The Digital Energy Footprint: A Field Full of Wrong Numbers
Digital energy use is a case where the instinct to quantify ran ahead of the ability to measure. The resulting numbers spread widely, were corrected quietly, and the originals are still in circulation. Establishing what is actually true here is unusually useful, because so much of the public material is wrong by one or two orders of magnitude.
The Number That Was Wrong
In 2020 a widely covered analysis estimated that half an hour of video streaming produced about 1.6 kilograms of carbon dioxide - comparable to driving several kilometres. It was reported everywhere and is still quoted.
The estimate contained errors that its authors subsequently identified and published a correction for, revising the figure down by roughly a factor of ninety. The main problems were using an outdated energy-per-gigabyte coefficient from a period when networks were far less efficient, and treating network energy as proportional to data transferred.
That second assumption is the more instructive error and it recurs constantly. Network equipment - routers, base stations, transmission gear - draws close to its full power whether or not traffic is passing through it. Energy per gigabyte is therefore not a meaningful coefficient: it falls as traffic rises, because the same fixed consumption is divided among more data. Multiplying a per-gigabyte figure by a quantity of data produces a number that looks precise and means very little.
The episode is worth recounting not to dismiss digital energy but because the corrected figure is genuinely useful. Streaming an hour of video in Europe is on the order of a few tens of grams of carbon dioxide - real, small, and about a hundredth of what many people believe.
Where the Energy Actually Is
For streaming, the dominant term is usually the screen. A large television consumes a hundred watts or more; a phone screen consumes a fraction of a watt. Watching the same programme on a television rather than a phone can change the total energy by more than an order of magnitude, which means the device chosen matters far more than the resolution or the length.
Data centres are the part that receives the attention and, for streaming specifically, they are efficient. Video delivery is heavily optimised, cached close to users, and the energy per stream is small. Data centres as a whole consume on the order of one to a few percent of world electricity, a number that has grown much more slowly than data volumes because efficiency improved in step - the subject of its own page.
Networks sit between and are the least well measured. Fixed-line networks are efficient per unit of data; mobile networks are substantially less so, and older generations more than newer ones. Watching over mobile data rather than fixed broadband is one of the few user choices that changes the network term significantly.
None of these is large compared with the things covered elsewhere in this section. An hour of streaming is a small fraction of what the same household spends on heating water in a day. Keeping the proportions straight is the point: digital energy is real, growing, and not where a household's consumption is decided.
The Device Is the Footprint
For a phone, a laptop or a tablet, manufacturing typically accounts for the large majority of lifetime energy - commonly cited as three quarters or more for a smartphone. Using the device is the small part; making it is the large one.
That inverts the intuitive advice completely. Charging habits are close to irrelevant, and how long the device is kept is close to decisive. Extending a phone from two years to four roughly halves its annualised embodied energy, which no amount of careful charging approaches.
The obstacles to keeping devices longer are mostly designed rather than physical: batteries that are difficult to replace, software support that ends before the hardware does, and repair costs set close to replacement prices. Right-to-repair legislation and minimum software support periods address exactly this, and they are energy policy even though they are rarely described as such.
The same logic applies to the network equipment in a home. A router, a set-top box and a smart speaker each draw a few watts continuously, and continuously is the operative word: a device drawing five watts all year uses more electricity than many things used intensively but briefly. This is the standby problem in a new form, and it is currently much less regulated.
What Is Actually Growing
While streaming was being blamed for an energy problem it did not have, a real one was developing elsewhere. Training large AI models consumes substantial energy, and serving queries from them consumes more per query than a conventional web search by a wide margin.
The aggregate is what matters and it is rising quickly. Data centre electricity demand, which had been roughly flat for a decade despite exploding data volumes, has turned upward in several regions, and grid operators in Ireland, Virginia and parts of the Nordics now treat it as a primary driver of load growth. This is a different situation from the streaming claims: the growth is measurable, the drivers are identifiable, and the projections disagree mostly about speed rather than direction.
There is a countervailing factor that deserves stating without being used as a dismissal. Efficiency per computation continues to improve, model architectures are getting cheaper to run, and specialised hardware is substantially more efficient than general-purpose processors. Whether efficiency or demand wins is the open question, and the historical pattern in computing has been that demand wins - which is a reason for caution rather than a prediction.
The practical conclusion for a reader is narrow and worth holding onto. Individual digital habits are not where a personal energy footprint is decided, and confident claims to the contrary have a poor track record. The infrastructure behind them is a genuine and growing load, and it is decided by siting, procurement and grid planning rather than by anything a user does. Those are different questions, and conflating them is what produced the wrong numbers in the first place.
Frequently asked questions
How much energy does streaming video actually use?
An hour of streaming in Europe is on the order of a few tens of grams of carbon dioxide. A widely reported 2020 estimate put half an hour at 1.6 kilograms; its own authors published a correction revising it down by roughly a factor of ninety, mainly for using an outdated energy-per-gigabyte figure.
Why is energy per gigabyte a misleading measure?
Because network equipment draws close to its full power whether or not traffic passes through it. Energy per gigabyte therefore falls as traffic rises, since the same fixed consumption is divided among more data. Multiplying such a coefficient by a data volume produces a number that looks precise and means very little.
What dominates the energy of watching a video?
The screen, usually. A large television draws a hundred watts or more while a phone screen draws a fraction of a watt, so watching the same programme on a television rather than a phone can change total energy by more than an order of magnitude - far more than resolution or duration.
What is the biggest part of a phone's energy footprint?
Manufacturing, commonly cited as three quarters or more of lifetime energy. That makes how long the device is kept close to decisive and charging habits close to irrelevant: extending a phone from two years to four roughly halves its annualised embodied energy.
Is AI a real energy problem or another exaggeration?
Real, unlike most of the streaming claims. Data centre electricity demand had been roughly flat for a decade despite exploding data volumes and has turned upward in several regions, with grid operators in Ireland, Virginia and parts of the Nordics treating it as a primary driver of load growth.
