The paper states plainly that AI companies’ 2030 net-zero aspirations cannot be met without “substantial reliance on highly uncertain carbon offset and water restoration mechanisms”.

AI’s Environmental Cost in the United States: A Critical Assessment

by ChatGPT-5

The accelerating growth of AI servers in the United States—fuelled by generative AI adoption and exponential increases in chip and server production—has created a sustainability challenge of unprecedented scale. According to a report, AI server deployment could generate 731–1,125 million m³ of annual water use and 24–44 Mt CO₂-equivalent emissions per year by 2030, depending on growth trajectories.

Even the lowest growth scenario shows a dramatic upward curve, exceeding previous forecasts for the entire U.S. data-center market, revealing the extent to which AI-specific server architectures differ from traditional data-center assumptions. The study also highlights the deep role of local grid composition, climate conditions and cooling methods, and demonstrates that the industry is unlikely to achieve its publicly stated “net-zero by 2030” ambitions without relying heavily on uncertain carbon offsets and water-restoration schemes.

Environmental Burden and Spatial Vulnerabilities

The report identifies that indirect water use—primarily the water consumed to generate electricity—accounts for 71% of total AI server water footprint. Direct cooling water accounts for the remaining 29%. The intensity of the footprint is highly dependent onlocation-specific grid water factors, grid carbon factors, and local climate. For instance, AI servers placed in California, Oregon, and Washington benefit from relatively low-carbon grids because of heavy hydropower use—but hydropower significantly increases water consumption through evaporation, producing a paradox where “green” locations for carbon are “red” for water scarcity.

The authors highlight that ten states face severe water scarcity risks, including California, Nevada, Arizona, New Mexico, Utah, Washington, Colorado, Wyoming, Oregon, and Montana. Ironically, many of these states host existing or planned AI data-center capacity. The greatest opportunity for sustainable expansion lies in Midwestern states—such as Texas, Montana, Nebraska, and South Dakota—which combine high renewable energy potential, lower water stress, and superior projected carbon and water intensities per unit of energy.

Efficiency and Technological Limits

The report presents analyses showing that operational efficiencies—measured through Power Usage Effectiveness (PUE) and Water Usage Effectiveness (WUE)—can deliver meaningful but limited relief:

• Best-case PUE improvements could reduce energy and carbon by ~7%

• Best-case WUE improvements could cut water use by ~29%

• Server Utilization Optimization (SUO) adoption reduces all environmental footprints by ~5.5% in best case, but increases them by ~7.3% in worst case

Even combining all best-case industry practices yields only 12% energy reduction, 32% water reduction, and 11% carbon reduction by 2030—far short of the net-zero threshold.

Furthermore, the report shows that innovations such as advanced liquid cooling (ALC) provide marginal improvements (1.7–2.4% reductions) and cannot offset macro-level demand growth. Efficiency gains in future chips (e.g., Blackwell, Rubin) may be negated by a rebound effect, where cheaper computational cost increases overall demand, especially under high-application scenarios.

Most Surprising Findings

1. Water footprint dominance – Indirect water use (71% of total) dramatically overshadows direct cooling water use, contradicting the public focus on “cooling water” alone.

2. Hydropower paradox – States celebrated for clean electricity (e.g., California, Oregon, Washington) are simultaneously the worst locations for water sustainability due to high evaporative losses from hydropower.

3. Net-zero is mathematically unachievable without massive offsets – Even with best-case improvements, residual carbon and water footprints remain so high that companies would need enormous volumes of offsets—28 GW of wind or 43 GW of solar to compensate the residual 2030 footprint in best case scenarios.

4. Worst-case scenario exceeds sustainable limits by orders of magnitude – Under worst practices, the U.S. AI server footprint would generate 71 Mt of residual yearly emissions and over 5.2 billion m³ of water consumption, which the authors call “nearly impossible to fully compensate”.

Most Controversial Findings

1. Industry net-zero claims are unrealistic – The paper states plainly that AI companies’ 2030 net-zero aspirations cannot be met without “substantial reliance on highly uncertain carbon offset and water restoration mechanisms”.

