Reference edition checked November 5, 2026 (ET). Reporting feeds update independently; reference facts retain their stated dates.
AI infrastructure turns computing equipment into a usable service. It includes data-center buildings, electrical connections, cooling, networks, storage and cloud operations. A chip purchase is only one part of that system. This page follows facility and energy evidence; AI hardware covers processors and semiconductor access, while AI economics explains financing and spending measures.
A facility has several milestones
A company can announce a site before obtaining permits, signing an interconnection agreement or installing servers. Construction, energization and commercial operation are separate milestones. Announced capacity is therefore not the same as electricity consumed or customer computing delivered. Readers should look for the facility’s location, project phase, counterparties and the date each milestone was confirmed.
The IEA’s Energy and AI report describes servers, storage, networks, cooling and backup equipment as parts of the data-center system. It also distinguishes different types of facilities. Some serve many conventional digital workloads as well as AI. Labeling every megawatt at a data center “AI electricity” would overstate what the evidence establishes.
Separate electricity measures and forecasts
Power capacity is measured in units such as megawatts; energy consumption measures power over time. An announced electrical connection is not an annual electricity-use figure. A projection should retain its reference year, scenario and assumptions rather than being described as a guaranteed outcome. The IEA explicitly uses scenarios to address uncertainty around adoption, efficiency and energy bottlenecks.
A report on grid pressure should identify which utility or system is affected and whether it concerns connection capacity, peak demand, annual energy or reliability. Cooling needs and local constraints can differ between sites. This page does not infer a project’s water use, emissions or reliability from a company’s headline computing-capacity claim.
Read cloud spending and availability carefully
A cloud service may advertise a model before offering it in every region or customer tier. Reserved capacity, available instances and physical equipment owned by the provider are different measures. Contractual access to compute is not necessarily ownership of the underlying hardware. Model and service availability should be checked in the relevant provider documentation.
Corporate capital expenditure can include assets outside AI and can use different accounting conventions. It should not be presented as a single standardized global AI-spending series without adjustments and a methodology. Existing TBR coverage below follows specific projects and company reports, while this guide sets out the distinctions needed to interpret those developments. The initial page is not a live map of data centers or a forecast endorsed by TBR.
Questions to ask before drawing a conclusion
- What project phase has actually been confirmed?
- Does the capacity describe a connection, installed servers or customer-accessible compute?
- Are power and annual energy being confused?
- Is the forecast labeled with its scenario and reference year?
- Do spending figures include leases, non-AI assets or previously announced commitments?
What the next facility record needs
An infrastructure register should preserve the site, owner or operator, utility connection, project stage and evidence date. Planned capacity, energized capacity and measured consumption belong in separate fields. An update should explain a change in scope or timing rather than treating all earlier announcements as still additive. Public planning and utility records can be important corroboration; a company rendering or general cloud-region announcement alone does not establish a completed facility.
Related TBR reporting
- OpenAI Offers to Cut Up to 1,000 Megawatts at Georgia Campus When the Grid Strains 2026-11-02
- Nvidia Weighs Guaranteeing Debt for Nscale’s Anthropic Data Center 2026-10-31
- Microsoft Pulls Out of Nscale’s Loughton AI Data Centre 2026-10-18
- Amazon Earnings: AWS Grows 39% but AI Spending Pushes Free Cash Flow to -$42.4 Billion 2026-10-30
From the central AI news archive
OpenAI Ends NDAs With Officials After Voters Reject AI Data Centers
Voters Rejected Data Centers in 37 of 38 Called Ballot Measures on Election Day
Supermicro Earnings: Sales Nearly Quadruple to $19.9 Billion, but Margins Thin
OpenAI Offers to Cut Up to 1,000 Megawatts at Georgia Campus When the Grid Strains