The internet does not feel like it has a carbon footprint. It is invisible. It does not produce exhaust fumes. You cannot smell it or see it. But behind every search, every video, every email, every page load, there is a vast physical infrastructure of servers, cables, cooling systems, and transmitters — all drawing electricity, most of it still generated from fossil fuels.
How much? The numbers, when you actually look at them, are significant. The internet as a whole now accounts for somewhere between 2% and 4% of global greenhouse gas emissions — a range that puts it in the same territory as commercial aviation, a sector nobody considers trivial in climate terms. And unlike air travel, most people have no mental model of digital carbon at all. They have never been asked to think about it.
This article is the data. All of it sourced, placed in context, and translated into something that connects to the real world rather than sitting as abstract terawatt-hours.
Key Statistics at a Glance
- The internet consumes an estimated 416 TWh of electricity per year (IEA, 2023)
- ICT accounts for 2–4% of global CO₂ emissions — roughly comparable to aviation
- Data centres use approximately 1% of global electricity, stable since 2010 despite exponential traffic growth (IEA 2024)
- The average web page weighs 2.5 MB and produces about 0.5g CO₂ per pageview
- Streaming video represents 60%+ of global internet traffic
- A standard email produces ~4g CO₂; one with a large attachment: ~50g
- A single Google search emits roughly 0.2g CO₂
- There are now 5.4 billion internet users globally (2024)
- Training GPT-3 released an estimated 552 tonnes of CO₂ — equivalent to 120 petrol cars driving for a year
- Cryptocurrency mining consumed roughly 110 TWh in 2023 — more than most countries' entire electricity supply
The Overall Picture: How Big Is the Internet's Carbon Footprint?
Let's start with the number that frames everything else. The IEA estimated that the internet — meaning all data centres, network infrastructure, and end-user devices considered together — consumed approximately 416 terawatt-hours of electricity in 2023. For reference, the United Kingdom's entire national electricity consumption runs around 300 TWh per year. The internet uses more electricity than the UK every twelve months.
Convert that electricity to CO₂ — accounting for the global average grid carbon intensity of approximately 494 grams per kilowatt-hour — and you get something in the range of 200 million tonnes of CO₂ equivalent annually from digital infrastructure. That is before you factor in the manufacturing of devices, which adds substantially more.
The frequently cited 2–4% of global emissions figure comes from several independent research groups, including researchers at Lancaster University, the Shift Project, and various IEA analyses. The range exists because the boundary conditions vary — some studies include only data centres and networks, others include device manufacturing, others include the full ICT sector including telecommunications hardware. The honest answer is that nobody knows the exact number to a decimal place, but the order of magnitude is not in serious dispute.
What makes the number striking is the comparison to aviation. Commercial aviation generates about 2.5% of global CO₂ emissions (around 915 million tonnes in 2023). Aviation is the sector politicians take international action on, the one that attracts carbon offset schemes, the one that individuals agonise over when deciding whether to take a transatlantic flight. The internet, in aggregate, emits as much or more — and receives a fraction of the scrutiny.
Data Centres: Less Catastrophic Than Expected, For Now
One of the genuinely surprising statistics in recent IEA reporting is that data centre electricity consumption has remained roughly stable at around 1% of global electricity use since 2010 — despite internet traffic increasing roughly tenfold over the same period. How is that possible?
Efficiency gains, mostly. Hyperscale cloud providers like Google, Amazon, and Microsoft have invested massively in data centre efficiency, achieving Power Usage Effectiveness (PUE) ratings around 1.1 compared to 2.0 or worse in older facilities. The shift from on-premise servers to consolidated cloud infrastructure means the same traffic gets served with far less hardware. And hardware itself — particularly server CPUs, memory, and storage — has become more energy-efficient per unit of compute delivered.
That said, the plateau is fragile. AI workloads are reshaping data centre demand in ways that efficiency gains may not offset. Training large language models and running inference at scale requires sustained high-utilisation GPU clusters that draw far more power per rack than traditional server workloads. Goldman Sachs estimated in 2024 that AI could drive data centre electricity consumption to 3–4% of global electricity by 2030 — a tripling of the sector's share. The 1% stability may be a historical footnote rather than a permanent feature.
