Generative AI's data center electricity mix
An excuse to play with maps!
The explosive rise of Generative AI has ignited worries about parallel growth in carbon dioxide emissions. “Artificial intelligence is booming — so is its carbon footprint,” cautioned Bloomberg in a March 2023 piece.
The concern over AI carbon dioxide emissions has attracted the attention of researchers at AI technology companies who have started to address the issue in their published articles. For instance, the recently released “Segment Anything” research paper from Meta contained an “Ethical considerations” section acknowledging the release of 2.8 metric tons of carbon dioxide while training the AI model. That’s about the same as 7,000 miles driven in an average gasoline-powered U.S.-sized car.
Today, we're delving into an intriguing aspect of AI carbon emissions: the electricity mix powering the data centers where AI models are trained and deployed.1 Electricity mix refers to the combination of different power sources, like wind or coal, within a region’s power grid.
Initially, I believed the electricity mix of a region’s power grid would play a significant role in Generative AI carbon emissions. But what I found is that individual data centers have some control over their own electricity mix, which is likely more important for carbon output than the mix of the local power grid.
Why are AI carbon emissions increasing?
As AI systems advance in complexity, their computational demands surge. This is evident not just in the number of AI applications but also in the more data-intensive training processes and the expanded size of the AI models themselves. The graph below illustrates the rising computation, measured in a unit called petaFLOPs, used to train notable AI systems. This growth has been exponential and more computation generally means more carbon emissions.
Yet, even two AI models of identical size trained with identical data on identical computer hardware can have vastly different carbon emissions. This disparity is due to the emissions characteristics of data centers where AI models are trained and subsequently deployed for use by customers. The electricity mix — the blend of various electricity generation sources such as wind or coal — that powers these data centers can have a direct bearing on their carbon emissions. We’ll talk more about electricity mix in a moment.
While AI training and deployment represent a growing share of data center operations, these centers also handle cloud storage, web hosting, database management, and other non-AI computer processing. Still, thinking about Generative AI carbon emissions using the data center as a proxy can be instructive. OpenAI, for instance, has publicly announced that their AI systems, like ChatGPT, run at Microsoft Azure data centers. AI company Anthropic is hosted by Google Cloud.
Not all Generative AI models rely on data centers for training, though. A notable exception is the BLOOM Model, which was trained on the Jean Zay supercomputer situated at the French National Centre for Scientific Research. But, even in such cases, the emissions from these alternative training sites are likely similar to typical data center emissions.
Globally, data centers account for about 1% of the world's total carbon dioxide equivalent emissions. To put that in perspective, their emissions are roughly half of what the aviation industry produces. A significant chunk of these data centers, more than a third, is located in the U.S.
Thanks for reading 96 layers! Subscribe for free to receive new posts and support my work.
What is carbon intensity?
Carbon intensity measures the emission rate of a pollutant relative to a specific activity, such as the amount of carbon dioxide released per unit of electricity produced. This metric aids in comparing the environmental effects of different fuels, activities, and industrial processes.
Carbon intensity is measured in grams of carbon dioxide and equivalent greenhouse gases — written “gCO₂eq” — divided by kilowatt hours, written “kWh.” Therefore, the full measurement of carbon intensity is written as gCO₂eq/kWh.
Carbon dioxide equivalents (CO2eq) are commonly utilized as a standard measure for comparing the global warming potential (GWP) of various greenhouse gases to carbon dioxide (CO2). As an example, methane has a GWP 25 times that of CO2. This means that it is equal to 25 CO2eq.
What is electricity mix and why might it matter for AI carbon intensity?
The carbon intensity of any activity depends on the method of electricity production used. As an example, according to the best available research, wind power has a carbon intensity of around 13 gCO₂eq/kWh, while coal has a much higher intensity of 1,000 gCO₂eq/kWh. This means that if an AI model is trained on a power grid predominantly fueled by coal, it would result in 75 times more carbon dioxide emissions compared to a grid powered totally by wind, assuming all other factors remain constant.
To draw an analogy with automobiles, a difference in emissions of 75 times is akin to comparing the carbon footprint of a drive from New York City to Stamford, Connecticut (40 miles) with that of a cross-country journey from New York City to San Francisco (2,900 miles).
