GPU hosting: when you need graphics cards in a server, and when you do not

“GPU” has become the word you attach to any quote to make it look modern. In reality, a server with a graphics card solves a narrow category of problems — very well — and is waste for everything else.

What a GPU actually does

An ordinary processor has a few dozen powerful cores, good at tasks that must happen one after another, with a decision at every step. A graphics card has thousands of far simpler cores, good at doing the same calculation on thousands of pieces of data at once.

The difference is like a head chef who can cook any dish versus a brigade of a thousand people who only know how to peel potatoes. For a full menu you want the chef. For a tonne of potatoes, the brigade.

Workloads that genuinely need a GPU

• Running artificial intelligence models. A language or image-recognition model performs millions of matrix multiplications — exactly what a graphics card is built for.
• Training your own models. This is where the gap is widest: what takes weeks on a processor can take hours on suitable cards.
• Video processing at scale. Transcoding, resizing, automatic subtitling across many files.
• 3D rendering. Product configurators, architectural visualisations, images generated on demand.
• Real-time image analysis. Counting objects, reading number plates, detecting events in a video stream.

Workloads that do NOT need one, whatever anyone says

• A brochure site or an online shop. However large, its work is reading from a database and sending pages.
• A management application, an ERP, a CRM. These are database applications; memory and fast disks help, graphics cards do not.
• A slow website. If it is slow, the cause is almost certainly a badly written query or a missing cache. A GPU does not fix that.
• Sending emails, generating reports, running exports. Input-output work, not computation.

The question that ends the discussion quickly

“Does my application do the same calculation across thousands of items in parallel?” If the answer does not arrive immediately with a concrete example, the answer is no.

If you genuinely need one: what matters when choosing

Card memory, before speed

With AI models, the first limit is not speed, it is memory. A model that does not fit in the card's memory simply will not run, however fast that card is. Check first how much the thing you want to run actually needs.

Rented or owned

The practical rule: if you need a GPU for a few hours a day or in bursts, renting makes more sense. If it runs continuously for months, owning the hardware becomes reasonable — and your data never leaves your own infrastructure, which matters if you work with personal or medical data.

What happens between runs

The most common waste: a GPU server left on around the clock for a task that runs two hours a day. If you go the rented route, check whether it can start and stop automatically.

The option few people consider

For many companies the right answer is neither a GPU server nor an ordinary processor, but an external service that runs the model and answers through an API. You administer nothing and pay only for what you use.

Your own infrastructure earns its place when the data is not allowed to leave the company, when volume is high and constant, or when you need a model trained on your own data.

If you want to know which category you fall into, write to us with what you want to run. It usually becomes clear in a single conversation.

← Back to the blog