A single hyperscale datacenter at full operational load typically consumes between 1 million and 5 million gallons of water per day — comparable to a town of 10,000 to 50,000 residents. For the largest AI training campuses being announced in 2026, the figures are higher still. As community pushback intensifies and drought-affected regions face increasing political pressure, water usage has become the single most-discussed environmental impact of the US datacenter buildout.
This post answers the literal question — how much water does a datacenter actually use? — and then breaks down the variables, the public disclosures, and where the industry is headed.
How Much Water Does a Datacenter Use?
The honest answer is: it depends. Water consumption varies dramatically based on cooling system design, climate, facility load, and how aggressively the operator has prioritized water efficiency.
Traditional evaporative-cooling datacenters (the dominant design through ~2020):
- ~1.8 million gallons per MW per year at full load
- For a 100MW facility: ~180 million gallons/year, or roughly 500,000 gallons/day
- For a 1GW campus (modern hyperscale scale): ~1.8 billion gallons/year
Air-cooled and adiabatic-cooled designs (increasingly common since 2021):
- ~0.05-0.3 million gallons per MW per year — a 6-30x reduction
- For a 100MW facility: ~30 million gallons/year
Closed-loop liquid cooling (emerging in AI-dense halls):
- Near-zero ongoing water consumption after initial fill
- A few thousand gallons per year for makeup and maintenance
These ranges matter because the spread between best- and worst-case designs is roughly 100x. A 100MW campus might consume 500M gallons/year (worst case) or 1M gallons/year (best case) of identical compute capacity. The design choice — frequently made years before public scrutiny arrives — locks in the water footprint for the life of the facility.
Hyperscaler Disclosures: What Operators Actually Report
The major hyperscalers have begun publishing water-use data in sustainability reports. The most useful disclosures (and the most-cited numbers in industry press):
Google: Reported 4.3 billion gallons of water used company-wide in 2023, with ~25 facilities listed individually in Google's environmental report. Google has committed to replenishing 120% of its operational water consumption by 2030 — the most ambitious public commitment among the hyperscalers.
Microsoft: Reported 1.7 billion gallons of water used at datacenter facilities in fiscal 2023. Microsoft's sustainability report documents a commitment to water-positive operations by 2030, replenishing more water than the company consumes.
Meta: Reported approximately 870 million gallons across operational facilities in 2023. Meta's sustainability page documents the company's commitment to be water-positive by 2030 — and Meta has been the most aggressive among the hyperscalers in retrofitting facilities to air-cooled and adiabatic designs.
AWS: Significantly less granular disclosure than the other three. AWS reports water use as part of Amazon's broader sustainability report but does not break out datacenter water usage as cleanly. Amazon sustainability disclosures include the framework but limited per-facility data.
Apple: Almost zero datacenter water use disclosed — Apple's iCloud infrastructure is small relative to the others, and the company has used air-cooled designs aggressively for its small US datacenter portfolio.
The WUE Metric (Water Usage Effectiveness)
Water Usage Effectiveness — WUE — is the industry's standardized water-efficiency metric, analogous to PUE for power efficiency.
WUE = liters of water consumed per kilowatt-hour of IT energy used.
Typical ranges:
- 1.8 L/kWh — traditional evaporative cooling, hot/humid climate. Most colocation and older hyperscale.
- 0.7-1.2 L/kWh — well-tuned evaporative cooling, moderate climate.
- 0.1-0.4 L/kWh — adiabatic / air-assisted cooling, modern design.
- <0.05 L/kWh — fully air-cooled or closed-loop liquid cooling.
Hyperscalers report WUE numbers in their sustainability reports. Google's 2023 fleet-wide WUE was 1.0 L/kWh; Microsoft reported approximately 0.49 L/kWh. The downward trend across years is one of the few unambiguously good stories in datacenter sustainability metrics — but new AI training facilities can drive WUE numbers back up if they default to traditional cooling.
Why AI Workloads Are Making This Harder
AI training clusters generate dramatically more heat per rack than traditional cloud workloads — 20kW to 80kW per rack versus 5-10kW for traditional cloud. That heat has to go somewhere. The most efficient cooling for AI density — direct-to-chip liquid cooling — actually has very low water consumption. But many AI clusters are being retrofit into existing facilities with evaporative cooling infrastructure already in place, which increases water consumption per compute-unit.
Newer purpose-built AI campuses (Stargate sites, Meta Prometheus and Hyperion, Microsoft's recent Wisconsin and Pennsylvania campuses) are being designed for liquid cooling from the start. The water-use trajectory for those facilities is much more favorable. The trajectory for AI workloads retrofit into older evaporative-cooled facilities is much worse.
The Drought Compounding Factor
Water consumption matters everywhere, but it becomes politically explosive in drought-affected regions. The most-discussed cases:
Utah / Eagle Mountain: Meta's campus has been the focal point of organized opposition by the Utah Rivers Council and others. Multi-year drought in the Great Basin has put residential water restrictions in place while datacenter water permits have continued to be approved.
Arizona / Phoenix: Salt River Project and other utilities are managing increasingly tight water budgets. Maricopa County is reportedly considering water-efficiency conditions for new datacenter approvals.
Nevada / Las Vegas, Reno: Lake Mead is at historic lows. Switch, Vantage, and other operators in the region have invested heavily in dry-cooling designs partly in response.
Central Oregon and Eastern Washington: Hyperscale clusters in Quincy, Hermiston, Boardman, and Umatilla draw heavily on Columbia River water rights — politically contested between agriculture, salmon habitat, and tribal interests.
How the Industry Is Responding
Three structural shifts are visible:
1. Cooling-design migration. New builds are overwhelmingly specifying air-cooled, adiabatic, or closed-loop liquid cooling rather than evaporative cooling — even where evaporative would be cheaper in absolute terms. The political cost of water consumption now exceeds the savings.
2. Water-replenishment programs. Google, Microsoft, and Meta have committed to net water-positive operations. The mechanism is generally watershed-restoration investments in the same geographic regions where the operators consume water — funded by direct investment rather than offsets.
3. Disclosure standardization. WUE per facility, not just fleet-wide, is becoming the industry standard. The next 12-24 months will likely see investors and regulators expecting per-facility disclosure as a default.
What This Means for Vendors and Site Selection
For cooling-equipment vendors, the water-efficiency conversation is reshaping which products win specifications. Sales reps for adiabatic-cooling, closed-loop liquid-cooling, and direct-to-chip-cooling products have a structural tailwind that didn't exist five years ago. Traditional CRAC/CRAH evaporative-cooling vendors are increasingly competing on water-efficiency claims rather than raw thermal performance. For developers and enterprise build teams, water-use risk is now a meaningful factor in site selection — particularly in drought-affected regions. Our Site Selection briefings include water-availability and water-use political-risk assessments for every candidate site.
For vendors selling cooling systems or working in the AI-density retrofit market, the buying signals are increasingly tied to water-efficiency upgrade cycles, not just raw capacity expansion. Our Vendor Sales briefings flag these signals at the facility level.
Sources & Further Reading
Google Environmental Report — Annual disclosure including per-facility water usage data and WUE figures.
Microsoft Sustainability Hub — Water-positive commitment framework and annual progress reports.
Meta Sustainability — Water-positive commitment and facility-level disclosures.
Washington Post: Data Centers' Hidden Costs — Investigative coverage of water and power impacts of US hyperscale infrastructure.
The Green Grid — Industry consortium that defines WUE and PUE metrics and publishes implementation guidance.