Start with the number that will travel. "Up to 239 percent" is the ceiling of a range across four unnamed AI and cloud providers [3], not the group average, and the published summary does not say which company hit it or how small its 2020 base was. Percentage growth off a low base is loud as a ratio and quiet in tonnes. What survives that caveat is direction: over the same four years the telecom cohort in the same report moved the other way [4], and the ITU attributes the divergence at the AI and cloud firms to surging energy needs and infrastructure expansion [17].
The two sector totals are less comparable than they look. Operational emissions of 301 million tonnes of CO2 equivalent are put at 0.8 percent of global energy-related emissions [5], while the electricity draw is about 1.7 percent of world consumption [6], a ratio of roughly 2.1 [5]. Part of that gap is renewable purchasing. Part of it is the denominator, since energy-related emissions include fuel burned in transport, buildings and industry, which makes it the larger base.
They also do not share a reporting perimeter. Amazon reported no electricity data for 2024 [7], so the sector figure is a floor rather than a measurement. Divide 301 million tonnes by 500 TWh and you get about 0.6 tonnes of CO2e per megawatt-hour [2], a figure that resembles a grid intensity and should not be read as one: the numerator includes fuel burned on site, some purchased-electricity reporting is market-based, and at least one very large consumer is absent from the denominator.
Concentration is the sturdier structural finding here. The ten largest consumers account for 269 TWh [9], roughly 54 percent of the sector's electricity [1], led by China Mobile at 63 TWh, with Alphabet and Samsung at 32 TWh each and Microsoft at 30 [8].
The thing this does not tell you is what the buildout costs upstream. The totals are operational, and WBA executive director Gerbrand Haverkamp said companies "need to engage suppliers and address emissions across the products and services they rely on" [19]. The scoring also cannot rank what it cannot see: 18 companies have no score at all [15], and the single perfect score went to Swisscom on a scale built from targets, data and performance [14], which rewards disclosure quality alongside tonnage. Among the companies that have set near-term targets, 66 are not yet assessed as on track [3], and only 81 have comprehensive plans for meeting their goals [12].
On this evidence I will commit to the narrow claim, that reported emissions rose at the firms doing the building while a cohort with flatter demand cut theirs, and I would not go further, because the dataset does not separate datacenter load from the rest of each company's business, which is the split that would attribute the rise to AI rather than to ordinary growth.
