Cloud Sustainability Watch analysis
The environmental impact of the cloud
Summary
Data centers have become essential to modern life
Since the rise of the internet created the data center, the facilities have grown in number, size, and compute intensity. The boom in cryptocurrency, the explosion of streaming and social media, and the spread of e-commerce have made data centers central to the modern economy. In recent years, cloud computing has accelerated data center expansion. Emerging use cases have included big data applications, internet-connected appliances and mobile devices, and data storage security. Moving forward, the high compute demands and fast adoption of generative AI are unknown but significant factors.
Most of the environmental impact that makes online life possible takes place invisibly, unless you live near a data center. Internet users may perceive online activity as a more environmentally friendly alternative to in-person events or physical media without awareness of the environmental impact of the data centers that make it possible. Individual internet users can tread more lightly in terms of GHG emissions and water impacts, but their mitigation steps are limited to the periphery of the problem.
The environmental impacts degrade our surroundings
In 2023, researchers in the telecommunications industry reported that data centers and data transmission networks constituted 0.6% of the world’s 47.5 gigatons of GHG (greenhouse gas) emissions for 2020. That comes to 285 megatons of GHG emissions — about the same as the world’s pulp and paper industry. If you add in end-user devices, it matches the airline industry. Data centers also consume large amounts of fresh water for cooling their equipment. A large facility uses 500,000 gallons daily.
Nonetheless, the hyperscaler cloud providers are markedly more efficient with power and water than on-premise or rented colocation computing systems. Professionally run data centers are the most power-efficient and carbon-efficient method of delivering the world’s computing at this time. McKinsey & Company’s 2023 report “Cloud-powered technologies for sustainability” estimates that in 2050, cloud usage will reduce greenhouse gas emissions by 1.5GtCO2e annually compared to staying with existing data center decisions.
Our position
Cloud computing’s data center operators and IT professionals are responsible for bringing the environmental impact of the cloud to heel. These groups built out the cloud computing industry, and their daily decisions steer its direction.
Foremost, these professionals have an opportunity to tackle GHG emissions by shifting to consumption of low-carbon electricity. With its large electricity purchases and employee base, the cloud industry can pressure utilities and politicians to facilitate siting, building, and interconnecting new renewable energy power plants.
In turn, cloud computing companies must offer more sustainable options for enterprise customers. The industry also needs to innovate ways to reduce the water needed for cooling microprocessors —- from hardware that is more heat tolerant or less heat-producing, to direct liquid cooling and using non-water liquids, to allowing facility temperatures to rise.
What to do immediately in your own work life to reduce emissions? Deloitte has a few suggestions here.
READ MORE
Why do data centers impact the environment?
IT equipment that powers the internet drives carbon emissions and water use
The data centers that store, compute, and deliver cloud services are vast buildings with extremely high-capacity internet connections running millions of pieces of information technology hardware and software.
Data centers consume large amounts of electrical power and water to operate and cool densely packed racks of IT equipment that gets replaced.
When they consume electricity from power plants burning fossil fuels, they contribute to greenhouse gas (GHG) emissions. When they pull millions of gallons of water a day from municipalities and aquifers, they cause alarm in their communities. And they generate electronic waste when equipment gets upgraded.
READ MORE
What is the trend for sustainability?
Cloud computing’s impact is growing
Cloud computing platforms let customers use data center resources remotely as their own virtual data center. While cloud users don’t operate the IT hardware, they still cause real-world data centers to emit GHGs from power consumption. They also contribute to the cloud data center’s use of 500,000 to 2+ million gallons of drinking water a day.
Use and demand are growing across the board. Consumers use streaming and social media, and businesses are putting their IT “up in the cloud” with large amounts of data processing and backups. Complex new cloud applications, such as generative AI (GenAI), add complexity with unknown implications.
GenAI models and applications use up electrical power in new and power-hungry ways. The market adoption rate and technology evolution of GenAI are not yet clear but show concerning potential to drive up data center GHG emissions and water use at an unprecedented rate. .
READ MORE
What should cloud customers do?
We recommend discontinuing on-premise data centers and finding sustainable cloud options
Every enterprise that discontinues running on-premise or rented colocation data centers and migrates to the cloud drops its greenhouse gas emissions.
Professionally run data centers are the most power-efficient and carbon-efficient method of delivering the world’s computing at this time. McKinsey & Company’s 2023 report “Cloud-powered technologies for sustainability” estimates that in 2050, cloud usage will reduce greenhouse gas emissions by 1.5GtCO2e annually compared to staying with existing data center decisions.
READ MORE
Referenced on this page
- The Cloud Is Material: On the Environmental Impacts of Computation and Data Storage · Blog: MIT Schwartzman College of Computing https://mit-serc.pubpub.org/pub/the-cloud-is-material/release/2#n93o7parpwu
- Cloud-powered technologies for sustainability | McKinsey: https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/cloud-powered-technologies-for-sustainability
- A new front in the water wars: Thirsty, giant data centers – The Washington Post: The Washington Post, “A new front in the water wars: Your internet use”, April 25, 2023, Shannon Osaka, https://www.washingtonpost.com/climate-environment/2023/04/25/data-centers-drought-water-use
- Cloud computing’s real-world environmental impact | TechTarget: https://www.techtarget.com/sustainability/feature/Cloud-computings-real-world-environmental-impact
- Understanding Carbon Footprint in the Cloud – Watershed: https://watershed.com/blog/carbon-in-the-cloud
- What Is a Data Center? | TechTarget.com: https://www.techtarget.com/searchdatacenter/definition/data-center
The leading cloud providers
Summary
Largest impact overall
The largest cloud computing brands that hold 2/3 of the market are Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure – The Big 3. The large environmental impact of the Big 3’s global data center infrastructure has drawn attention from internal stakeholders and outside organizations, including Greenpeace, the EU government, and university researchers. These companies’ success or failure in reducing their environmental impact affects climate change worldwide. They lead among US corporations in setting voluntary targets for emissions reduction, and they are the largest corporate purchasers of renewable power. The enormous pace of demand from their customers may be outstripping the progress of their sustainability programs.
