Commitments to massively expand infrastructure for artificial intelligence (AI) have accelerated significantly within the past year. The January 21st announcement of Project Stargate — a four-year, $500 billion push to scale AI infrastructure in the US by OpenAI, Microsoft, NVIDIA, SoftBank, and others — represents an unprecedented scale of infrastructure investment in a single technology sector. 

It’s likely that power-hungry AI infrastructure will continue to grow and will need to be served by electricity generation in some form, existing or new, renewable or fossil-fueled, on-site or across town. If managing the climate impacts of this growth is a priority, then emissions impacts should be considered when evaluating the options for how to meet the growing AI demand. Already, AI-driven electricity demand has increased the emissions of large companies like Google and is threatening their climate goals. 

Forecasts for data centers’ ballooning electricity demand

Lawrence Berkeley National Laboratory (LBNL) recently projected that data centers could consume between 6.7% and 12% of US electricity by 2028, a 2-3x increase from 2023. The corresponding load growth from data centers alone is in the range of 145-400 TWh, which may require some 33-91 GW of new generation capacity to be built by 2028. That's a massive amount of new electricity generation. That's also just for the US, a leader — but far from the only — player in the AI landscape.

The LBNL report is one to take seriously. It was created by researchers, some of whom have been loudly skeptical of AI, to fulfill a 2020 request from Congress. These numbers are in line with or relatively low compared to other recent projections. Meanwhile, Chinese AI startup DeepSeek’s announcement last week that it managed to produce a powerful model with a fraction of the compute compared to leading AI companies is a reminder that efficiency gains are likely. These numbers are constantly in flux.

Estimating the emissions implications of data center load growth

The implications of how new electricity is supplied are massive. If all LBNL’s projected U.S. data center load growth through 2028 were served exclusively by natural gas generation (the most prevalent source of generation in the US currently), it would result in about 180 million tonnes of additional CO2 emissions annually. If it were a country, that would make it the 45th largest emitter. 

The emissions impact of this massive buildout depends entirely on choices made today about where to site these facilities and how to power them. A particular computing load, depending on time and location, could cause coal to be burned in Wyoming, or gas to be burned in California, or could cause no emissions at all if it absorbs surplus wind power in Kansas or Texas.

The speed demanded by AI development timelines creates pressure to choose quick solutions over optimal ones. While co-located gas-fired power might seem like an expedient and economically wise choice, particularly given current political signals or to avoid long interconnection queues, it also comes with future fuel cost risk and creates long-term emissions lock-in that will be increasingly difficult to unwind as climate pressures mount.

Ultimately, the emissions impacts are dependent on which power plants will serve the new electrical demand of these facilities. When a new grid-connected data center is switched on, one or more existing power plants will ramp up to meet the demand. In some cases, new power plants will be built to make sure enough generation capacity is available to serve the new load. 

The emissions caused when existing power plants respond to changes in load (or new plants are built) are measured by marginal emissions rates. We can use these marginal emissions rates of electricity grids to compare the climate outcomes of different data center scenarios — both the emissions caused by which data centers get built where, and also the emissions consequences of associated impacts on marginal emissions by that load and any new power plants that get built.

Siting new data centers to cause fewer emissions

Let's compare and contrast two significant data center hubs: Northern Virginia's Data Center Alley (in the PJM grid) and Texas's emerging AI corridor (in ERCOT), where the first Stargate data center is being constructed in Abilene. The induced emissions impact of a grid-connected 100 MW data center operating at 95% capacity differs substantially between locations.

A 100 MW data center in Northern Virginia would result in about 463,000 tonnes of CO2 emissions annually, while the same facility in Texas would produce about 386,000 tonnes (17% lower).

These, of course, aren’t the only places where new data centers could get built, and in fact, neither location represents the optimal case from an induced emissions perspective.

For example, the same 100 MW facility built in Kansas (in SPP) would produce about 358,000 tonnes of CO2 annually (23% lower than in Virginia). Further, building it in Northern California (CAISO) would produce about 309,000 tonnes (an even greater 33% reduction vs. Virginia). What Kansas and California have in common is an oversupply of clean and renewable energy for many hours of the year — wind in one and solar in the other.

These calculations assume constant operation near maximum capacity throughout the year, typical for large data centers with critical workloads. While actual emissions would vary based on specific operating patterns and grid conditions, these numbers illustrate the massive emissions implications of siting decisions for new AI infrastructure.

Siting data centers in grids with lower marginal emissions rates can cut the potential induced emissions by up to half. That’s massive.