The geographic distribution matters too. Data centres in Iceland (near-zero-carbon geothermal grid) have a different actual carbon impact than equivalent facilities in China or Poland (high-carbon coal-dependent grids). The IEA data combines them all into an aggregate figure that obscures these differences. A data centre running on Norwegian hydropower genuinely does not contribute meaningfully to CO₂ — but the global average includes a lot of facilities that do.
Websites: The Part You Can Actually Control
The average web page, according to HTTPArchive's 2024 Web Almanac, weighs approximately 2.5 MB in total transferred data on desktop. Apply the Sustainable Web Design Model v4 (the formula behind Carbon Badge and most credible carbon tools), and a 2.5 MB page on a non-green host produces roughly 0.24g of CO₂ per pageview.
The real-world average cited widely — including by WebsiteCarbon.com — runs closer to 0.5g per pageview, because popular high-traffic sites tend to run heavier than the median. JavaScript frameworks, unoptimised product photography, multiple analytics and advertising pixels: these accumulate quickly. A site at 0.5g CO₂ with 100,000 monthly pageviews emits about 50 kg of CO₂ per year — roughly the equivalent of driving a petrol car 200 kilometres, every year, from people just loading your pages.
Scale matters enormously here. A major news site with 50 million monthly pageviews at even 0.3g CO₂ per pageview emits 15,000 kg (15 tonnes) of CO₂ every month. That is 180 tonnes per year — the average annual footprint of 18 European citizens from all activities combined.
The good news: websites are one of the few digital carbon sources where individual decisions have a measurable, immediate, and compounding effect. The introduction to website carbon footprints explains what drives individual site emissions, and the full reduction guide covers the practical techniques that can cut page emissions by 50–80%. The sustainability audit guide walks through the complete audit process if you want a systematic approach.
For measurement, the Carbon Badge scanner applies SWDM v4 to your URL and returns a letter grade in seconds. No account required.
Email: The Carbon Cost of Your Inbox
Mike Berners-Lee — whose book How Bad Are Bananas? remains one of the most cited references for everyday carbon calculations — estimated the carbon footprint of email at approximately 4 grams of CO₂ for a standard email and up to 50 grams for an email with a large attachment. A brief, one-line reply costs roughly 0.3g.
These figures account for the energy used by the sender's device, the email servers processing and storing the message, and the recipient's device reading it. The attachment figure is high because large files must be stored on multiple servers (often redundantly) for potentially years — every copy of a 10 MB PDF sitting in thousands of recipients' inboxes is consuming storage energy indefinitely.
Put those numbers against real-world email volume: the world sends approximately 330 billion emails per day in 2024. Even if most are spam that gets filtered before storage, the email infrastructure as a whole contributes meaningfully to global digital carbon. One estimate suggests email generates around 50 million tonnes of CO₂ equivalent per year globally — comparable to the aviation emissions of a mid-size country.
The practical implication for individuals is modest but real: unsubscribing from newsletters you do not read, avoiding unnecessary reply-all chains, and questioning whether that file needs to be attached (or could be linked from a cloud document instead) are all genuine, if small, carbon reductions. For large organisations with tens of thousands of employees each sending dozens of emails daily, the aggregate adds up to something worth measuring.
Search: Google's Footprint Per Query
Google has been relatively transparent about its search carbon figures compared to other technology companies. The company's own estimates put a single search query at approximately 0.2 grams of CO₂. Google processes around 8.5 billion searches per day, which means the global search traffic from Google alone generates roughly 1,700 tonnes of CO₂ every single day — over 600,000 tonnes per year.
That 0.2g figure has been relatively stable over years, largely because Google has been running its data centres on matched renewable energy for many years and continues investing in efficiency. Other search engines operating on less efficient infrastructure or less green energy would produce higher per-query figures.
