The story of the next five years of AI infrastructure has stopped being told in GPU shipments. It is being told in water rights filings and substation queues.
In Utah this month, the developers of a data center called Stratos quietly withdrew a water rights application that would have shifted irrigation water to industrial use near the Great Salt Lake. The withdrawal followed nearly three thousand nine hundred public protests, in a region drought-stressed enough that the lake itself is shrinking. The project did not get cancelled. The water request did. The build will have to find its cooling somewhere else.
A week earlier, a different story broke in Georgia. A data center operator drew roughly twenty-nine million gallons of municipal water across fifteen months of construction without authorization. The local utility found out because residents complained about water pressure. Officials are not fining the operator. Twenty-nine million gallons, for concrete and dust control, on a six-point-two million square foot facility — and the regulatory consequence is paperwork.
These are the same story. The constraint on AI infrastructure is no longer compute. It is power and water, in that order, and the constraint is local.
According to recent industry tracking, more than half of the twelve gigawatts of AI data center capacity announced for 2026 will not be built this year — roughly seven gigawatts has been cancelled or delayed. The reasons are grid bottlenecks, electrical component shortages, the second-order effects of Chinese tariffs, and a category that did not exist on these slide decks two years ago: community opposition. Between March and June of last year alone, ninety-eight billion dollars of project value was blocked or delayed by local action. At least twenty-five projects were cancelled outright.
Texas has spent the past year arguing about a single county. Hays County, in extreme drought, fielded requests for one million gallons of cooling water per day. California's legislature came close, in late 2025, to passing data center disclosure rules. The bill was reduced to a study requirement after lobbying from firms that publicly champion responsible siting.
The vendor pitch has not caught up with any of this. Decks still show capacity as a smooth supply curve, indexed to GPU manufacturing and not to substation queues, water tables, or county commissioners. Most procurement teams are still optimizing for unit economics on a curve that no longer applies.
What changes for an operator who is not a hyperscaler? Two things, both quiet.
The first is that "the region" is no longer a single variable. Until recently you picked a region the way you picked a coffee — close enough, fast enough, priced by the cup. The region is now four overlapping decisions: whether the substation will sign the load letter, whether the county will sign the water permit, whether the community will accept the construction, and whether any of the three can survive a drought year. A vendor who names a region without naming a substation is selling you a promise they cannot keep.
The second is that the build calendar has uncoupled from the announcement calendar. A megawatt announced in 2025 is now landing in 2028 if it lands at all. The contracts being signed today against capacity that has not been permitted are the contracts that will get repriced when the permits stall. We have written before that the grid has the last word on what gets built; the water table is the second word, and it is being spoken louder than the procurement teams are listening.
For builders sitting outside the bidding war for North American grid capacity — including those of us working from places with cooler air and lower population density per megawatt — the constraint is also an opening. The places that build the next decade of compute will be the places where someone has already done the unglamorous work of water modeling, transmission queueing, and community consent. Not all of those places are in the countries currently holding the contracts.
The compute era has had a long honeymoon with the assumption that physical inputs are infinite at the price quoted. Stratos and the Georgia meter are the same story, told by the lake and by the tap respectively. The honeymoon is ending where it always ends, which is at the county line.
The short of it.
A withdrawn water filing in Utah and an unmetered Georgia tap, both reported this week, are the surface of the same shift: power and water — not GPUs — are now the binding constraints on AI build-out, and the binding happens at the county level. Half of 2026's announced capacity is already delayed. Operators should treat "region" as four separate gates (substation, water, community, drought) rather than a single dropdown.