AI’s Environmental Impact: Calculated and Explained

Large Language Models (LLMs) represent a breakthrough in Artificial Intelligence (AI). AI is transforming everything, from online information retrieval to business operations and decision-making processes. Pioneers like OpenAI, with ChatGPT, alongside Google's Gemini, Meta's Llama, Anthropic's Claude, and others, are at the forefront of this AI revolution. 

‍

But here’s what no one’s talking about: AI can be pretty terrible for the planet. 

‍

Let’s understand AI’s environmental impact and see what we can do to reduce emissions from AI.

‍

Why do AI emissions matter?

AI’s backbone consists of vast data centers that house thousands of machines that require continuous power. The energy consumed by AI systems worldwide could surpass the annual consumption of entire countries like Denmark or Portugal.

Below, I use a high-level example to illustrate the carbon impact of ChatGPT, the most widely used large language model (LLM). 

What is the carbon impact per ChatGPT query?

Each query sends out roughly 4.32 grams of CO₂e (MLCO2), which may seem trivial on its own but adds up millions of queries a day, and you're looking at a staggering daily output.

‍

To put that into perspective, 1 million messages sent to ChatGPT is equivalent to (EPA):

• 11,001 miles driven by an average gasoline-powered passenger vehicle
• 4.3 acres of carbon absorbed by U.S. forests in one year
• 349,258 smartphones charged

‍

The industry's response to environmental concerns

AI leaders are taking significant steps to mitigate this impact. OpenAI, for instance, partners with Microsoft to leverage Azure's carbon-neutral capabilities, aiming to curb emissions related to their vast AI operations. 

Their collaboration reflects a growing industry trend toward sustainability:

• Azure: Azure's data centers are committed to becoming carbon-negative by 2030. To reduce their carbon footprint, they use renewable energy and offset programs.
• AWS: Strives to achieve 100% renewable energy usage by 2025 and offers sustainability tools to help clients calculate and manage their emissions.
• Google Cloud: With its sustainable data centers, Google Cloud matches 100% of the energy used by its operations with renewable energy purchases.

‍

How to measure the environmental impact of AI

Here are five simple steps to estimate the carbon emissions from your AI workloads. 

Step 1: Figure out what you’re actually tracking

Measuring AI emissions begins with setting clear boundaries. Consider where AI is used in your organization and what activities you want to measure.

‍

Which do you fit into?

1. Are you training your own AI models on internal infrastructure?

If you’re training your own LLM models using your company’s owned servers, the emissions fall under Scope 1 and 2, as they stem directly from your controlled operations.

‍

2. Are you licensing AI capabilities externally, like OpenAI’s models or using APIs?

If you’re paying for server costs from external providers like AWS or Azure, those emissions shift to Scope 3 since the carbon impact is added under Scope 3 Category 1: Purchased Goods and Services.

‍

Step 2: Track how much compute power you’re using

AI runs on computation, and computation runs on electricity. To calculate your emissions, you need to track how much compute you're using. 

‍

For Internal AI Models

If your company builds and trains AI models in-house, you’ll need to calculate the energy usage of all hardware. This includes tracking energy consumption (in kWh) in stages such as the training, finetuning, and inference phases of the AI lifecycle.

‍

For External AI Models

If you’re licensing an AI model from external providers (e.g., OpenAI, AWS, or Azure), you won’t have direct access to their computing power or the servers running behind the scenes. Instead, focus on usage metrics and count the number of API calls or inference requests.

‍

For example, each time your app queries OpenAI's API, that action consumes compute resources.

‍

💡 Pro Tip

Most cloud providers offer dashboards or billing reports that list usage by compute hours, storage, and networking. Use those as your starting point.

‍

Step 3: Find your energy source

Electricity powers your AI, but not all electricity is created equal. Depending on whether your data centers (or cloud provider servers) are in areas powered by coal, natural gas, or renewables, the carbon impact can vary a lot.

‍

Figure out the location of the servers running your workloads.

• If your servers are based in a region with clean energy grids (like Norway or parts of Canada), emissions will be much lower.
• On the other hand, data centers operating in coal-heavy regions will produce significantly more emissions.

‍

Use Our World in Data’s tool to determine the carbon intensity of the energy grid in your region.

‍

Step 4: Use emission factors to calculate the carbon impact

Once you know how much electricity your workload consumes and where the servers are located, use emission factors to calculate how much CO2e (carbon dioxide equivalent) is produced per unit of energy.

‍

For example:

• Let's assume a typical data center in the U.S. emits 0.4 kg of CO2e per kWh of electricity used.
• If you know your AI workload used 1,000 kWh, calculate as follows:
• 1,000 kWh × 0.4 kg CO2e/kWh = 400 kg CO2e

‍

Step 5: Quantify total emissions using a carbon calculator

‍

Here are two reliable options:

1. Cloud Provider Emissions Calculators: AWS, Microsoft Azure, and Google Cloud all offer tools to estimate emissions for their customers.
2. Self-serve Carbon Accounting: Platforms like Arbor specialize in bottom-up, value-chain-specific carbon footprints, empowering companies with precise sustainability data.

‍

But remember—calculators are only as good as the data you put in, and actual improvement comes when you use your own data to optimize your value chain.

How can we reduce the environmental impact of AI?

Addressing the environmental challenges posed by AI requires a shift towards sustainable practices. This includes:

• Investing in energy-efficient hardware and algorithms: Developing AI models and hardware requiring less energy can significantly reduce the overall environmental impact.
• Powering AI operations with renewable energy sources: Transitioning to renewable energy sources like solar and wind power can mitigate carbon emissions associated with AI.
• Optimizing data center locations and cooling systems: Strategically locating data centers in regions with access to renewable energy and implementing efficient cooling systems can minimize environmental impact.
• Developing AI models focusing on resource efficiency: Designing resource-efficient AI models can help reduce energy and water consumption.
• Promoting responsible e-waste management: Proper e-waste disposal and recycling programs can prevent environmental contamination.

‍

The future of AI and the environment

Industry leaders must work toward sustainable solutions to curb AI's environmental footprint. This includes investing in energy-efficient technology, using renewable energy, improving data center operations, designing resource-efficient AI models, and responsibly handling e-waste.

AI and its systems will become more efficient over time, balancing their environmental impact with their positive benefits.

Measure your emissions with Arbor.