The Ocean Might Be Next Frontier For Data Centers - But There's A Catch

The mention of an AI boom just five years ago would've elicited little more than a quizzical look among most people who heard it. These days, however, Artificial Intelligence (AI) can provoke strong reactions. With the arrival of companies like OpenAI and Anthropic, and with giants like Google and Microsoft pouring billions into the technology, AI is finding its way into everything from daily personal use to big business and government. But the transformative technology requires vast resources to operate, potentially impacting the environment and pushing up people's utility bills, raising an important question about how to power data centers more efficiently. 

One of the major challenges is managing the huge amount of heat generated by AI servers as data centers handle a constant stream of inquiries from users around the world. The issue requires serious attention because AI's growing need for electricity — some of which is used to cool the centers — puts pressure on power grids and, when the electricity comes from fossil fuels, adds to greenhouse gas emissions. 

A potential solution is to place the data centers under water in the ocean to take advantage of its natural cooling properties. In fact, Microsoft has already tested the idea, operating a sealed data center on the seabed off the Scottish coast for several years, starting in 2018. But while the Project Natick experiment was deemed a success from a technical perspective, Microsoft eventually ended the effort and decided not to pursue its undersea system as a commercial product. A Reuters report said the computer giant abandoned the project due to a lack of customer demand and cost concerns, a decision that highlighted some of the challenges facing ocean-based data centers.

Notable drawbacks of ocean-based data centers include maintenance and upgrades. If a server suffers an issue in a land-based data center, it can be easily accessed and even repaired on-site. But if it fails underwater, technicians can't access the server so easily, so repairs would have to be carried out remotely or the equipment brought to the surface. Similarly, large-scale upgrades for land-based data centers are relatively straightforward because they're readily accessible, whereas with sealed ocean-based facilities, it may be more cost-effective to replace the entire installation than to replace the chips in individual servers. This matters especially for AI because the hardware for it is evolving quickly, and so data center operators may want to upgrade chips and servers on a regular basis.

There are environmental concerns, too, with ocean-based data centers. For example, although the seawater may be helping to cool the equipment, the heat the facility generates is being absorbed by the nearby water, raising its temperature and potentially affecting oxygen levels and marine life. With a data center along the lines of the one used in Project Natick, the water just a few meters downstream of the data center may get a few thousandths of a degree warmer than its normal temperature. That relatively small increase may sound reassuring, but the impact could be more notable if multiple data centers were deployed in a small area.

The ocean environment itself also presents challenges for data centers. Saltwater, for example, could potentially corrode the external equipment, while marine creatures like barnacles can accumulate on external heat exchangers, interfering with heat transfer. It seems that while the ocean has the potential to solve AI's cooling problem, it creates another set of problems at the same time. So are there any other alternatives?

A more ambitious solution is to deploy data centers on satellites in low-Earth orbit and, eventually, on or around the moon, where they'd be able to harness solar power while avoiding many of the cooling restraints faced by facilities on Earth. Data would be sent between the satellites and Earth using high-speed laser links, allowing the servers to process information and send the results to users back on land-based servers. Google is already exploring the idea and, as part of Project Suncatcher, plans to launch the first of three prototype satellites this year. Another company, Starcloud, is also developing an AI delivery system for space and recently announced a lunar-orbit demonstration with spaceflight company Firefly that will attempt to validate the technology for a future data center around the moon.

SpaceX is also looking to build an orbital AI network made up of satellites in low-Earth orbit as part of its Starmind initiative. It's reportedly aiming to demonstrate the technology in 2027. The company arguably has a head start in this area thanks to its extensive experience launching rockets and operating its burgeoning Starlink satellite network, both of which will play a role in the proposed orbital computing network.

But the idea of space-based data centers faces some major hurdles. Jensen Huang, the CEO of chip-maker Nvidia, a company that's currently riding high on the AI boom, has pointed to the high cost of sending the necessary equipment into orbit as well as the challenge of dispersing heat in the vacuum of space. Radiation could also damage the AI chips and other equipment. And if repairing and upgrading data centers in water is difficult, doing the same in space would be an even greater challenge.