Artificial intelligence is using as much power as a small city, and some now think the solution is to launch the computers into space.
The International Energy Agency projects that data centers could consume more than 1,000 terawatt-hours of electricity by 2026, roughly equal to Japan’s total power use.
The reason is simple: GPUs, the chips that drive AI, are energy-hungry. Each NVIDIA H100 draws roughly 700 watts at full load, and a single rack of them can consume several megawatts, enough to strain local grids that were never designed for this kind of load.
The Space Gambit
That pressure is driving bold ideas. One of the most ambitious comes from Starcloud, a California startup that wants to move compute infrastructure beyond the atmosphere.
Instead of drawing from terrestrial grids, its satellites would run on continuous solar power and dump excess heat into the vacuum of space. The company says its orbital data centers could be ten times more energy efficient than those on Earth and eliminate the need for water-based cooling entirely.
Starcloud, backed by NVIDIA and In-Q-Tel, calls its approach “data centers in space.” Its first satellite, Starcloud-1, carries an H100 GPU and Google’s open-source Gemma model. It’s expected to enter a sun-synchronous orbit in 2026 to test whether large-scale AI workloads can run reliably above Earth.
To some, the plan is a futuristic fix for AI’s growing footprint. If land-based infrastructure can’t keep up, why not take advantage of endless sunlight and natural cooling in orbit? But the idea hasn’t convinced everyone.
That kind of reaction captures a wider skepticism among engineers and researchers: that orbital compute is a flashy solution to a solvable problem. Behind the excitement is a real constraint, though. Regions like Virginia, Dublin, and Singapore have slowed or paused new data-center projects because of grid strain. The industry’s anxiety is understandable, but the physics of efficiency are catching up faster than the panic.
The Efficiency Fix
AI’s energy footprint is climbing, but its efficiency is improving even faster. The IEA says chip performance per watt has doubled roughly every three years, and that a modern AI processor now uses 99 percent less energy to perform the same computation as one made in 2008.
On the ground, engineers have slashed cooling overheads: modern hyperscale facilities have reduced cooling power from 30-40 percent of total electricity use to around 7 percent, thanks to liquid and immersion systems. Operators are pairing these with renewable power sources and scheduling heavy workloads to coincide with peak solar or wind availability.
Hardware and software improvements are advancing in parallel. NVIDIA’s next-generation Blackwell architecture promises up to 25-times better performance per watt than the A100 generation. Model compression, sparsity, and mixed-precision training are cutting the energy cost per token at similar rates. Together, these shifts show that the fastest way to make AI sustainable is to make the work more efficient here.
This rethinking of infrastructure is already underway. Cisco, for instance, argues that “the edge is the new data center,” as power scarcity and latency push AI workloads away from hyperscale campuses and closer to where data is created: factories, hospitals, and retail floors. The company expects that by 2027, about 75 percent of enterprise data will be created and processed at the edge, underscoring how compute is decentralizing on Earth.
The limits of relocation have already been tested.
In 2018, Microsoft ran its Project Natick experiment by submerging a data center off Scotland’s coast to use ocean cooling. The project worked (failure rates were lower than land-based systems) but it was too costly to maintain and scale. The company confirmed to Techerati this year that underwater facilities are “no longer an area of interest.” Orbital data centers face the same hurdles at higher altitude: launch costs, radiation, and no possibility of repair once deployed.
A McKinsey report this year estimates AI-ready data-center capacity is growing about 30 percent annually, yet most added power demand can still be met with better chips, cooling, and localized renewables. In energy-constrained regions, companies are building microgrids, installing large-scale batteries, and reusing server heat for district heating systems. These approaches are practical, measurable, and already cutting real emissions.
The Ground Reality
Efficiency isn’t exotic. AI doesn’t need solar panels in orbit to become sustainable. It needs smarter workloads, smaller models, and better hardware. Fatih Birol, executive director of the IEA, summed it up earlier this year: “AI is one of the biggest stories in the energy world today, but until now, policy makers and markets lacked the tools to fully understand the wide-ranging impacts.” Those tools are finally emerging in the form of transparency standards, model efficiency benchmarks, and regional caps on new high-density compute.
Starcloud’s vision might prove technically impressive, but it remains a high-cost experiment chasing a low-cost problem. Energy efficiency has always been solved through better design, not higher altitude. Space might one day offer perfect sunlight and cooling, but the smarter bet is still on Earth, where GPUs can learn to think harder while wasting less.
