The Memphis Sand Aquifer supplies drinking water to roughly one million people across the greater Memphis area. On its single highest draw day recorded so far, xAI's Colossus facility pulled 812,502 gallons from that same aquifer. That is the daily water consumption equivalent of approximately 2,500 American households drawn from a municipal drinking water source by one GPU cluster.
In March 2026 alone, xAI drew more than 25 million gallons total from the aquifer. The facility is operating 50 portable methane-fired generators to power its compute load while permanent grid infrastructure catches up. The combination of fossil-fueled generation and municipal aquifer draw at that scale is the kind of operational profile that turns a permit into a political liability.
Memphis Light, Gas, and Water CEO Doug McGowen negotiated the arrangement. The rate his utility extended to xAI: 19 cents per 100 gallons. Standard Memphis customers pay 32 cents. That is a 40 percent discount on water from an aquifer that supplies the city's taps, extended to a company running a facility that Elon Musk described as the world's largest supercomputer.
The wastewater treatment facility built to handle Colossus's discharge was originally budgeted at $80 million. The final cost landed at $200 million. That gap, $120 million in overrun on a facility that processes the water a data center exhausts, reflects what happens when thermal infrastructure is scoped before the full compute load is specified. The water draw was almost certainly not modeled at anywhere near 812,000 gallons per day when the treatment plant was designed.
This is a recurring failure mode in hyperscale builds: the cooling infrastructure is sized against the initial compute spec, and then the compute spec doubles. The treatment plant cannot be retroactively redesigned mid-construction. The budget absorbs the delta. At Colossus, the delta was $120 million.
Anthropic is paying xAI $1.5 billion per month for compute access through 2029. At that revenue rate, 25 million gallons of water at 19 cents per 100 gallons costs xAI approximately $47,500 for the month. The water bill is a rounding error on the invoice. McGowen's below-market rate does not move the needle for Elon Musk's P&L. It moves the needle for every Memphis resident drawing from the same aquifer.
That asymmetry is why preferential utility arrangements at hyperscale facilities produce community opposition. The financial irrelevance of the discount to the operator is obvious to the people who pay the standard rate and share the water source. The perception problem writes itself.
Daily water draw per GPU cluster is the number that matters for siting decisions, permit applications, and community relations. Not annual totals. Not comparisons to global consumption. The per-day figure, drawn from a specific source, measured against that source's municipal capacity.
At 812,502 gallons on a peak day, xAI Memphis gives the industry a reference point. Microsoft's buildout in Wyoming draws from a water system already under pressure from agriculture and energy extraction. The Memphis situation adds a second case study where the compute load and the municipal water supply are in direct tension, and where the tension was not resolved before the facility came online.
The methane generator problem runs parallel. Fifty portable generators are a temporary solution to a permanent power problem. They are loud, they emit, and they generate heat. That heat has to go somewhere. Colossus is running evaporative cooling on a building powered by fossil fuel backup units drawing municipal drinking water at a preferential rate. Each element of that sentence is documented and public. Together they describe a facility whose cooling and power strategy will face increasing pressure as the legal and political environment tightens around AI infrastructure buildouts.
Several states and municipalities have already moved to tighten data center water use agreements. The federal transparency push on data center water and energy disclosure is active. When the xAI Memphis numbers circulate through state utility commissions and city councils evaluating their own large-user agreements, McGowen's 19-cent rate becomes a comparison point no municipal water authority wants to defend publicly.
The argument for below-market rates on municipal utilities has always been economic development: tax base, employment, anchor tenant for regional infrastructure investment. That calculus held when data centers employed large on-site workforces. A GPU cluster needs cooling technicians, security, and facilities staff. The ratio of jobs to water draw at a facility like Colossus is nothing like the ratio at a manufacturing plant or distribution center that would be the traditional beneficiary of a utility rate discount.
Municipalities are beginning to notice that arithmetic. The ones that notice it before signing the utility agreement are in a better negotiating position than the ones that notice it after the facility is operational and drawing 25 million gallons a month.
There are no published specifics on what cooling systems xAI is running inside Colossus at the rack level. At the densities required for Grok training infrastructure, the assumption is direct-to-chip liquid cooling with rear-door heat exchangers or CDU-based loops. The water drawn from the aquifer is almost certainly for the building's cooling towers, not in-rack liquid cooling circuits, which typically run closed-loop.
That distinction matters for the policy conversation. Closed-loop direct-to-chip cooling at the rack level consumes virtually no water. The cooling tower supporting the overall building envelope is where the consumptive evaporation happens. Facilities that eliminate the cooling tower, using dry coolers or adiabatic backup, cut the aquifer draw to near zero. That technology exists. It carries a higher upfront cost and typically a higher PUE at peak ambient temperatures. But for a facility in Memphis drawing 800,000 gallons a day from a municipal drinking water source, the cost comparison looks different than it does in a temperate climate with no community opposition.
The national framing of AI water use is frequently overstated. The Colossus case is the local version of the problem: a specific facility, a specific aquifer, a specific daily draw, a specific preferential rate, a specific community that drinks from the same source. That is the argument that moves city councils and utility commissions. The math here does not need inflation to be persuasive.
The facilities that get built next in markets with constrained water supplies will face tighter terms than the ones built during the permissive window xAI took advantage of in Memphis. The 812,502-gallon peak day is now on the public record. Every municipality reviewing a large-user water agreement for a GPU cluster will know that number exists.