Nvidia reduces water consumption in data centers but ignores usage in power plants

What happened?

Nvidia introduced a new cooling system for data centers that significantly reduces water consumption within facilities. The company claims its direct-to-chip liquid cooling (DLC) solution can eliminate up to 100% of water use for cooling in certain configurations. However, according to a TechCrunch analysis, this measure does not address the largest water consumption related to artificial intelligence: the water used in thermoelectric power plants (coal, natural gas, and nuclear) to generate the electricity that powers servers. This announcement comes amid a context where water demand for AI is growing exponentially, and tech companies face increasing pressure from investors and regulators to improve their water sustainability.

Why is it important?

AI's water consumption is a growing problem. A study by the University of California, Riverside, estimated that queries to models like GPT-3 can consume up to 500 ml of water per 20 questions, considering the entire energy supply chain. Most of that consumption does not occur in the data center but in power plants, where water evaporates in cooling towers. By ignoring this link, Nvidia's solutions only address a fraction of the problem. For readers, this means that the sustainability promises of big tech companies may be incomplete if all indirect impacts are not considered. Historically, the tech sector has faced similar criticism with carbon emissions: companies reported reductions in their direct operations but ignored supply chain emissions (Scope 3). Now, water follows a similar pattern, where transparency is key to assessing real impact.

What consequences will it have?

In the short term, Nvidia's measure will help reduce water use in water-stressed locations but will not significantly change AI's total water footprint. Companies adopting these systems will be able to report improvements in water efficiency, but critics will point to greenwashing if energy generation is not addressed. In the long term, it could pressure tech companies to invest in renewable energy (which does not require water for cooling) or dry cooling technologies for power plants. It could also incentivize transparency in measuring water consumption across the value chain. An example is Google, which reported a 20% increase in water consumption in 2022 due to AI expansion and has faced criticism for not breaking down indirect consumption. Nvidia's solution, while a technical advance, may not be enough to meet long-term sustainability goals if not accompanied by changes in energy sources.

What should readers know?

• Water used in data centers represents only a fraction of AI's total water consumption; most is due to electricity generation. According to the International Energy Agency, thermoelectric plants account for about 40% of total water consumption in many countries.
• Nvidia has not provided data on total water savings in the energy supply chain, only within the data center. TechCrunch notes that the company has focused on direct use, ignoring indirect use.
• Direct-to-chip liquid cooling systems can save water but require more energy to pump the liquid, which could increase emissions if electricity comes from fossil fuels. This trade-off must be carefully evaluated.
• Transparency and standardized measurement of water use across the entire chain are essential to assess real impact. Currently, there is no widely accepted standard for reporting indirect water consumption in the tech industry.

“You can't solve AI's water problem just with data center improvements; you need a transformation in energy generation,” notes the TechCrunch analysis.

Broader context

Companies like Google and Microsoft have already reported significant increases in water consumption due to AI expansion. Google, for example, saw a 20% increase in 2022, while Microsoft reported a 34% increase in 2021-2022. Nvidia's solution, while a positive step, must be framed within a broader effort that includes locating data centers in water-abundant regions, using renewable energy, and adopting cooling technologies that do not evaporate water. Comparatively, the data center industry has advanced in energy efficiency with metrics like PUE, but water lacks an equivalent widely adopted metric. Nvidia's initiative could help establish a standard for water efficiency in data centers, but the larger challenge remains energy generation. In this regard, investment in renewables like solar and wind, which do not require water for cooling, is crucial. Additionally, strategic placement of data centers near renewable energy sources or in regions with lower water stress can mitigate impact. However, as long as AI demand continues to grow, total water consumption will likely keep increasing unless all links in the chain are addressed.