AI data centers face water wars and cooling crossroads

The initial spotlight on AI data centers' water consumption emerged in early 2026 with a pivotal study by UC Riverside and UT Arlington researchers revealing that GPT-3's operation equated to consuming a 500 ml bottle of water for every 10 to 50 medium-length responses. This academic finding catalyzed public debate, which intensified dramatically when Google’s data center in The Dalles, Oregon, was exposed for using roughly a third of the city’s water supply after a protracted 13-month legal battle. The city’s attempt to expand water sourcing from the Mount Hood National Forest further amplified environmental concerns, illustrating the tangible local tensions between industrial growth and natural resource preservation.

Early discourse often simplified AI data centers' water footprint by comparing it to familiar water-intensive activities like fast food restaurants or golf courses, as seen in SemiAnalysis’s analogy equating xAI’s Colossus 2 data center to 2.5 In-and-Out burger outlets. However, these comparisons risk obscuring the nuanced environmental calculus, since water use intertwines with carbon emissions and varies widely depending on cooling technologies—such as evaporative cooling used by Google in Oregon versus closed-loop or air cooling systems elsewhere—and local climate conditions. This complexity challenges straightforward assessments and calls for a more sophisticated understanding beyond catchy metrics.

Transparency issues have compounded public confusion and debate, as companies often treat water consumption data as proprietary, exemplified by Google’s initial refusal to disclose its usage in The Dalles. This secrecy hampers accurate evaluation of environmental impacts, especially since water withdrawal figures alone do not capture critical factors like whether water is consumed or returned, peak demand timing, infrastructure strain, or who bears associated costs. The controversy underscored that community rights to understand shared resource allocation must be balanced against commercial sensitivities, a tension that shaped the broader discourse on AI’s environmental footprint.

By early 2026, the scale of AI’s water demand was already recognized as significant and rapidly growing, with the AI economy consuming 23 cubic kilometers annually and projections indicating more than doubling by 2050. Training large models like GPT-3 alone could consume millions of liters of water, including hundreds of thousands of liters of clean freshwater evaporated during data center cooling. Yet, despite these large numbers, AI data centers currently use far less water than other sectors such as golf courses, highlighting that the emerging concern centers more on the rapid growth and geographic concentration of data centers in water-stressed regions than on total volume alone.

By mid-2026, mounting community opposition and legislative scrutiny have propelled data center water use into the regulatory spotlight, with states like Texas holding hearings that revealed significant transparency gaps—only 17% of data centers responded to mandatory water surveys, complicating oversight and fueling mistrust. Lawmakers such as Rep. Brad Buckley lamented, 'We’re making huge decisions on bad data,' while citizens called for moratoriums on new construction until environmental impacts are fully assessed, underscoring a growing demand for accountability and community engagement.

In response to these pressures, tech giants like Google have proactively introduced comprehensive water stewardship frameworks aiming to set industry standards by 2030, including commitments to return more water to local watersheds than consumed, avoid water-intensive cooling in stressed regions, and annually disclose water usage. This initiative seeks to rebuild trust through transparency and encourage widespread adoption of sustainable practices amid escalating public and regulatory scrutiny.

Despite these efforts, transparency struggles persist, exemplified by The Dalles, Oregon, where a 13-month legal battle ended with the city revealing that Google's data center consumed roughly a third of its municipal water supply. This case highlights the tension between corporate confidentiality and public interest, prompting commitments to ongoing disclosure and fueling calls for clearer water usage reporting to empower communities and policymakers alike.

At the local and state levels, regulatory responses remain fragmented and evolving; municipalities like Longmont and Boulder have enacted moratoriums on hyperscale data centers to address water concerns, while states such as New South Wales impose stricter water usage reporting and power efficiency caps, as seen when OpenAI and NEXTDC scrapped recycled water cooling plans due to infrastructure and permitting hurdles. Meanwhile, Colorado’s repeated legislative failures to balance economic incentives with environmental safeguards illustrate the complex policy landscape confronting AI data center water management.

By early 2026, closed-loop water cooling systems became the cornerstone of sustainable AI data center operations, with industry leaders like OpenAI and Microsoft championing their water-saving benefits. These systems circulate coolant internally to absorb GPU heat and expel it via radiators without relying on external water sources or evaporation, dramatically cutting fresh water withdrawals—Microsoft alone reports saving over 125 million liters annually per site. Oracle’s innovation in direct-to-chip, closed-loop, non-evaporative cooling further pushes efficiency boundaries by slashing server-fan power consumption by up to 80%, signaling a major leap in both sustainability and operational cost reduction.

Nvidia’s Rubin AI data center design exemplifies the next evolution in liquid cooling by operating coolant at unprecedentedly high temperatures—up to 45°C at intake and 55°C at outlet—enabling the use of outdoor dry coolers and eliminating the need for traditional water-intensive cooling towers. This closed-loop system, using a 75% water and 25% propylene glycol mix, not only reduces water consumption to near zero but also cuts cooling energy costs significantly, with projected savings exceeding $4 million annually for a 50-megawatt facility. Nvidia’s approach also removes noisy fans, allowing for higher server density and quieter operation, addressing both environmental and practical challenges posed by the Vera Rubin GPUs’ intense power demands.

