Data center capacity growth is key to the U.S tech industries' development of AI and cloud computing technologies. Related increases in power consumption and emissions appear incompatible with many tech companies' carbon-reduction goals.

Why it matters:  S&P Global Ratings' estimates that U.S data center power demand will increase at 12% per year until the end of 2030. That could double the tech sectors' current carbon emissions as constraints on renewable generation growth, coupled with data centers' requirement for stable power, mean about 60% of new demand in the U.S, to 2030, could be met by natural gas.

What we think and why:  Increased carbon emissions from data centers are unlikely to be a material credit risk to the operators in the short term given their key role in supporting AI technologies and economic growth. The aggregate increase in annual carbon emissions from data centers is relatively small compared to other high-emitting sectors, or even some companies--our estimate of this emissions addition by 2030 is equivalent to about 2%-4% of 2023 U.S. power-related emissions, or an average U.S. oil companies' direct (Scope 1) greenhouse gas (GHG) emissions in 2023. But as one of the few sectors rapidly expanding power use, headwinds and challenges could emerge.

Major U.S. Tech Companies' Emissions Will Likely Increase

As cloud services have become a mainstay of how businesses and consumers use the internet, the energy use associated with them has increased. Services such as social media, AI, entertainment, banking, and business applications increasingly rely on data centers and their infrastructure. Continuous data collection, processing and storage, ongoing cloud migration, and AI will be key drivers of increased demand for data centers over the coming decade.

Carbon emissions from data centers mainly relate to the high power demands of their processors and data storage systems. Cooling is also a major source of power usage--data centers generate a lot of heat that needs to be dissipated to maintain optimal performance. We understand cooling can account for 25%-40% of data center energy use, though performance can vary greatly depending on the location (and thus ambient temperatures) and the cooling methods used.

Estimates suggest that data centers and their associated networks were responsible for about 170 terawatt hours (TWh) of U.S. electricity demand in 2023. That equates to about 4% of total demand, representing about 75 million tons of CO2.

Tech Sector Renewable Use May Not Provide Net Benefits

Tech companies have been amongst the most proactive in setting targets that aim to address their environmental impacts. Many target substantive reductions of direct carbon emissions (Scope 1) and indirect emissions from purchased electricity (Scope 2) by 2030. Others have been even more ambitious, aiming to become net zero or carbon neutral in the same timeframe, including across their whole value chain (Scope 3), albeit using mechanisms such as carbon credits or removals to support these goals (see table 1). Most also have specific targets for renewables or other low-carbon power use.

Table 1

Tech companies are investing heavily in sourcing low-carbon energy and in efficiency projects, but how these translate to emissions depends on perspective.   Using the market-based accounting method, which can account for actions such as specific renewable energy contracts, emissions linked to power use at the top U.S. tech companies have been flat since 2019 despite overall growth in power demand. Using the location-based approach, emissions have scaled with power use, which has increased at major companies at a compound annual growth rate (CAGR) of about 19% since 2019 (see chart 1). A description of the two accounting methods is provided in an appendix to this article (see "Appendix 1: The Complexities Of Power-Related Emissions Accounting").

The market-based approach is more commonly used to define company targets as it can more clearly reflect individual company actions. Also, while our analysis of major tech companies (see chart 1) suggests that market-based emissions have been flat, significant differences exist between the companies included in our sample.

Chart 1

The tech sector's increasing power demands could contribute to increased emissions even if the sector secures low-carbon power.   The utilities sector is investing in renewable capacity, notably encouraged by policies such as the Inflation Reduction Act. But all things being equal, every MWh of additional renewable power that is secured by a tech company is renewable power not available to other consumers. Since utilities are typically obligated to provide power, if they can't meet rising demand with low-carbon generation they are forced to use fossil fuels. So, while tech companies' emissions balance sheets might look good, increased demand can put pressure on utilities' own emissions as reflected in the location-based approach (see chart 1).

Tech Companies' Growth Complicates Their Emissions Goals

No matter which accounting method is used, increases in emissions could put targets under pressure. Many major tech companies have set targets of a 25%-50% reduction in power-related emissions, which they aim to meet despite increased power demand and competition for renewable capacity. That could prove difficult, especially if gas power is needed in the short term.