Building new clean power where it can avoid more emissions

As new data centers increasingly look to bring their own clean power (or procure it), this also opens the question of where that new clean generation should get built to not just meet data center load growth but also avoid the most fossil emissions. (In practice, which power plants get built where has many influences, including the capacity needs of a specific balancing area, interconnection queues, transmission constraints, and other factors. But for now, let’s assume total freedom to choose your location.)

With this in mind, siting new data centers on grids with lower marginal emissions rates is only half the story. The electricity generation supply side of the equation is the other.

There are two dominant ways new power plants could get built for data centers: a) co-located within the same grid balancing area as the data center itself, and b) siting the new clean power on grids with higher marginal emissions rates, and thus where new wind or solar could avoid more fossil emissions.

Continuing our earlier example of would-be new data centers in either Northern Virginia’s Data Center Alley or Texas’s emerging AI corridor, let’s look at a few scenarios for emissions implications, depending on where new wind or solar capacity gets built in association with the new data center load. For example:

• Building a new data center in VA or TX, and co-locating new renewables there, too
• Building a new data center in a grid region with lower marginal emissions rates, such as Northern CA or KS, and co-locating new renewables there
• Building a new data center in a grid region with lower marginal emissions rates, and siting new renewables in other grid regions with higher marginal emissions rates and where wind or solar would avoid more emissions (such as Colorado-Wyoming or Kenya)

We see a clear pattern. When renewables sized to 100% of data center load are procured from within the same grid as the data center, those renewables have a more-modest avoided emissions effect relative to the data center load’s induced emissions. On the other hand, siting data centers on grids with lower marginal emissions rates — and then investing in new renewable capacity on grids with higher marginal emissions rates (where wind and solar displace more fossil fuel generation) — can generate substantial net reductions in total emissions. This approach to building renewable energy in the most impactful places regardless of where data centers are built is already being used by Amazon, Meta, Apple, and Salesforce.

The role of compute load shifting to further reduce emissions

While siting decisions have the largest impact on emissions, there's also potential to reduce emissions of data center use through smart load management. Data centers, particularly those running AI training workloads, require extremely reliable, constant power. Once a training run starts, interruptions can waste days or weeks of compute time. This makes them less flexible than other types of new electricity demand like EV charging, where timing can be shifted to match clean energy availability. 

But many data center workloads, like batch processing and cooling, are timing flexible, so shifting that energy use to times of the day when marginal emissions are lower, like when renewables are being wasted, can achieve large emissions reductions.

While load shifting alone won't solve data center emissions, it represents another tool for reducing emissions impact, particularly for facilities that handle workloads beyond AI model training that are flexible. These techniques are already being used by Microsoft, UBS, and other members of the Green Software Foundation.

Data centers will require massive amounts of energy, even if we don’t know precisely how much. And the combination of high reliability requirements and constant load patterns means careful planning is crucial — rushed infrastructure decisions could lock in unnecessarily high emissions for decades. 

Conclusion 

At a moment when electrification is accelerating across the economy, from vehicles to buildings, the surge in AI infrastructure presents both a challenge and an opportunity. By making smart decisions now about where to build data centers, how to power them, and how to operate them, we can ensure a revolution in compute drives rather than hinders the clean energy transition.

The unprecedented scale of AI infrastructure investment — from Stargate's $500 billion commitment to the broader industry — represents the largest concentrated buildout of computing power in history. Every decision about where and how to build this infrastructure matters more than ever.

image source: iStock | Gerville

For companies and other organizations investing in new renewable energy projects, two main strategies guide their procurement:

1. 24/7 Carbon-Free Energy (24/7 CFE): focuses on hour-by-hour megawatt-hour (MWh) matching of renewable generation’s timing and a corporation’s electricity demand load profile, with the clean energy procured from the same grid region where the electricity load is located
2. Emissionality: an emissions-first approach that targets the dirtiest grids globally, procuring clean energy from wherever the new renewable capacity will have the greatest avoided emissions benefit by displacing generation from the most-polluting fossil-fueled power plants

These two strategies have important implications for where clean energy investment will flow and where new renewable capacity will get built, and consequently, on how much (or how little) climate benefit those projects will ultimately have.

In this analysis, we use marginal emissions data from the United Nations Framework Convention on Climate Change (UNFCCC) to gain insights into these questions, especially: Does location matter for the avoided emissions benefit of a new renewable energy project? And if so, where should new renewable energy projects get built to have the greatest overall climate benefit?

Tapping the UNFCCC’s marginal emissions data

To better understand the beneficial impact of new renewable projects across the globe, the UNFCCC — the UN body that oversees the Paris Agreement — developed a methodology for estimating the long-term impact on grid emissions in each country around the world. UNFCCC’s combined margin emissions factors take into account both operating margin and build margin, giving a sense for renewable energy’s climate benefit in the nearer and longer terms.

