Carbon Footprint of AI: Data Centres, Training and Inference Explained

The carbon footprint of artificial intelligence has emerged as one of the defining environmental challenges of the current technological era. Electricity demand from AI-focused data centres surged by around 50% in a single year ^[1], and projections from the International Energy Agency (IEA) indicate that global data centre electricity consumption could approach 950 terawatt-hours by the end of the decade ^[2]. For organisations deploying AI across their operations, understanding this impact is no longer optional. This article explores the three main drivers of AI’s carbon footprint: data centres, model training and inference, and examines what responsible AI adoption looks like in practice.

Understanding the Carbon Footprint of AI

The carbon footprint of AI encompasses all greenhouse gas emissions generated throughout the lifecycle of an artificial intelligence system: from the manufacturing of hardware components and the construction of data centres, to the energy consumed during model training, fine-tuning and real-time inference. The total impact is the sum of these operational and embodied emissions, expressed in tonnes of CO2 equivalent.

Estimates of the aggregate climate impact of AI systems are significant. Research cited by MIT places the carbon footprint of AI systems globally at between 32.6 and 79.7 million tonnes of CO2 equivalent, with associated water consumption reaching hundreds of billions of litres ^[3]. These figures span a wide range precisely because AI is deployed across enormously varied contexts, from low-energy text completions to computationally intensive image generation and scientific modelling.

Three factors primarily determine an AI system’s carbon footprint: the energy intensity of the hardware running the computation, the carbon intensity of the electricity grid powering that hardware, and the scale of deployment (how many queries or tasks are processed). All three factors are shifting simultaneously, which makes the aggregate climate trajectory of AI difficult to predict with precision.

For organisations seeking to understand their total carbon footprint, including the growing contribution of digital tools and AI services, TheGreenshot’s guide to carbon audits for media productions provides a practical framework applicable across production types.

Data Centres: The Physical Infrastructure Behind AI

Data centres are the physical foundation of AI. They house the servers, networking equipment and cooling systems that process AI workloads, and their energy consumption is the primary driver of AI’s direct carbon footprint.

According to the IEA, electricity consumption from data centres amounted to approximately 415 terawatt-hours (TWh) in a recent baseline year, representing around 1.5% of global electricity consumption ^[2]. The IEA projects this figure will roughly double to around 950 TWh by 2030, at which point data centres would account for approximately 3% of global electricity demand ^[2]. AI workloads are projected to constitute 35 to 50% of total data centre power use by that date, up from around 5 to 15% in recent years ^[4].

The carbon intensity of this electricity consumption depends heavily on geography. Data centres located in regions with high renewable energy penetration, such as Iceland or parts of Scandinavia, produce significantly lower emissions per kilowatt-hour than those powered predominantly by coal or gas. As of the IEA’s most recent central scenario, CO2 emissions from electricity generation for data centres are projected to peak at around 320 million tonnes before entering a gradual decline driven by grid decarbonisation ^[2].

Beyond direct operational electricity use, data centres also carry embodied carbon: the emissions associated with manufacturing servers, memory chips, networking hardware and cooling systems. These supply chain emissions can be substantial and are often omitted from headline figures, making the true lifecycle footprint of AI infrastructure larger than operational energy data alone suggests.

Training vs. Inference: Where AI Emissions Come From

Within the AI lifecycle, emissions are unevenly distributed between training and inference, the two primary computational phases.

Model training

Training a large AI model requires vast computational resources concentrated over a finite period. The energy cost of training can be substantial: estimates suggest that training a large-scale language model of the generation currently in commercial deployment consumed approximately 50 gigawatt-hours of electricity, enough to power a major city for several days ^[5]. Training is also a one-time cost per model version, meaning that a single training run amortises across the billions of inferences the model subsequently serves.

Inference

Inference refers to the act of running a trained model to generate a response, an image, a transcription or any other output. While each individual inference consumes far less energy than training, the cumulative scale of inference operations now represents the dominant share of AI’s ongoing energy demand. Current estimates indicate that inference accounts for 80 to 90% of total AI computing power consumed ^[6]. As AI tools become embedded in everyday professional and consumer workflows, inference volumes will continue to grow.

