← All demos
MLOps & Cost
GPU Memory & Cost Calculator
Size the VRAM any LLM needs, and find the point where self-hosting beats an API.
Production proof: The interactive version of my most-read LinkedIn post (5,100+ impressions) and the build-vs-buy analysis behind a 5.4× cost cut. Read the case study →
How it’s engineered
- VRAM is estimated with M = (P × 4 bytes / (32 / Q)) × 1.2: parameters × bytes-per-param at the chosen quantization, plus ~20% overhead for activations and KV cache.
- Quantization is the biggest lever: dropping from 16-bit to 4-bit cuts memory ~4×, which is what lets a 70B model fit on a single 48GB GPU.
- Build-vs-buy is an architecture decision, not a vibe. The break-even compares amortized GPU cost against per-token API pricing at your real monthly volume.
- This is exactly the reasoning that justified self-hosting at ~5.4× lower cost than an external API at scale in my production work.