Lowest impact per customer
The Big 3 run the most efficient data centers in the public cloud computing industry, so anyone using on-premise data centers and wishing to reduce their carbon footprint should migrate to the Big 3. They also buy enormous amounts of low-carbon electricity to power their data centers. We recommend migration to a Big 3 location that offers low-carbon services over less carbon-efficient alternatives.
Local harms felt in communities
Large data center campuses operate around the world, and in each location local environmental impacts exist. Water, power, noise, and landscape are all negatively affected, and residents are opposing projects. The Big 3 face difficulties siting their facilities in locations that can support their power and water needs. These facilities can take as much as a quarter of local drinking water supplies and as much electrical power as one million US households. Their campuses much go water positive immediately where their campuses are located in drought regions such as the American Southwest and the regions of Chile where citizens have so little drinking water that their taps flow with brackish water and the government distributes two liters of fresh water daily to low-income residents.
Our position
The size, growth and influence of AWS, Azure, and GCP create an obligation to invest as much money and attention as possible into reducing environmental impact in the face of exploding market demand. The local environmental impacts of their data centers are often severe, and the Big 3 must spend what it takes to dramatically reduce impacts for existing campuses and going forward site facilities in areas with adequate power (ideally low carbon) and an absence of drought conditions.
The three companies vary substantially in their disclosure of climate information, with Amazon notably trailing the other two. We see an imperative for these organizations to share transparent and meaningful information about their environmental impact and to achieve their public targets for reduced GHG emissions and become water positive by 2030.
The Big 3 brands are the market leaders in global cloud computing services. Their hyperscale infrastructure is the pace-setting leader in energy efficiency, access to low-carbon electricity and facilities operation. We believe the size, growth and influence of AWS, Azure, and GCP creates an obligation to invest as much money and attention as possible into reducing environmental impact in the face of exploding market demand. The three companies do not offer complete disclosure of climate information, with Amazon trailing the other two in that regard, likely their enormous global retail business makes it more difficult versus the other two. We see an imperative for the Big 3 to share transparent and meaningful information about their environmental impact and to achieve their public goals for reduction and elimination of emissions.
Public targets and sustainability progress
Notes: Alphabet and Google Cloud Platform
A high percentage of Alphabet’s reported Scope 2 (electricity) GHG emissions is likely to be from its data centers because the company does not match its competitors in the size of its retail and equipment businesses. Electrical power is a rough proxy for data center GHG emissions. The Scope 2 GHG emissions are equal to Microsoft’s, and we are unsure if the comparison is apples to apples.
Notes: Amazon and Amazon Web Services (AWS)
Notes: Azure and Microsoft
- Corporate-reported scope 2 GHG emissions from Microsoft and Google are the same. Our analysis shows that Google has significantly fewer data centers than Azure, which is to be expected since Azure has twice the market share. We assess that Azure is not 2x better than GCP at emissions mitigation, and we conclude that our figures are not apples to apples.
- The reporting of the number of data centers by the companies varies by factors we could not identify. The definition of what constitutes a single data center can be variable. The outside sources DataCenters and DataCenterLocations do not distinguish individual centers from data center campuses.
Hyperscaler Sustainability Scorecard
READ MORE
The Big 3 dominate cloud computing
Market leaders grow and so does their impact
The Big 3 cloud computing leaders hold 67% market share: AWS 31%, Azure 25%, and GCP 11%. Their businesses are growing at a staggering 20+% annually. They are growing in large part because more customers are migrating to the cloud and choosing them. They also grow in response to increasing demand among existing customers for additional storage and compute power.
The Big 3 keep expanding their global infrastructure of large and data centers in many geographic locations to meet customer needs.
All the new construction and equipment purchases raise the Big 3’s embodied carbon emissions. In May 2024, Microsoft announced that new data center construction had increased its carbon footprint by 30% since 2020— even though its GHG emissions had gone down 6% in the same period. We can assume similar trends for AWS and GCP.
READ MORE
The hyperscalers are the lowest-carbon choice
The top companies invest in renewables and efficiency, and their scale yields results – but challenges persist
The hyperscaler cloud providers run their data centers with the lowest rate of GHG emissions of the cloud marketplace. (And are vastly more emissions-efficient than on-premise data centers, which we discuss in “Environmental Impact”.)
Their hyperscale infrastructure of up-to-date data centers does achieve industry-leading power and water efficiencies. They also sometimes run on low carbon energy. Their efforts to get a handle on water use are less successful.
Over time, the Big 3 (AWS, Azure, and GCP) have held down the growth rate of their carbon footprint in comparison to the growth rate of their businesses. Their capacity to invest in efficiency monitoring, upgrades to optimal-performance equipment and renewable power have been essential to keeping their carbon footprint under control. And their workforce is focused and trained for the companies’ main business.
But their business is growing quickly and it will be a mounting challenge to contain their environmental impact as they meet customer demand for computing, storage and new applications.
READ MORE
Siting data centers pits the cloud companies against localities
Environmental impacts are unpopular in nearby communities
A difficult fact of siting data centers is that nearby communities suffer reduced supplies of power and water, exposure to noise from cooling systems, and the eyesore of the facility. Community members often object to data center construction while local governments are frequently welcoming.
Data center siting decisions incorporate low land cost, which makes for large infrastructure being built in hot, dry locations with fresh water supply problems. In the American Southwest there are many data centers that use the plentiful solar energy to reduce GHG emissions while consuming unsustainable amounts of scarce drinking water. Yet their water sourcing, treating, transporting and disposal generates GHG emissions, and increases the campus’ carbon footprint.
In recent years, citizens opposed data center construction in Valencia County, New Mexico, US; Fairfax County, Virginia US; Canelones Dept., Uruguay; Zeevolde, Netherlands; and Cerrillos, Chile (where a court intervened on behalf of the aquifer).
READ MORE
We recommend advocating for cloud company sustainability
- We see an imperative for Amazon, Google, Microsoft, and other cloud providers to share transparent and meaningful information about the environmental impact of their data center fleets and to follow through on their public goals for reducing and eliminating emissions.