The advent of AI-powered search — where queries trigger large model inference rather than a simple index lookup — is changing this picture significantly. AI search responses require substantially more compute than traditional search results. Early estimates suggest AI search may cost 10× the energy of a traditional search query. If AI search becomes the dominant mode (as Microsoft, Google, and others are actively pursuing), the energy profile of the global search market could shift substantially upward.
Streaming Video: 60% of the Traffic, Concentrated Carbon
Video streaming — Netflix, YouTube, TikTok, Disney+, Twitch — accounts for more than 60% of all internet traffic globally, according to Sandvine's annual Global Internet Phenomena report. It is by far the largest single category of internet data transfer.
The Shift Project generated significant attention in 2019 with estimates suggesting video streaming accounted for up to 300 million tonnes of CO₂ per year. Those estimates were later revised downward — significantly — by other researchers, including George Kamiya and others at IEA, who pointed out methodological issues in the original analysis. The revised estimates put video streaming at something more like 100 million tonnes per year, which is still substantial but not the apocalyptic figure originally reported.
The key variable is device type. Watching Netflix on a large OLED TV at 4K resolution consumes dramatically more energy than the same content on a smartphone screen. The display accounts for the majority of energy consumption in video streaming, not the data transfer itself. A smart TV running 4K HDR content at peak brightness can draw 150–200W. A smartphone watching the same stream draws 3–5W. The difference is an order of magnitude.
Resolution choices matter too. Streaming at 1080p rather than 4K roughly halves the data transfer and the server-side encoding energy. Most content is visually indistinguishable at 1080p on screens under 55 inches. This is one of the few areas where individual viewing behaviour choices have a meaningful impact on personal digital carbon.
Artificial Intelligence: The New Energy Demand
AI is the fastest-growing source of digital carbon emissions, and the numbers are significant enough to warrant separate treatment.
Training large language models is extraordinarily energy-intensive. Researchers at the University of Massachusetts Amherst calculated that training GPT-3 released approximately 552 tonnes of CO₂ — equivalent to the lifetime emissions of roughly five average American cars, or the annual carbon footprint of 120 petrol cars driving for a year. That was a single training run. OpenAI has trained multiple versions of GPT since, and GPT-4 is widely believed to be substantially larger than GPT-3, implying higher training costs. Google's Gemini Ultra, Anthropic's Claude models, Meta's Llama series — the frontier of AI represents enormous one-time training investments in energy terms.
Training, however, is not the largest ongoing cost. Inference — running the model to generate responses to real queries — is the persistent energy burden. Every ChatGPT conversation, every Copilot code completion, every AI image generation request consumes energy. Goldman Sachs estimated in mid-2024 that AI inference could require the equivalent of several new nuclear power plants' worth of electricity capacity over the next decade to support projected demand.
The water consumption dimension is also significant and less discussed. Training and running large AI models requires substantial cooling, much of which is water-based. Microsoft disclosed that its water consumption increased 34% in 2022 relative to the prior year, attributed in significant part to AI infrastructure. Google reported similar increases. For data centres in water-stressed regions, this is a separate sustainability concern from carbon alone.
The semiconductor manufacturing required for AI hardware (GPUs and custom ASICs like Google's TPUs) has its own embedded carbon from the highly energy-intensive chip fabrication process. NVIDIA's A100 GPU — the workhorse of AI training until recently — requires significant energy to manufacture before it processes a single training token.
The SWDM v4 methodology article discusses how AI workloads are changing data centre energy profiles — relevant context if you want to understand why the 1% electricity stability figure may not hold.
Mobile and 5G: The Network Transition
Mobile now accounts for approximately 60% of global web traffic by pageview count, up from under 10% in 2010. This shift matters for digital carbon because mobile networks — particularly 4G LTE and 5G — are substantially less energy-efficient per gigabyte than fixed-line broadband.
Fixed-line broadband (fibre, cable) uses approximately 0.04 kWh per gigabyte for the access network. Mobile 4G LTE uses roughly 0.10 kWh per GB — about 2.5 times more energy. The reason is fundamental to radio physics: base stations must maintain coverage continuously regardless of traffic load, so their energy consumption is partly fixed overhead rather than purely proportional to data transmitted.
