The biggest risk in artificial intelligence today is not a lack of investment; it is where that investment is going. Technology companies are pouring vast sums into AI infrastructure. Microsoft, Amazon,
and Alphabet are racing to build more data centres, faster and at a larger scale than ever before. Global spending is expected to approach $1 trillion this decade. But the problem is no longer capital; it is the physical world that the capital depends on—electricity, land, water, and industrial supply chains. Those limits are no longer theoretical. They are already visible.
In Northern Virginia, one of the world’s largest data centre hubs, these facilities now consume roughly a quarter of the region’s electricity. In Ireland, they account for about 20% of the national power demand, prompting regulatory pushback. Singapore temporarily halted new data centre approvals after reaching its energy limits. Across parts of Europe, projects are being delayed not because of financing issues but because grid connections, transformers, and supporting infrastructure are not available in time. In simple terms, we are not running out of money. We are running out of the ability to use it efficiently.
At the same time, AI systems are becoming far more demanding. Training advanced models today requires enormous energy—closer to running industrial operations than traditional software. The next generation of AI, which will operate continuously and at a global scale, will only intensify that demand. This creates a growing imbalance. The need for computing power is rising rapidly, but the infrastructure required to support it—power grids, land, and cooling systems—takes years, often a decade, to expand.
This is not a short-term bottleneck. It is a structural constraint. And when constraints harden, economics begins to change. The question is no longer whether we can build more data centres on Earth; it is how costly it becomes to keep doing so.
That is where a once-distant idea begins to enter serious discussion: moving some parts of AI infrastructure into space. This does not mean replacing Earth-based computing; it means dividing the system more intelligently.
Certain functions—like real-time applications, financial systems, and everyday consumer services—must remain on Earth because they depend on speed. But other functions, such as training large AI models, running simulations, and processing massive datasets, are less sensitive to location. They depend on raw computing capacity rather than proximity. These are also the most energy-intensive parts of the AI stack. They are the ones most likely to move first.
The logic is straightforward. On Earth, every new data centre competes for limited resources: electricity, land, and water. In space, some of these constraints change. Solar energy is more consistent, cooling does not rely on water, and expansion is not limited by land availability or local regulation.
For decades, this idea remained impractical because the cost of launching equipment into orbit was too high. That barrier is now falling. Reusable rocket systems have significantly reduced launch costs, and next-generation vehicles are expected to push them lower still. As these costs decline, the economics begin to shift—especially when compared with the rising expense of building large-scale infrastructure on increasingly constrained land.
This does not make space an easy solution. There are serious challenges.
Data transmission between Earth and orbit remains expensive and limited. Equipment must survive radiation and operate reliably with minimal maintenance. Orbital space itself is becoming crowded, increasing the risk of collisions and debris. And global rules for large-scale industrial activity in space are still underdeveloped. These are real constraints, but they are different constraints.
Economic transitions do not require perfect solutions; they begin when one option becomes more efficient at the margin. Even a small shift of high-energy AI workloads into space could have significant consequences on Earth.
Today, data centres are among the largest drivers of demand for electricity, land, and industrial construction. Entire regional strategies are being built around attracting them. Governments offer incentives, utilities expand grids, and private capital flows into infrastructure based on the expectation of steady, long-term growth.
That expectation rests on a simple assumption: that most computing will continue to happen on Earth. If that assumption weakens, even slightly, the financial implications are substantial. This is not a collapse scenario, but it is expensive.
Infrastructure that appears essential today may turn out to be overbuilt if future demand shifts elsewhere. The risk is not that data centres will disappear, but that their projected growth will become too linear in a world that is becoming increasingly non-linear. That is how trillion-dollar mistakes are made—not through failure, but through misallocation.
There is also a deeper shift underway. If computing expands into space, control over it may become more concentrated. Access to orbital infrastructure depends on launch capabilities, in-space operations and logistics—areas currently dominated by a small number of companies and nations.
In that sense, the competitive advantage in AI may not just come from better algorithms but from control over where and how computing can scale.
For decades, computing reduced the importance of geography. AI reversed that trend by tying digital systems back to physical resources like power and land. Space may change it once again—not by removing constraints but by relocating them to a new domain.
The mistake today is not technological; it is strategic.
Too much capital is being deployed on the assumption that the future of AI infrastructure will look like the past: more data centres, more land, and more electricity, all on Earth. That assumption is already under strain. The future of computing will not be entirely in space. But it may not remain entirely on Earth either. And when that shift begins, its first signals will not come from orbit; they will appear in the investments on the ground that assumed demand would never change.
Nishant Sahdev is a physicist at the University of North Carolina at Chapel Hill, United States.