- The Big 3 cloud companies have made public pledges to reverse the harm of their greenhouse gas emissions, and their parent companies have made “water positive” pledges. Generative AI will make it harder to fulfill these pledges, because it uses computing power and heats up data centers in drastic ways.
- Customer pressure will affect decisions the Big 3 will make during the pivotal years as generative AI’s impact on data centers takes shape.
Referenced on this page
- Microsoft 2024 Environmental Sustainability Report, https://www.microsoft.com/en-us/corporate-responsibility/sustainability/report
- Google 2024 Environmental Report, https://blog.google/outreach-initiatives/sustainability/2024-environmental-report/
- Amazon 2023 Sustainability Report, AWS Summary: 2023-aws-summary.pdf
- Harvard University, Meta, Arizona State University, “Chasing Carbon: The Elusive Environmental Footprint of Computing”, Udit Gupta, Young Geun Kim, Sylvia Lee, Jordan Tse, Hsien-Hsin S. Lee, Gu-Yeon Wei, David Brooks, Carole-Jean Wu, https://arxiv.org/pdf/2011.02839
- Greenpeace, “Clicking Clean”, 2017, Gary Cook, https://www.greenpeace.org/static/planet4-international-stateless/2017/01/35f0ac1a-clickclean2016-hires.pdf
- Datacenter Knowledge, “AWS, Microsoft, Google Continue to Invest in Cloud as Demand Grows in AI Era”, May 2, 2024, Sean Michael Kerner, https://www.datacenterknowledge.com/cloud/aws-microsoft-google-continue-to-invest-in-cloud-as-demand-grows-in-ai-era
- Synergy Research Group, “Huge Cloud Market Sees A Bounce in Growth Rate for the Second Consecutive Quarter”, April 30, 2024, no author named https://www.srgresearch.com/articles/huge-cloud-market-sees-a-strong-bounce-in-growth-rate-for-the-second-
consecutive-quarter - The Washington Post, “A new front in the water wars: Your internet use”, April 25, 2023, Shannon Osaka, https://www.washingtonpost.com/climate-environment/2023/04/25/data-centers-drought-water-use
- SDX Central, “Hyperscaler sustainability scorecard: Rating Google’s, Microsoft’s and AWS’ environmental strategies”, July 20, 2023, Emma Chervek, https://www.sdxcentral.com/articles/analysis/hyperscaler-sustainability-scorecard-rating-google-microsoft-and-aws/2023/07/
- Algorithm Watch, “With Google as My Neighbor, Will There Still Be Water?”, April 6, 2023, Interview with Dr. Sebastián Lehuedé, https://algorithmwatch.org/en/protests-against-data-centers/
- Fairfax Times, “Community rallies against county’s data center expansion”, June 7, 2024, Jimmy Henderson, https://www.fairfaxtimes.com/articles/fairfax_county/community-rallies-against-county-s-data-center-expansion/article_97db5e00-2419-11ef-b7b5-7b7a8975d46f.html
- AP News, “Google says it will rethink its plans for a big data center in Chile over water worries”, September 17, 2024, no author, https://apnews.com/article/chile-google-data-center-water-drought-environment-d1c6a7a8e8e6e45257ac84fb750b2162
- Tech Times, “Google Data Center Plan Angers Drought-Stricken People of Uruguay”, July 11, 2023, John Lopez, https://www.techtimes.com/articles/293657/20230711/google-data-centre-plan-angers-drought-stricken-people-uruguay.html
- CNBC, “Microsoft’s carbon emissions have risen 30% since 2020 due to data center expansion”, May 15, 2024, Spencer Kimball, https://www.necn.com/news/business/money-report/microsofts-carbon-emissions-have-risen-30-since-2020-due-to-data-center-expansion/3235611
The contest for electrical power
Summary
Data centers consume more electricity than France
Data centers in the US make up 4% of the nation’s 2022 electricity consumption, according to the International Energy Agency. According to IEA estimates, data centers, cryptocurrencies, and artificial intelligence consumed about 460 TWh of electricity worldwide in 2022, almost 2% of total global electricity demand. 460 TWh is more than the electricity consumption of France in 2021.
The world’s electricity supply hasn’t increased yet
While electricity production has recently grown enormously in China, it has stayed flat in the EU, US, and Japan over the past 20 years. Building new generation capacity is complex and slow, and vast improvements in efficiency throughout the economy have enabled electrical utilities to plan for little or no change in demand. The fast explosion in power demands from data centers and other areas of the economy have changed this picture entirely. Utilities are planning and building generation and transmission. But turning the ship of electricity production to growth after decades of mostly flat demand has meant re-learning infrastructure expansion and permitting, so delivering the power data centers need is slower than required.
The Big 3 are on the hunt for electrical power
The cloud computing boom has set off a race for the world’s finite supply of electrical power. The Big 3 continue to expand their data center infrastructure, with new campuses underway that consume the electricity of a midsize city with a population of 1 million. While they are investing in projects to bring new power online with renewables, they rely on utilities to supply all the power they require. In some cases, they have reversed plans to decrease fossil fuel use.
Our position
Cloud service providers require additional electricity for their core business to survive, and it should be low-carbon. The Big 3 are well-positioned to get new low-carbon power plants built, and they need to resist pressure from utilities or politicians mandating gas plants as the only path forward.
Operating their data center infrastructure on low-carbon power is the only way the Big 3 can meet net-zero pledges in the face of exploding demand for cloud services. Utilities use the need for fast supply expansion to justify returning to more familiar gas plant construction. However, renewable power plants have an advantage in siting because they don’t pollute the air or water, and don’t require noisy trucks or unpopular pipes to deliver a nonstop supply of fossil fuel. And renewables win on the economics of a power plant’s lifetime: they don’t require the expense of purchasing fossil fuels.
Despite interconnection queues, political considerations, and sizeable capital expense, the Big 3 should commit and stick to this path.
READ MORE
The hyperscaler cloud providers are managing exploding electricity requirements
More demand means large new facilities that need power
A global boom in cloud computing means the world’s technology companies are seeing fast growth in their cloud platform divisions.