2. Optimal siting conflicts with U.S. infrastructure reality – Texas and Midwestern states offer the best sustainability profile but lack current infrastructure to support hyperscale AI deployments, raising concerns about reliability, connectivity, and security.

3. Renewable energy adoption may worsen water scarcity – High renewable penetration (LRC scenario) helps carbon reduction but can increase water risks if it depends too heavily on hydropower; conversely, low renewable adoption (HRC scenario) worsens carbon by +20% and water by +2%.

4. Huge public-private partnerships are recommended – The authors propose that AI companies should co-invest in monitoring air quality (e.g., PM2.5 sensors) and water systems with local governments—suggesting a role for tech companies far beyond traditional boundaries of corporate responsibility.

Most Valuable Findings

1. Holistic modelling across energy–water–carbon nexus – The report integrates PUE, WUE, local grid projections, and climate data for every U.S. state, creating an actionable spatial allocation strategy.

2. Clear, quantified regional guidance – Midwestern states—Texas, Montana, Nebraska, South Dakota—are identified as highest-value sites for sustainable AI expansion, balancing water, carbon, and renewables availability.

3. Quantification of rebound effects – Unlike previous studies, the report explicitly models scenarios where efficiency gains increase AI usage, altering environmental impacts.

4. Sensitivity analysis identifies key uncertainties – Server lifetime, manufacturing capacity, U.S. allocation ratio, idle power ratio, and training/inference mix each produce up to 40% variance in footprints, framing the limits of predictability.

Recommendations for Regulators

Based strictly on the report’s findings, regulators should:

1. Introduce mandatory environmental disclosure for AI server operators

• Require reporting of PUE, WUE, electricity source, and grid water factors.

• Ensure independent verification to counter the opacity highlighted by the study.

2. Establish national limits on AI data-center siting in water-scarce regions

• Restrict or regulate new capacity in states identified as high-scarcity/high-water-footprint regions (CA, NV, AZ, NM, UT, CO, OR, WA, MT).

3. Incentivize low-impact locations through federal tax credits

• Align incentives with states offering best combined water–carbon profiles and renewable potential.

4. Mandate integrated water–carbon planning in permitting processes

• Add water scarcity and grid water intensity assessments into environmental impact review.

5. Develop real-time environmental impact monitoring systems

• Use AI-driven dashboards to track energy, water, and carbon in real-time, as the report suggests.

6. Limit dependence on offsets through strict additionality requirements

• Enforce standards for carbon and water offsets to avoid “virtual compliance” pathways.

Consequences of Non-Regulation

If regulators fail to act, the report warns of severe systemic consequences:

1. Water scarcity crises in major U.S. regions

• Increased risk to municipal water supplies affecting millions, especially in western states.

2. Significant contribution to national emissions

• AI server emissions alone could reach 44 Mt CO₂e annually by 2030—similar to the annual emissions of a mid-sized industrial nation.

3. Uncontrollable growth of offset markets

• Companies would be forced into unprecedented reliance on offsets, risking credibility, transparency, and market distortion.

4. Infrastructure congestion and grid instability

• Local grids, especially in Texas and other renewable-rich states, may face congestion and reliability challenges requiring huge upgrades.

5. Rebound effects causing runaway demand

• Efficiency gains without regulation will accelerate AI workloads, pushing total environmental burdens to unsustainable levels.

6. Loss of public trust and increased environmental conflicts

• Local communities will face air quality impacts, water stress, and land use conflicts, leading to backlash and resistance to AI infrastructure expansion.

Conclusion

The report makes clear that AI’s environmental footprint is not a distant concern but an urgent and rapidly escalating crisis. Its integrated modelling across energy, water, and climate dimensions reveals structural contradictions in current AI deployment strategies and demonstrates that, without substantial regulatory intervention, the AI industry cannot achieve its stated sustainability goals. Regulators must act decisively to coordinate siting, efficiency standards, grid decarbonization strategies, and transparency requirements. Otherwise, the U.S. risks locking in a trajectory where unregulated AI growth accelerates environmental degradation faster than technological progress can mitigate it.