TL;DR
- The gist: A new peer-reviewed study estimates that AI systems will consume up to 23 gigawatts of power in 2025, generating carbon emissions comparable to New York City.
- Key details: The research projects AI water consumption could reach 764.6 billion liters, exceeding global bottled water demand.
- Why it matters: The findings challenge corporate sustainability claims by highlighting the massive “indirect” water and carbon costs hidden in electricity generation for AI infrastructure.
- Context: Tech giants are scrambling for power, pivoting to nuclear and gas turbines, while critics argue current reporting standards obscure the true environmental toll.
The artificial intelligence boom has generated a carbon footprint in 2025 comparable to the entire annual emissions of New York City, according to a new study.
Published Wednesday in the journal Patterns, the research estimates that AI systems alone will consume up to 23 gigawatts of power by year’s end, nearly half of global data center demand.
Even more significant is the water impact: the industry’s thirst has now surpassed the world’s total annual consumption of bottled water, driven largely by the hidden costs of electricity generation.
Promo
NYC Equivalent Carbon Footprint
Driving this unprecedented surge is the rapid expansion of AI infrastructure, which is now projected to consume nearly 50% of the world’s total data center electricity, excluding cryptocurrency mining.
According to the study, carbon emissions associated with this demand will fall between 32.6 and 79.7 million tons of CO2 equivalent in 2025. To contextualize the upper bound of this estimate, it rivals the annual carbon output of New York City, which typically hovers between 50 and 52 million tons.
Calculating these figures remains challenging due to limited corporate transparency. Alex de Vries-Gao, a PhD candidate at VU Amsterdam and the study’s author, highlighted the difficulty of obtaining precise data, noting that “there’s no way to put an extremely accurate number on this, but it’s going to be really big regardless… In the end, everyone is paying the price for this.”
Despite these measurement hurdles, corporate disclosures confirm the upward trend. Google’s own data reveals a 27% year-over-year increase in data center electricity consumption, reaching 30.8 TWh in 2024, as detailed in the 2025 Environmental Report.
A major point of contention in the industry remains the accounting method used to report these emissions. Tech giants often prefer “market-based” metrics, which allow them to subtract renewable energy purchases from their total footprint, effectively lowering their reported impact.
Critics argue that “location-based” metrics, which reflect the actual carbon intensity of the local grids where data centers operate, provide a more accurate picture of environmental cost. Franz Ressel, lead researcher at the Kairos Fellowship, previously criticized this practice, stating that “market-based emissions are a corporate-friendly metric that obscures a polluters’ actual impact on the environment.”
Big Tech Environmental Disclosures vs. Reality
The Hidden Water Crisis: Beyond Cooling
While direct cooling often garners the most headlines, the study reveals a far more extensive water footprint hidden within the supply chain. Total AI-related water consumption is projected to reach between 312.5 and 764.6 billion liters in 2025.
Surpassing the entire global annual demand for bottled water, which stands at approximately 446 billion liters, this figure highlights the scale of the issue. Emphasizing the magnitude of this consumption, the research compares it to other major urban centers.
According to the Cell Patterns study:
“The carbon footprint of AI systems alone could be between 32.6 and 79.7 million tons of CO2 emissions in 2025, while the water footprint could reach 312.5–764.6 billion L. To put this into perspective, this is in the same range as the carbon footprint of New York City (52.2 million tons of CO2 emissions in 2023). Similarly, the water footprint of AI systems may be in the same range as the entire global annual consumption of bottled water (446 billion L).”
Most corporate sustainability reports focus on direct water withdrawal for cooling servers, but this accounts for only a fraction of the total. Indirect water consumption (the water used to generate the electricity that powers the chips) is approximately four times higher than direct use.
Meta stands out as an outlier for disclosing this specific metric. In 2024, the company reported 72.2 billion liters of indirect water consumption, a figure that dwarfs its direct usage.
This broader view of water consumption contrasts sharply with recent industry narratives. In the Gemini transparency report, Google claimed a median AI prompt uses just “five drops” of water. However, this figure notably excluded the indirect water consumption associated with power generation.
Jeff Dean, Google’s Chief Scientist, previously framed the impact as minimal, suggesting that “…it’s actually equivalent to things you do without even thinking about it on a daily basis, like watching a few seconds of TV or consuming five drops of water.”
AI Infrastructure vs. Real-World Benchmarks (2025 Estimates)
The Infrastructure Scramble and Transparency Gap
Compounding the issue is the industry’s reliance on fossil fuels to meet immediate energy needs. Physical constraints on the grid are forcing a “dash for gas,” creating a shortage of turbines with wait times exceeding three years.
Simultaneously, tech giants are pivoting to nuclear energy to secure baseload power. Microsoft recently secured a $1 billion federal loan to support the Three Mile Island restart, while Google signed the $3 billion hydropower deal to guarantee clean energy supply.
Despite these high-profile investments, the study argues that current reporting practices obscure the true environmental cost of AI.
The study explicitly notes the limitations caused by this lack of transparency:
“The lack of distinction between AI and non-AI workloads in the environmental reports of data center operators means it is possible to assess the environmental impact of AI workloads only by approximating them through data centers’ general performance metrics. The environmental disclosure of tech companies is, however, often insufficient to determine even the total data center performance of these companies.”
Without granular data on specific AI workloads, researchers must rely on aggregate company-wide metrics, making it difficult to isolate the impact of artificial intelligence from broader cloud services.
Donald Campbell, Director of Advocacy at Foxglove, criticized this dynamic, arguing that “this is yet more evidence that the public is footing the environmental bill for some of the richest companies on Earth.”
Concluding their analysis, the Patterns study authors argue that voluntary corporate reporting is insufficient to address the scale of the problem. Without mandated disclosure of specific metrics like indirect water use and location-based emissions, the true cost of the AI boom will remain hidden from public view.
Alex de Vries-Gao called for a broader conversation on these externalized costs, stating that “we can really ask ourselves, is this how we want it to be? Is this fair? We really need to have that transparency, so we can start having that discussion.”