Geographical topology, economic resources, and governmental policies all contribute to the variation in electricity sources across different countries and regions. For instance, areas with access to rivers can tap into hydroelectric power, while other regions may rely more heavily on fossil fuels. Consequently, carbon intensity differs significantly based on the electricity mix in each region. Countries that incorporate renewable electricity sources alongside fossil fuels demonstrate lower carbon output per unit of electricity (see map below). For instance, consider the contrast between Norway and South Africa, with average carbon intensities of 28 gCO₂eq/kWh and 716 gCO₂eq/kWh, respectively. This translates to a staggering 25-fold difference in average carbon intensity.
Considering the importance of data centers in AI model training and deployment, the location of these facilities seems crucial in assessing the overall carbon-equivalent output. The choice of a data center powered by renewable electricity versus one heavily reliant on fossil fuels could have a substantial influence on the environmental footprint of AI activities.
The complexity of power distribution extends within national boundaries. Some countries do not rely on a single nationwide power grid with consistent carbon intensity. Take the United States as an example. It has more than 11,000 power plants spread across the country, employing various methods of electricity production (see map below).
Some of these power plants serve vast regions through a network of long-distance transmission lines. Take the Bonneville Dam, for instance, located in Northwestern Oregon. Despite its physical location, it provides power to eight western states, including Washington, Oregon, Idaho, Montana, Wyoming, Utah, Nevada, California, and even parts of Canada. Its reach extends as far as Los Angeles, about 1,000 miles to the south.
Diverse power generation methods, combined with short- and long-distance power transmission, along with the involvement of federal, state, and local utility oversight, contribute to the blending of power sources into about 50 different administrative zones across the United States. Each of these zones exhibits its own quantifiable carbon intensity.
These zones are shown in the map below, created using the wonderful Electricity Maps website, which tracks both real-time and historic electricity metrics.
The regional variation in carbon intensity within the United States can be as significant as the differences observed between countries. For instance, the average carbon intensity over the past 12 months is as low as 24 gCO₂eq/kWh in parts of the Pacific Northwest,2 while it reached as high as 660 gCO₂eq/kWh (with occasional spikes above 900 gCO₂eq/kWh) in South Carolina.3 This amounts to a difference of 27-fold.
Shifting to an example of intranational power regions outside the United States, Italy has six distinct zones, with carbon intensities ranging from 264 gCO₂eq/kWh in Central North Italy to 615 gCO₂eq/kWh on the island of Sardinia, while the Balkan countries each have a single zone per country.
Utilizing this variation at both the country and regional levels, we can estimate the carbon intensity associated with different data centers worldwide, allowing each center to inherit the carbon intensity characteristic of its respective parent region. It’s important to note that this method used to calculate data center carbon intensity is approximate.
As an example, let's consider selected data centers from Microsoft's Azure Cloud offering. Using this simplified methodology the data center located in Gävle, Sweden, would be anticipated to have an average carbon intensity of around 20 gCO₂eq/kWh. In contrast, the data center situated in Johannesburg, South Africa, would be expected to exhibit an average carbon intensity of approximately 700 gCO₂eq/kWh. This signifies a substantial disparity of 35 times in terms of carbon intensity.
But we can still go one level deeper, to the individual data center…
Why power grid electricity mix might not matter that much
Carbon intensity starts to look more optimistic if we go one level deeper to the electricity mix of an individual data center. That’s because there are numerous ways data centers can optimize electricity usage. This can include upgrading hardware, using improved AI algorithms that reduce computational intensity, and using already trained AI models to avoid the carbon footprint of retraining.4 Data centers can also make improvements to core building systems like lighting and cooling. This reduces the so-called power usage effectiveness (PUE), the ratio of total facility energy usage to energy usage dedicated to IT equipment.
On a bigger scale, reducing data center carbon intensity can include augmenting, or even replacing, the electricity provided from the local power grid. For instance, despite India being in the top quintile for carbon emission intensity, certain data centers from the provider CtrlS have implemented solar technologies and other electricity-saving techniques. As a result, they have achieved LEED Platinum Certification for several Indian data centers, a recognition for buildings that strongly demonstrate environmental performance and sustainability.
Given the scrutiny and criticism they face, major technology companies have ramped up sustainability initiatives in data centers, too. Amazon AWS, Google Cloud, and Microsoft Azure have all taken steps to mitigate their data center carbon emissions. AWS is already powered by 90% renewable energy. Google Cloud aims to run on carbon-free electricity by 2030, irrespective of the energy mix of the local power grids serving its data centers.