As more companies migrate to the cloud and new applications drive usage and demand, cloud providers are banking on continued rapid growth. Generative AI (GenAI) multiplies the scope and pace of expansion. Learn more in “GenAI“
GenAI, with its specialty hardware and cooling requirements, requires almost completely new data centers. It could be tough for smaller cloud companies to meet these needs, making GenAI support a potential differentiator for the Big 3 cloud brands (AWS, Azure, Google Cloud). They have to move quickly to lock in their advantage.
The leading data center operators are building new and upcoming facilities that are enormous, with power draw on par with cities. This exceptional power demand, along with less well-known excessive water consumption, is making it hard to find locations where they can get the electricity they need.
READ MORE
Data center companies are racing for supply
Data center companies and hyperscaler cloud providers are scouring the landscape for new spots to build bigger data center campuses. For customers that want minimal lag time when they use their cloud instance, sites have to be located near population centers with expensive land and constrained power and water. Siting new facilities has to trade off distance from customers with availability of adequate land and resources.
As a result, data centers cluster in favorable locations, and those locations are facing an electrical power crunch. The 80+ data centers in the Republic of Ireland consume about 28% of the country’s electricity — more than all its residential demand; leading the government to tighten criteria for permitting new projects. In 2022–2023, Dominion Energy in Virginia’s “data center alley” paused new data center connections to the power grid for three months while it caught up in transmission build-out.
READ MORE
Utilities’ plans include new gas plants
Electrical utilities are exploring solutions to the problem of how to meet spikes in energy demand and overall demand growth in the near term. Some utilities are working on shifting the time or level of electricity use to reduce peak demand. And for example California’s Silicon Valley Clean Energy has a program working with manufacturers in its service area to implement energy efficiency improvements at appealing cost savings.
Demand response programs that reward companies and households for reducing their electricity use during demand peaks are growing quickly, and data center operators are starting to participate. Even more importantly, data centers — notably Google — are using their power storage capacity to become good neighbors by delivering power to the grid during crunch times, becoming virtual power plants.
But some utilities located in data center clusters have no choice but to build out new generation facilities. And many of them are appealing to legislators to roll back mandates for low-carbon power transition. They plan on building new gas plants, familiar infrastructure whose construction timeline is well understood.
A new gas power plant will operate, emitting greenhouse gas and pollution for decades — and cloud computing will be accountable for a share of that damage.
READ MORE
Low GHG emissions is becoming a requirement
Public awareness of data center emissions has drawn regulatory attention
Regulators are examining data centers and instituting regulations around greenhouse gas emissions, power and water use. The European Union introduced new reporting requirements under its Corporate Sustainability Reporting Directive (CSRD) for data center energy use and greenhouse gas emissions. There is also a data center energy efficiency reporting requirement.
The EU Climate Neutral Data Center Pact is a voluntary agreement under which operators pledge to make their data centers climate neutral by 2030. The US SEC’s issued draft CDR rules that are in review and would standardize corporate climate disclosures and apply to data center operations. In Canada, the Net-Zero Emissions Accountability Act means that data centers will need to complete climate-related reporting. Japan also has financial reporting requirements around climate.
READ MORE
Cloud providers: expanding renewables or using up existing supply?
It’s hard to tell which low carbon cloud locations are reducing global emissions and which are shuffling existing renewable supply
The largest tech companies are the largest purchasers of renewable electrical power in the world. They have spent a great deal of money on long term contracts with existing renewables suppliers, and they have partnered with renewable energy developers and utilities to build new generation capacity.
These companies are using Power Purchasing Agreements (PPAs) to point to the new clean power generation facilities they induce to be built by guaranteeing they will buy enough power to make the facilities economically successful. It may not always be the case that they induce construction that would not have otherwise been done, but technology companies are among the biggest funders of renewables expansion.
The technology companies that fund renewable generation with PPAs account for the low carbon power against their corporate greenhouse gas emissions in their disclosures, to support their emissions reductions reporting. The power plants under PPA agreements may not be generating power in a location that allows for an explicit matching of the renewable power production to their data centers’ power consumption. And it is impossible for clean power to match every hour of a data center’s consumption because renewables don’t generate electricity 24 hours every day, so 40 – 70% of emissions match demand. A solution is to add to a PPA an energy storage system such as batteries.
The leading cloud providers are deploying their influence as large players in local economies and large power purchasers to urge utilities to move faster building out clean energy, such as by striking deals with private clean power developers. They are reported, as in The Wall Street Journal, to be exerting pressure on regulators and developers to strike deals with independent power producers.
In some locations cloud data center operators are building microgrids of renewable power from solar or wind paired with battery storage directly on their data center campuses. These power backups replace high emission diesel generators, and are increasingly becoming “good grid citizens” that deliver power back to the grid during high demand periods.
A data center’s power consumption will wax and wane and the pattern of its power consumption is not aligned with the power supply from the PPA-contracted plant. The result is peak demand in a fossil fuel grid that exceeds any delivery from the renewable plant elsewhere. Google is leading its cloud service category in its progress working toward “27/7 matching” to trace and prove that their data centers’ power consumption is consistently connected to their contracted renewable power supply. In some cases they have built data centers directly adjacent to clean power plants.
