Cloud, Data Centres and Compute Infrastructure

Chips do nothing until someone puts them in a building with power, cooling and a network, and rents or uses the result. That layer, the clouds and data centres, is where most of the AI industry's capital goes. The four largest US platform companies (Microsoft, Alphabet Google, Amazon AWS and Meta) spent roughly $230 billion on capital expenditure in 2024 and around $400 billion in 2025, and guided to roughly $700 billion or more for 2026 (company guidance, approximate; definitions differ). Amazon alone expected ~$220 billion of cash capex in 2026 and reported negative trailing free cash flow of ~$42 billion in October 2026 (company). Around the hyperscalers grew a second tier of neoclouds, GPU-specialist clouds such as CoreWeave, Nebius, Lambda and Crusoe, funded heavily with debt secured on GPUs and on contracts with a few customers; a third tier of colocation landlords such as Equinix; a fourth of infrastructure financiers; and a fifth of governments buying sovereign compute. The business has turned software companies into utilities: capital-intensive, power-constrained and dependent on long contracts. A compute business is therefore an answer to five questions: who is the end customer and how creditworthy is it, how long is the contract relative to the life of the chips, what does a unit of capacity cost to build and power, how fully is it used, and how is it financed? This section maps the layer, explains hyperscaler capex and cloud revenue, the physical data centre, neocloud economics, GPU pricing and utilisation, the depreciation debate, financing structures, colocation, power, sovereign compute and margins, then closes with cases, pitfalls and how practitioners think. A worked example models a stylized hyperscaler's capex, depreciation and free cash flow against its AI revenue, and a checklist sets out diligence questions for data-centre and neocloud businesses.