The Hidden Costs of AI: Energy, Water, and Community Impact

The public is divided on artificial intelligence, with some envisioning a future of solutions to global problems and others fearing an AI-driven apocalypse. However, the majority of people fall somewhere in the middle, feeling uncertain and uneasy about its implications.

Despite the magnitude of the questions surrounding AI, concrete answers about its real-world impacts on the planet, including energy, water, land, and surrounding communities, are difficult to ascertain. This video aims to provide a shared truth by consulting scientists and researchers who study these issues daily, rather than relying on social media claims or tech CEOs' statements.

Concerns about computers consuming vast amounts of electricity are not new. In 1999, predictions suggested the internet would use half of all US electricity. One commentator claimed a wireless Palm Pilot's networking power was equivalent to a refrigerator, an overestimation by a factor of 2,000.

In reality, the entire information technology industry in the late 1990s accounted for approximately 3% of US electricity demand, a significant but not critical amount for the grid. Early data centers consumed only about one-tenth of 1% of the total US electricity.

Between 2010 and 2017, internet usage increased, but data center power consumption grew at a much slower rate. This was due to engineers and scientists improving computing efficiency, enabling more tasks with less energy, a concept some refer to as Koomey's law.

Today's AI technology has introduced an unprecedented electricity demand through hyperscale AI data centers. These facilities are giant industrial operations dedicated to providing massive computing power 24 hours a day, filled with tens to hundreds of thousands of powerful GPU chips.

A single rack of AI servers can consume five to 30 times the power of a conventional server rack. A GPU rack, about the size of a home refrigerator, can power up to 65 houses. Hyperscale data centers in the US require an electric power draw exceeding 100 megawatts, enough to power approximately 80,000 American homes.

Currently, 68 hyperscale data centers are either built or under construction in the US, with an additional 267 approved or planned. This rapid proliferation means that by 2024, compute data centers accounted for about 4.7% of US electricity use, roughly one-seventh of the electricity used to power US homes annually.

Gathering valid and reliable information on the impacts of data centers is incredibly challenging due to the heavy shroud of secrecy, non-disclosure agreements, and a general lack of transparency surrounding these projects. This opacity extends to details like water and electricity consumption, and job creation figures.

The AI industry is projected to invest over $1 trillion in data centers and infrastructure this year. Forecasts suggest data centers could account for 7% to 12% of total US electricity use by 2028, though these are just predictions. The Jevons Paradox, where increased efficiency leads to greater overall resource consumption, suggests that even with more efficient processing chips, total electricity use could rise due to increased computing activity.

Unlike a national problem, data center electricity use is a highly localized phenomenon. Data centers connect to specific local electrical grids, meaning that national totals can mask the severe strain placed on particular areas.

Connecting a hyperscale data center requires significant infrastructure, including high-voltage connections, large transformers, additional transmission lines, and potentially a dedicated power substation. These upgrades can cost millions of dollars and take years to implement.

The costs of these infrastructure upgrades may be passed on to ratepayers, leading to higher local electric bills. In areas with older grids or limited excess energy generation, data centers can severely strain local resources.

Around 25% of announced data center construction plans include building their own power plants, although only about 2% to 3% have been completed to date. Satellite analysis from the International Energy Agency indicates that roughly a quarter of these planned private power plants have already begun construction.

An example is the 'GW Ranch' in West Texas, which plans to install 35 gas turbine generators capable of producing over 7 gigawatts of peak power. If completed, this would make it the largest power plant in the United States, consuming more than 1% of the entire US natural gas output for a single data center.

While it is entirely possible to power data centers with renewable energy, such as solar panels and batteries, this approach is largely not being adopted. Some data centers, like one in Nevada, currently utilize renewable energy sources, demonstrating its technical feasibility.

Critics highlight the paradox of companies claiming no time to wait for renewables yet accepting seven-year timelines for gas turbine installations. The urgency often drives companies to choose familiar and readily available fossil fuel options rather than cleaner alternatives.

Companies prioritize rapid development, and if local grids lack the capacity for connection, on-site generation becomes one of the few options. However, for society as a whole, there is no real urgency for AI development to occur within a year or two versus three, implying a less hurried approach could allow for more sustainable solutions.

Currently, US data centers account for approximately 0.5% of America's total greenhouse gas emissions, a fraction compared to sectors like transportation or agriculture. However, due to the Jevons Paradox, where increased efficiency drives greater overall use, total AI use is expected to increase, leading to higher emissions.

Estimates suggest that AI carbon emissions could reach 44 megatons by 2030, which is double the industry's emissions in 2025. This increase is roughly equivalent to adding 10 million gas-burning cars to US roadways, though it would still constitute only about 1% of the US's total CO2 emissions. Predicting future AI emissions remains largely a guessing game.

Carbon emissions are a global problem, as the atmosphere is indifferent to the location of CO2 release, with impacts being the same everywhere. Beyond carbon, data centers also generate significant amounts of heat as a direct consequence of power flowing through components like copper wires and GPUs.

