As I See It: Spacing Out

September 8, 2025 Victor Rozek

Computer rooms have always been somewhat problematic to build and maintain: Raised floors, 24/7 air conditioning, a reliable power supply, and, for larger installations, a backup power source. These requirements were spendy but manageable at least until the rise of AI. To put it bluntly, AI is a resource hog, addicted to excess. Its drugs of choice are land and energy.

The future of datacenters is unfolding in Abilene, Texas. There, Project Stargate is under construction, and the mega-scale OpenAI datacenter can hardly be called affordable. Its estimated cost is $500 billion. The facility includes a whopping eight million square feet of buildings, with plans to expand to 20 locations beyond Abilene. To put the scale of the project in context, it is equivalent in size to Central Park in New York City.

The availability of space is one factor, but it has to be paired with the availability of power. AI’s intense computational demands dramatically increase power requirements per rack. The extraordinary volumes of data processed by AI, and the increased use of Graphics Processing Units (GPUs) are the drivers. GPUs are designed to speed up the process of altering and manipulating memory when creating visual content. Beyond designing and editing images and videos, GPUs are necessary for tasks that require parallel processing of large amounts of data, such as 3D rendering, and machine learning. All of which boosts the need for vigorous cooling systems and failsafe backup power. Facilities of this size require gigawatts of power and can negatively affect aging power grids. The herculean scale and the appetite for energy makes these facilities undesirable neighbors.

As the demand for massive AI datacenters grows, pressured communities are pushing back at the resource-intensive, multi-million square foot facility in their midst.

Environmental and quality of life issues are chief concerns. Datacenters that rely on water for cooling are an enormous drain on local aquifers, consuming as much as 5 million gallons per day. Local residents report shortages, sedimentation in their water, and decreased water pressure.

In places where power plants rely on fossil fuels to generate electricity, increased greenhouse gas emissions and air quality are a concern, as are noise and light pollution. Cooling systems, generators, pumps, transformers, and other equipment produce considerable noise, particularly annoying at night. Likewise, the lights emanating from 24/7 operation deprives neighbors of darkness. Sleep disruption is a common complaint. Local governments have been forced to adopt stricter zoning laws to mitigate negative impacts.

There are approximately 5,400 AI datacenters in the United States. As AI workloads and computational demands increase, so do requirements for power. And although the current administration is promising to fast-track datacenter building permits, finding large tracts of available land in urban areas is challenging. Meanwhile, AI’s energy needs are growing faster than grid capacities. What to do.

One solution, as improbable as it is ingenious, is moving datacenters into space. Why struggle squeezing AI into the cloud, when you can spread it across the cosmos? It would potentially solve a number of problems, notably land constraints and energy limitations.

There’s a lot of space in space and you don’t have to buy it or lease it. You do have to get to it, however, and if Elon Musk’s SpaceX failed attempts to get off the launch pad are any indication, reliable and affordable space transport is a work in progress. Having the building blocks of datacenters blowing up on launch is, at best, an inelegant attempt at distributed data.

But assuming SpaceX, or other private carriers, can function more like FedEx, there are advantages to be had. Solar energy is abundant, reliable, and free. The vacuum and extreme cold of space allows for efficient heat dissipation eliminating the need for energy-intensive cooling systems, an advantage particularly useful in Quantum computing. For applications requiring global data transmission, space-based datacenters could significantly decrease latency.

There is also a growing interest in space manufacturing essential to future exploration and lunar resource extraction. On-demand fabrication and repair would eliminate costly launches from Earth.

There are a number of companies investing in space-based IT. According to Google, Avalanche Technology, for example, is developing “high-density, high-performance STT-MRAM (Spin-Transfer Torque Magnetoresistive Random Access Memory) for space datacenters, enabling faster and more reliable data storage and processing.”

Lonestar Data Holdings “is working on a lunar datacenter, initially testing data transfer and secure data protocols.”

HPE’s Spaceborne Computer Project “demonstrated the feasibility of using commercial off-the-shelf hardware in space, paving the way for more advanced datacenter solutions.”

And, finally, the aptly named Starcloud is “developing in-space data processing capabilities, initially focused on providing GPU compute to other satellites, and later addressing the broader needs of AI and other data-intensive applications.“

Although in early stages of development, there is enough interest and investment to ensure mini datacenters will be orbiting the globe in the not-too-distant future, especially if launch costs continue to decline and launch reliability continues to improve.

For those of us who started working in IT decades ago, the idea of managing a fleet of orbiting datacenters is fantastical at best. But by the time classic datacenters become outdated, the very notion of “managing” will long have been redefined. Advanced AI will be doing the managing – and the programming, and whatever else needs to be done. What mere humans will be tasked with is not yet clear.

Maybe Scotty of the original Star Trek was thinking about progress and its effects on humankind when he uttered his famous line: “We cannot take much more of this, Captain.”

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search