Cloud and artificial intelligence (AI) services typically account for under 1% of most consumer goods companies’ Scope 3 greenhouse gas (GHG) emissions, according to The Consumer Goods Forum’s Climate Transition Coalition. However, the infrastructure behind AI is expanding rapidly. The International Energy Agency (IEA) reports that data centers used 415 terawatt-hours (TWh) of electricity in 2024, about 1.5% of global consumption, and that use has grown 12% per year over the last five years.
While the impact is a small share of an individual company’s total footprint today, it is forecast to grow across the global economy. This gives companies an opportunity to develop an approach to manage AI’s impact appropriately through measurement and communication practices to reduce risk, build resilience, and support growth.
Where the footprint comes from matters
AI use carries 70% of the GHG impact across the life cycle. This includes amortized training, which involves building AI’s intelligence and developing models, and inference or user queries, prompts, and outputs. Since most businesses do not train their own models (that’s the role of large tech and AI companies such as Google, OpenAI, and Anthropic), a company’s AI footprint comes mainly from everyday use, such as drafting, product search, and service chat. McKinsey and Company projects that running AI queries will grow from 21 gigawatts (GW) to 93 GW of global data center demand between 2025 and 2030, as total usage rises from 82 GW to 219 GW.
A few key factors are critical when considering AI’s impact: the model being used, the data center location, how much or how often it is used, and what it is used for.
Google reports that its median Gemini text prompt uses 0.24 watt-hours (Wh) of energy, emits 0.03 grams of carbon dioxide equivalent (CO2e), and consumes 0.26 milliliters of water, based on Google’s average fleet-wide data after its clean energy purchases. It is estimated that these emissions could be up to three times higher if the actual local grid mix were used. OpenAI reports about 0.34 Wh for an average ChatGPT query. Further, the selection of the model level within a large language model (LLM) (e.g., Gemini) can result in a range of energy use from 0.1 Wh to 1.03 Wh across Gemini models alone.
U.S. data center power has a carbon intensity 48% higher than the national grid average, largely because many data centers are built in regions where the grid relies heavily on fossil fuels. Location matters for water, too: how a data center is cooled and whether it sits in a water-stressed region can shape its impact as much as the amount of computing does.
Common business considerations with AI impact
Though AI is a small part of a business’s footprint, stakeholders, such as employees, are noticing it, so a company should understand its own AI impact to determine its materiality and the appropriate level of action, such as communicating about it and mitigating its impacts.
1. Account within existing scopes.
Purchased cloud and AI tools sit in Scope 3, Category 1 (purchased goods and services). Server purchases fall in Category 2 (capital goods), and owned data centers fall under Scopes 1 and 2. For most businesses, AI therefore sits in Scope 3.
The GHG Protocol, the Science Based Targets initiative (SBTi), and the Corporate Sustainability Reporting Directive (CSRD) do not yet have AI-specific accounting rules. Across the frameworks and standards, AI sits in existing scopes and gets included when it’s material by size, influence, or risk. Companies with material emissions from cloud computing and AI may want to track updates from GHG Protocol and SBTi, such as the GHG Protocol’s Scope 2 revision and how version two of SBTi’s Corporate Net-Zero Standard may have implications for emissions mitigation across supply chains over which an individual company may have little control.
In addition, there are no widely industry-accepted AI emission factors yet. In the absence of standard AI emission factors, companies should document their assumptions now and expect methodology and guidance to keep evolving. Spend-based estimates are a practical place to start.
However, as AI emissions become more material or need closer tracking, Watershed’s open framework offers a way to standardize estimates and connect them to decisions. Covering the full system, from model training and inference to data center overhead and hardware, the framework measures emissions in kilograms (kg) of CO2e per million tokens, linking emissions to the token usage companies already track for costs, and offers three calculation tiers matched to data quality. Because it also reports electricity use, companies can see which choices drive their footprint, such as the model, the region, and how prompts are written. As providers disclose more data, this method of calculating emissions can also help companies compare vendors, factor emissions into procurement, and track efficiency gains.
2. Communicate to empower change.
Accounting for the impact of AI should help a company understand how material it is to the business and what other considerations it raises. From there, the company can decide which stakeholder groups to communicate with, and how, based on their level of interest and the issue’s relevance to them.
Because AI’s environmental impact is rarely material today, few companies report it externally. Companies that do report on AI tend to focus on governance, though a few go further. For example, Ingka Group (IKEA) has published a Digital Ethics and AI stance that commits to sustainable AI, while Etsy breaks out cloud computing emissions, including AI, as a separate line within Scope 3, Category 1.
An important stakeholder group to engage with is upstream vendors. Currently, few AI and cloud vendors provide information on environmental impacts. Google and OpenAI publish per-query impact data, but even these say little about a customer’s own footprint. Going forward, we recommend asking vendors for account-level energy, emissions, and water data, ideally as part of the procurement or contracting process. These requests signal that customers value lower-impact services, which can encourage vendors to measure and reduce their impacts.
Employees are also looking to understand how material AI impacts are to the business and, importantly, what they can do to help reduce those impacts. Even if individual AI use is small, engaging with employees on this can prompt discussions about the company’s broader sustainability work and how they can support progress. At a minimum, companies should provide responsible-use guidance that goes beyond data protection and privacy to help users select the right-sized model for each task, driving greater efficiency within the company.
3. Mitigate impact.
Mitigation should align with materiality and start with governance and efficiency. Companies can consider environmental impact when deciding where to use AI, skip uses that add little value, and encourage employees to pick the simplest AI tool that gets the job done. For example, running AI more efficiently—using processes such as caching (reusing work the AI has already done) and batching (processing non-urgent requests together)—and designing AI software and hardware to run together can reduce the energy each query uses. Many AI providers offer caching and batching as options, often at a lower price.
Over time, companies can favor providers and hosting regions with cleaner power and less water stress. Google’s Nevada data centers, for example, draw on geothermal power from Fervo’s plant, which has fed the local grid since 2023 and, unlike solar and wind, runs day and night. A utility tariff lets Google pay for 115 more megawatts (MW) without raising other customers’ costs. Clean power at one site lowers a company’s footprint only if its workloads actually run there, so companies can ask providers which regions and power sources serve them and, where possible, choose cleaner regions. Even when they don’t, investments like this lower a provider’s overall emissions and signal a broader commitment to reducing impact.
AI can also support sustainability goals, from demand forecasting that reduces overstocking of perishables to extracting emissions and sourcing data from supplier documents. Walmart uses AI forecasting, inventory tracking, and sensors to flag spoilage early, and reported a 3.45% reduction in operational food loss and waste intensity in 2025 toward its goal of a 50% cut by 2030 against a 2016 baseline. These benefits are not designed to offset AI’s own emissions but rather point to ways AI could support sustainability aims.
Next steps
In another Pure Strategies blog, we covered the importance of understanding AI use across a business and within a sustainability function, and of governing it appropriately and managing potential risks. The next step is to screen AI impacts for materiality and manage them accordingly: account for them within existing scopes, ask cloud and AI suppliers for energy, emissions, and water data, make sure responsible-use guidance covers efficiency, and mitigate where impacts are greatest. While AI accounts for a small share of impact today, acting now gives companies a head start as AI use and its footprint grow.
