CIOs need to start accounting for AI carbon footprint

AI's energy thirst

The International Energy Agency estimates that data centers consumed approximately 415 terawatt-hours of electricity in 2024, around 1.5% of global electricity consumption. It forecasts this could more than double to approximately 945 TWh by 2030. Accelerated servers, principally driven by AI adoption, account for almost half of that increase.

This does not mean that AI is poised to consume all the world’s electricity. Data centers are still projected to represent less than 3% of global demand in 2030. However, the impact is concentrated geographically and natural gas and coal are projected to supply more than 40% of the additional electricity required by data centers up to 2030. The IEA expects data center electricity emissions to peak at around 320 million tons of CO2 by 2030.

These are substantial numbers, but they do not tell an individual CIO what using Microsoft Copilot, ChatGPT or an internally developed AI agent does to their own carbon inventory.

The dangerous simplicity of “per prompt”

There have been several attempts to calculate the footprint of an individual AI query. Google estimates that the median text prompt submitted to its Gemini applications in May 2025 used 0.24 watt-hours of electricity, emitted 0.03 grams of CO2 equivalent and consumed 0.26 milliliters of water.

Using those figures, one million prompts would consume approximately 240 kWh and produce 30 kg of CO2e. That sounds surprisingly modest.

Unfortunately, it's not a universal conversion factor. The result changes according to several factors, including: the model; the length of the input and output; whether the system is generating text, images or video; the hardware operations; the data center location and the carbon intensity of the electricity being used. A complex reasoning task -- or an AI agent making dozens of calls to complete one employee request -- cannot reasonably be equated with a short chatbot response.

Google also reports that the energy consumption of its median prompt fell 33-fold in 12 months. Efficiency is improving remarkably quickly, but total demand can still rise if usage grows even faster. This is the familiar rebound effect that making something cheaper and more efficient encourages much more use.

The biggest organizational risk, therefore, may not be one extravagant AI model. It may be millions of apparently inconsequential interactions embedded invisibly into everyday processes.

Start by recognizing what you do not know

CIOs should resist the temptation to create a wonderfully precise spreadsheet based on gloriously imprecise assumptions. The more useful first step is to establish visibility.

Create an inventory of sanctioned AI services, internally hosted models, API consumption and significant AI features embedded in existing software. Include shadow AI where it can reasonably be identified. Usage should be measured by model, application, business unit and workload type, rather than simply counting licenses.

CIOs should then ask cloud and AI suppliers for energy, carbon and water data attributable to customer usage. The suppliers should explain whether figures include cooling, idle capacity, networking and embodied emissions from manufacturing hardware. Google Cloud already provides carbon data by project, product and region, while Microsoft offers emissions reporting for its cloud services. The information is imperfect, but considerably better than pretending outsourced computing has no footprint.

For most organizations, emissions from externally supplied AI services are likely to sit within Scope 3 as purchased goods and services. Internally operated infrastructure may create Scope 2 electricity emissions as well as Scope 3 emissions associated with hardware and suppliers. Sustainability and finance teams should agree to use an accounting treatment rather than leaving AI consumption outside the reporting boundary.

Carbon needs its own FinOps discipline

The practical mitigations look remarkably similar to good cost management.

Use the smallest AI model capable of completing the task. Route simple enquiries to smaller models and reserve expensive reasoning models for problems that justify them. Limit unnecessarily long prompts and responses, reuse results where appropriate and prevent autonomous agents from entering loops or making unlimited calls.

Development teams should test energy and carbon alongside cost, latency and accuracy. Workloads that are not time-critical may be directed to regions or times with cleaner electricity, subject to data residency and operational requirements. Redundant models, experimental environments and unused GPU capacity should be retired.

Procurement also has a role. Requests for proposals should require suppliers to disclose their measurement methodology, regional energy mix, water usage and efficiency improvements. “Powered by renewable energy” is not enough on its own, CIOs should distinguish contractual renewable-energy claims from the carbon intensity of the electricity physically consumed.

Finally, AI’s footprint should be considered against the emissions it helps avoid. An optimization system that reduces transport mileage or energy consumption may deliver a net environmental benefit. A tool generating meeting summaries that nobody reads probably will not.

The objective is not to discourage AI adoption. It is to make carbon another visible unit of consumption, alongside tokens, API calls and pounds spent. An organizations can challenge what it measures. What remains hidden in the cloud will simply continue to grow.