READ MORE
We recommend reducing emissions from power consumption
- Discontinue using inefficient on-premise and cloud services for your organization’s computing and storage
- Migrate to a hyperscaler cloud provider with the best PUE you can find
- Ask your cloud provider how they are reducing power consumption in their data center fleet
- Ask about and select low carbon locations for your cloud service, balancing latency and requirements
- Use virtual servers as much as possible
- Develop capabilities for running batch jobs at times or in locations with the most low carbon power
- Advocate in your industry for computing sustainability
- Advocate against utilities building new fossil fuel power-powered generation facilities
- Advocate for the Big 3 to continue building partnerships to develop additional renewable power plants
Referenced on this page
- Data Center Dynamics, “Global data center electricity use to double by 2026 – IEA report”, January 6, 2024, Matthew Gooding, https://www.datacenterdynamics.com/en/news/global-data-center-electricity-use-to-double-by-2026-report/
- International Energy Agency, “Electricity 2024 Analysis and forecast to 2026”, May 2024, Eren Çam, Zoe Hungerford, et.al.,, https://www.iea.org/reports/electricity-2024
- Statista, “Total gross electricity consumption in France from 2013 to 2023”, no date or author given, accessed June 2024, https://www.statista.com/statistics/460064/france-electricity-consumption/
- International Energy Agency, “Data Centres and Data Transmission Networks”, July 11, 2023, Vida Rozite,https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks
- Our World in Data, Data Page: Electricity demand”, part of the following publication: Hannah Ritchie, Pablo Rosado and Max Roser (2023) – “Energy”. Data adapted from Ember. Retrieved from https://ourworldindata.org/grapher/electricity-demand
- PWC, “The global economy has become much more energy efficient. But are there further gains to come?”, undated, Barret Kupelian, https://www.pwc.com/gx/en/research-insights/economy/global-economy-watch/energy-efficient-global-economy.html
- The Wall Street Journal, “There’s Not Enough Power for America’s High-Tech Ambitions”, May 12, 2024, David Uberti, https://www.wsj.com/business/energy-oil/data-centers-energy-georgia-development-7a5352e9
- Ernst & Young, “What’s next for Japanese sustainability disclosure standards”, October 13, 2023, Kyle P. Lawless, Keiichi Ushijima, et.al., https://www.ey.com/en_jp/insights/sustainability/whats-next-for-japanese-sustainability-disclosure-standards
- Black & Veatch Blog: “Water Efficiency Opportunities for Data Centers”, no date given, Mikeal Vincent, https://www.bv.com/perspectives/water-efficiency-opportunities-data-centers
- Uptime Institute, “Five data center predictions for 2024”, January 2024, no author given, https://uptimeinstitute.com/uptime_assets/ebdfbf14591d2e8bfe05a0e4faddf024af34ab7ef7cfe8cb21182774b490e58a-five-data-center-predictions-for-2024.pdf
- European Commission: Energy Efficiency Directive, no author or date given, accessed June 12, 2024, https://energy.ec.europa.eu/topics/energy-efficiency/energy-efficiency-targets-directive-and-rules/energy-efficiency-directive_en
- Securities Exchange Commission press release, “Press Release
SEC Adopts Rules to Enhance and Standardize Climate-Related Disclosures for Investors”, March 6, 2024, https://www.sec.gov/news/press-release/2024-31 - Techerati, “Data Centre Sustainability Regulation: The timeline that requires immediate action”, November 14, 2023, Vanessa Moffat, https://www.techerati.com/features-hub/data-centre-sustainability-regulation-the-timeline-that-requires-immediate-action
Water and cooling are large considerations
Summary
Equipment cooling is a water problem
The Big 3 and the data center industry use water-based methods to cool IT equipment instead of air conditioning because water costs less than electrical power. Over the years there have been few consequences for using a lot of it. Municipal utilities provide potable water to data center facilities, in many other cases their water is pulled from underground aquifers, drawing down water supplies without verification of how much is being taken. Taking large amounts of water from drinking supplies is a harmful business practice.
Scrutiny of water usage has intensified
Municipalities and citizens groups track data center water consumption and discharge. Communities object when they discover that as much as 25% of their drinking water goes to a data center in their area, and when the facilities get advantageous pricing from water utilities makes conservation dispensable. Water usage has been undisclosed and the Big 3 have fought revealing water use at the facility level. They now disclose unevenly the water they use, and the quantity can rise as high as several million gallons a day.
Costly and complex improvements curtail water use
Expensive and far more efficient cooling methods exist. One is placing tubes directly on the hot parts of the equipment to circulate chilled water. The other is immersing the equipment, chassis and all, in a tank of low-boiling-point liquid. Generative AI equipment has to be cooled via these new methods because it runs too hot for conventional methods.
Our position
The Big 3 must improve their water conservation. Upgrading to low-water cooling systems and restoring wastewater to the level of quality needed for reuse. There are paths toward innovation in equipment design and cooling technology that will advance efficiency further. Battling disclosure that allows communities to understand their resources are being degraded must yield to full transparency at the level of availability zone of facility campus. A benefit will be customers who demand low carbon and low water options will find and use them.
READ MORE
Water is used for cooling IT equipment
Air conditioning cools inefficiently
Data centers require potable water for their operations and usually pull from groundwater more than from municipal suppliers in the same manner as other industrial water users. Data center siting decisions incorporate low land cost, which makes for large infrastructure being built in hot, dry locations with fresh water supply problems. Water use drives up carbon footprints because sourcing, treating, transporting and disposing of water generates emissions. The regulations and programs such as CSRD and CDR address water use because fresh water is a critical environmental concern.
The importance of cooling improvement is large enough that the US National Renewable Energy Laboratory and the US Department of Energy’s Advanced Research Projects Agency created a $40M prize for cooling breakthroughs. New cooling systems and better analysis models for how to cool are an evergreen research area. Water costs less than electricity, so the Big 3 and the cloud industry maximize their PUE by using water-intensive cooling methods instead of air conditioning.
It can take a lot of power to cool the racks of hardware because the silicon microprocessors at the heart of the IT equipment heat up when they are in use. The capability to cool racks of servers limits how densely servers can be placed in a data center and in turn how much capacity can be sold to customers. The Big 3 are very good at cooling their equipment using a range of technologies, and through monitoring and investing in it, they get their PUE numbers very low. They often (ironically) utilize power-hungry artificial intelligence to determine how to reduce the power used by cooling systems.
The Green Grid coined the term “water usage effectiveness” to track how wastefully or well a data center uses water. Combined with PUE, Power Usage Effectiveness, WUE is a good description of the tradeoff between consuming power vs water. The calculation is the number of liters of water used per kilowatt hour. The formula abstracts the consumption and makes it less easy to understand the large quantity of water a data center uses.
