Data Centers: the Good, the Bad and the Ugly
Just because you don’t see something doesn’t mean it isn't happening.
My grandpa joked about how my aunts and uncles used his car. According to him, they never checked the tires, the motor oil, or the breaks – they just hopped in and drove away. I think many people use AI in a similar way. People just come up with a smooth prompt and let AI work its magic without thinking much about what happens “under the hood.”
A common problem set in microeconomics classes involves maximizing a profit function subject to a budget constraint. The focus is on firms’ profit maximization and very little attention, if any, is given to externalities.1 In general, when applying this to the real world, we end up in situations where firms—and their shareholders—focus primarily on profits.2 Firms’ profits can be expressed as a function of revenues and costs.3 To keep it simple, unlocking greater revenues requires growing market shares or introducing a new (profitable) product or service. Costs can be cut by becoming more efficient, reducing workforce, being cheaper with labor and other inputs, or paying less taxes or interest on debt.
Let’s think about artificial intelligence (AI) as the new product or service that helps firms increase their revenue. While AI has been around for some time now4 , its popularity blew up when ChatGPT went viral in November 2022. As AI’s popularity grew, companies saw an opportunity to charge for their AI tools, software or premium upgrades. Generative AI appears to be the most popular mode among my acquaintances, though these days you can find AI products pretty much everywhere online, and newer phones and computers have AI features built-in by default.
Let me introduce data centers now. Data centers date back to the 1940s.5 There are different types of data center facilities catering to business needs and workloads. Hyperscale data centers are those designed to handle AI-related workloads and are thus the focus of this Substack. Hyperscale data centers are facilities equipped with a vast amount of computing resources and capability to handle massive workloads. They provide the infrastructure needed to train and deploy complex machine learning models and algorithms. In simple terms, data centers are the libraries where AI goes to read, learn, and train mental gymnastics.
As AI software and products become more sophisticated and ambitious, they require greater computing power. In the physical world, this is reflected in more, more potent and bigger data centers. However, data centers do not exist in a vacuum – they need land, electricity to run the computers, and cooling systems to prevent overheating.
Data centers represent a huge market. To many firms, this is a new source of revenue that complements that of AI. A new source of revenue means more profits.
There are currently 1,140 hyperscale data centers in the world, with 54% operating in the U.S. and 15% in Europe.6 McKinsey projects that, by 2030, data centers will require $6.7 trillion globally to keep up with the computer power demanded by AI. Only two economies exceed this volume in size (in nominal terms): the U.S. ($29.2 trillion) and China ($18.7 trillion).7 This is a staggering and highly optimistic figure. However, it evidences the sector’s potential. In fact, some are labeling these developments as an arms race. AI is hot and investors are hungry for high returns on investment.
The trillions of dollars will never be redirected towards universal health care, universal basic income, or ending hunger and poverty. Facing this reality, I think it is important to find ways to make AI and its ecosystem (which includes data centers) work for us and our planet, and not against us. It is understandable that people are finding ways to colonize Mars, but let’s not forget about the numerous issues we have here.
In this Substack I provide an overview of data centers, following a good, bad and ugly format to make it direct and easy to digest. At the end of the article I reflect briefly on similar periods of history (i.e., those of deep, structural changes) that can teach us how to face this new revolution with minimum downside.
The Good – Investment, productivity, temporary jobs, and economic growth
Data centers embody one of the most significant investments of our time. By superficial means, more investment (in this case, gross fixed capital formation) contributes directly to GDP growth. If McKinsey’s 2030 estimate is right and materializes, the $6.7 trillion of new investment in data centers will add a significant boost to the global economy, with countries hosting the bulk of this investment expected to grow at faster rates. Investing in data centers will enable the continued, uninterrupted and more ambitious use of AI, which some argue will have positive spillovers across the board. Key benefits include greater productivity and innovation, which enable faster growth.8
The construction of data centers creates jobs. This report estimates that a “large” data center creates about 1,700 local jobs during a 18-24 construction phase.9 Most jobs in construction do not require college degrees, so they are accessible to most individuals that fit a certain profile (e.g., age). Construction workers usually spend money locally on food, entertainment, etc., which provides an additional stimulus to the local economy, albeit temporarily. In addition, data centers are made from a combination of materials, ideally sourced sustainably, creating direct links with various industries (e.g., steel, concrete, glass, wood, etc.). These linkages support the development of other industries.
As infrastructure, data centers are pivotal to the success of the digital economy. Just like the Øresund Bridge connects Denmark and Sweden, data centers connect us, humans, with digital services.10 Connectivity unlocks new opportunities. Because of data centers, we can enjoy streaming services, seamless online gaming (i.e., low ping), and public/private clouds.11 Data centers also unlocked web hosting, enabling many small firms to have a digital presence and entrepreneurs to launch their own businesses. In their most powerful form, (hyperscale) data centers enable all those AI platforms we see advertised on LinkedIn and beyond to exist.
