When does buying an on-premise AI server beat cloud APIs? Hardware tiers and payback

The author argues that most companies should start with a cloud API, since for short queries it stays cheaper than a GPU workstation for roughly a year and a half. An on-site machine becomes worthwhile when models must process large volumes of internal documents per request or when data cannot leave the company, with a single-GPU system potentially paying back in about four months. Hardware sizing depends mainly on GPU memory, since models must fit entirely in VRAM, with a rule of thumb of about 0.5 GB per billion quantized parameters plus headroom. The article compares hardware tiers and Russian market prices as of October 2026, from PCs with RTX 5090 cards for smaller models to Mac Studio systems for the largest open models, alongside GPU rental and cloud API options. A laptop benchmark shows that larger models generate text much more slowly and that prompt processing suffers sharply when memory is tight, so buyers should plan for 1.5 to 2 times more memory than the model requires.