Ascentis AI · Resources
Guides
Long-form reference material on how AI actually works and what it costs to run. Written to be read rather than skimmed, kept current, and free of registration walls.
Start with Understanding AI in 2026 if you want the whole picture. The comparisons below answer narrower questions and can be read on their own.
Reference · 34 sectionsUnderstanding AI in 2026How language models actually work, and when the answer is not a language model at all. Tokens, attention and context windows; the four layers around the model; the nine RAG patterns, embeddings, hybrid retrieval and rerankers; agents, evaluation and sovereignty. The longest thing on this site and the one to read first.
Build guideSelf-Hosted AI for UK SMEsA real RTX 5090 build at UK prices, with VRAM tables, measured throughput and honest cost-per-seat arithmetic against frontier APIs. The numbers are ours, not a vendor’s.
ComparisonSelf-Hosted AI vs Cloud AIConfidentiality, intellectual property, cost, compliance and vendor lock-in, compared for engineering and manufacturing businesses. Where each model wins, and where the honest answer is a mix.
ComparisonOff-the-Shelf AI vs Bespoke AIA practical build-versus-buy guide: where a subscription is entirely sufficient, and where it quietly stops being. Includes the cost crossover most vendors do not publish.