5G is more nuanced. On a per-bit basis at high network utilisation, 5G is more efficient than 4G — newer antenna technology and more efficient signal processing reduce energy per byte transferred. However, 5G base stations consume significantly more absolute energy than 4G base stations (typically 2–3 times more), and current 5G networks operate at low utilisation in most markets. The result is that 5G networks in their current early-deployment state are often less energy-efficient overall than mature 4G networks, with efficiency benefits expected to materialise as utilisation increases over time.
The implication for website developers: the growing share of mobile traffic in high-mobile-utilisation markets means page weight reduction is not just a carbon concern but an energy justice concern. Users in developing markets with predominantly mobile and often 3G or early-4G connectivity bear the energy cost of overweight pages more acutely than broadband desktop users in wealthy countries.
Cryptocurrency: The Uncomfortable Comparison
Cryptocurrency mining exists in a different category from other digital carbon sources, but it provides a useful reference point for scale.
Bitcoin mining consumed approximately 110–150 TWh of electricity in 2023, according to Cambridge Centre for Alternative Finance estimates. For context, that is more electricity than Argentina's entire annual national consumption. It is roughly 25–35% of the internet's total electricity use, concentrated in a single application.
Ethereum's transition from Proof of Work to Proof of Stake in September 2022 — the Merge — reduced Ethereum's energy consumption by approximately 99.95% overnight. That single technical decision eliminated roughly 30–40 TWh of annual energy consumption. It is one of the most dramatic carbon reductions in the history of any technology system, and demonstrates that protocol-level decisions can have enormous aggregate environmental consequences.
Bitcoin, which has no current plans to adopt Proof of Stake, remains the dominant energy consumer in cryptocurrency. The environmental argument about Bitcoin is contested — proponents argue that miners often use surplus renewable energy or stranded gas that would otherwise be flared, while critics argue that mining creates economic incentives to build or maintain fossil fuel capacity. The data on actual Bitcoin energy mix remains difficult to verify independently.
Cloud Computing: The Efficiency Paradox
Cloud computing — the shift from on-premise servers to shared infrastructure at Amazon Web Services, Google Cloud, Microsoft Azure, and similar providers — is generally cited as a net positive for digital carbon emissions. And the IEA data supports this for the transition period: hyperscale cloud achieves PUE around 1.1–1.2 versus 1.5–2.0 for typical enterprise data centres.
But cloud computing also enables things that previously did not happen at scale. The classic Jevons Paradox: efficiency improvements can increase total consumption by making a service more accessible and affordable. Cloud computing has enabled a vast expansion of digital services that would not exist at their current scale without cheap, elastic compute. Each of those services generates emissions that would not otherwise exist.
The major cloud providers have published sustainability commitments: Google has been carbon-neutral since 2007 and aims for 24/7 carbon-free energy by 2030. Microsoft committed to carbon-negative operations by 2030. Amazon aims for net-zero by 2040. These commitments are meaningful — and contested by critics who question whether carbon offsets and Renewable Energy Certificates represent genuine additionality. The trajectory is directionally positive; the pace relative to demand growth is the open question.
If you want to understand how cloud hosting choices affect individual website carbon scores — and what verified green hosting means in practice versus self-reported claims — the SWDM v4 deep-dive explains the Green Web Foundation verification process used in website carbon calculations.
Context: Real-World Analogies
Numbers are easier to hold when they connect to something tangible. Here are the statistics from this article rendered in more graspable terms:
416 TWh annual internet electricity consumption is equivalent to running the entire United Kingdom's electricity grid for one year and three months. Or powering every household in France, Germany, Italy, and Spain simultaneously for a year.
552 tonnes CO₂ for a single GPT-3 training run is roughly equivalent to flying from London to New York and back about 1,500 times. Or 120 average petrol cars being driven for a full year.
