Microsoft built an AI data center in Iowa. The town's water supply paid the price.

The tech industry has spent the last decade confidently making ambitious, highly publicized pledges to achieve "Net-Zero Emissions." This is a corporate or governmental pledge to achieve a balance between the greenhouse gases put into the atmosphere and those taken out, often heavily reliant on purchasing carbon offsets. You have seen the press releases promising carbon negative operations, water positive status, and zero-waste campuses. These promises formed the bedrock of modern cloud computing's pristine, ethereal branding. The "cloud" sounds weightless and clean, abstracting away the industrial reality of server farms.

However, the resource-intensive rise of generative AI over the past few years has fundamentally altered the physical math of these operations. We are rapidly transitioning from an era of relatively low-cost data retrieval to an era of high-cost data generation. Despite public commitments from major tech companies to reach net-zero emissions by 2030, the unprecedented resource demands of training and running generative AI models are measurably increasing global fossil-fuel reliance and straining local water supplies. This shift renders current corporate sustainability targets physically unachievable under existing infrastructure trajectories. The scale of deployment has simply outrun the physical limits of local electrical grids and municipal water basins.

The Thirsty Data Centers in the Heartland

When you query a large language model, the computation does not happen in the ether; it happens in places like West Des Moines, Iowa. When Microsoft built a massive data center cluster in this Midwestern city to power its cloud infrastructure and train OpenAI's flagship models, the local municipal water supply allegedly paid the price. We have the receipts on this usage. According to AP News, the supercomputing clusters required to train models like GPT-4 caused a highly noticeable spike in local water consumption.

In July 2022, the month before OpenAI announced it had completed training GPT-4, Microsoft pumped approximately 11.5 million gallons of water to its Iowa data centers. This represented roughly six percent of all the water used in the district, a heavy burden for a single corporate customer. These facilities operate at high temperatures and require millions of gallons of water simply to prevent the servers from melting down under the immense computational load documented during these training runs (AP News).

A cooling tower works by evaporating water to remove heat from the data center's internal air conditioning systems. The servers generate intense thermal energy, which is transferred to a closed-loop water system. This heated water is then pumped to the exterior cooling towers, where it is exposed to outside air. A fraction of the water evaporates, cooling the remaining liquid, which is then cycled back inside. This evaporative process is highly effective at managing thermal loads, but it directly consumes the water, meaning it cannot be recycled back into the local municipal system.

When temperatures spike in the summer months, the evaporation rate increases dramatically. This places the greatest strain on municipal reservoirs precisely when community drought conditions are at their worst.

Other tech giants have faced similar localized reckonings across different municipalities. Google’s facilities in The Dalles, Oregon, drew public scrutiny after records revealed their data centers accounted for more than a quarter of the city's total water use. Across the broader corporate landscape, Google reported a 20 percent increase in its total water consumption in 2022 alone. This increase was driven largely by its artificial intelligence operations, as detailed by The Washington Post. Microsoft’s global water consumption spiked 34 percent during the same period, highlighting a sector-wide trend.

This localized strain is just a microcosm of a much broader, continent-wide surge in resource demand. North American data center power requirements grew from a baseline of 2,688 megawatts in 2022 to 5,341 megawatts in 2023. They effectively doubled in a single year, as documented by MIT News. This rapid deployment is outstripping the capacity of sustainable infrastructure. "The demand for new data centers cannot be met in a sustainable way," researcher Noman Bashir explained to MIT News. The pace at which these corporate monoliths are building new data centers means that "the bulk of the electricity to power them must come from fossil fuel-based power plants" (MIT News).

The Thermodynamics of Generative AI

To understand why this technology requires such vast resources, we have to look at the thermodynamics of how it functions. Training just one early foundational model, OpenAI's GPT-3, consumed an estimated 1,287 megawatt-hours of electricity and generated 552 tons of carbon dioxide, according to MIT News. That is merely the baseline cost to bring a legacy model into existence. For newer, exponentially larger models, the energy costs scale aggressively in proportion to the parameter count.

Training relies on tens of thousands of specialized graphics processing units running at maximum capacity for months. A single Nvidia H100 GPU, the current industry standard, has a maximum power consumption of 700 watts. When deployed in clusters of 100,000 units, the baseline electrical draw becomes equivalent to heavy industrial manufacturing. All of this electricity inevitably turns into heat, and that heat requires water or intensive air conditioning for dissipation.

Once trained, the models enter the phase of "Inference." This is the process of using a trained AI model to make predictions or generate text or images on new data, distinct from the initial, highly energy-intensive training phase. Inference operates at a staggering scale across hundreds of millions of daily active users. A single ChatGPT query consumes approximately five times more electricity than a traditional Google web search, as documented by MIT News.

