Your AI server, part by part.
Pick each component: the machine assembles in 3D, and VRAM, power draw, speed and price are recalculated live.
What this machine can do
The models it loads, the expected generation speed, and when buying beats renting.
Loadable models
8k-token context, one user. Click a model for its page.
Performance
Buy or rent?
How much VRAM do you need?
Roughly the model’s parameters × the bytes per parameter of the chosen precision (2 in FP16, about 0.6 in Q4), plus the context cache, which grows with the number of tokens and simultaneous users.
Why bandwidth sets the speed
For one user, every generated token reads the whole model from memory. Speed therefore depends on memory bandwidth far more than on compute — and MoE models only read their active parameters.
Buy or rent the GPUs?
Rackprint compares the hourly price of cloud GPUs with buying the server plus electricity, to show after how many months owning the machine pays off.