Imagine a town hall meeting in a sleepy suburb of Virginia. A local resident, who just wants to know why their electricity bill jumped 15% in a year, is staring at a glossy slide deck presented by a representative from a Big Tech firm. The slides are filled with images of green leaves and “sustainable infrastructure,” but the resident is looking at the massive, windowless concrete block that just appeared three miles from their house—a building that consumes as much power as a small city and drinks millions of gallons of water a day just to keep a few thousand H100s from melting. The air in the room is thick with a specific kind of tension: the realization that the “cloud” isn’t a cloud at all, but a very loud, very thirsty neighbor.
For years, the industry has treated the “cloud” as a magical, ethereal place where compute just exists as a utility, like air. It’s a convenient fiction we’ve all agreed to maintain. In reality, the cloud is a series of loud, hot, thirsty warehouses that require an absurd amount of civil engineering to maintain. We’ve spent the last two years arguing over context windows and quantization, but the real bottleneck isn’t a software bug or a lack of high-quality tokens. It’s the fact that you cannot simply “scale” a power grid by wishing it into existence. (Assuming the utility companies are even being honest about their remaining capacity, which is a huge assumption).
The current trajectory is unsustainable. As The Verge points out, the fight is moving from the boardroom to the backyard. This is essentially the 19th-century fight over where to lay railroad tracks, except instead of steam engines, we’re arguing about liquid cooling and substation capacity. The labs are pretending that their “Net Zero” promises are a shield, but those promises usually rely on carbon offsets—a financial shell game that does nothing to stop a local aquifer from drying up because a cluster of GPUs needs to stay at 30 degrees Celsius. Who actually believes a carbon offset saves a local river?
The industry is now pivoting toward nuclear power and Small Modular Reactors (SMRs) because they’ve realized that the existing grid is a joke. But building a nuclear reactor isn’t like deploying a new version of a model; it takes a decade, not a sprint cycle. There is a massive mismatch between the speed of AI software development and the speed of concrete and steel. We are essentially trying to run a Formula 1 race on a dirt road that’s currently being washed away by a rainstorm. Or maybe I’m being too cynical—perhaps the SMRs actually work on a timeline that matches the H100 refresh cycle. But I doubt it.
This friction will eventually force a reckoning that changes where the money flows. We’ll see a shift where compute is no longer centralized in a few “AI alleys” but is forced to migrate to wherever the power is cheapest and most stable—likely meaning a lot more data centers in places where nobody wants to live, far from the political noise of the suburbs. Within 12 months, we’ll see the first major US municipality implement a total ban on new AI data center permits based specifically on water scarcity and grid instability. The era of the “invisible” data center is over; the physical footprint is now the primary constraint. It’s a reminder that no matter how “intelligent” the software gets, it still needs a plug in the wall and a lot of cold water.
The cloud is just a very expensive way to hide a plumbing problem.