We think power usage by data centers could rise by 150 to 250TWh per year between now and 2030 (see " Data Centers: Surging Demand Will Benefit And Test The U.S. Power Sector," Oct. 22, 2024). Based on our assumption that about 40% of this demand could be met by wind and solar, and the remainder of by natural gas, this could add another 40 million to 67 million tons of CO2 emissions by 2030 (see "Appendix 2: Emissions Calculations For Additional Power In 2030"). Those additional emissions would represent about another 2-4% of the U.S.'s 2023 power-related emissions and could slow the overall pace of decarbonization in the grid.

The main constraint on the growth of new data center facilities will be utilities' ability to support new grid connections and to deliver power (see "Data Centers: Welcome Electricity Growth Will Fall Short Of U.S. Data Center Demand," Oct. 22, 2024). Given these constraints, we expect existing and under-utilized fossil fuel or nuclear generation to fill the short-term supply gap, adding to both tech companies and utilities' emissions. An uptick in natural gas, or other fossil fuel-powered, generation risks raising overall grid emissions and therefore the emissions associated with all power users.

Data Center Operators' Environmental Challenges

Chart 2

We have identified five key environmental challenges that could prove to be risks for data center operators.

Challenge 1: Securing stable, low-carbon power

Data centers' growing demand for stable power will likely be prioritized over clean power, in the short term.   To date, tech companies have relied mainly on power purchase agreements (PPAs) with renewables generators to deliver clean power. Tech companies compete with other industries for new capacity, and currently dominate in securing PPAs, according to S&P Market Intelligence, a division of S&P Global (see chart 3). That is partly because, larger, investment-grade hyperscalers have the financial means to secure such contracts. Recent announcements from large tech companies on agreements to source nuclear power, and in particular on the development of small modular reactors, are emblematic of the search for significant volumes of clean and reliable power, though that will not meet their short- to medium-term needs.

Chart 3

Data center power demand growth is likely to outstrip low-carbon generation growth in some markets, leaving gas and oil-fired generation to fill the gap.   The addition of new renewable and nuclear capacity is a long-term process, and has already resulted in "connection queues" of many years in most major markets (see "Gridlock: Interconnection Queue Backlog Adds Risks For U.S. Not-For-Profit Power Sector, Oct. 8, 2024). For example, Virginia, which is home to the most data center capacity in the U.S., has seen commercial energy demand grow at over 14,000 gigawatt hours (GWh) per year since 2017, yet increased its low-carbon energy generation (nuclear, wind, solar, and biomass) by less than 4,500 GWh per year in the same period, according to the US Energy Information Agency. Meanwhile, in Texas, renewable deployment rates have far exceeded growth in commercial demand, meaning data centers' increasing needs can be absorbed more readily through, for example, more desirable physical PPAs with clearer emissions benefits (see chart 4).

Chart 4

On-site, low-carbon power for data centers remains niche, even amongst the most advanced companies.   Data centers requirement for 24/7 stable energy means any deployment of traditional renewables (solar or wind) needs to be supported by energy storage and other solutions that limit their potential. Deployment of small nuclear reactors has been cited as a potential solution, but is yet to be proven at scale and comes with long development timelines. For example, Google, Microsoft, Amazon, and Equinix have all announced that they aim to use the technology to provide continuous, stable power, but also signaled that the first units won't come online until 2030. Small-scale nuclear deployment also comes with increased cost uncertainty compared to more mature technologies, though that might not be a significant barrier to hyperscalers with ample financial resources.

Challenge 2: Securing further energy efficiency gains through innovation

Power's position as a key cost component of data center operations has made efficiency a key focus. Progress has been made across many operational aspects, including through the use of more efficient chips, better cooling systems, workflow optimization, and due to benefits of scale (see chart 5 and chart 6).

Chart 5

Chart 6

Efficiency gains have slowed.   The overall energy efficiency of data centers is often measured in terms of power usage effectiveness (PUE), often cited in regulations. While not a perfect indicator, given often complex ownership and operating responsibilities, it is a useful measure of the ratio of the total facilities power demand to the power demand of the facilities' IT equipment alone. A PUE close to 1.0 means almost all power contributed to useful output (processing of data). Global data centers' PUE improved significantly from 2007 to 2014, but has since remained steady at around 1.5 (meaning that for every unit of power required for the IT equipment, 0.5 of a unit is required for the rest of the facility), according to a study by The Uptime Institute (see chart 5). Further research suggests that hyperscalers are typically the most efficient operators, with data center PUE values close to 1.2 (see chart 6).