The Bad – Energy hungry, water thirsty, and community backlash
Data centers have substantial electricity demands and thirst for water. These facilities are used to train and run the deep learning models of generative AI.12 The intense computing power required to fine-tune your AI tool of choice requires significant electricity to run the computers and water to cool down the equipment. With more applications for AI, the demand for high-performance computer hardware has increased. To put it simply: more AI applications = more data centers for AI = more demand for resources.
Among data center types, hyperscale AI-focused data centers have the highest power demand, starting at 100 MW each.13 This is equivalent to the annual electricity consumption of 100,000 households and comparable with power-intensive factories such as aluminum smelters.14 Some data centers under construction will consume 20 times as much. A data center’s main energy consumers are the IT equipment (45%) and the cooling equipment to keep the processors and servers at optimal temperatures (38%).15
Unlike industrial facilities, data centers tend to be geographically concentrated and close to urban areas. Their electricity demand competes directly with that of households. In countries like the U.S. where the power grid is strained, this can result in power shortages or more expensive power generation, which is ultimately transferred to households.
The U.S. has the highest concentration of data centers worldwide and their electricity consumption is projected to double in the coming years. This report released by the Department of Energy found that data centers consumed 4.4% of total U.S. electricity in 2023.16 This proportion is expected to increase to between 6.7% and 12% by 2028, driven by AI.17 More than half of the electricity used to power data centers nationwide comes from fossil fuels. The national grid is strained – building more fossil fuel plants to fulfill this new demand will increase carbon emissions. Hyperscale developers (Amazon, Meta, etc.) promise that clean energy will power their newer data centers, but they should be held accountable for this.
Data centers require significant amounts of water to cool the hardware used for training, deploying, and fine-tuning generative AI models. Water consumption varies widely, with some data centers consuming 100 million gallons of water annually (equivalent to about 1,000 U.S. households) and others 1.8 billion gallons annually (equivalent to towns of up to 50,000 people). A more relatable estimate is that each 100-word AI prompt “drinks” one 16.9oz bottle of water.18 However, it is well known that water usage reporting is not a common practice among data center operators (see here, here and here).
Data centers’ water intensity can strain regional water sources and disrupt local ecosystems.19 This is particularly acute in water-scarce territories, where data centers compete directly with local communities and agriculture. Water used in data centers is often treated with chemicals to prevent corrosion and bacterial growth, making it unsuitable for human consumption or agricultural use.20
In the U.S., two-thirds of new data centers built or in development since 2022 are in locations with high water stress. Five states stand out: California, Arizona, Texas, Illinois, and Virginia. Beyond the U.S., Uruguay, Chile, Spain, Saudi Arabia and the UAE are examples of countries targeted by tech giants to expand their data centers, despite their struggles with droughts or arid weather.
Data centers can get water from various sources, including blue sources (e.g., surface water and groundwater), piped sources such as municipal water, and gray sources (e.g., purified reclaimed water). Using recycled or non-potable water to meet a data center’s cooling needs is a well-established practice to conserve limited potable water resources, particularly in dry or drought-prone areas.21
Communities are being negatively affected by the development of data centers. Unlike mining and industrial facilities, data centers are often located near urban areas – this means close to where people live, and kids go to school. Communities are exposed to the health risks stemming from air pollutants emitted by data centers. This research paper examines various sources of air pollutants, with the one directly affecting neighboring communities being emissions from the maintenance and operation of backup generators. Communities also complain about the noise pollution created by the cooling systems.22
People are mobilizing to block or delay the construction of data centers. It would be naïve to think that communities will not react to data centers causing energy bill spikes, impacting the quality and availability of water locally, disrupting their sleep, or creating possible health risks.23 Notable cases are in Ireland, the Netherlands, Chile, Uruguay, and the United States.
The Ugly – Short-term profits over long-term sustainability, and data oligopoly
The explosive growth of data centers requires tradeoffs. By enabling such growth, we are prioritizing short- and medium-term lucrative ventures in exchange for long-term sustainability. In addition to pressures on the electricity grid and water resources for urban areas, data centers have significant land requirements. Data center campuses are only getting bigger.24
As urban areas become saturated, developers are increasingly looking at rural zones. Data centers will become a well-financed competitor for farmland. This implies sacrificing future production of food items or carbon captures to satisfy the growing demand for data centers from the AI boom. Building data centers in farmland presents other challenges in terms of planning and water scarcity. This article summarizes some of those challenges.
