Self-hosted LLM engineering and economics — production numbers, not benchmarks.https://romavarnava.com/en-ushttps://romavarnava.com/blog/sample-first-hardware/https://romavarnava.com/blog/sample-first-hardware/What a product assistant actually needs — one GPU, a 2GB embedder, an OpenAI-compatible API — and what you should not buy.Tue, 14 Jul 2026 00:00:00 GMT<blockquote> <p>SAMPLE POST — replace body with the real article before launch.</p> </blockquote> <p>Models with reliable tool calling already fit on a single card.</p> <h2>What you actually need</h2> <ul> <li>One workstation-class GPU</li> <li>A 2GB embedder</li> <li>An OpenAI-compatible API</li> </ul> <pre><code>def capacity(input_tokens: int, output_tokens: int) -&gt; float: return input_tokens / (input_tokens + output_tokens) </code></pre> <h2>Limits of applicability</h2> <p>All numbers here are N=1: one card, one production, one stack.</p>