Hybrid cooling systems like those deployed by Isambard AI demonstrate a pragmatic balance between closed-loop dry cooling and evaporative methods, minimizing water loss by operating primarily in dry mode and only resorting to water use during peak heat conditions. These systems achieve remarkable energy efficiency, with each cooler managing up to 2 megawatts of heat using just 16 kilowatts of electricity, and annual water consumption comparable to that of 21 typical UK households. Additionally, Isambard’s liquid cooling enables repurposing waste heat for building heating, a sustainability advantage over traditional air cooling where heat dissipates unused into the atmosphere.

Emerging technologies like Accelsius NeuCool® two-phase direct-to-chip liquid cooling push the envelope further by operating efficiently at facility water temperatures up to 54°C and even 59°C with increased flow rates, surpassing single-phase cooling capabilities. This advancement aligns with Nvidia’s vision for global free cooling, offering approximately 4% annual energy savings per 1°C temperature increase and enabling AI infrastructure to thrive in hotter climates. Meanwhile, pioneers like BDC have scaled direct-to-chip liquid cooling to meet soaring rack densities—now exceeding 100 kilowatts per rack and projected to surpass 200 kilowatts—highlighting liquid cooling’s critical role in supporting next-generation high-density AI hardware deployments.

By early 2026, the surge in AI workloads has pushed rack power densities from traditional air cooling limits of 10–20 kW to 50–130 kW, rendering conventional cooling architectures inadequate and forcing data centers to reconsider their thermal designs. This escalation has transformed cooling from a manageable operational expense into a critical architectural constraint that determines a facility’s ability to attract premium AI tenants. Simultaneously, water use for cooling has become a binding constraint, especially in water-stressed regions, with U.S. data center water consumption projected to nearly double from 17 billion gallons in 2023 to up to 33 billion gallons annually by 2028, raising the risk of regulatory curtailments that could disrupt operations.

Data center operators face a complex trade-off between conserving water and increasing energy consumption, as cooling already accounts for 30–40% of total energy use and this share grows with rising densities unless thermal architectures improve significantly. To maintain competitiveness, investors must decide whether to overhaul mechanical cooling systems with advanced methods such as direct-to-chip, immersion, or hybrid cooling, or risk losing high-value AI tenants to better-equipped facilities. Enterprises are increasingly adopting a future-proofing strategy by incorporating scalable liquid cooling options, as exemplified by providers like Aligned Data Centers, allowing them to defer costly investments until workloads demand higher density cooling beyond 35 kW per cabinet.

The operational and environmental trade-offs inherent in cooling choices are starkly illustrated by recent high-profile cases such as OpenAI and NEXTDC’s abandonment of recycled water cooling for their 612MW Sydney data center due to pipeline and regulatory hurdles. This pivot to more energy-intensive air cooling eliminates water use but increases electricity demand and potentially worsens power usage effectiveness (PUE), highlighting how infrastructure limitations and regulatory environments can force compromises between sustainability goals and operational feasibility. NEXTDC CEO Craig Scroggie emphasized that while recycled water remains preferred where infrastructure permits, current constraints in Western Sydney made it unviable for this project.

Google’s decision to employ air cooling at its Botetourt County data center further underscores the nuanced sustainability trade-offs between water and energy consumption. While air cooling reduces water use, it increases electricity demand, and as Eric Sjostedt notes, 'which method is less costly to the environment depends on the location of the data center and the specifics of the facility.' Moreover, local utility infrastructure capacity often dictates cooling choices, as some systems cannot support the high water demands of evaporative cooling. Beyond energy and water metrics, air cooling introduces additional environmental concerns such as noise pollution and heat island effects that can impact local ecosystems, as highlighted by Kathryn Hatam’s observations on low-frequency sounds affecting pollinators and surrounding wildlife.

By early 2026, the AI data center industry and investors have grown acutely aware that traditional underwriting models no longer suffice as AI workloads push rack power densities to unprecedented levels of 50–130 kW, creating new cooling and design risks. Facilities unable to support these premium kW/rack profiles face stranded-capacity risks, where data centers may appear fully utilized by megawatts but generate diminished revenue due to cooling constraints, fundamentally reshaping investment strategies and asset valuations.

Transparency around water use and cooling risks is improving, with major players like Google introducing comprehensive water usage guidelines aiming for industry-wide adoption by 2030, including commitments to return more water to local watersheds and avoid water-intensive cooling in stressed regions. Credit and risk frameworks have begun pricing exposure to these constraints, reflecting growing economic impacts on communities and investors, while companies such as OpenAI and Microsoft emphasize closed-loop cooling systems that significantly reduce water withdrawal, signaling a shift toward sustainable operations.

Technological advances, particularly Nvidia’s Rubin AI data center design unveiled mid-2026, demonstrate a transformative future outlook by employing hotter closed-loop liquid cooling systems operating at up to 45°C, which drastically cut water and energy consumption while enabling siting in arid regions. This innovation, alongside Microsoft’s water-free cooling commitments and Amazon’s plan to use reclaimed wastewater at over 120 data centers by 2030, exemplifies an industry trend toward near-zero water use, although experts caution that zero consumption remains unrealistic and upfront costs remain high.

Despite efficiency gains lowering per-query costs, total AI data center water and energy demand is projected to surge, with hyperscale water withdrawals potentially reaching 16–33 billion gallons annually by 2028 and up to 6.6 billion cubic meters by 2027—comparable to the consumption of small nations. This growth underscores the urgency of sustainable cooling solutions, while community engagement efforts highlight the economic benefits data centers bring through job creation, tax revenues, and grid infrastructure investments, even as operators strive to combat misinformation about water use and environmental impacts.