The AI stack, layer by layer

AI runs on electricity: every answer comes from inference on chips that turn power into heat, so a data centre needs a grid connection, cooling and often water. In 2025, data centres worldwide used about 485 TWh, around 1.5% of global electricity, and the IEA expects about 950 TWh by 2030. Each request gets more efficient while total use keeps rising, a rebound effect. In many regions, waits of 5 to 10 years for grid connections now set the pace of AI expansion.

Tap an underlined term for a short explanation.

AI calculates with huge tables of numbers. A GPU or another AI chip does thousands of these calculations at once, for training as well as inference. Memory matters just as much: HBM feeds the data fast enough. The most advanced chips are made by a few foundries, above all TSMC in Taiwan, from raw materials that often come from a handful of countries.

Tap an underlined term for a short explanation.

A data centre turns electricity into computing: thousands of GPUs with liquid cooling and a fast interconnect work as a single machine. Global capacity stands at about 103 to 161 GW in 2026, depending on what is counted. Most companies rent this capacity through cloud computing from hyperscalers and neoclouds, or use a model through an API. That choice sets cost, data residency and lock-in.

Tap an underlined term for a short explanation.

Models learn from training data counted in tokens: a filtered web crawl, books, code and, increasingly, synthetic data. Open models disclose 15 to 36 trillion tokens for pre-training, and careful filtering with deduplication clearly improves results. In post-training, people rate answers through RLHF, data labelling paid anywhere from cents per task to $50 an hour depending on the country. Courts, licences and signals such as robots.txt decide what may be used.

Tap an underlined term for a short explanation.

A model is a huge set of learned numbers, the parameters, arranged mostly as a transformer that reads and writes tokens. It learns in pretraining by predicting the next token and is shaped in post-training with human feedback and checkable tasks; reasoning models also spend extra compute while answering. Closed models are used via API, while open-weight models can be downloaded and trail the closed frontier by about four months. Hallucinations and contaminated benchmarks mean you choose with your own evals.

Tap an underlined term for a short explanation.

An AI agent is a language model that calls tools in a loop until a goal is reached. This layer makes agents manageable inside a company: orchestration, connections through MCP and A2A, evals, human approvals and, for every agent, an identity with limited permissions. Prompt injection remains unsolved, so the architecture has to limit the possible damage. According to McKinsey, 40% of large companies already scale agents; in a Gartner survey, only 13% of companies consider their governance for them adequate.

Tap an underlined term for a short explanation.

An application is the product people use; the model is the engine inside. Around the model, good applications build a system prompt, context, retrieval of own data, tools, guardrails and evals. In mid-2026, 18.8% of the world's working-age population used generative AI, and official statistics in the EU and the US count about one in five firms. The EU AI Act gives providers and deployers their own duties.

Tap an underlined term for a short explanation.