This generated heat must be managed to prevent damage to expensive computer chips. While there was a previous claim that asking an AI question uses one bottle of water, a 2023 study by Dr. Shaolei Ren indicated that asking OpenAI's GPT-3 model 30 medium-sized questions used about 500 milliliters. However, these numbers are quickly becoming outdated in the rapidly evolving AI industry, and the majority of this water usage is indirect, for electricity generation.

While individual AI models may become more efficient in their water use per question, the overall consumption of water is likely to increase because people are asking more questions more frequently, illustrating the Jevons effect once again.

This phenomenon means that even if the systems become more efficient to train and run, the net environmental impact of AI can still be larger due to increased deployments and usage of these models.

In 2023, US data centers directly used approximately 17 billion gallons of water annually for cooling. When indirect water use for electricity generation is included, this figure rises to 228 billion gallons, about two-thirds of New York City's annual drinking water. Projections suggest this could skyrocket to over 840 billion gallons annually, exceeding the combined drinking water use of New York City and Chicago, which is concerning given that new water cannot be created.

Data centers manage heat dissipation through various cooling methods. One option involves giant outdoor radiators with fans, which use less water but require more electricity to move the heat, indirectly increasing water demand for power generation.

Another method is evaporative cooling, which is more energy-efficient but consumes significantly more water, especially during summer months. This method's effectiveness decreases when both indoor and outdoor temperatures are high, coinciding with periods of acute water scarcity in many regions.

For instance, data centers around Phoenix, Arizona, currently use 385 million gallons of water for cooling annually, excluding electricity generation. This figure is projected to increase to 3.7 billion gallons per year as more data centers come online, exacerbating an existing water crisis where the Colorado River, a primary source for Phoenix, has lost about 25% of its flow due to drought and overuse.

Companies investing in AI recognize that increasing fossil fuel power plants and water consumption are not inherently beneficial. However, they anticipate that AI will eventually become powerful enough to solve these problems, with industry figures claiming AI will self-finance its emissions and could reduce global emissions by 10% by 2030.

The host and an expert counter that these claimed benefits remain unmeasured, while communities are already experiencing the environmental impacts. The promised future benefits, such as enhanced air travel efficiency, improved crop prediction, or reduced food waste, are currently speculative.

The narrative that AI must be deployed rapidly for climate problem-solving serves in part to justify actions like building private power plants. However, the costs are immediate and tangible, with little evidence of the promised benefits materializing as yet.

Data center development is characterized by speed, secrecy, and opacity, effectively excluding community members, regardless of their background, from the conversation. Developers frequently use shell companies to establish a presence.

For example, in West Memphis, Arkansas, 'Willowbend Capital' and 'Forgelight Ventures' paved the way for a Google data center. Meta utilized 'Winner' and 'Wurldwide' to secure multimillion-dollar tax breaks for data centers in Fort Worth and El Paso, Texas. Non-disclosure agreements often prevent officials from sharing information with citizens about potential data centers; in Pine Island, Minnesota, city officials knew of a proposed data center for two years before residents.

This lack of transparency often leads to feelings of abandonment among residents, especially by local government officials who collaborate secretly with the tech industry. While data centers can generate substantial tax revenue, as seen in Loudoun County, Virginia, which reported a billion dollars in revenue from taxing data centers, the benefits often come with significant trade-offs.

Despite data centers' potential for significant tax revenue and the tech industry's claims of AI's immense value, 38 US states offer various tax incentives to attract these operations. This practice is described as a 'race to the bottom,' where states reduce taxes to entice development, effectively becoming a net transfer of money to tech companies.

With the tech industry spending $700 billion this year on AI data centers, the question arises as to why such high-value investments require state subsidies. Critics argue that these tax revenues come with serious strings attached, including other costs, and that small communities struggle to negotiate with billion-dollar corporations.

Furthermore, once operational, data centers do not generate many local jobs, as they are largely automated. This challenges the notion that these facilities provide substantial community benefits beyond tax revenue, which is often diluted by tax breaks.

For communities to make informed decisions about data center projects, they need access to crucial information, including projected water usage, energy sourcing, and expected community benefits. Unfortunately, this information is often not readily available, despite existing.

Local utilities possess data on water and power consumption for each data center but are typically restricted from releasing it due to secrecy agreements. This lack of transparency prevents communities from assessing the full costs and impacts of these technologies. It is argued that for such profitable infrastructure, companies should be held accountable for following rules and paying the full societal cost of their actions, as imposing costs on others is both inefficient and unfair.

The scientific community is currently unable to measure the net effects of AI data centers. While ideally, policy decisions should weigh costs against benefits, the comprehensive data needed to quantify this balance does not yet exist.

The future of artificial intelligence will ultimately be shaped by human intelligence, but this requires a clear understanding of reality versus speculation. Regardless of individual perspectives on AI's importance or its cost, the full implications of an AI-powered future are being determined now.

Society needs to accurately assess the real benefits and costs of this technology to decide whether specific applications are truly worthwhile, underscoring the necessity for wise choices.