Operational Emissions

Operational emissions encompass the greenhouse gases emitted from an organisation’s own technology infrastructure and operations. These are categorised into three groups - On-Prem, Cloud, and Generator emissions (O, C, and G) - based on the level of ownership and control an organisation has over the assets.

Understanding operational emissions allows organisations to quantify, monitor, and benchmark the climate impacts of their technology infrastructure and business operations. It enables setting emission reduction targets and strategies and facilitates compliance with current and emerging climate disclosure legislations.

Emissions are further classified into direct and indirect sources:

Direct Emissions

Direct emissions are a result of the organisation’s direct consumption of grid-supplied electricity (Category O) or combustion of fossil fuels like diesel or natural gas for owned power generators (Category G). The organisation can directly measure and account for emissions from owned assets.

On-premise

Emissions associated with the actual operation and use of devices owned by an organisation (usage carbon).

Servers and Storage

The energy consumed by on-premise servers and data centres. This includes those dedicated to AI operations and associated cooling systems, and other infrastructure necessary for maintaining AI workloads.

Machine Learning training and fine-tuning

Foundation Models

This section focuses on organisations that develop and train foundation models, or fine-tune third-party models. In these scenarios, organisations must account for the significant computing resources required throughout the process.

These operations typically demand intensive use of specialised hardware including Graphics Processing Units (GPUs) and Central Processing Units (CPUs), along with associated infrastructure such as high-performance storage systems, large amounts of RAM and cooling equipment.

Assessing the efficiency of hardware and algorithms can help reduce unnecessary energy consumption. Techniques such as model distillation, quantisation, energy-aware pruning and low-precision arithmetic operations can further lower the computational and energy requirements of models.

It should be acknowledged however, that not all model development requires vast amounts of computing power: while Large Language Models are resource-intensive and versatile, smaller models (SLMs) trained on more modest datasets are designed to be more compact and efficient, requiring less computational power and memory.

Self-hosting

When self-hosting a foundation model, organisations must account for the energy consumption of inference operations, including any open-source interfaces and the specialised hardware required to support these services. This includes not only the direct energy use of GPUs/CPUs during inference but also the overhead of maintaining on-premises or private cloud infrastructure, such as cooling, power distribution, and data storage.

Click here to see a worked example of estimating usage emissions for a server.

Networking Devices

In the information on networks, networks are split into four categories - the operational network, the wired backbone internet, the wireless internet and your consumer’s network. Of these, the first three need to be considered in Operational Emissions.

Within the operational network, consider the running cost of the connected devices. These include, but are not limited to:

routers
switches
bridges
Wi-Fi access points
firewalls
modems
hubs
repeaters

Research¹ has shown that not all networking equipment will use power in line with the data it is transmitted over it. These devices will have a baseline load (i.e. the energy they consume for simply being switched on, regardless of the processing they do), a maximum load and a typical load. Most wired networking equipment, such as routers and switches, will consume at least 60% of their maximum power as a baseline going up to 90% of their maximum power under a heavy load². For accurate numbers, power monitors on the power socket are an excellent way of gathering data. As the power usage of networking equipment has very little variance while it is running, attributing the power consumption to specific applications or processes is less important.

Data consumed from a supplier’s API, that is external to your organisation, is accounted for in the indirect category. Public wired and wireless internet usage does not need to be measured in the Direct Emission category. Data sent from the service to end users is accounted for in the downstream emissions category.

With the increasing adoption of AI-powered applications by organisations, efficient networking systems are essential to handle high volumes of data transfers, reduce latency and meet the high bandwith demands of AI workloads. One example is NVIDIA’s H100 tensor core GPU, a product specifically designed for accelerated computing that includes a dedicated transformer engine to speed up the training of trillion-parameter language models.

For organisations treating content and data as a commodity, emissions from network infrastructure, including content delivery networks (CDNs) and edge servers to cache content and reduce the load on central servers must be taken into account. These systems reduce central server strain but add their own energy overhead.

Onsite Employee Devices

The energy consumed by personal computers (desktops, laptops), smartphones, higher resolutions screens (4K, HDR), monitors, printers and other devices used by employees in offices. This is energy directly used by the business, and all of the emissions related to this energy are their responsibility.

When content and data are consumed as a commodity, it is important to account for the emissions generated by employees devices to stream, download or create content. While individual device use may seem small, the cumulative effect can be substantial.

Special consideration also needs to be given to the increased use of LLMs. As their size and complexity continues to grow so do the hardware requirements needed to run them effectively. If employees in your organisation require higher-spec devices to develop or interact with these applications then it’s important to account for the increased energy consumption associated with more powerful processors, GPUs, and memory configurations.

Click here to see a worked example of estimating usage emissions for a laptop.

Generators

Any fossil fuel-powered generators, solar PV, wind turbines, or other systems installed on-site to supply electricity to technology equipment.

Category G (Generators) can be related to GHG Protocol Scope 1.

Indirect Emissions

Indirect emissions result from the organisation’s purchase of cloud computing services, software, or managed IT services from third-party technology providers.

As the organisation does not own or control the IT assets delivering these services, their associated emissions cannot be directly quantified. But use of these services can contribute substantially to the organisation’s overall technology emissions footprint. Emissions estimates for third-party services can be obtained from providers to account for these indirect, supply chain impacts.

Category C (Cloud) emissions can be related to GHG Protocol Scope 3.

Cloud Services

The emissions associated with cloud platform services like compute, storage and networking. Services are backed by computing hardware with associated upstream and operational emissions. The proportion of such emissions attributable to an organisation will vary based on service, server instance types, and region.

If your organisation is developing an AI product — whether you’re building a large language model from the ground up or leveraging an existing foundation model — it’s highly likely that it will be hosted on a Cloud platform, whether using Model as a Service (Maas) or self-managed Cloud infrastructure for custom deployments. This is due to the substantial computational resources required to run AI systems efficiently.

SaaS

The emissions associated with Software-as-a-Service (SaaS) applications, delivered primarily through cloud computing to accommodate for the rapid increases in usage, represent a growing share of digital carbon footprints. While service providers may report these emissions, a portion should also be attributed to the organisations using these services, in line with shared responsibility models like those outlined in the GHG Protocol.

This challenge becomes even more pressing with the rise of AI-powered applications. Organisations increasingly interact with large language models (LLMs) directly through their chatbot platforms (e.g., Claude, Le Chat, ChatGPT), API providers (e.g., Anthropic, OpenAI, Google), and AI-driven tools (e.g., Copilot, AI-enhanced search engines). As companies embed these AI-driven SaaS solutions into their workflows, the share of emissions attributable to their usage must be identified and included in organisational carbon accounting.

Managed Services

Managed Security Operations, IT support and data backup services.

Offsite Employee Devices

The emissions incurred by computers (desktops, laptops), monitors, printers, and other devices used by employees, when used out-of-office. Since these are required by the business, they are accounted for, but as with other indirect groupings, the business does not control them directly. Accordingly, they come under GHG Protocol Scope 3, and so are a separate category.

Offsite employee devices are typically accounted for with an attribution factor, representing the proportion of time that their use is dedicated to business purposes.