This record on the Oxford University Research Archive (ORA) holds Mytton and Ashtine’s 2022 study, which audits where the data center energy numbers researchers and journalists cite actually come from. Read it for that audit rather than for a verdict on any one figure: it reviews data center energy estimates generally, drawing on 46 publications from 2007 to 2021, and so predates today’s AI-driven buildout.

The authors trace 258 energy estimates back to the 676 sources behind them. Only 31% were peer-reviewed; 38% were non-peer-reviewed reports, and many carried no clear methodology or data provenance. The review also flags heavy reliance on private data from IDC (43% of sources) and Cisco (30%), plus practical tracing problems: 11% of sources had broken web links, and 10% were cited with too little detail to locate at all.

That makes it a useful reference for checking how reliable a data center energy claim actually is, including ones you'll meet in AI-era coverage elsewhere, since the sourcing problems it documents don't vanish just because the topic moved on. The trade-off: this is an audit of source quality, not a fresh calculation, and its own figures predate today's AI-driven demand growth.

ORA lists the article as published and peer reviewed, and the full text is open access there as two deposited PDFs: the main paper and a supplementary file.