0.5g CO₂ per page load sounds trivial — until you remember that websites with millions of monthly visitors are multiplying that number millions of times. A site at 1g CO₂ per pageview with 500,000 monthly visits produces 6 tonnes of CO₂ per year. That is the annual footprint of an average European citizen from all sources combined.
4g CO₂ per email: a typical office worker sends and receives around 100 emails per day. That is 400g CO₂ per person per day from email alone — about 146 kg per year, or equivalent to driving around 600 kilometres in a petrol car, just from email.
110 TWh for Bitcoin mining: the electricity consumed by Bitcoin in 2023 could have powered all the electric vehicles in the world for approximately two years.
Year-Over-Year Trends
The trajectory matters as much as the snapshot. A few trend lines worth tracking:
Internet traffic roughly doubled between 2018 and 2022 and is projected to double again by 2027, according to Cisco's Visual Networking Index projections. Data centre electricity consumption has not doubled over this period — which reflects genuine efficiency gains — but the buffer is narrowing, particularly under AI workload pressure.
Median page weight has increased year-over-year in HTTPArchive data: from 1.9 MB in 2021 to approximately 2.5 MB in 2024. We are building heavier pages despite better tools for compression and optimisation being widely available. The dominant driver is JavaScript framework adoption and the proliferation of third-party marketing scripts.
Renewable energy in data centres is growing substantially. The proportion of data centre electricity consumption matched by renewable energy has increased from under 10% in 2015 to over 40% for the largest cloud providers today. This is meaningful carbon reduction even if the absolute energy demand is growing.
Mobile data consumption per device has grown substantially as streaming resolution has increased and social media platforms have shifted toward video. Average mobile data consumption per user roughly tripled between 2018 and 2023.
What You Can Actually Do About It
Data without action is just anxiety. Here is where individual decisions genuinely connect to the statistics above.
Measure your site. Run it through the Carbon Badge scanner. If your site has significant traffic — even tens of thousands of monthly pageviews — the emissions are real and measurable. You cannot improve what you have not measured.
Optimise images first. They account for roughly 50% of median page weight. Converting to WebP or AVIF, adding responsive srcset attributes, and implementing lazy loading for below-the-fold images typically delivers 30–50% page weight reduction with no visible quality change. The complete reduction guide covers every technique with real numbers.
Audit your JavaScript. Chrome DevTools' Coverage panel shows what percentage of your JS is actually executed on a given page. Dead code — scripts loaded but never run — is extremely common and entirely wasteful. Third-party scripts (chat widgets, A/B testing tools, marketing pixels) often account for 30–50% of JavaScript weight on commercial sites.
Switch to a verified green host. In the SWDM v4 formula, a host verified by the Green Web Foundation as renewable-powered delivers a 24.3% automatic reduction in calculated CO₂ — with zero code changes. Infomaniak, Hetzner, and Google Cloud are among the verified providers.
Email hygiene. Unsubscribe from newsletters you do not read. Avoid large attachments by linking to cloud-hosted files. The aggregate effect across an organisation is real, even if any single email is trivially small.
Video resolution. If you produce or distribute video, the resolution setting affects data transfer significantly. 1080p rather than 4K roughly halves bandwidth. For content where the quality difference is imperceptible at typical viewing distances and screen sizes, this is a straightforward reduction.
Internal advocacy. The largest impact any individual technical practitioner can have is often not their own site but the organisations they influence. Making the case for sustainability audits, green hosting procurement, and lean development practices where you work amplifies your reach substantially. The website sustainability audit guide provides a structured framework for exactly this kind of internal proposal.
Display your grade. A carbon badge on your site creates accountability — and visibility. It signals to users and stakeholders that digital emissions are something you take seriously enough to measure and publish. That transparency is harder to walk back from than a press release, which is the point. See the pro tier for automated monitoring that catches regressions as they happen.
The internet's carbon footprint is large enough to matter and distributed enough that individual technical decisions compound across millions of sites and billions of users. That is not a reason for paralysis. It is a reason to start with what you can measure and control directly — which, if you have a website, is something you can act on today.