Getting started with measuring AI’s carbon footprint

Continuous measurement

While AI is driving exponential growth in computing, Benjamin Brial, founder of Cycloid, points out that most organisations never architect their compute usage to control the huge growth demanded by AI.

“Sustainability is still treated like a compliance checkbox, something reviewed after the fact in a quarterly report,” he says. “Treating sustainability as a separate reporting line only makes the problem worse.

“When GreenOps lives in dashboards disconnected from developer workflows, organisations effectively hard-code a silo of waste into their infrastructure. Unless a company is willing to staff, fund and operationalise that silo continuously – and most are not – it becomes theatre rather than control.”

According to Brial, real sustainability only works when cost and carbon signals are part of the same platforms developers use to build, deploy and scale software. Otherwise, inefficiency is not an accident, it is an architectural choice.

Sustainability is often misunderstood as doing less. In reality, it is about consuming better

So, what’s the answer? According to Brial, real sustainability shouldn’t start with slowing AI adoption or restricting experimentation, but giving teams platforms that make consumption visible and manageable from the start.

“Without that foundation, AI simply magnifies inefficiency. With it, teams can innovate confidently, knowing the impact of their choices before they commit to them,” he says.

As Brial notes, cost and carbon are driven by the same infrastructure decisions made every day by development teams. “In the cloud, financial and environmental impact cannot be separated. Instance sizing, storage choices, data movement and how long services are left running all influence both spend and emissions,” he adds.

These are not strategic decisions made once a year; they are small, frequent choices made at build and deploy time. “When teams lack visibility into those impacts at the moment the decisions are made, optimisation becomes slow, frustrating and often ignored as it becomes a blocker due to the initial bad implementation,” says Brial.

He states that developers need more oversight into their environmental impacts – if only for purely fiscal reasons. They already optimise for performance, reliability and delivery speed because those signals are visible and immediate. But in Brial’s experience, cost and carbon are usually more opaque until weeks later, buried in reports that arrive long after the code is in production.

“Sustainability is often misunderstood as doing less. In reality, it is about consuming better. Rightsizing, elasticity and automation reduce idle resources and unnecessary workloads. That improves delivery speed and reliability as much as it reduces waste and it allows more money to be spent on initiatives that work or deliver more. So, it really isn’t about squeezing innovation but making it delivery-focused,” Brial adds. 

And when platforms handle optimisation continuously, developers spend less time firefighting and more time building. Brial says the most effective organisations treat GreenOps and FinOps as outcomes of good product design, not as standalone initiatives. When cost and carbon are signals inside developer platforms, sustainability stops being a clean-up exercise and becomes part of how software is delivered every day.

According to Brial, teams that invest in this approach will move faster, waste less and scale responsibly, not because they were told to, but because the platform makes it the easiest path forward.

Needless waste

One area often overlooked when considering optimisation is the storage of data. Soham Mazumdar, co-founder and CEO of WisdomAI, says IT leaders should consider the waste that occurs when data is duplicated unnecessarily.

“Most organisations have three or four copies of every meaningful dataset: an original system of record; a derivative copy created through extract, transform and load operations [ETL] for analytics or reporting; one or more test or experimental versions; a production copy feeding dashboards and models or downstream applications,” says Mazumdar. “Each copy consumes storage, compute and operational effort.”

While a dataset may be critical during a product launch, a forecasting cycle or an AI experiment, its value drops once that window closes yet the data persists.

“Storage feels cheap and compute feels elastic, so the data stays rarely accessed, rarely validated and almost never deleted. That’s not good for global emissions,” adds Mazumdar.

In his experience, engineers focus on the immediate problem: moving data, transforming it, connecting systems. He says: “Nobody rewards garbage collection. In the cloud, creating resources is easy. There are few incentives to clean them up.”

When you understand which data is alive ... you reduce waste, lower environmental impact and create a healthier foundation for analytics and AI

According to Mazumdar, Google teams have explicit quotas for storage and compute. While these quotas may be large, they still exist, which forces prioritisation. “If a dataset or pipeline no longer mattered, it had to justify its continued existence. This produced healthier systems with fewer forgotten assets,” he adds.

He says manual accountability no longer scales. While the instinctive response is to demand more accountability from engineers, he feels that doesn’t work anymore. This is because in the age of AI, accountability moves in the opposite direction. Teams experiment, prototype and connect data to new models as fast as possible. Temporary pipelines and datasets proliferate. All of this means that manual processes cannot keep up.

Mazumdar recommends putting in place automation that tracks liveness. This includes systems that identify datasets not accessed in months, pipelines that no longer produce outputs and compute services that receive no traffic. “These signals should trigger action: archiving, tiering to cold storage or elimination,” Mazumdar says.

“Our mission now is to move from aspiration to operational reality. The path forward isn’t austerity – it’s visibility, liveness tracking and automated discipline built into data systems. When you understand which data is alive, which is dormant and which is genuinely needed, you reduce waste, lower environmental impact and create a healthier foundation for analytics and AI.”

Component-based metrics

Visibility starts with understanding the environmental impact of each component in the AI infrastructure stack from datacentre hardware through to software usage and the eventual disposal of equipment. Gartner’s How to measure and mitigate AI’s impact on environmental sustainability report, which was published in July 2025, recommends that IT leaders should prioritise the use of component-based measurements where possible as that is the most accurate methodology for measuring the environmental impact of AI. 

Gartner states positions component-based measurement as one of the more granular ways to measure AI’s carbon impact. It is based on breaking down the component parts of AI infrastructure and measuring these individually. These components cover physical IT infrastructure used for training and running AI models along with software operating systems, programming languages and the AI-enabled applications using the models and frameworks.

Gartner says the component-based approach measures these computational resources used by AI models, specifically quantifying the underlying hardware (primarily GPUs and CPUs), the time duration of the training process, the idle power draw of servers and the PUE of the datacentres where these computations occur.

With a component-based approach to calculating AI’s carbon footprint, Gartner says the carbon emissions associated with training and deploying AI models are then calculated by multiplying the total energy consumed by the carbon intensity of the electricity grid in the specific geographical region where the AI infrastructure is located. The electricity used by the hardware and the energy required for cooling datacentres must also be taken into account, as well as the electricity needed for storage of data used for training and AI inference.

For a full calculation, Gartner recommends accounting for the life cycle of datacentre equipment. This includes the manufacturing, deployment, operation and eventual disposal of all the components.

Sustainability is the roadmap for AI

Although the main focus of the tech sector has largely been about delivering more powerful AI models that can make the most of the latest developments in hardware, there has been less emphasis on achieving this in the most sustainable way. As powergrids become strained by the power requirements of AI factories and GPU-heavy datacentre facilities, planning for these sites is increasingly being put under the spotlight.

If the prediction from the Nvidia CEO shows the direction of travel the tech sector is taking, the efficiency of these facilities will need to improve exponentially. And for IT decision-makers, there is going to be much more focus on the efficiency of AI over its outright performance.