Every AI conversation runs in a data centre, and data centres run on power. The money is available. The wires, substations, transformers and generators take years to build.
Spending is doubling
The four largest US cloud companies spent about $166 billion on property and equipment in April to June 2026, about 1.9 times the same quarter a year earlier. Amazon spent $54.2 billion, Alphabet $44.9 billion, Microsoft $35.8 billion and Meta $31.1 billion. Each company defines this differently (Microsoft's figure excludes finance leases, Meta's includes them), so compare with care.
The limit is the grid
The International Energy Agency puts data-centre electricity use at about 415 TWh in 2024, around 1.5% of world electricity, and projects about 945 TWh by 2030.
US data centres used 4.4% of US electricity in 2023 and could reach 6.7% to 12% by 2028 (Lawrence Berkeley National Laboratory).
Data centres used 23% of Ireland's metered electricity in 2025, up from 5% in 2015.
Great Britain's grid operator logged about 92.9 GW of demand requests in March 2026, 50.8 GW of it for data centres. Only 32% of data-centre projects had secured a customer, so queues overstate what will be built.
The PJM capacity auction for 2028/29 in the eastern US cleared at its price cap and still fell 6,831 MW short of its reliability target.
Each answer is getting leaner
Google reports that its median text prompt used 0.24 Wh in 2025, 33 times less energy than a year earlier. That is a company's own figure for one type of prompt. Total demand still rises because volume grows faster.
Six questions to ask your AI supplier
Where does it run? Which region processes prompts, transcripts and recordings, and where does the model itself run?
Is capacity reserved for you? What happens in a surge, and does the failover region stay inside your data-residency boundary?
What holds the price? Ask for a price-hold period or a published rate card, with notice before any change.
What is the energy per conversation? Ask for energy and carbon per 1,000 conversations, with the method stated.
What powers it? Which grid, how much is matched with renewables or nuclear, and what data you get for your own reporting.
Is the model right-sized? Smaller models handle most routine conversations for a fraction of the energy and cost.
Read the full report
Our industry report The power behind AI: data centres, the grid and the cost of every conversation has the chart, the myths and every source.
View the full report on our Resources page →
This article summarises published research and company filings as of 23 October 2026 and is for general information only. Alphabet, Amazon and Microsoft had not reported Q3 2026 results when it was written.