For organisations reporting on Scope 3 emissions, purchased AI services consumed through cloud platforms represent an indirect value chain emission. Accurately attributing these emissions requires providers to disclose the energy consumption and carbon intensity of their inference operations, a practice that remains inconsistent across the industry. For a broader view of how emission scopes apply to media and production organisations, TheGreenshot’s breakdown of emission scopes in audiovisual production offers a useful sector-specific reference.

Can AI Reduce Its Own Carbon Impact?

One of the more complex dynamics in AI’s environmental story is the simultaneous increase in deployment scale and improvement in energy efficiency. Hardware designed specifically for AI inference, such as purpose-built accelerator chips, has dramatically reduced the energy required per computation compared to general-purpose processors. Software-level optimisations, model compression and quantisation techniques have further improved efficiency at the application layer.

The scale of these improvements is striking in some cases. Data published by one major AI provider showed a reduction in the median energy consumption per AI prompt by a factor of more than thirty over a single twelve-month period, alongside a comparable reduction in per-prompt carbon intensity ^[6]. These efficiency gains reflect both hardware improvement and model architecture optimisation.

However, efficiency improvements are being outpaced by the growth in total usage. The IEA notes that power consumption per AI task is declining rapidly, but more users are adopting AI and more energy-intensive applications, such as autonomous AI agents and multimodal generation, are proliferating ^[2]. The net effect is that aggregate emissions continue to rise even as per-task efficiency improves, a pattern well documented across the history of computing.

The most effective levers for reducing AI’s carbon footprint combine efficiency improvements at the hardware and software layers with decarbonisation of electricity grids, strategic location of data centres in regions with high renewable energy penetration, and demand-side awareness among organisations choosing between AI service providers.

AI and Carbon in the Media and Entertainment Sector

The media and entertainment (M&E) sector sits at an interesting intersection with AI’s carbon trajectory. On one hand, the industry is a growing consumer of AI-powered tools, from automated subtitling and colour grading assistance to generative visual effects and AI-driven production scheduling. On the other hand, the sector faces mounting pressure to reduce its own substantial environmental footprint.

AI tools in film and television production

Production companies and studios are adopting AI tools at accelerating pace across the editorial, post-production and distribution pipeline. Each of these tools carries a carbon cost through the cloud inference services they invoke. While individual queries are low-energy, a production team running AI-assisted review, transcription and effects processing across a feature-length project can accumulate meaningful digital emissions that belong within Scope 3 reporting under the purchased services category ^[7].

At the same time, AI-enabled workflows can deliver avoided emissions. Remote AI-assisted pre-visualisation reduces the need for exploratory physical shoots. AI-powered scheduling tools optimise crew movements, compressing shoot schedules and reducing the transport emissions that typically account for the largest share of a production’s carbon footprint. Research in the adjacent field of synthetic video production suggests that AI-generated corporate video content can be dramatically more carbon-efficient than equivalent physical production, though methodology and scope assumptions vary significantly between studies ^[8].

Events and live productions

For live event producers, AI tools are increasingly used for crowd management modelling, real-time energy monitoring and logistics optimisation. The operational carbon savings from better logistics coordination can outweigh the AI inference emissions associated with running these tools, particularly at the scale of large festivals and multi-venue events. The key for producers is to account for both sides of this equation in their carbon reporting rather than assuming AI tools are automatically carbon-neutral.

For production teams and event organisers seeking to measure the full scope of their digital and AI-related emissions, TheGreenshot’s overview of carbon calculators in the audiovisual sector provides a framework for selecting tools suited to the sector’s specific reporting requirements.

Conclusion

The carbon footprint of AI is substantial, growing and unevenly distributed across training, inference and data centre infrastructure. The IEA’s projections make clear that data centre energy demand will continue to expand significantly, even as per-task efficiency improves. For organisations integrating AI into their operations, this is increasingly a material emissions question, not merely a technology procurement one.

The path forward involves several converging responses: hardware and software efficiency at the AI provider level, grid decarbonisation at the electricity system level, and informed procurement choices at the organisational level. For the media and entertainment sector specifically, understanding the carbon footprint of AI tools sits alongside the broader challenge of measuring and reducing emissions across the full production and event lifecycle.

As reporting frameworks evolve and carbon disclosure requirements extend further into Scope 3 value chains, the emissions associated with AI services consumed through the cloud will increasingly require explicit accounting. Organisations that begin building this capability now will be better positioned as these requirements crystallise.

FAQ

Learn more with TheGreenshot