Ways of improving power use at the expense of WUE deploy water-based cooling in the place of air conditioning, and the water is used in large volumes. Water-cooled chillers directly cool the air passing over them with thermal heat exchange, and evaporative cooling towers do it indirectly. More cooling comes from methods of spraying water directly into the air at key points in its circulation. Data centers also use water to ensure the right humidity for IT equipment.
READ MORE
Water consumption is high
The totals are staggering
Big 3 infrastructure around the world consumes enormous amounts of potable water. Their fleet of data centers require it for their operations and likely pull from groundwater along with municipal suppliers in the same manner as other industrial water users. None of the Big 3 disclose their water sources for their data center fleets.
Most data centers store emergency water and facilities often have 10 million gallons in storage unavailable for public consumption.
Microsoft disclosed that its data center water use grew 34% over 2021 to 1.69 billion gallons. Google disclosed their data center fleet consumed 5.2 billion gallons. AWS disclosed only a fleet-wide WUE.
Google revealed the water consumption for 23 data centers along with its total. Per-facility usage ranged from 100,000 to 896 million gallons per year.
The WUE that AWS provided tells us that their data center fleet uses 1.8 liters of water for every kilowatt of power. For a 100MW nameplate datacenter, that figures out to about the same as Google.
This amount of water equates to the water consumption of three hospitals, according to a 2015 Wall Street Journal article.
READ MORE
Data center water consumption draws more notice
Communities are fighting technology companies and their own city governments
The large amounts of water taken out of drinking supplies is a harmful business practice. That level of consumption causes problems for communities, and local residents have protested losing their drinking water supply.
A town in Oregon only learned how much of their city water went to a Google data center campus after the local newspaper sued the municipal government and won. The disclosure revealed that Google was taking 29% of their water.
And a residents’ group called Valencia Water Watch in Los Lunas, New Mexico, are protesting expansion of a Meta data center with, among other things, a tractor protest to city hall. (They previously won the fight keep out a water-thirsty bottling company.)
At a national level, environmental reporting organizations are defining water use as a factor in data center sustainability. And WUE is becoming standard in tech company reports. Out in front for government regulators is the European Commission. It released its “Data Centers in Europe” reporting requirements and included water use disclosure among them. The regulations and programs such as CSRD and CDR also address water use because fresh water is a critical environmental concern.
READ MORE
Advanced liquid cooling for the age of GenAI
GenAI equipment runs too hot for common cooling techniques.
The GPU microprocessors required for generative AI (GenAI) applications run so hot that air cooling cannot work. GenAI’s iterative computations require equipment racks to be packed tightly, so air cooling isn’t adequate. These situations require retrofitting and investment in either direct-to-chip liquid cooling, in which coolant or water flows in tubes to a cold plate directly on the equipment, or newer liquid immersion cooling, in which computing equipment is placed in a tank of low-boiling-point liquid.
These types of liquid cooling are between 50 and 1,000 times more efficient than air cooling. Liquid cooling is also complex and expensive. It is also far quieter than air cooling – which impacts the willingness of municipalities to accept siting of datacenters – and it uses little water or in some methods none at all.
The hyperscaler data center operators titrate the optimal liquid-air cooling balance to target liquid cooling only where it is required. Even where the Big 3 deploy direct liquid cooling, there is still plenty of water-intensive cooling. And the extremely large quantities of water consumed by data centers are an environmental problem.
READ MORE
We recommend reducing your usage to reduce your water consumption
- Use green software methodology and tools to reduce unnecessary compute and storage for your applications and for your customers who use them
- Ask your cloud provider about their efficiency programs
- Use virtual servers and maximize server utilization
- Review if a GenAI application is the right tool for the job
- Develop the most efficient applications for your organization’s needs
- When choosing a cloud platform and an availability zone or location, ask about:
- WUE
- Water reuse and replenishment program
- Water source
- Types of water-based cooling
Referenced on this page
- WP#35 – Water Usage Effectiveness (WUE): A Green Grid Data Center Sustainability Metric | The Green Grid: https://www.thegreengrid.org/en/resources/library-and-tools/238-Water-Usage-Effectiveness-%28WUE%29%3A-A-Green-Grid-Data-Center-Sustainability-Metric-
- How to manage data center water usage sustainably | TechTarget: https://www.techtarget.com/searchdatacenter/tip/How-to-manage-data-center-water-usage-sustainably
- Data Center Water Usage: A Comprehensive Guide – Dgtl Infra: https://dgtlinfra.com/data-center-water-usage/#Water_Usage_Effectiveness_WUE
- Businessman proposing data center on banks of Columbia River has history of troubled deals – Columbia Insight: https://columbiainsight.org/exclusive-businessman-proposing-data-center-on-banks-of-columbia-river-has-history-of-troubled-deals/
- Data centre water consumption | npj Clean Water (nature.com): https://www.nature.com/articles/s41545-021-00101-w
- Data Centers and Hidden Water Use – WSJ: https://www.wsj.com/articles/data-centers-1435168386
- US National Renewable Energy Laboratory blog, “NREL Joins $40 Million Effort To Advance Data Center Cooling Efficiency”, December 14, 2023, no author given, https://www.nrel.gov/news/program/2023/nrel-joins-effort-to-advance-data-center-cooling-efficiency.html
- Datacenter Hawk, “How to Measure the Data Center Market – Data Center Fundamentals”, March 21, 2021, Mike Netzer, https://datacenterhawk.com/resources/hawkpodcast/how-to-measure-the-data-center-market-data-center-fundamentals
- Google data centers blog: “Efficiency”, no date or author given, accessed July 2024, https://www.google.com/about/datacenters/efficiency
- CGDirector, “Graphics Card Temperature Guide”, November 25, 2022, Alex Glawion, https://www.cgdirector.com/gpu-temperature-guide
Generative AI’s growing impact
Summary
Generative AI (GenAI) infers the meaning behind large amounts of information to create from scratch its responses to inquiries
Generative artificial intelligence is a software innovation that makes it easier and faster to make use of large amounts of data. It studies trillions of data ranging from text to media, and makes sense of it by understanding the underlying context. Thereafter it answers queries by generating a new response — something that didn’t previously exist — using its inferences about the data it trained on. It does not look up data directly on a server but instead “thinks” from its body of learning.
