Every time you ask an AI chatbot a question, generate an image, or use a voice assistant, you are indirectly consuming a small amount of freshwater. While the digital world feels intangible, the physical infrastructure powering the AI boom is alarmingly resource-intensive. As we push toward artificial general intelligence, we are also pushing the planet’s water and energy grids to their breaking point.
The irony is stark: to cool down the “brains” of the digital revolution, we are boiling the planet’s most precious resource. This article looks at why water is the hidden ingredient in the AI arms race, how much energy and water we are really using, and why engineers are turning to the ocean, specifically underwater data centers, as the most radical fix to the crisis.
Why Do AI Data Servers Need Water?
To understand the thirst of AI, you have to look at the chips. For the last decade, data center cooling was fairly simple. You could use fans (air cooling) to whisk away heat from processors running at 100 to 150 watts. However, AI workloads, especially the training of Large Language Models (LLMs), run on Graphics Processing Units (GPUs) like NVIDIA’s H100 or Blackwell chips. These are not your standard computer parts; they are power-hungry machines.
Current GPUs consume between 700 and 1,200 watts each. Roadmaps for the coming years show chips reaching 2,000 watts, with industry experts preparing for 5,000-watt chips in the near future. When you pack thousands of these chips into a single rack to train a model like GPT-4, the heat generation becomes almost biblical. A single AI rack can now draw 100 kW or more, an order of magnitude higher than traditional server racks.
Air has a specific heat capacity roughly four times lower than water and is 800 times less dense. Simply put, air can’t carry the heat away fast enough. If you tried to cool a modern AI cluster with fans alone, the chips would melt in seconds.
This is where water comes in. Most hyperscale data centers today use evaporative cooling or closed-loop water systems. Water is circulated through “cold plates” placed directly on the chips, absorbing heat and carrying it to external cooling towers. However, in many systems, this water is evaporated into the atmosphere to dissipate heat. That water is gone forever, a sharp contrast to the “intangible” nature of the cloud.
How Much Energy Do We Consume? The Jaw-Dropping Numbers
The energy consumption of AI is no longer a niche concern; it is becoming a driver of global electricity demand. The numbers from recent analyses are staggering and show an industry racing toward a sustainability wall.
According to research firm Gartner, global data center electricity consumption is projected to hit 565 terawatt-hours (TWh) in 2026. To put that in perspective, that’s more than the total electricity consumption of many entire industrialized nations. By 203, that figure is expected to exceed 1,200 TWh.
More critically, the split of this energy is changing. In 2026, AI-optimized servers will account for 31% of this consumption, but by 2027, they will surpass conventional servers, consuming an estimated 258 TWh annually. This surge is so intense that major analysts suggest grid supply will be insufficient to meet the power demands of future data center construction.
Even more alarming is the water consumption. It is one thing to use electricity; it is another to drain the water. Research published in ScienceDirect (Water Research) estimates that AI’s global water footprint could reach 4.2 to 6.6 billion cubic meters annually by 2027. Training a single model like GPT-3 required around 700,000 liters of clean freshwater. To visualize this, generating just 10 to 50 medium-length AI responses consumes the equivalent of a standard 500 ml water bottle.
The industry uses a metric called Water Usage Effectiveness (WUE). In water-stressed regions, where two-thirds of post-2022 data centers are located, facilities relying on evaporative cooling can report WUE values exceeding 3 liters per kilowatt-hour.
The “Underwater” Solution: Putting Servers in the Sea
So, if evaporation is the problem, what if we stopped evaporating water? What if we used the ocean itself as a massive heat sink? The idea of underwater data centers is moving from science fiction to commercial reality, driven by the urgent need to lower the Power Usage Effectiveness (PUE), the ratio of total building energy to IT energy.
The standard industry average PUE hovers around 1.5, meaning 50% of the energy is used just for cooling and overhead. If you submerge servers in water, you don’t need chillers, you don’t need fans, and you don’t need evaporation ponds.
Case Study: Project Natick and the Chinese Commercial Leap
Microsoft tested this concept through Project Natick. They deployed a data center cylinder off the coast of Scotland. The results were promising: the underwater environment was not only naturally cool but also consistent. Plus, the absence of oxygen and humidity inside the sealed vessel drastically reduced corrosion and hardware failure rates.
However, China has taken the baton and run with it. In 2026, China launched the world’s first commercially operated underwater data center located 35 meters deep near the Lingang Special Zone in Shanghai. This $226 million facility houses nearly 2,000 servers, including GPU clusters for AI tasks.
The results have been impressive. The facility achieved a PUE below 1.15, a major improvement over land-based rivals. It is powered largely by offshore wind turbines, creating a near-self-sufficient bubble of computing power at sea.
Why Don’t We Put All Servers Underwater?
If it works so well, why aren’t we dropping every server into the ocean right now? The answer lies in physics, politics, and maintenance.
- The Maintenance Nightmare: Hard drives fail. RAM throws errors. Cables corrode. In a land data center, a technician can swap a broken drive in 5 minutes. At the bottom of the ocean, you have to bring the entire heavy cylinder to the surface. Operators must rely entirely on redundancy (having three backups for every one part) and remote management, which is costly.
- Corrosion and Biology: Saltwater is hostile. It eats through metal, and sea life loves to attach to submerged structures. This is known as biofouling. Even with advanced coatings, the long-term durability of subsea cables and heat exchangers in a saltwater environment remains a technical challenge.
- The Heat Just Moves: While underwater data centers save freshwater and electricity, they don’t “destroy” the heat; they dump it into the ocean. Environmentalists worry about “thermal pollution”, raising the local temperature of the seabed, which could harm marine ecosystems. We are trading air pollution for ocean heating.
The Verdict: A Hybrid Future
The reality is that we cannot simply pick one solution. The “underwater server” idea is smart for specific locations, coastal cities with expensive land, tropical climates where air cooling is inefficient, or regions suffering from drought.
However, for most of the AI boom, the future is Direct-to-Chip Liquid Cooling and Immersion Cooling (dunking the server in non-conductive oil or fluid on dry land). These systems recycle the same fluid in a closed loop, using heat exchangers rather than evaporation. This can cut cooling energy use by up to 50% compared to traditional methods, eliminating water waste while keeping the servers accessible for repairs.
As the data from the Department of Energy and various academic surveys shows, we are entering an era where computing power is no longer limited by silicon, but by energy availability and water rights. The companies that win the AI race will not just be the ones with the best algorithms; they will be the ones who figure out how to keep their chips cool without draining the planet dry.
Whether that means sinking servers to the bottom of the South China Sea or recycling boiling dielectric fluids in a desert warehouse, one thing is clear: The future of AI is liquid.
by LILA WHITE