Enhancing the electricity mix of data centers often means establishing a direct power supply nearby. For instance, Google collaborated with a wind-turbine company to construct 23 wind turbines in Chile’s Biobío region, allowing the adjacent data center to operate with 80% carbon-free energy. Likewise, nearby solar farms can be built to provide data centers with electricity. Apple even used biomass to power one of its data centers in Viborg, Denmark, though biomass remains impractical in most cases. Finally, there is speculation that Microsoft will set up micronuclear plants to power its data centers.
Another way to mitigate data center carbon output is to simply purchase carbon offsets. Microsoft Azure data centers have been carbon neutral since 2012 if carbon offsets are included in the calculation.6 Microsoft also aims to have data centers shift to 100 percent renewable energy by 2025.7 (However, note that despite its carbon neutral commitments, Microsoft Scope 3 CO₂eq emissions increased between 2021 and 2022 as outlined in their latest Environmental Sustainability Report. Scope 3 emissions contain many data center activities).8
A longer-term approach involves “power purchasing agreements” (PPAs). In these arrangements, major data center providers commit to buying energy from power companies, often for periods ranging from 5 to 20 years.9 Such agreements are meant to fund the capital and infrastructure expenses needed to build renewable energy. PPAs thereby promote the gradual greening of a region's electricity mix.
These measures, combined with other strategies like more efficient computer hardware, have helped curb data center carbon emissions worldwide. Despite the massive growth in computing workloads in data centers — a staggering 550% increase between 2010 and 2018 — the electricity consumption of these centers has only gone up by 6%.
It’s important to remember that while renewables have numerous advantages over traditional sources of electricity, challenges still exist. For starters, they require vast tracts of land. In countries with densely populated regions, like Taiwan or the Netherlands, finding the necessary space can be problematic. Remote locations are often needed, which can necessitate the construction of long-distance transmission lines to deliver power to data centers, introducing further complications. For instance, these power lines and their associated infrastructure can disrupt forests, wetlands, and other pristine environments. In places with rich biodiversity, like parts of Brazil, there are concerns about potential habitat disruption.
Additionally, the manufacture and eventual disposal of solar panels — termed ewaste — in countries with lax environmental regulations can lead to chemical pollution. Though, note that there are programs underway to recycle solar panels, including in lower-middle-income countries such as India.
Moreover, these energy sources are weather-dependent. Solar panels, for example, rely on sunlight, which can be inconsistent during cloudy days or during certain seasons. Wind turbines, on the other hand, need wind, which can be unpredictable. This irregularity necessitates the need for energy storage via lithium-ion batteries in order to ensure steady capacity. Though creative solutions are on the horizon, storing energy in batteries has pitfalls. Battery production involves mining for materials like lithium and cobalt, which can lead to environmental degradation and labor violations.
Still, when zeroing in on just the electricity mix, the ability of data centers to be directly powered by renewables allows them to counterbalance high carbon-intensive regional grids. The electricity mix of a regional power grid can absolutely matter for data center carbon emissions. But what matters more is the electricity mix of an individual data center.
The “Segment Anything” paper notes that they open sourced the model to avoid other teams having to retrain it.
About Amazon the report noted, “In the meantime, there is neither explicit clarity on the coverage of its target — including whether it just refers to carbon dioxide emissions or to all greenhouse gases — nor on the extent to which it plans to achieve the target through delivering actual emission reductions, as opposed to procuring offset credits.” About Google the report noted, “The carbon neutrality claim is derived through the procurement of renewable energy and offset credits, and covers only selected emission scopes. Major Scope 3 emission sources that accounted for 60% of the company’s GHG emissions in 2020 are omitted from the carbon neutrality claim. For the emission scopes that are covered by offsets, the environmental integrity of the offset credits is highly contentious. The scope of the carbon free by 2030 target is unclear, but this claim appears mostly alongside Google’s description of its plans for renewable electricity, and may only apply to renewable electricity generation and procurement.”
This is noted in Microsoft’s 2021 Environmental Sustainability Report (emphasis added): “2021 was a year of both successes and challenges. While we continued to make progress on several of our goals, with an overall reduction in Scope 1 and Scope 2 emissions, our Scope 3 emissions increased year over year, due in substantial part to significant global datacenter expansions and the growth in Xbox sales and usage as a result of the pandemic.”