What Washington Did Not Ask the Watershed

A citizen's framework for siting American AI infrastructure before the basins decide for us.

The question Washington is asking is where do we put the data centers. The question Washington should be asking is what does the watershed permit.

I write from the upper Rio Grande, where ditches older than the republic still move water by gravity and consent — so I have seen what it looks like when a watershed is asked.

That inversion is the entire argument. Right now, private capital picks a site, signs a non-disclosure agreement with a county that lacks the technical capacity to evaluate it, and asks the aquifer to negotiate after the concrete is poured. The result is visible in three places at once: the Columbia River basin, where two decades of hydropower abundance ended in a temporary natural gas plant approved by the Grant County PUD to keep the data centers running; Loudoun County, Virginia, where the state legislature's own auditors project average residential electricity bills rising by up to $37 a month by 2040 to underwrite infrastructure for facilities that employ fewer than 150 people each; and the Colorado River basin, where the most water-stressed major watershed in North America continues to receive new hyperscale proposals as if the river were not in structural collapse.

None of these outcomes were inevitable. All of them were chosen — by absence. The federal government has not produced a national siting framework for data center infrastructure. In its absence, capital sites itself.

The scale

The Federal Energy Regulatory Commission confirmed in its March 2026 State of the Markets report that US data center capacity exceeded 50 gigawatts at the end of 2025, representing 24% compound annual growth since 2020. Wood Mackenzie projects expansion to 100 GW by 2030. Goldman Sachs Research forecasts a 165% increase in global data center power demand over the same period. By Wood Mackenzie's estimate, data centers will account for roughly 68% of US electrical load growth through 2030, and approximately 40% of total US electrical equipment demand.

This is the largest discretionary infrastructure decision the country will make this decade. It is being made without a map.

The structural insight: not all compute is the same

Frontier AI training — the process of building new foundation models — requires centralized hyperscale clusters with internal bandwidth that cannot be distributed across small sites. This is the workload most associated with the current build-out, and it is roughly 30% of total demand. The remaining 70% — inference, cloud services, storage, video processing, enterprise computing, scientific workloads, archival — can be distributed across mid-scale modular facilities of 2 to 10 megawatts, paired with closed-loop cooling and waste heat reuse.

This split changes the conversation. Of the 100 GW projected by 2030, roughly 30 GW must be hyperscale. Roughly 70 GW need not be. At an average node size of 5 MW, the distributed portion translates to approximately 15,000 mid-scale facilities — comparable in count to the national wastewater treatment system and a fraction of the country's K-12 school footprint. This is infrastructure the United States has built before, at this scale, multiple times.

The policy question is therefore not whether to permit AI infrastructure. It is which 30 GW should be sited where, and on what terms, with the remaining 70 GW directed toward a distributed model that captures benefits locally rather than concentrating strain.

Six filters for hyperscale siting

Any framework that pretends a single criterion — cheap power, cool climate, fiber proximity — is sufficient has already failed. A defensible site satisfies all six of the following, not any subset:

1. Climate suited to free-air or closed-loop cooling. Cooling currently accounts for roughly 40% of a typical data center's electricity consumption. Sites north of the 43rd parallel, or with consistent low ambient temperatures, materially reduce both energy and water demand at the source.

2. Existing surplus low-carbon generation with documented headroom. Hydroelectric, geothermal, wind, or nuclear capacity that is currently curtailed, exported below cost, or operating below firm capacity. Promised future capacity does not qualify. The Quincy precedent — where hydro abundance was exhausted and replaced with new gas — must not be repeated.

3. Watersheds with positive recharge. Annual surface flow or aquifer replenishment that exceeds projected withdrawal, including indirect water consumed in upstream electricity generation. Sites that depend on stressed basins should be disqualified regardless of available cooling technology.

4. Grid capacity that does not require new fossil generation. If a project's power plan depends on new combined-cycle gas to meet its load, the project fails the carbon test by definition. This filter alone disqualifies the majority of currently proposed sites.