Innovation in computation, power supply, and cooling will be needed for further efficiency gains.   Computation components design is increasingly focused on specific applications that could improve efficiency. For example, inferencing engines (which can use learning from model training in an application without needing to frequently retrain the model itself) could reduce data center components by over 80%, compared to standard data centers, and would also run cooler. Liquid cooling, which has already been adopted by some companies, could reduce energy demand by replacing traditional air-cooling systems.

Challenge 3: Evolving environmental regulations, but lack of consensus for now

There is little specific U.S. regulation covering data centers, though California has, since 2015, used building regulations to enforce some efficiency requirements. Regulation is, however, emerging in other regions that could affect U.S.-based companies that operate globally. Germany, for example, has laws requiring both new and existing data centers to meet certain PUE levels over next four years (see table 2). Other nations have energy efficiency requirements for companies that exceed a certain amount of power usage and on commercial buildings.

Table 2

Regulations targeting existing facilities could necessitate new investment.   Regulations and initiatives in place today often target a PUE of 1.5 or less, while about half of operational data centers are estimated to have a PUE value higher than that, according to 451 Research. Generational changes in computational efficiency will help, particularly where they reduce heat load and thus cooling needs. Some older and smaller facilities may need to upgrade cooling and energy recovery systems to meet the PUE target, potentially incurring costs and downtime. Stronger regulatory requirements could affect retail data center operators hardest since these tend to be generally less efficient than larger scale wholesale and hyperscale facilities.

Challenge 4: The cost and risk associated with carbon credits and removal

Major U.S. tech companies have launched initiatives exploring the use of carbon offsets in the expectation they won't be able to avoid all of their emissions.   Based on our assumptions of emissions growth in 2030, using carbon credits to offset all the future increases in emissions would cost $600 million to $1 billion a year (assuming a carbon credit cost of $15/tCO2). The use of technology-based removal-based credits (so-called negative emissions solutions) in the place of offsets would be significantly more costly, while supply of such solutions couldn't meet demand.

Microsoft and Google have backed negative emissions solutions, including direct air capture and storage. Direct air capture projects currently cost at least $320 per ton of CO2 captured, according to S&P Global Commodity Insights, making them a very expensive solution. Carbon removed and stored using this solution is clearly measurable, supporting companies' claims that a real benefit has occurred. Other removal solutions, such as biochar (black carbon used as a carbon store) are also expensive at around $130 per ton of carbon captured. The scale and availability of these types of credits is very small, with even the most recent projects only able to deliver carbon credits at a scale of tens of thousands of tons per year.

Chart 7

Other types of carbon credits, such as nature-based schemes, bring potential reputational risks.   These are cheaper (about $5-$15 per ton) but are often closely scrutinized by stakeholders questioning if they deliver real benefits. Several controversies have highlighted those issues in recent years, including surrounding "avoided deforestation projects" which may not have delivered claimed benefits or might have happened anyway. Standards around carbon credits are evolving, and companies are increasingly adopting offsetting policies or principles that focus on improving the reliability and transparency of the carbon credit market. At the same time the carbon credits contribution to emission reductions is increasingly being scrutinized by standards setters keen to avoid over-reliance on offsets ( see "Carbon Capture, Removal, And Credits Pose Challenges For Companies", published June 8, 2023).

Challenge 5: Management of other environmental exposures

Many data centers use large amounts of water for cooling, while the location of data centers and their communication networks can have local effects on biodiversity.   Many data centers use large amounts of water for cooling, while the location of data centers and their communication networks can have local effects on biodiversity or be exposed to climate hazards. Water use is an area of increasing focus and could become more relevant to data centers choice of location and technologies, with many operators exploring cooling solutions that use less water. As fixed assets, they could also become increasingly exposed to climate hazards, such as droughts and floods, as could the networks they rely on. There could be also be tradeoffs. Liquid cooling solutions tend to be more energy efficient than air cooling, according to a study by chipmaker NVIDIA and Vertiv, an IT infrastructure support group. Lessons on management could be learned from industries, such as semiconductor manufacturers (foundries), which use very large amounts of water and power (see "Sustainability Insights: TSMC And Water: A Case Study Of How Climate Is Becoming A Credit-Risk Factor," Feb. 26, 2024).