Data centers will consume far more power and water to serve GenAI’s computing and cooling demands
GenAI impacts data centers because it runs computers for a very long time while it learns information, and runs them again every time it generates query responses. Data networks will also get more use from larger amounts of data in a larger number of iterations. GenAI’s intensive processing requires specialized chips that generate 10x more heat than standard chips, and in turn require far more cooling.
Reducing wasteful GHG emissions depends on advanced application design
Customers should specify the GenAI applications that are targeted to their needs and economical. Their applications should be trained on a data set honed to their specified subject area and designed for high quality results with the least compute time. Programming project plans should be designed to minimize compute time during development.
How quickly and broadly GenAI becomes widespread dictate the path of its environmental impact
Steep growth forecasts of GenAI’s growth may oversimplify its future based on the rush of customers jumping to use it. Customers’ experience with GenAI’s expense, complexity, and legal developments will cause some to pause and search for smaller alternatives. The landscape of general-use GenAI applications will mature, and GenAI application companies will seek profits in targeted market segments.
Our position
The Big 3 are striving to lead in GenAI offerings and must manage the environmental impacts rigorously. GenAI-capable data centers must be the ones running on low carbon power. They must have the best power efficiency, especially with cooling specialized for GenAI equipment. The Big 3 should offer targeted GenAI applications that meet customers’ needs and don’t exceed requirements. Targeted and higher quality training data sets will make for more accurate responses and lower compute time.
READ MORE
GenAI requires a lot of power and water
Custom-generated query responses cause equipment to work hard
GenAI applications have spurred significant increases in Big 3 power consumption, along with extensive equipment disposal caused by switching in specialized equipment.
GenAI makes it easier and faster to make use of large amounts of data. The software studies a large amount of information and makes sense of it by understanding the underlying context. It then generates query responses using its inferences about the data it trained on, rather than looking up data stored elsewhere.
Generative AI models run on trillions of data points. User queries make the entire model go to work and draw a lot of power during its compute time. There is also a lot of power consumed while the model trains on an enormous data set. The founder of OpenAI estimated that it cost more than $100M to train ChatGPT-4. And periodically the models have to repeat the long compute process of re-training, because performance degrades with use and its original training gets out of date.
GenAI’s power needs are also amplified by its specialized computer microprocessors called GPUs that run parallel computing tasks at once and consume 1,000+ watts compared to older CPUs that use 100 watts. GPUs also run markedly hotter than older microprocessors and require new cooling systems.
Goldman Sachs Research estimates the US data center power consumption from GenAI was 4TWh, which makes the calculation of greenhouse gas emissions for GenAI in the 1.5 million metric tons using USEIA’s average grid power carbon intensity. By 2028, they estimate AI to represent about 19% of data center power demand, but EPRI calculates that GenAI is already using 15-20% of data center power consumption.
We at Cloud Sustainability Watch estimate that GenAI in US data centers consumed 23.76 billion gallons of water, which was 0.02% of all US water supply that year.
READ MORE
Market adoption has been fast
Early rush to buy GenAI could cool as customers and suppliers re-assess
GenAI is driving innovation in all sectors including healthcare, military, education, and media; and it is also raising data center power consumption. There is also promising potential for GenAI to create operational improvements and energy efficiencies in data centers.
The growth of GenAI application usage depends on customer satisfaction and willingness to buy more. The explosive rush to buy and use GenAI applications has informed forecasts of continued high growth. Statista’s GenAI market forecast gives 46% annual growth until 203 when market volume will be $356B(6). That is the same as the forecast value of Australia’s raw materials exports in 2025-26(7).
Customers are re-evaluating. Corporate users will need to change business processes to build GenAI into work flows and make the most of its benefits. And making money after the enormous cost of training a GenAI model is uncertain.
The May 2024 Wall Street Journal article “The AI Revolution Is Already Losing Steam”(20) analyzes roadblocks to customers pouring money into GenAI market. These are:
- Legal exposure, data privacy violations from training data
- Inaccuracy of results
- Business process changes
READ MORE
Big 3 data centers feel the heat
High power consumption and equipment replacement
To deploy GenAI applications, the Big 3 must update their infrastructure with increased power delivery to serve new microprocessors in computers that are more tightly grouped on equipment racks with upgraded networking connections and complex new cooling systems.
The computing microprocessors for the recent generation of GenAI applications generate 10x more heat than regular computing chips(19). With tighter, dense equipment racks. water-chilled plates and air flows have to be replaced with direct contact liquid cooling systems such as direct-to-chip tubing configurations or immersion of computing equipment in tanks of low-boiling-point liquid.(15)
GenAI equipment also requires rapid connections between storage and computing equipment for the iterative process of generating new content. It demands significantly greater density equipment racks than traditional machine learning workloads. The dense racks require seven times more electrical power than traditional data centers are built to sustain.(19).
The Big 3 are working to reduce the expense of offering GenAI services. There are delays getting GenAI GPUs from the companies that sell them, and prices are high while demand and supply are mismatched. And the infrastructure costs of accommodating GPUs’ heat and power are leading the Big 3 to design and manufacture proprietary chips with better efficiency.
READ MORE
The Big 3 are responsible for defining low-carbon GenAI
Market power, funds and reputational benefits
The Big 3 have invested what is necessary to lead in the race for GenAI profits and market share. Amazon spent $4B to buy a GenAI company (8) while Google spent $12B in just one quarter(17), and Morgan Stanley reports that Microsoft spent $16B on GenAI data center infrastructure.(9) The next step is being leaders in GenAI offerings that feature valuable reductions in power consumption and GHG emissions. Microsoft Phi is an early example.(18)
The high-emissions GenAI infrastructure should be built out in the low carbon data centers first. Customers may make the trade-off between latency and environmental impact for the sake of their own Scope 3 targets. Low-energy liquid cooling systems and the most efficient GPUs should be a priority, and reuse of disposed equipment and water should be incorporated into planning.
Building greener GenAI models reduces their emissions in training, use and platform. Programmers can design project plans that reduce how much compute is required at each step.