5. Legal infrastructure for community and tribal consent. Water rights regimes with prior-appropriation transparency, county zoning authority with technical staff, utility commissions empowered to enforce large-load tariffs, and meaningful consultation with affected tribal nations. The pattern of redacted utility filings — visible currently in Montana, where NorthWestern Energy has provided the Public Service Commission letters of intent with most information blacked out — is incompatible with this filter.

6. Latency tolerance. Frontier training does not require coastal access. Filter 6 simply confirms that sites satisfying filters 1–5 can serve their actual function.

The map

Applying all six filters, approximately 60 hyperscale sites can be defensibly located across four regions:

• The Upper Midwest hydro-and-cold belt (northern Minnesota, Upper Peninsula Michigan, northern Wisconsin): roughly 12 sites, leveraging Manitoba hydro corridors and recharging Great Lakes-basin groundwater under closed-loop conditions.

• The Pacific Northwest interior (eastern Washington, northern Idaho, Snake River basin): roughly 6 sites, conditional on firm prohibitions against fossil backfill.

• The Northern Great Plains wind corridor (eastern Montana, western North Dakota, northern South Dakota): roughly 10 sites, contingent on new transmission and mandatory closed-loop cooling given Missouri headwaters and Ogallala constraints.

• Northern New England (northern Maine, northern New Hampshire, upstate New York along the Hydro-Quebec intertie): roughly 5 sites, the most underrated region on the map.

• 
  

An additional 15 sites can be defensibly paired with existing nuclear capacity in the Tennessee Valley Authority service area, southwest Michigan, and Maryland's Calvert Cliffs corridor — using carbon-free generation already paid for rather than chasing new build. A reserve of roughly 9 sites should be held for small modular reactor pairing in Wyoming, North Dakota, Idaho, and the Tennessee Valley, contingent on commercial deployment that does not yet exist.

The no-build list

The harder document, and the one no federal agency has produced:

• The Colorado River basin — Arizona, southern Nevada, southern Utah, southern California. The river is in structural collapse. Every hyperscale facility sited here takes water from a downstream user who will not receive it back.

• The Rio Grande basin — the whole of it. The lower basin is in chronic shortage and worsening; the Supreme Court's final decree in Texas v. New Mexico this May ordered New Mexico to retire roughly 5.9 billion gallons of annual groundwater pumping below Elephant Butte. The upper basin is fully appropriated, and its acequia-served communities hold first-priority water rights that predate state water law by two centuries and federal water law by three. There is no latitude on this river below which the water stops being spoken for.

• The Loudoun County corridor and the broader Virginia/DC region. Grid and water systems are already past defensible limits.

• The Southeast — Georgia, the Carolinas, Florida, central Texas. Heat, drought cycles, and evaporation losses disqualify the region even under closed-loop conditions.

• The Ogallala aquifer footprint — west Texas, eastern New Mexico, western Kansas, the Oklahoma panhandle, eastern Colorado. The aquifer is collapsing on a timeline measured in decades. Federal policy should not accelerate it.

• 
  

One international observation

China's "East Data West Computing" strategy, launched in 2022, geographically separates compute from population centers and routes AI training workloads toward inland regions with surplus renewable generation and cold climate. The policy is imperfect and serves multiple political functions beyond environmental ones, but it demonstrates a point relevant here: the question of where can be asked at the national level, before the construction, rather than after. The United States has not asked it.

Three asks

To the Department of Energy: Publish a national data center siting framework that incorporates watershed recharge, grid carbon intensity, and community legal capacity as binding filters for federally supported projects, including those receiving tax incentives or transmission upgrades.

To the Federal Energy Regulatory Commission: Require transparency in large-load utility contracts exceeding 100 MW, including disclosure of water consumption projections, fossil generation dependencies, and tribal consultation records, as a condition of interstate transmission approval.

To Congress: Direct Treasury and the Department of Energy to study and report, within the year, a tax structure that flips the current incentive — subsidizing mid-scale distributed facilities with mandatory heat reuse, and applying graduated cost recovery to hyperscale facilities sited outside the framework above.

The watershed has not been asked. It is still possible to ask it. The window is roughly two years.

The map is not the watershed — but the map is where the watershed gets defended or sold.