2026-10-17 · AI in production
Ten GPU hours a month: why we did not build a GPU cluster
Our own AI pipeline for vehicle photos worked as a prototype. Next question: on which hardware, if the data must stay in the EU and the rest of the application runs on Cloudflare.
In AI projects the infrastructure decision is often made by feel: “images, so GPU, so cluster”. I did the maths first.
GPU hours first
Only the frames of the 360° rotation need the heavy segmentation, not every shot. With the measured time per frame that comes to about eight to ten GPU hours a month. Even the peak hour loads a single mid-range card to about ten percent. So the question was not “how many GPUs” but “where do I keep one machine warm, and where may the data live”.
Six options, three gone at once
Two AI models independently assessed the same six options: dedicated GPU server, serverless GPU, CPU only, an office Mac, a GPU pool in managed Kubernetes, Cloudflare.
CPU only was out: 40 seconds per frame means a queue of up to an hour at peak. Kubernetes with GPUs is the most expensive at such a small volume. Cloudflare does not run your own models on GPUs and stays what it is good at: storage, queues, tunnels.
The decision
A single dedicated GPU server for about 200 euros a month, with the pilot running on it directly, alongside the existing service and without affecting the running export. Serverless GPU in an EU region for backup and peaks — with a provider whose control plane is in the US, only links to files go there, never the photos themselves.
At small inference volumes the price is set not by compute but by what it costs to keep a machine ready and where the data may live. So measure first, then pick the provider.