Governments offer generous tax exemptions and preferential treatment to attract data centers.25 This creates a race to the bottom dynamic, where these incentives become the decisive factor for companies to invest at a given location. While it is true that there are some economic spillovers from data centers, relying on their contribution to the local economy can be ill-advised. There is a high risk of developing “Dutch disease,” where the development of one sole industry suppresses the development of others.
In addition, when a location’s only differential is tax incentives, another government can offer more generous incentives to poach developers. Developers can pack the equipment and locate it elsewhere with specialized logistics (recall that these are not steel mills or OEM facilities), leaving behind a government and community whose revenues and livelihoods relied on data centers.
Hyperscalers are becoming too big. All AI organizations need data centers of the kind provided by hyperscalers.2627 Amazon Web Services, Microsoft Azure, Google Cloud Platform and Meta are the largest providers and operators of data centers. There is a good chance that the data you input when doing your daily web-browsing is stored in one of these companies’ data centers and potentially used to train their machine learning algorithms. Many of these companies store data for their public sector clients, including the U.S. government.
The high concentration of data and power in a handful of tech giants suggests non-negligible cybersecurity implications. You have most likely received an email saying that “some of your saved passwords were found in a data breach from a site or app that you use.” Data breaches happen often and at massive scales. This company tracks them.
Learning From the Past
The so-called AI revolution, along with the trillion-dollar investment needs of data centers, got me thinking about the state of the world. Going back to my grandpa’s car anecdote, it does not seem like many people think about data centers and their hidden costs when submitting a prompt to their AI tool of choice. I am cautious about AI because it can be wrong or imprecise – and this creates more work for the user. Today I do think more about externalities, although I do recognize the benefits and economic impact of data centers.
I wanted to write this Substack to flesh out the key good, bad and ugly elements of data centers. Hyperscale data centers embody the growth of AI, and both are here to stay. In a different article I reflected on how AI can contribute positively to both society and our planet. However, it would be a mistake to turn a blind eye to its environmental and societal implications, particularly regarding the data center boom.
It is useful to compare this moment with the First and Second Industrial Revolutions. These two critical moments of our history saw unprecedented improvements in living standards, productivity, and growth. However, these were accompanied by poverty, wider inequalities, and environmental degradation. To this date, there does not seem to be a consensus on whether the benefits outweighed the costs. What is unambiguously clear is that the global economy changed drastically during the industrial revolution, going from agriculture to manufacturing to mass production.
A very similar story will unfold with AI. There will be more profits on the table, economic spillovers, and productivity improvements, but at what cost? There will be winners and losers, but how can those losses be mitigated? This revolution is giving us an opportunity to think harder about balancing innovation with ethics and responsibility, ensuring that the benefits of AI are not at the expense of our environment or societal well-being.
Externalities are properly covered in more advanced classes.
Shareholders are becoming increasingly aware of aspects other than profits, like environmental impact, labor practices, governance, etc.
More revenues and lower costs are good for firms’ profits.
Think about playing against the computer in the SNES, the paper clip from old MS Word, or Siri/Alexa.
In total, there are 11,800 data centers in the world. Countries with the most data centers include the U.S. (45.6% share of total), Germany (4.4%), and the UK (4.4%). Source: Ranked: The Top 25 Countries With the Most Data Centers.
The EU’s GDP is also greater ($16.4 trillion).
In my analysis of a popular academic paper on economics and AI, I explain how AI can drive productivity and innovation.
The report also provides an estimate for operations jobs created. However, this figure is small when considering the magnitude of what the report considers a “large” data center. It is also unclear whether they can actually source all the required, more skillful jobs from the local labor force. The number of permanent jobs created by data centers is a subject of debate.
This is made possible by the internet as well.
International Energy Agency. (2024). What the data centre and AI boom could mean for the energy sector. Link.
Environmental and Energy Study Institute. (2025). Data Center Energy Needs Could Upend Power Grids and Threaten the Climate. Link.
In Europe, data centers account for slightly less than 2% of the region’s electricity consumption. In China, this share is 1.1%. Source: Energy and AI report by the IEA.
Olson, Eric, Grau, Anne, Tipton, Taylor. (2024). Data centers draining resources in water-stressed communities. Link.
National Public Radio. (2025). Why more residents are saying ‘No’ to AI data centers in their backyard. Link.
Not to mention the extra noise and traffic caused during construction phase.
In general, tax incentives are used to correct for market failures which imply issues of low human capital, poor connectivity, or high interest rates. When they are used as a sole factor to lure investors, they become problematic and often trigger a race to the bottom, where governments compete with each other to see who offers the most generous incentives.