With the first wave of mammoth GenAI models done, programmers can evaluate how much and what quality of data to use for training. Smaller data sets with higher quality refined data reduce compute time and deliver faster and more accurate results.
READ MORE
We recommend using GenAI applications and facilities that serve your needs economically
- Purchase your GenAI deployments using availability zones or locations that draw low-carbon electrical power
- Balance a detailed assessment of speed and requirements against reduced environmental impact. Regions are too large for a claim of low carbon to be verifiable.
- Specify the GenAI applications that are targeted to their needs and economical to run
- Train models on a data set honed to the subject area required and designed for precise results with the least compute time
- Consider carefully the deployment of GenAI applications to minimize the inferences and compute time efficienctly and to reduce the frequency of use
- Review if a GenAI appication is the right tool for the job
- Select a small and specialized GenAI model to ensure lower compute and store needs — smaller model sizes can be one-sixth the cost
- If you are developing GenAI applications in house, use green programming approaches and power consumption targets can bring down GHG emissions
- Green programming techniques encompass smaller and higher quality training data sets, development project planning that minimizes compute time, and selection of low emissions platforms.
Referenced on this page
- Data Center Knowledge, “AWS, Microsoft, Google Continue to Invest in Cloud as Demand Grows in AI Era”, May 2, 2024, Sean Michael Kerner https://www.datacenterknowledge.com/cloud/aws-microsoft-google-continue-to-invest-in-cloud-as-demand-grows-in-ai-era
- BBC News, “Electricity grids creak as AI demands soar”, May 20, 2024, Chris Baraniuk, https://www.bbc.com/news/articles/cj5ll89dy2mo
- ACM FAccT ’24 conference presentation, “Power Hungry Processing: Watts Driving the Cost of AI Deployment?”, November 28, 2023, Alexandra Sasha Luccioni, Yacine Jernite, Emma Strubell, https://arxiv.org/abs/2311.16863
- Yahoo Finance, “The cost of training AI could soon become too much to bear”, April 24, 2024, David Meyer, https://finance.yahoo.com/news/cost-training-ai-could-soon-101348308.html
- Techstrong, “Hyperscalers Invest in Proprietary Semiconductors to Counter Nvidia”, May 30, 2024, Nathan Eddy, https://techstrong.ai/articles/hyperscalers-invest-in-proprietary-semiconductors-to-counter-nvidia/
- Statista, “Generative AI – Worldwide | Statista Market Forecast”, date and author not given, accessed June 2024, https://www.statista.com/outlook/tmo/artificial-intelligence/generative-ai/worldwide
- Smallcaps.au, “Australian resource earnings forecast to return to steady growth as commodity prices stabilise”, July 3, 2024, Colin Hay, https://smallcaps.com.au/australian-resource-earnings-forecast-steady-growth-commodity-prices-stabilise/
- Amazon Blog, “Amazon completes $4B Anthropic investment to advance generative AI”, May 27, 2024, author not given, https://www.aboutamazon.com/news/company-news/amazon-anthropic-ai-investment
- Investors Business Daily, “Microsoft Stock: AI Investment Returns Questioned”, July 10, 2024, Patrick Seitz, https://www.investors.com/news/technology/microsoft-stock-msft-ai-investment-returns-questioned/
- Innodata Blog, “How to Manage Model Drift in Generative AI”, no date or author given, accessed July 2024, https://innodata.com/how-to-manage-model-drift-in-generative-ai/
- Equinix Blog: “How to Do AI Sustainably”, January 26, 2023, Ram Bala, https://blog.equinix.com/blog/2023/01/26/how-to-do-ai-sustainably/
- Forbes Magazine, “The Difference Between Generative AI And Traditional AI: An Easy Explanation For Anyone”, August 23, 2023, Bernard Marr, https://www.forbes.com/sites/bernardmarr/2023/07/24/the-difference-between-generative-ai-and-traditional-ai-an-easy-explanation-for-anyone/?sh=5de07b9a508a
- Nuclear Newswire, “AI and data center growth equal power demand”, April 3, 2024, Ken Petersen, https://www.ans.org/news/article-5872/ai-and-data-center-growth-equal-power-demand/
- Goldman Sachs, “AI is poised to drive 160% increase in data center power demand”, May 14, 2024, no author credited, https://www.goldmansachs.com/intelligence/pages/AI-poised-to-drive-160-increase-in-power-demand.html
- ACQ2 Podcast, “Saving the Planet with Better AI Data Centers with Crusoe CEO Chase Lochmiller”, August 14, 2023, Ben Gilbert, David Rosenthal, https://www.acquired.fm/episodes/saving-the-planet-with-better-ai-data-centers-with-crusoe-ceo-chase-lochmiller
- Columbia University, “Artificial Intelligence for Climate Change Mitigation Roadmap”, December 1, 2023, David B Sandalow, Colin McCormick, et.al., https://academiccommons.columbia.edu/doi/10.7916/cfn0-xj39
- NPR, “AI brings soaring emissions for Google and Microsoft, a major contributor to climate change”, July 12, 2024, Dara Kerr, https://www.npr.org/2024/07/12/g-s1-9545/ai-brings-soaring-emissions-for-google-and-microsoft-a-major-contributor-to-climate-change
- he Wall Street Journal, “For AI giants, smaller is sometimes better”, July 6, 2024, Tom Dotan, Deepa Seetharaman https://www.wsj.com/tech/ai/for-ai-giants-smaller-is-sometimes-better-ef07eb98
- The Wall Street Journal Explains, “AI Is Fueling a Data-Center Boom. Can the Power Grid Keep Up?”, April 15, 2024, author not given,https://www.wsj.com/video/series/wsj-explains/ai-is-fueling-a-data-center-boom-can-the-power-grid- keep-up/A7059830-C06B-47A8-97CD-D08031EF84CC
- The Wall Street Journal, “The AI Revolution Is Already Losing Steam”, May 31, 2024, Christopher Mims, https://www.wsj.com/tech/ai/the-ai-revolution-is-already-losing-steam-a93478b1