Maria Gonzalez's ranch is already losing about two feet of groundwater annually. She just learned it's been contracted to cool seven gigawatts of computing power - computers she'll never see. Her family has survived droughts before, but this one feels different. She says she isn't against progress, she just wants to know whose progress this is for.
Whenever I ask AI a question, the reply arrives in mere seconds but the energy fuelling AI no longer comes from a grid in Virginia or Dublin but from space: satellites in constant daylight with panels catching sunlight eight times stronger than Earth’s, processors cooled by vacuum, lasers trading data faster than any cable.
No land cleared, no rivers diverted, no diesel backups. Just light feeding intelligence. But even in orbit, intelligence still runs on metal, fuel and power drawn from somewhere, some finite reserve we have stopped counting.
If every AI model carried a label, like food or fuel, what would one for a space-based model say? Tonnes of rocket fuel per launch? Litres of water per terabyte? Square kilometres of mined earth per chip? We already struggle to make supply chains visible; now we are launching them beyond sight.
Google’s Project Suncatcher is a genuine attempt to think differently. Instead of adding more data centres to an already strained planet, it asks whether intelligence could operate beyond our terrestrial limits. Satellites powered by constant sunlight, cooled by a vacuum, form a self-contained network in orbit. It is an elegant, even hopeful, proposition. In theory, it avoids the land, water and grid pressures that define computing on Earth. Suncatcher is a research concept filled with variables and technical challenges still to be solved. If the economics hold - launches under $200 a kilogram and prototypes flying by 2027 - it could mark a real breakthrough in sustainable infrastructure.
What counts as ‘sustainable’ depends entirely on what and who you choose to count. Each launch burns hundreds of tonnes of fuel. Each chip still relies on metals mined from scarred provinces. When the satellites fail, and all satellites do, they become debris circling the planet at 17,000 mph. Suncatcher offers a brilliant idea for reducing stress within Earth’s surface systems; yet it does so by exporting that stress upward, into a commons none of us can clean. The label would not read zero emissions; it would read relocated impact.
Back on Earth, OpenAI’s Stargate pursues the opposite dream: rebuild the planet as a power station. In a letter to the White House, the company warned of America’s ‘electron gap’, the shortfall between AI’s appetite and the grid’s capacity. Six planned sites across Texas, New Mexico, Ohio and Wisconsin will draw seven gigawatts and $400 billion. One of them sits near Maria’s ranch. The contracts were settled elsewhere, long before she knew what was at stake.
OpenAI wants the US to add 100 gigawatts of new energy each year, to ‘unlock electrons and economic opportunity’. Google lifts the problem upward; OpenAI digs deeper into Maria’s water level. No one asked her permission. No one was obliged to.
Suncatcher is a future moonshot - a speculative experiment in powering intelligence without touching the planet’s surface. OpenAI’s Stargate, by contrast, is firmly of the present. Announced only in recent weeks, both projects reveal how quickly AI’s energy race is accelerating. One reaches for orbit, the other for the grid, both chasing abundance while ignoring the question: what kind of intelligence are we really building and who pays the price?
In Texas, the grid powering AI farms already strains during heatwaves. In orbit, cooling means slow radiation into vacuum darkness. Physics does not negotiate, and neither geography nor altitude changes the equation: intelligence costs matter.
And then there is the question of who will account for the energy debt we have already accrued? The same governments that have yet to meet their own climate pledges? The same companies now building the next grid?
The atmosphere already carries our backlog of emissions and rivers bear the heat of today’s data centres. We are still trying to solve yesterday’s costs with tomorrow’s technology. In AI with a Label, I argued that seeing the energy bill changes behaviour. The bill is already here: rising seas, shrinking water levels, collapsing ecosystems. Before we engineer new energy for intelligence, perhaps we should learn to pay for the energy we have already spent.
The International Energy Association’s Energy and AI report, published this year, makes the numbers clear. Data centres already emit around 180 million tonnes of CO₂ each year, roughly half a percent of all combustion emissions, and their electricity use could double by 2030. Even if every known AI efficiency were adopted, total energy emissions would fall by only 4 percent and that is in a best-case scenario.
AI advocates argue the trade is worth it: medical breakthroughs, climate modeling, scientific discovery. They're not wrong. ChatGPT can write my grocery list. AlphaFold can predict protein structures that cure diseases. But nowhere in that calculus is the assumption that every use deserves equal access to limited resources. We have not built a system that can distinguish between curing cancer and generating images of cats wearing hats - and until we do, both draw from Maria's water table with equal claim.
Since intelligence has consequences, accountability starts with a simple label: This query used 0.5 litres of water and 0.3 kWh of electricity. Make the cost visible before the answer appears. Make timing matter - cheaper rates when wind is blowing, higher costs when the grid strains. Make AI companies publish not just energy use, but energy source: solar, gas, coal. This requires governance that connects AI regulation with energy and space policy. Only then can we start to see the true cost of intelligence.
The other night, I looked up and wondered how crowded the sky was going to become. The old map of stars is slowly giving way to a web of moving lights, more infrastructure than constellation.
I still ask AI questions. But I ask fewer, not out of guilt, but out of respect - for the electron, the river, the miner, the rancher, and everyone whose world may bear the cost of my curiosity.
Maria Gonzalez is still waiting for someone to ask her a question. About her water. About her future. About whether the computing power produced just a mile from her home is worth what it will cost her land.
I don't think AI can answer that one.
Note: Names and situations such as Maria Gonzalez are illustrative composites, used to represent real dynamics facing communities affected by data-centre expansion and resource extraction.
