Society
The environmental cost of digital technology
The sector-level numbers are real and rising, but the per-activity figures that circulate most widely are modelled, and the most famous one was a unit-conversion error.
Digital technology's measured environmental cost sits in two places: the electricity running data centres and networks, and the hardware manufactured and thrown away. Data centres consumed about 415 TWh of electricity in 2024, roughly 1.5% of global consumption, growing around 12% a year since 2017 (International Energy Agency, 2025). That much is real. Almost everything else here is weaker than it looks: per-activity figures for an email, a streamed film or a chatbot query are modelled from industry averages rather than metered, and the best-known one was wrong by roughly two orders of magnitude. Virtually every headline number covers operational electricity only and excludes the emissions of manufacturing the equipment. For a workplace-oriented example of technology measurement in practice, see this reference.
Data centre electricity is measured badly and moving fast
The most reliable sector figures come from the IEA, built from partial operator disclosures and shipment modelling because no hyperscaler publishes site-level consumption. The revision between two consecutive editions is the most useful thing in the data.
Data centres used 485 TWh of electricity in 2025, about 2% of global electricity demand, a 17% rise on the previous year.
Direction: Increase. Strength of evidence: Strong.
Caveat A one-year revision this large means any figure older than about 18 months should be treated as superseded; the boundary for crypto mining and enterprise server rooms also shifts between studies.
National figures are firmer where a government has funded the work. The United States has the most detailed public accounting, and its authors are candid about what they cannot see.
US data centres used 176 TWh in 2023, 4.4% of total US electricity, projected to reach 325-580 TWh by 2028, or 6.7% to 12.0% of US electricity.
Direction: Increase. Strength of evidence: Strong.
Caveat The authors state that "the lack of direct energy data" limits the analysis; the 2028 range turns on unknowable GPU shipment volumes, utilisation rates and cooling choices.
Forward-looking figures are scenarios. The IEA projects consumption roughly doubling to around 950 TWh by 2030, about 3% of global demand, and its 2035 cases span 700 to 1,700 TWh around a Base Case near 1,200 TWh (IEA, 2025 and IEA, 2026). That 2.4-fold spread is the state of knowledge; most coverage discards it for the middle number. Anyone reading a forecast about what infrastructure will be built should treat the range as the finding.
The whole-sector share is a range, not a number
"The internet is X% of global emissions" is the most repeated claim in this subject and the least stable. The defensible peer-reviewed estimate is a range whose authors argue it is too low.
Peer-reviewed estimates put the whole ICT sector at 1.8-2.8% of global greenhouse gas emissions, rising to 2.1-3.9% once fuller supply-chain and lifecycle accounting is included.
Direction: Increase. Strength of evidence: Strong.
Caveat The authors argue existing estimates understate the true figure by up to 25% because scope-3 emissions are poorly captured, so even this range is a judgement about missing data, and it predates the AI build-out.
The commonly cited 3.7% figure is not a competing measurement. It traces to The Shift Project's 2019 Lean ICT modelling — the same exercise that produced the streaming claim below.
The Netflix figure was a unit-conversion error
This is the clearest documented correction in the subject, and the original still circulates.
The widely reported claim that watching 30 minutes of Netflix emits 1.6 kg of CO2 was roughly 90 times too high; the IEA's estimate is about 0.018 kg.
Direction: Decrease. Strength of evidence: Strong.
Caveat The error was traced to a unit conversion — 3 Mbps read as 3 MBps, inflating the assumed bitrate eightfold — and The Shift Project revised its figure down by a factor of eight in June 2020, but the original remains in wide circulation.
An eightfold bitrate error compounded with other over-assumptions produced a result about 90 times the corrected value: the difference between "streaming a film is like a long drive" and "streaming a film is negligible next to the drive to the shop". It has been corrected by an intergovernmental agency and revised by its originator, and it circulates anyway, because it was more quotable than its replacement. That is why checking where a figure came from matters more here than in most subjects.
The rest of the per-activity family is no better. Per-query AI energy and water figures are modelled from industry-average cooling efficiency rather than metered data: the best-known water estimate, putting the freshwater directly evaporated in training GPT-3 at around 700,000 litres, is a preprint built on public averages whose authors note that AI water use is "kept a secret" (Li, Yang, Islam and Ren, 2023).
E-waste is the part that is counted, and it is going the wrong way
Hardware is where measurement is strongest, because waste is weighed rather than modelled.
The world generated 62 million tonnes of e-waste in 2022, projected to reach 82 million tonnes by 2030, with only 22.3% documented as formally collected and recycled.
Direction: Increase. Strength of evidence: Strong.
Caveat The 22.3% is documented collection only; informal-sector recovery in low-income countries is real but unmeasured, so the reported rate says as much about record-keeping as about practice.
Generation is rising roughly five times faster than documented recycling, at about 2.6 million tonnes a year, and the documented rate is projected to fall to 20% by 2030 (UNITAR/ITU, 2024). Regions with mandatory targets do better and still miss them: the EU collected 11.6 kg per inhabitant in 2023 against 30.2 kg placed on the market, a rate of 37.5% against a 65% target that only Bulgaria, Latvia and Slovakia met (Eurostat, 2023 reference year).
Since 20 June 2025, EU ecodesign rules have required smartphone and tablet makers to supply spare parts for seven years after end of sales, five years of operating system upgrades, and batteries retaining 80% capacity after 800 charge cycles (European Commission, Regulation (EU) 2023/1670). That is a legal requirement, not an outcome: whether it lengthens the life of a phone, or of the cheap connected devices that fill a house, is unmeasured.
The short version
- Data centres used about 485 TWh of electricity in 2025, roughly 2% of global demand, 17% up in a year (IEA, 2026).
- The peer-reviewed range for the whole ICT sector is 1.8-3.9% of global greenhouse gas emissions, and its authors think the true figure is higher (Freitag and colleagues, 2021).
- The claim that half an hour of Netflix emits 1.6 kg of CO2 was about 90 times too high, the result of a bits-versus-bytes conversion error, and has been formally revised.
- E-waste reached 62 million tonnes in 2022 with 22.3% documented as recycled, and is growing about five times faster than documented recycling (UNITAR/ITU, 2024).
- The limits are severe: no operator publishes site-level data, 2035 projections span 700 to 1,700 TWh, and manufacturing emissions sit outside almost every headline figure.