Just off the coast of Shanghai, beneath the churn of offshore wind turbines, China is launching a new kind of digital infrastructure—one that could reshape how the world powers its artificial intelligence boom.

Encased in watertight pods and cooled by ocean currents, the world’s first commercial underwater AI data center is now coming online. The facility, built by the Chinese tech firm Hailanyun, is designed to handle the kinds of intensive computational tasks that are quickly becoming the lifeblood of modern economies, from training large language models to powering real-time complex simulations.

And it does so with almost no freshwater, no bulky air conditioning units, and nearly zero carbon emissions.

“China’s ambitious approach signals a bold shift toward low-carbon digital infrastructure, and it could influence global norms in sustainable computing,” Shabrina Nadhila, an analyst at the energy think tank Ember, told Live Science.

The Cooling Crisis Brought Upon by AI

Today’s artificial intelligence is a hot topic, literally. Servers in modern data centers perform trillions of calculations each second, generating intense heat that must be constantly removed to prevent damage and preserve performance.

Nearly 40% of the energy used by a typical land-based data center goes to cooling. Most centers rely on freshwater, either sprayed, evaporated, or chilled and pumped around the machines. But this comes with a steep price: every day, these systems consume hundreds of thousands of gallons—water that might otherwise irrigate crops, flow through rivers, or fill drinking taps.

Some of the world’s biggest tech companies, including Google and Meta, have placed their data centers in dry, arid climates like Arizona or southern Spain, where the air’s low humidity helps protect sensitive electronics. But this tradeoff—better hardware conditions in exchange for even more water demand—has drawn criticism as regions face worsening droughts and growing competition over natural resources.

A 2023 study from UC Riverside estimated that every 20 to 50 AI queries—such as asking ChatGPT a question—can require the evaporation of half a liter of freshwater. Training large models can be even worse: one estimate suggests GPT-3 may have consumed over 700,000 liters of water during training, although newer models are more efficient.

Now, China is trying a radically different approach: skip the land entirely.

Beneath the Waves, Beyond the Grid

Instead of using chillers and fans, Hailanyun’s underwater data center takes advantage of the ocean’s natural thermal stability. The facility uses sealed pipes to pump seawater across radiators attached to its server racks, absorbing heat and carrying it away in a slow, constant flow. According to internal assessments conducted with the China Academy of Information and Communications Technology, this method slashes electricity use by at least 30% compared with land-based systems.

The new center is also powered almost entirely by a nearby offshore wind farm, which provides 97% of its energy, according to Hailanyun spokesperson Li Langping.

The first operational pod contains 198 server racks—enough to run between 396 and 792 AI-ready servers. Hailanyun says this setup is powerful enough to train a model like GPT-3.5 in a single day. But it’s still a small step compared to the scale of China’s AI ambitions. Traditional mid-size data centers in the country can house 3,000 racks or more. Superscale facilities exceed 10,000.

Even so, researchers are taking note. Hailanyun transitioned from a pilot project launched in Hainan in December 2022 to full commercial deployment in under 30 months, “something Microsoft’s Project Natick never attempted,” said Zhang Ning, a postdoctoral researcher at the University of California, Davis, as per Live Science.

Project Natick was Microsoft’s own attempt at ocean-based computing. In 2018, the company sank a sealed server pod off the coast of Scotland. After two years, they found it had lower hardware failure rates than land-based centers—thanks in part to its nitrogen-filled, human-free environment. But Microsoft has since shelved the project, noting only that it remains “a research platform to explore, test, and validate new concepts.”

China, by contrast, is moving fast—and at scale.

Not Without Consequences

Despite the environmental promise, ocean-based data centers raise difficult questions. Warm water discharge can lower oxygen levels in the surrounding seawater, especially during marine heatwaves. Aquatic organisms already stressed by rising ocean temperatures may find it harder to survive if nearby outflows push temperatures higher or disrupt local chemistry.

Microsoft researchers found their prototype caused water temperatures to rise only by “a few thousandths of a degree” in nearby currents—but that was under controlled, cooler North Atlantic conditions.

Hailanyun, for its part, says its own tests showed a rise of less than one degree Celsius in surrounding waters—“virtually no substantial impact,” according to Li Langping.

Security is another concern. A 2024 study from the University of Florida found that certain sound frequencies, such as those transmitted by underwater speakers, could potentially damage server systems. Researchers have since developed machine learning tools to detect and counter these threats early, but the vulnerability highlights how underwater infrastructure faces unique risks.

And then there’s repair. What’s routine on land, such as swapping out a failed hard drive or checking a connection, becomes slow, expensive, and dangerous when it involves divers or robotic subs.

Still, the ocean offers benefits that land can’t match: no dust, no seismic vibrations, and no temperature swings. All of which can extend the life of sensitive hardware.

A Blueprint for the Future?

For now, China’s experiment is unique. But interest is growing elsewhere. South Korea has announced plans for its own undersea facilities. Japan and Singapore are considering floating data centers—moored above the water, but still tapping the ocean for cooling.

Zhang, the researcher at UC Davis, believes the trend could spread if countries can resolve the regulatory, ecological, and supply-chain issues that China is already tackling.

And as AI continues to scale, those pressures won’t go away.

After all, there’s no shortage of seawater. What’s lacking is a sustainable vision for how to feed an AI future without burning through our most precious resources.

By sinking its servers into the deep, China may have surfaced an answer.