compatible API endpoint on fully isolated European infrastructure. With the A40 you can serve models in the 13B - 34B parameter range comfortably, with no shared resources and no data leaving your environment.

At the new price point, A40 is a strong option whether you are running GPU instances directly or building on top of Dedicated Inference.

[Learn more about A40 (GPU3) instances](https://www.exoscale.com/gpu/gpu3/)