Cooling these massive banks of data center servers typically requires about two liters of water for every kilowatt-hour of energy consumed, per MIT News. Some researchers estimate that when a user engages in a standard 20-to-50 prompt interaction with a large language model, the system essentially uses a 500-milliliter bottle of water. When billions of queries are processed daily, the municipal water burden scales linearly. This shifts the environmental cost directly onto local communities.

Extending the Hardware Lifecycle Burden

Beyond electricity and water, the physical footprint of generative AI includes the manufacturing and disposal of the hardware itself. The servers powering these systems are not permanent fixtures that run indefinitely. The intense thermal loads and rapid pace of hardware deprecation mean that AI clusters often require replacement every three to five years. This introduces a massive embodied carbon footprint.

Scope 3 emissions account for the vast majority of a tech company's total carbon footprint. While Scope 1 covers direct emissions and Scope 2 covers purchased electricity, Scope 3 encompasses everything from the mining of raw materials to the final disposal of the hardware. The rapid deployment of generative AI has caused these supply chain emissions to skyrocket. When companies publish their sustainability reports, they often emphasize their success in mitigating Scope 1 and Scope 2 emissions through renewable energy credits, while acknowledging that Scope 3 emissions remain a massive, unresolved challenge. The sheer volume of steel, concrete, and silicon required to build out these 100-megawatt facilities guarantees a massive carbon debt before the servers are even powered on.

Manufacturing a high-performance server requires extracting rare earth metals, refining silicon, and executing complex global shipping operations. According to research published in Nature, the embodied emissions—the greenhouse gases emitted during the creation and transport of the hardware—can rival the operational emissions over the lifespan of the equipment. We are trading long-term hardware durability for short-term computational gains.

When these specialized servers reach the end of their brief operational viability, they contribute to the rapidly growing problem of electronic waste. Unlike general-purpose CPUs which might be repurposed for lower-intensity tasks, the highly specialized nature of AI accelerators limits their secondary market utility. The industry has largely obscured this hardware churn behind the clean abstraction of cloud scaling. As Scope 3 emissions rise, the true cost of hardware deprecation becomes harder to hide.

Delaying Fossil Fuel Retirements

The most direct threat to corporate net-zero pledges is the unavoidable reality of base-load power. Renewable energy sources like wind and solar are intermittent; they generate power when the wind blows or the sun shines. Data centers, however, require constant, uninterrupted power 24 hours a day, 365 days a year. To guarantee this uptime, grid operators must maintain reliable base-load generation.

In many regions, this base-load power comes from fossil fuels. Because the pace of data center construction vastly outstrips the timeline required to permit and build new renewable energy sources, the energy gap is inevitably filled by exactly the infrastructure we are supposed to be retiring. Purchasing renewable energy credits does not change the physical fact that coal or natural gas is being burned locally to keep the servers online.

For example, in states like Virginia, Utah, and Kansas, utility companies have explicitly cited data center growth as the reason for extending the lifespans of aging coal and natural gas plants. As Bloomberg reported, power companies are delaying the planned retirement of fossil-fuel facilities simply to ensure the grid does not buckle under the new industrial demand. In Ireland, the situation became so severe that state grid operator EirGrid imposed a de facto moratorium on new data center connections in the Dublin area, citing grid stability concerns (Reuters).

From Crypto Mining to LLMs

This is not the first time an opaque digital asset class has plausibly threatened to strain regional infrastructure. Cryptocurrency mining surges in previous years similarly strained local power grids. They drew massive public backlash over their environmental footprint (MIT News) before generative AI became the main driver of global compute demand.

However, AI enjoys a level of institutional and corporate backing that crypto never achieved, meaning its infrastructure integration is far more deeply embedded in the modern web. Cryptocurrency miners often sought the cheapest, most remote power sources, regardless of latency. They could set up shipping containers in rural areas next to hydroelectric dams or flared gas wells. If the power went out, the mining simply paused without disrupting consumer services.

Large language models require a fundamentally different architectural footprint. The shift from retrieving indexed data to generating novel tokens on the fly necessitates high-bandwidth interconnects and proximity to major network backbones (MIT News). Every generated summary or synthetic image requires a cascading sequence of matrix multiplications happening near major population centers. As Elsa A. Olivetti explained to MIT News, "When we think about the environmental impact of generative AI, it is not just the electricity you consume when you plug the computer in. There are much broader consequences that go out to a system level."

Industry Defense: The Efficiency Argument

Faced with mounting scrutiny, the tech industry has mounted a defense centered largely on optimization and future breakthroughs. Tech companies confidently argue that while total data center footprint is growing, proprietary hardware and optimized routing have made the per-inference environmental cost negligible. They contend that the models themselves are becoming vastly more efficient, keeping them on track for their sustainability goals.

