Every serious argument in AI infrastructure is being conducted in a shared vocabulary that isn’t actually shared. Six of the terms that decisions turn on carry fundamentally different operational meanings in 2026 depending on who is using them — and the gap between what a buyer means and what a vendor means is where mispriced contracts and stalled projects are born. This is the piece that fixes the vocabulary before the money moves.

This is the decoder the rest of the collection is written against. It is entry-point agnostic by design: you do not have to read it first or read it linearly, but every analytical dossier in the collection cross-references it. It maps the four architectural traditions — HPC, cloud, data platforms and AI systems — that quietly collide inside every “AI infrastructure” conversation and are the reason two experts can use the same word and mean incompatible things.

The cost structure is the most misunderstood part of the field, and the misunderstanding drives bad capex. The instinct that “training is the cost” misprices multi-year infrastructure: training dominates Year 1, but inference overtakes it — and the decisions made on the wrong curve are expensive to unwind. We lay the curve out plainly.

Geography is not a footnote to the architecture — it is part of the architecture. The jurisdictional fault line produces distinct deployment layers, and every European deployment has to pass a sequence of four filters — energy, compliance, integration and procurement — before “global innovation” becomes “commercial adoption in Europe.” Naming those filters is what turns a vague sense of “it’s harder here” into a checklist you can act on.

This is a reference you keep open, not a report you read once. It is priced as a foundation, not a forecast: it makes no dated predictions, because its job is to make the predictions in the other pieces legible. If the analytical dossiers are the arguments, this is the grammar — and it is the natural entry point to the whole collection.