Carbon Footprint of Computing
Also known as: digital carbon footprint, ICT emissions, computing emissions, tech carbon footprint
The carbon footprint of computing refers to the total greenhouse gas emissions associated with the technology sector: from the energy used to run devices and data centres to the emissions embedded in manufacturing hardware.
The ICT sector accounts for roughly 2-4% of global greenhouse gas emissions, a share comparable to aviation. Individual contributions include device manufacturing (responsible for the majority of a device's lifetime carbon), electricity consumption during use, and end-of-life disposal. Understanding where computing emissions come from is the first step to reducing them.
Carbon Footprint of Computing, explained at three levels
At home
The carbon footprint of computing is all the greenhouse gas pollution caused by making and using technology. This includes the electricity used by your phone and laptop while you use them, the energy used by data centres that run websites and cloud services, the emissions from manufacturing all those devices (which are actually the biggest share for many devices), and the pollution from recycling or landfilling old technology. Your laptop produces most of its lifetime carbon emissions before you even turn it on for the first time.
At work
For businesses, understanding computing's carbon footprint matters for sustainability reporting, supply chain due diligence, and IT procurement decisions. Most large organizations include technology-related emissions in their Scope 1-3 carbon reporting. Electricity for an organization's own servers and offices is Scope 2, while cloud services bought from a provider fall under Scope 3 (supply chain emissions). Hardware manufacturing emissions are Scope 3 for the organization buying the hardware. Data centres can provide carbon intensity reports to corporate customers.
At home
The carbon footprint of a device has three main components: embodied carbon (the emissions from mining raw materials, manufacturing components, assembly, and shipping: typically 60-80% of a device's lifetime footprint for smartphones and laptops), operational carbon (the electricity consumed during use), and end-of-life carbon (emissions from disposal or recycling). Extending device lifetime is the single most effective action for reducing device-level emissions because it amortises the large embodied carbon over more years of use.
At work
Data centre carbon footprints depend on two factors: the efficiency of the data centre (measured by Power Usage Effectiveness, or PUE: the ratio of total facility energy to IT equipment energy) and the carbon intensity of the electricity grid serving the data centre. A very efficient data centre running on coal-powered electricity can have a higher carbon footprint than a less efficient data centre running on 100% renewable electricity. Location of data centre matters enormously for carbon intensity.
At home
The environmental impact of AI is significant and growing. Training large AI models consumes substantial compute resources: GPT-3's training was estimated at 552 tonnes of CO2e; larger models and repeated training runs scale significantly beyond this. Inference (running models in production) may ultimately have larger aggregate emissions than training because it happens continuously at scale. Some AI developers (such as Google, Meta, and Mistral) have published training or per-query footprints; many have not, making sector-wide accounting difficult.
At work
From a practitioner standpoint, the most impactful actions for reducing computing's carbon footprint are: extending device lifetime (refusing to upgrade unnecessarily), selecting cloud providers with renewable energy commitments and efficient data centres, choosing efficient hardware, turning off or hibernating equipment when not in use, and including carbon considerations in software architecture decisions (efficient code running fewer server-side computations reduces energy consumption at scale). Carbon-aware scheduling (scheduling intensive compute workloads to run when and where the grid is greenest) is an emerging technique for cloud-native applications.
Real-world examples
At home
Sophie read that her three-year-old laptop was still meeting her needs but she had been considering an upgrade. She researched that manufacturing her next laptop would produce roughly 300kg of CO2e. She decided to keep her current laptop for at least another two years, reducing her personal computing carbon footprint meaningfully.
At work
The sustainability team at a media company included digital infrastructure in their annual carbon report for the first time. They discovered that their cloud computing bill implied roughly 400 tonnes of CO2e annually (based on their provider's carbon intensity report). They moved three major workloads to the provider's lowest-carbon availability region, reducing digital infrastructure emissions by 35%.
Common confusions
- Streaming video is not the biggest source of computing emissions: manufacturing is. Popular news articles have overstated streaming's carbon footprint. The largest source of emissions in most individuals' computing is the manufacturing of their devices, not their electricity consumption during use. Reducing device turnover rate has a much larger impact than reducing streaming hours.
- Cloud computing is not automatically lower-carbon than on-premises computing. A cloud provider with renewable energy commitments and efficient data centres may have lower carbon intensity than your on-premises servers. But a cloud provider in a coal-intensive region may have higher carbon intensity. The carbon footprint depends on where and how the data centre is powered, not just whether it is cloud or on-premises.
- Energy efficiency improvements do not automatically reduce absolute emissions. While computing has become dramatically more energy-efficient per unit of compute over decades, the total amount of computing has grown even faster: producing higher absolute energy consumption despite higher efficiency. This 'Jevons paradox' effect means that efficiency improvements alone do not reduce total emissions without also constraining total compute growth.
For analysts
In BA artefacts, computing carbon footprint requirements surface in sustainability reporting systems, green IT procurement, and software architecture decisions for carbon-sensitive organizations. When documenting requirements for cloud-based systems, include carbon considerations as NFRs: 'The system should preferentially deploy to regions with carbon intensity below [threshold] gCO2e/kWh, using the provider's carbon intensity API to guide scheduling of non-real-time workloads.' For procurement requirements, include: 'Hardware purchases must include the manufacturer's stated product carbon footprint (PCF) in the procurement comparison criteria.'
History & origin
Concern about computing's energy consumption began in the 1990s when data centres became significant electricity consumers. The first rigorous lifecycle analyses of personal computing and internet infrastructure emerged in the 2000s. A 2007 Gartner report estimated that ICT accounted for approximately 2% of global CO2 emissions: a figure frequently cited as analogous to aviation. As cloud computing, streaming, AI training, and cryptocurrency mining scaled dramatically in the 2010s and 2020s, the sector's emissions grew in absolute terms even as efficiency improved per unit of compute. The environmental impact of AI training runs (GPT-3's training was estimated to emit 552 tonnes of CO2) attracted specific research attention from 2019.