The construction of new data centers and the equipment within them are part of the emissions burden for data center operators, and are growing.   Like many other sectors, disclosures of such emissions by the companies that finance data centers are improving. The split of emissions between different activities is driven by individual companies activities (see chart 8)--for example, Amazon's disclosures include its extensive vehicle fleet and construction of warehousing facilities. Emissions beyond power usage, such as those linked to cement and steel use, could become more relevant to assessment of total emissions and are often harder to abate (see "Decarbonizing Hard-To-Abate Sectors: Credit Quality Implications And Six Key Observations" published 25 June, 2024).

Chart 8

Can Tech Companies Continue On Their Current Path?

The combination of bullish growth projections and ambitious decarbonization goals means tech companies have mapped out a challenging path that will increasingly draw attention from customers, investors, and global regulators.

The sector has been proactive in voluntarily setting goals and securing low-carbon power, notably using its financial resources to advance plans. However, we expect that short-term constraints on low-carbon power will require greater innovation and more aggressive efforts to deliver solutions if tech companies are to meet their 2030 decarbonization goals. Regulatory reform, particularly outside the U.S., also suggests that the bar for energy efficiency could be raised and become harder to clear in the coming years, especially in places where data centers constitute a significant share of power demand.

Editor: Paul Whitfield.

Appendix 1: The Complexities Of Power-Related Emissions Accounting

There are two ways of accounting power related emissions: the location-based approach and the market-based approach. The approaches' differences result in varying outcomes and best practice likely involves the application of both methods.

The location-based approach:

• All power consumption within a grid (or territory) is assigned the same emission factor, which is based on the typical mix of technologies supplying the grid in a calendar year.
• The result then reflects a proportional impact, based on the broader power system and power consumption of all consumers, but doesn't necessarily reflect an individual company's actions.

The market-based approach:

• Emission factors are assigned to specific energy purchases, such as a particular energy contract. This means that if an entity agrees a renewable energy contract, or similar, it can claim the right to the lower emissions associated with the supplied power.
• The result can robustly reflect a low-carbon claim, for example where a physical PPA represents a direct connection between a supplier and consumer, especially with regards to a new power project.
• Yet the link can also be less direct, for example where renewable energy credits (RECs), or similar, are purchased separately from the actual power used (so called un-bundled certificates) and come from different power markets. That means a company could claim benefits, even where a REC doesn't result in any tangible reduction in emissions at the point of use. Unbundled RECs also come with an increased risk of double counting of benefits when they are traded between different entities.

For more on the accounting method for purchased emissions, see "Purchased Energy Emissions In Second Party Opinions And ESG Evaluations," March 23, 2023.

Appendix 2: Emissions Calculations For Additional Power In 2030

Table 3 is based on assumed total additional power of 150TWh to 250 TWh by 2030, and 0.269kg of CO2 per kilowatt hour (kWh).

Total assumed carbon emissions at the lower are thus 40.3 million tons of CO2 (tCO2) (150 TWh x 0.269kgCO2/kWh = 40.3 million tCO2). Total assumed carbon emissions at the higher end of the scale are 67.2 million tCO2 (250 TWh x 0.269kgCO2/kWh = 67.2 million tCO2). For more on our assumptions see "Data Centers: Surging Demand Will Benefit And Test The U.S. Power Sector," Oct. 22, 2024."

Table 3

Related Research

• Data Centers: Computing Risks And Opportunities For U.S. Real Estate, Oct. 22, 2024
• Data Centers: Welcome Electricity Growth Will Fall Short Of U.S. Data Center Demand, Oct. 22, 2024
• Data Centers: Surging Demand Will Benefit And Test The U.S. Power Sector, Oct. 22, 2024
• Data Centers: More Gas Will Be Needed To Feed U.S. Growth, Oct. 22, 2024
• Gridlock: Interconnection Queue Backlog Adds Risks For U.S. Not-For-Profit Power Sector, Oct. 8, 2024
• Decarbonizing Hard-To-Abate Sectors: Credit Quality Implications And Six Key Observations, June 25, 2024
• Sustainability Insights: TSMC And Water: A Case Study Of How Climate Is Becoming A Credit-Risk Factor, Feb. 26, 2024
• Carbon Capture, Removal, And Credits Pose Challenges For Companies, June 8, 2023
• Purchased Energy Emissions In Second Party Opinions And ESG Evaluations, March 23, 2023