- AI's growing power and cooling requirements are forcing the data center industry to explore unconventional deployment models
- Industry experts say engineering, connectivity, maintenance and commercial viability remain significant hurdles before these alternative AI data centers can scale
- Rather than replacing traditional hyperscale campuses, distributed, marine and space-based data centers are likely to complement conventional infrastructure as AI computing demand accelerates
As AI drives a massive surge in computing demand, the data center industry is beginning to explore unconventional locations, including distributed AI servers installed in homes to underwater facilities cooled by seawater and even orbital data centers powered by solar energy.
The interest stems from a growing challenge confronting cloud providers worldwide. AI workloads require significantly more computing power and electricity than conventional applications, putting increasing pressure on power grids, water resources and land availability. This has prompted startups and infrastructure companies to examine alternative deployment models that could complement conventional data centers rather than replace them.
“The main issues with AI compute developments in relation to infrastructure design are the power and water/cooling requirements. We are seeing data centers being built in places that are much more remote than traditionally, mainly to meet power and/or water needs,” Tony O'Sullivan, CEO of RETN, told Fierce Network, an independent network service provider for IP Transit and bandwidth services.
Additionally, there are issues related to permitting and community opposition. Yesterday, New York enacted a one-year construction ban on data centers that use 50 megawatts or more. Earlier this month, Blackstone terminated its planned digital gateway data center project in Virginia because of local opposition.
“I see permitting and community opposition as closely linked: power and water shortages often trigger both. In reality, there are plenty of places with sufficient cooling and the ability to generate power with the right investments; in these cases, permitting is far less likely to become the bottleneck. Made the right way, these developments will also improve local power grids and infrastructure, as well as provide well-paid, skilled jobs to more remote communities,” added Sullivan.
These challenges are prompting the industry to explore alternative deployment models that could distribute computing capacity more efficiently.
Distributed AI faces practical hurdles
However, locating data centers in remote areas is easier said than done. It can potentially lead to several network challenges, including interconnection. “As these [locations] are often many hundreds of kilometers from the closest interconnection hubs and often major cities and towns, this creates unique challenges when it comes to ensuring resilient network connectivity. Sometimes, we have to build long-distance, large-scale, resilient connectivity infrastructure from scratch that serves only these AI data centers; this is both time-consuming and expensive,” said Sullivan. This may lead to some operators “cutting corners,” leading to risks.
One emerging concept is distributed residential infrastructure. Several companies such as California-based Span and UK-based Heata, are experimenting with small AI servers that can be installed in homes and small businesses. The systems not only contribute computing capacity to cloud providers but also recycle waste heat to warm homes and water, potentially lowering household energy bills while reducing strain on centralized infrastructure. However, the concept remains at a very early stage.
While distributed residential infrastructure appears attractive because electrical power is already available in homes and businesses, Vik Malyala, Chief Business Officer at Supermicro believes the commercial model remains uncertain. "The main challenge will be the financial model and incentives for homeowners. Also, a question will be whether the bandwidth is available for server at one residence to talk to and work with a server in another residence," he said.
Unlike traditional hyperscale data centers, where thousands of servers are connected through ultra-low latency networking, distributed computing depends on residential broadband infrastructure that was never designed for AI workloads.
Even so, Malyala of Supermicro believes distributed computing could gradually scale because additional capacity can be added as more households agree to host cloud-owned servers. "Power is already distributed to homes and businesses, so this is less of an issue than building one massive facility," he said. "While the aggregate power consumption may be similar to a large data center, capacity could grow gradually as more locations sign up."
Noise generated by enterprise-grade servers can also become an issue in certain residential environments, although this will largely depend on the computing density of individual systems, he added.
Underwater computing has moved beyond theory
Among the alternative deployment models, underwater data centers arguably have the strongest technical validation.
Microsoft's Project Natick demonstrated that sealed underwater data centers could operate reliably for two years beneath the North Sea. The company concluded the concept was technically feasible and could deliver environmental and operational benefits, including significantly lower hardware failure rates compared with conventional facilities. Another instance is Keppel’s, the Singapore-based global asset manager and operator, floating data center project, which is under development and is likely to go live in 2028.
Still, commercial deployment remains challenging. "For marine deployments, fewer architectural changes would be required than for orbital computing," Malyala said.
He noted that surrounding water could provide highly effective cooling, but several engineering issues remain unresolved. "The cooling liquid could certainly be used to cool the servers, but a significant issue would be filtering the seawater before it enters cooling loops."
Engineers would also need sophisticated thermal modeling to understand how seasonal variations affect operating efficiency. Serviceability presents perhaps the biggest obstacle.
Unlike terrestrial facilities, where faulty hardware can be replaced within minutes, underwater systems are expected to operate as sealed units for years, making component reliability and remote management far more important.
Space remains the longest-term vision
Perhaps the most ambitious and exciting concept involves moving AI infrastructure into orbit. Starcloud, Google and SpaceX are some of the firms that have launched initiatives recently to make orbital data centers a reality.
However, Malyala believes the technical barriers remain substantial. "Significant engineering challenges remain, including radiation and the cost of sending racks of servers into low Earth orbit," he said. Even if launch costs continue to decline, maintaining operational data centers in space introduces an entirely different set of engineering problems. Radiation hardening, maintenance and connectivity are some of the other challenges.
The general consensus is that these alternative deployment models are complementary to conventional hyperscale campuses rather than direct replacements.
"There are merits to each approach," Malyala said. "As demand for computing continues to expand, different types of infrastructure will likely coexist depending on the application."
For now, however, the industry's immediate focus remains on solving practical engineering problems rather than replacing the data center as we know it.
Fierce Network related content on data centers
Oracle does damage control after New Mexico data center backlash
Ciena’s Gary Smith: Telecom is underestimating AI’s network demands