Using data from the IEA’s Global Energy and Climate Model (previously known as the World Energy Model) — which underpins IEA’s annual World Energy Outlook — UNFCCC experts calculated the marginal emissions rates resulting from predicted new generation across both traditional firm energy sources (e.g., fossil fuels, nuclear, geothermal) as well as clean energy technologies (e.g., solar PV, wind, tidal). The IEA model incorporates information on existing energy sources, economics, and policies in 26 large countries and regions, with additional regression modeling for other countries, making it comprehensive in breadth and scale.

Mapping renewable energy’s avoided emissions potential

Looking at a global heat map of marginal emissions rates for new renewable energy sources, the greatest avoided emissions potential based on UNFCCC data is primarily located in the Global South, in countries spanning Asia, Africa, and Eastern Europe (red shades on the map). These countries’ grids tend to rely on heavier-polluting sources of generation, such as coal-fired power plants.

Conversely, the lowest avoided emissions potential is primarily in the Global North, in countries spanning the EU and North America, as well as select countries elsewhere around the world where hydropower (and sometimes, nuclear) provides a dominant share of electricity generation (blue shades on the map).

Multiplying the avoided emissions benefit of renewable energy investment

For any organization deciding where to invest in new renewable capacity, using marginal emissions estimates like these from UNFCCC can lead to much larger reductions in overall global emissions.

In Annex I countries (which largely overlaps with the Global North), the average avoided emissions rate (weighted by total electricity generation) is 345 g CO2/kWh. Meanwhile, countries in the top 50% of most-polluting power grids have an avoided emissions potential of 702 g CO2/kWh, and countries among the top 10% of most-polluting electricity generation have an avoided emissions rate of 979 g CO2/kWh. These heavier-polluting power grids are predominantly throughout the Global South.

In other words, investing in renewable energy projects across the Global South can yield 2x to nearly 3x greater climate benefit vs. renewables projects in the Annex I countries of the Global North. Organizations considering siting renewable energy — whether bilateral national agreements now being drafted under the revised Article 6 framework of the Paris Agreement, or voluntary corporate actors using the GHG Protocol — may want to consider UNFCCC’s data when deciding where to invest in renewable energy projects.

How procurement approach influences renewable energy’s potential

24/7 CFE proponents argue that it encourages the buildout of renewables that would generate during “off” hours, helping power grids move closer to 100% clean energy around the clock. This may be partly true. But it also amounts to massive investment aimed at “squeezing the last drop” of emissions from grids that have already significantly decarbonized. This misses opportunities for major larger global decarbonization by building renewables in other places where they’d have greater avoided emissions benefits and where coal-fired generation still dominates the grid mix.

Moreover, some of 24/7 CFE’s biggest proponents are major tech companies whose operations and data centers are overwhelmingly located in the EU, US, and other Global North locations. These are regions that have already seen large investment in new wind and solar capacity, especially. Meanwhile, Global South locations — the same places where UNFCCC marginal emissions data show there are the greatest avoided emissions opportunities — have seen chronic underinvestment in clean energy technologies, according to IEA data.

From a global climate action perspective, it’s far less impactful to inch California or Texas (where wind and solar have already made huge gains) closer to 100% carbon-free energy than it is to invest in new renewables in a place such as India, where coal still contributes more than 70% of the nation’s electricity generation and clean energy investment in 2024 was just one-fifth of what it was in the US.

On top of this compelling climate argument, there’s also the crucially important humanitarian component, too. Investing in renewable energy in Global South countries will also bring economic and health benefits by expanding energy access and reducing air pollution in the places that also have the worst air quality. Globally, 1 in 8 deaths are now attributed to air pollution, predominantly in countries with the most-polluting electric generation, since the same power plants spew both carbon dioxide and PM2.5.

Conclusion

The UNFCCC model is not the only model for estimating marginal emissions rates. One key difference from WattTime’s MOER model is that the UNFCCC model does not account for imports. This can significantly affect rates when low-emission countries border high-emission ones, as is the case with Sweden and Finland. In terms of long-run build margin, the UNFCCC also lacks many of the more-detailed features of other models such as Cambium, GenX, and PyPSA. In particular, it does not consider variance in emissions rates within a country, which can be great in large countries such as the US or China. But it is one of the only existing models that covers the entire globe, which is a critical consideration when evaluating emissions reductions. 

More research is needed to evaluate these different modeling approaches and to develop more detailed models across the globe, so that renewable energy investments can be targeted at the location where they have the most impact. For now, one thing is clear: data is increasingly pointing to the Global South as a critical focus for the world’s future renewable investment.

image source: iStock | rvimages