Artificial Intelligence, Water Consumption and the Trillion-Radish Conundrum

The volumes of water consumed by artificial intelligence may be hundreds of times smaller than other parts of the economy, like agriculture. But this rapidly expanding sector, which uses as much water as Norway and Sweden combined, is a ticking time bomb.

Structural realities about freshwater availability, infrastructure and geopolitics mean that AI will have an outsized impact on water insecurity. The people who suffer the most will be those living close to the vast infrastructure powering this technological revolution.

The International Energy Agency estimates that the global data center capacity needed to train and run AI models has nearly doubled every five years since 2015.

Tech companies are tapping into precious water reserves to cool the servers behind machine learning models. The volumes of water involved are large. Cornell University researchers estimate that by 2027, AI-related water withdrawals could total more than 6 billion cubic meters per year — roughly equivalent to the total annual water use of New Zealand.

Numbers come with health hazards in this sector. That calculation is likely an underestimate because cooling needs vary depending on who operates a data center, where it’s located and the type of tasks the AI is asked to conduct. Many tech companies treat infrastructure data as proprietary and sensitive, and transparency across the sector is limited. As AI facilities scale rapidly in number and size, the true water footprint is likely higher.

Still, some disclosures are trickling through into public forums. In July, the French large language model (LLM) start-up Mistral AI published one of the first environmental impact reports in the industry.

According to the company, 91% of their water use across the supply chain is taken up by model training and inference (when an LLM uses its training to answer questions or to make decisions). That water is primarily used as a coolant. Mistral estimated that producing a single page of text using their model uses about as much water as growing a ‘small pink radish’.

Applied to OpenAI’s more popular ChatGPT, that calculation would make CEO Sam Altman the equivalent of a farmer harvesting a trillion radishes.

Despite this, AI’s water consumption still pales in comparison to other sectors, like the global agriculture industry, which is responsible for70% of all freshwater withdrawals. That’s about 3 trillion cubic meters of water each year — or six hundred times more than Cornell’s estimate of AI’s water requirements.

At the present rate of expansion, the AI industry would need around half a century to catch up with agriculture’s water consumption. That would be impossible, with the limiting factor being global and local food security. But what happens in the choppy transition period as more water is depleted to keep the AI boom afloat?

While there is certainly an argument that we should address water insecurity issues by starting with industrial agriculture, AI should not be brushed under the table.

There are three reasons why AI will have an outsized impact on efforts to build water resilience: availability, location and politics.

The availability of freshwater is already a concern globally. By 2030, demand for freshwater will outstrip supply by 40% — which will mean those reliant on water will become increasingly dependent on the extraction of non-renewable or unreliable water sources.

In essence, the data centers needed to power the AI revolution are already drinking from a drying tap.

Some tech accelerationists argue that AI isn’t the cause of this crisis. They may point to other thirsty sectors or, optimistically, novel technologies like closed-looped and non-evaporative cooling designs which increase the water retention of data centers by stopping superheated water from entering the atmosphere. The World Economic Forum suggests that water efficiency savings of between 50-70% are possible with circular management systems.

These technologies are promising — but they are not a panacea.

Which brings us to location. To reduce the risk of equipment corrosion from humidity or saltwater, many companies are building data centers in dry, inland regions. The result? Water-hungry facilities in places already under stress.

Bloomberg estimates that around two-thirds of data centers either built or in development since 2022 in the United States are in areas of high-water stress, such as Illinois, Arizona, California and Virginia.

And these data centers can be enormous. Meta has recently announced an AI lab (named Hyperion) in Louisiana the size of Manhattan.

To quench digital mega-facilities, companies often drill deep wells to tap groundwater. What can follow is the loss of access to clear, running water for local communities as pools of groundwater dry up and sediment from the industrial activity pollutes supplies.

And while communities might find the flood of data centers more palatable if they provide jobs for locals like other industries, including agriculture, such opportunities have so far failed to materialize.

There will be global environmental impacts too. Unprecedented depletion and drying of our water resources on land has contributed more to sea level rise than the thawing of polar regions.

Many of the largest tech companies have committed to becoming “water positive — meaning they plan to return more freshwater resources to the environment than they consume. But it’s unclear whether these efforts will restore water in the same locations where it was extracted.

Civilians losing access to water should worry politicians — they need only look as far as Bolivia’s Cochabamba Water War, where roadblocks and riots followed the privatization of the municipal water supply and the government was forced to backpedal on the reforms, to see the social conflicts that can arise when access to essential resources is restricted. But today, where AI users, companies and infrastructure often span national borders, political accountability is murky at best.

The publication of the Council of Europe’s Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law is the world’s first legally binding AI international treaty — but it fails to mention water consumption or broader environmental impacts of AI, including greenhouse gas emissions. Even if it did, it is difficult to imagine that this would be worth the paper it was written on in a Trump 2.0 world.

In the United States, the outlook is less promising at the federal level — where less (regulation) is more (acceleration). In July, the White House published its latest AI Action Plan on “Winning the Race.” Water is mentioned three times in the strategy. All three references take aim at the federal Clean Water Act which is seen as a barrier to AI roll-out.

Politically, perhaps the purported rapid economic gains to be made through developing and adopting AI and digital infrastructure are more appealing than the slowly evolving environmental crises that is welling up behind it.

And so many of the world’s most water-scarce countries are looking to onshore digital infrastructure. It would be a remarkable economic turnaround should previous importers of virtual water — or the hidden water that goes into products — begin to export it in AI products and services.

In June, Microsoft launched “Chile Central” — a sovereign data center that they say will bring $3.3 billion of investment into Chile. The country has been suffering from a 15-year drought.

The Saudi Public Investment Fund has channeled over $40 billion into AI ventures as part of its grand economic strategy Vision 2030. It timed the launch of its own artificial intelligence company — HumAIn — with US President Donald Trump’s visit to the Kingdom. The World Resources Institute’s Risk Atlas highlights Saudi Arabia as one of 25 countries facing extremely high water stress and the Kingdom plans to invest nearly $80 billion in desalination in the next decade as part of the same strategy.

It is understandable that countries want to enter the race for AI supremacy. But significant questions remain around whether countries are trading resilience for clever bits, increasing dependency on other countries for water and food security.

These are troubled waters for democratic politicians to operate in. Their role is to balance the environmental needs of the constituencies that voted them in, with the kind of corporate and political party optimism that is inflating an AI-shaped bubble.

Environmentalists might find unlikely allies in influential local politicians. To quote far-right Rep. Marjorie Taylor Greene (R-GA), after the greenlighting of a 150-acre AI data center in her constituency, “While I understand the many promised benefits of AI, I remain committed to protecting state rights, human jobs, human lives, human rights, our environment and critical water supply.”

This hints at a more long-term strategy to govern the trillion-radish conundrum — a fallback on state-led accountability and flexing of power. This is already the case in California, where updated climate laws will require major tech companies to disclose their greenhouse gas impacts.

Given the immense momentum behind AI, local water scarcity is likely to get worse before it gets better. Sooner or later, governments and companies will eventually have to act. They might do sooner if a trillion radishes were delivered to agency or company headquarters — but for now they seem content to talk while treading water.