Google, for instance, published a detailed analysis in the Google Cloud Blog. They estimated that a median Gemini Apps text prompt uses just 0.24 watt-hours of energy, emits 0.03 gCO2e, and consumes 0.26 milliliters of water. They equate the energy cost of a single prompt to watching television for nine seconds (Google Cloud Blog).

Furthermore, defenders point to improvements in "Power Usage Effectiveness" (PUE). PUE is a metric used to measure overhead energy efficiency in a data center, calculated by dividing the total amount of power entering a facility by the power used to run the IT equipment within it. A PUE of 1.0 is theoretical perfection. Modern hyperscale data centers often achieve numbers around 1.1 or 1.2, meaning very little energy is wasted on overhead cooling and lighting.

Industry advocates also routinely argue that AI will ultimately accelerate climate solutions. They suggest that large language models will optimize supply chains, discover new materials for better solar panels, and manage smart grids with unprecedented efficiency. In this view, the short-term spike in energy use is an investment that will pay planetary dividends down the road. They view the carbon emitted today as the necessary cost of admission for the climate-saving technologies of tomorrow.

Microsoft's recent investments in restarting the Three Mile Island nuclear facility demonstrate a corporate pivot toward securing massive, dedicated zero-carbon base-load power. As The Verge reported, this represents a long-term strategy to decouple AI growth from fossil fuels. Defenders argue these mega-projects prove the industry is taking the base-load problem seriously.

This defense, however, falls victim to Jevons paradox. This economic principle states that as technological progress increases the efficiency with which a resource is used, the rate of consumption of that resource actually rises due to increasing demand. Making models more efficient per query merely allows companies to deploy them in more places, lowering the cost of generating infinite synthetic text and ultimately driving up total power usage.

As researchers documented in MIT News, this macro-scale reality forces grid operators to rely heavily on fossil fuels to absorb the massive base-load increase, rendering the per-prompt efficiency moot in the context of total global emissions. The industry wants you to focus on the efficiency of the single drop of water, rather than the fact that they are draining the entire reservoir. Companies like Microsoft, Meta, and Google all still publicly pledge to reach net-zero carbon emissions and become "water positive" by 2030, despite their rapidly rising usage logs (NPR).

The 12 Percent Forecast

Looking ahead, the long-term implications of this trajectory are deeply concerning for global infrastructure. According to the International Energy Agency, global electricity consumption by data centers already reached 460 terawatt-hours in 2022. It is currently projected to approach 1,050 terawatt-hours by 2026. This is roughly equivalent to the entire electricity consumption of Japan.

The regional forecasts within the United States are even more stark. The Lawrence Berkeley National Laboratory forecasts that U.S. data centers could consume up to 12 percent of the nation's total electricity by the year 2028, as reported by NPR. As we have established, the timeline to permit, fund, and connect new solar or wind farms is measured in years. The timeline to deploy a new data center is measured in months.

As Benjamin Lee noted on NPR's Short Wave podcast, the corporate narrative is beginning to fracture under the weight of these physical realities. "I think before generative A.I. came along in late 2022, there was hope among these data center operators that they could go to net zero," Lee told NPR. "I don't see how you can, under current infrastructure investment plans, you could possibly achieve those net zero goals."

The disparity between ambition and physical reality is widening. A pledge to be "water positive" means the company promises to put more water back into the environment than it consumes. However, funding wetland restoration in one state does not alleviate the immediate strain on a drought-stricken community where the servers are actively drawing municipal water. The local utility still has to pump, treat, and deliver that water, and the local residents still bear the physical infrastructure costs. The spreadsheet balances, but the local watershed remains depleted.

Reconciling AI Innovation With Physical Limits

The physical reality of generative AI exposes a deep contradiction at the heart of Silicon Valley's operations. Returning to the thesis, the evidence clearly shows that while individual models and localized PUE metrics may become marginally more efficient per query, the sheer scale of the AI infrastructure rollout has fundamentally outrun the capacity of renewable energy grids. The timeline of technological expansion is categorically out of sync with the timeline of infrastructure transition.

The 34 percent spike in water consumption documented by Microsoft, the projected reliance on fossil fuels to meet the 1,050 terawatt-hour demand by 2026, and the delayed retirements of coal plants all validate the premise. Current corporate sustainability targets are physically unachievable under the current growth curve. Optimization at the micro-level cannot mathematically outpace expansion at the macro-level.

Until the industry acknowledges that its 2030 net-zero pledges are intrinsically incompatible with doubling regional power requirements and tapping municipal water supplies, their environmental accounting will remain fiction. The models might be generating impressive poetry, but their most lasting output may simply be a logged regression in global climate goals. The math does not support the marketing, and the grid cannot sustain the load.