Behind every search query, every streamed film, every email attachment and video call, sits a physical infrastructure of extraordinary scale: millions of servers drawing continuous power, cooling systems running around the clock, undersea cables spanning ocean floors, and device supply chains consuming rare materials across multiple continents. The internet is not a cloud. It is a very large collection of buildings, most of them hot, most of them energy-intensive, and almost none of them visible to the people whose daily lives depend on them.
Understanding the environmental footprint of digital technology requires confronting a figure that most people find genuinely surprising: the information and communications technology sector is currently responsible for approximately 2 to 4 percent of global greenhouse gas emissions. That range reflects genuine uncertainty in measurement methodologies rather than disagreement about the underlying reality. The upper bound puts it roughly on par with the aviation industry — an industry that attracts sustained environmental scrutiny and whose passengers are frequently reminded of their carbon footprint. Digital technology, somehow, largely escapes this framing.
The footprint breaks down across three distinct layers of infrastructure, each with its own emissions profile and its own trajectory.
Data centres
The physical heart of the internet: warehouses of servers processing and storing the world's digital activity. Data centres consume roughly 200 terawatt-hours of electricity annually — about 1% of global electricity demand. Cooling alone accounts for 40% of that consumption. The growth of AI workloads, which are dramatically more compute-intensive than conventional web tasks, is accelerating demand faster than efficiency improvements can offset it.
~45% of sector emissions
End-user devices
Smartphones, laptops, tablets, and televisions together account for the largest share of the sector's lifecycle emissions — primarily through manufacturing rather than use. The production of a single smartphone generates roughly 70kg of CO₂-equivalent, the majority of which is emitted before the device is ever switched on. Global smartphone shipments run at approximately 1.2 billion units per year. The emissions embedded in the manufacture of that volume of devices are substantial and largely invisible to the consumers who drive demand for them.
~50% of sector emissions
Network infrastructure
The cables, routers, base stations, and exchange points that carry data between devices and data centres. Mobile networks are significantly more energy-intensive per unit of data than fixed-line broadband — a gap that 5G is expected to narrow over time but has not yet closed at scale. The expansion of connectivity into underserved regions, while socially valuable, adds to the sector's overall energy demand.
~5% of sector emissions
Training a large language model consumes energy at a scale that has no good intuitive comparison. Estimates for training a single frontier AI model range from hundreds to thousands of megawatt-hours — equivalent to the annual electricity consumption of dozens to hundreds of average households. Inference — running the model for users after training — adds ongoing demand at a scale that scales with adoption.
The major AI developers are aware of this and have made commitments to renewable energy procurement. Whether those procurement commitments translate into genuine additionality — actual new renewable capacity added to grids, rather than certificates attached to existing generation — is a question the industry has not yet answered clearly. The distinction matters significantly for real-world emissions outcomes.
Forces reducing the footprint
Hardware efficiency has improved dramatically — modern data centre servers deliver vastly more compute per watt than their predecessors. Renewable energy procurement by major cloud providers has expanded significantly. Hyperscale data centres operate at power usage effectiveness ratios that were considered impossible a decade ago. Liquid cooling is replacing air cooling in the most energy-intensive facilities.
Forces expanding it
Global data volumes continue to grow exponentially. Video streaming now accounts for over 60% of internet traffic. AI inference is being embedded into more applications, adding persistent compute demand. Device replacement cycles, though lengthening slightly, still generate hundreds of millions of new devices annually. The expansion of connectivity to previously unserved populations adds genuine new demand.
The net direction of these pressures is genuinely uncertain, and honest analysts disagree about the trajectory. What is clear is that the assumptions underpinning earlier optimism — that efficiency improvements would keep pace with demand growth indefinitely — are being tested by the scale and intensity of AI adoption in ways that the models underpinning those assumptions did not anticipate.
The most significant gap in the current state of the debate is not technical — it is informational. Consumers have almost no visibility into the environmental cost of their digital behaviour. A flight booking confirms the associated carbon footprint. A streaming platform does not. A search query does not. An AI-generated image does not. The information asymmetry makes individual behaviour change essentially impossible, and makes corporate accountability difficult to enforce in the absence of mandatory disclosure requirements.
Several jurisdictions are moving toward mandatory scope 3 emissions reporting for large technology companies — a development that would, for the first time, create standardised, comparable data on the actual emissions associated with digital services. The EU's Corporate Sustainability Reporting Directive is the furthest advanced. Its implementation will test whether transparency alone is sufficient to drive meaningful change, or whether transparency is a necessary but insufficient condition for the kind of structural response the scale of the problem may ultimately require.
In the meantime, the infrastructure of the internet continues to grow, continues to consume, and continues to be invisible — which is, in environmental terms, exactly the problem.
What percentage of global carbon emissions does the internet produce?
The information and communications technology sector is currently responsible for approximately 2 to 4 percent of global greenhouse gas emissions, according to research published by institutions including the
International Energy Agency and peer-reviewed academic studies. The upper end of that range is roughly equivalent to the emissions of the global aviation industry. Measurement methodologies vary, which accounts for the range rather than disagreement about the underlying reality.
How much energy do data centres use?
Data centres consume roughly 200 terawatt-hours of electricity annually — approximately 1% of global electricity demand. Cooling systems account for around 40% of that consumption. The growth of AI workloads, which are significantly more energy-intensive per computation than conventional web tasks like search or streaming, is placing new pressure on data centre energy demand that efficiency improvements have not yet fully offset.
What is the carbon footprint of AI?
Training a single frontier AI model can consume hundreds to thousands of megawatt-hours of electricity — equivalent to the annual energy use of dozens to hundreds of households. Inference (running the model for users) adds ongoing demand that scales with adoption. Major AI developers have made renewable energy commitments, but the degree to which these represent genuine additionality — new renewable capacity rather than certificates attached to existing generation — varies and is not consistently disclosed.
What is being done to reduce the internet's carbon emissions?
Responses operate at several levels. Hardware efficiency improvements and renewable energy procurement by major cloud providers have helped contain the growth of data centre emissions. Regulatory frameworks including the EU's Corporate Sustainability Reporting Directive are moving toward mandatory emissions disclosure for technology companies. Critics argue that voluntary commitments and disclosure alone are insufficient given the scale of demand growth, and that structural regulatory intervention is likely required.
