How Data Centers Are Powered & Cooled

Where does the power come from?

Data centers are among the most power-intensive land uses in existence. A single large campus can draw 360–1,200 megawatts continuously — comparable to the output of a mid-sized nuclear plant. Where that power comes from determines not just the facility's carbon footprint, but how much it stresses local water supplies, the ERCOT grid, and the transmission infrastructure that serves residential customers.

There are five primary power source configurations seen in Central Texas proposals:

• Grid-only: the facility draws entirely from ERCOT. No on-site generation; carbon intensity tracks the grid mix (~320 gCO₂e/kWh estimated in 2025, down from ~384 gCO₂e/kWh in 2022 due to wind and solar growth).
• On-site natural gas: the developer builds a combined-cycle gas plant (NGCC) or gas turbine directly on or adjacent to the campus to supply all or part of its power. The highest-profile local example is CloudBurst Data Centers at its Francis Harris Lane campus in Hays County, with an Energy Transfer pipeline permitted to deliver up to 450,000 MMBtu/day.
• On-site nuclear (SMR): small modular reactors co-located on the data center campus, providing firm, carbon-free baseload. Announced by Last Energy (contract with a European hyperscaler) and Oklo (Microsoft agreement). No U.S. commercial SMR has yet operated.
• On-site solar + storage: solar PV arrays paired with battery energy storage systems (BESS) to provide daytime power and some nighttime coverage. Rarely sufficient as the sole power source for large campuses due to land requirements and storage costs.
• Hybrid: a mix of grid power, on-site generation (gas, solar, or storage), and long-term renewable PPAs. Most announced hyperscale facilities are some form of hybrid.

Grid vs. on-site generation: what changes locally

The grid/on-site distinction matters far beyond carbon accounting. When a data center draws from ERCOT, the water used to generate its electricity is spread across hundreds of power plants — many distant from Central Texas, drawing from the Colorado River, the Gulf Coast, or West Texas. That water use is real but diffuse.

When a developer builds an on-site gas plant, all the water used to cool those turbines is consumed locally — from the same Crystal Clear SUD, Guadalupe-Blanco River Authority allocation, or Edwards Aquifer pumping rights that serve nearby residents. An on-site combined-cycle plant at the scale of the proposed CloudBurst campus can add 3–12 million additional gallons per day to the local water footprint, on top of the direct cooling load.

On-site generation also means on-site air emissions. A gas plant requires a TCEQ New Source Review (NSR) or Title V air permit, which caps annual NOₓ, PM₂.₅, CO, and VOC emissions. These air quality impacts are in addition to the water and carbon impacts.

Small modular reactors: promise and timeline

SMRs are being announced as co-located power sources for data centers at an accelerating pace. The appeal is real: nuclear power has lifecycle carbon emissions of approximately 12 gCO₂e/kWh — lower than any other firm (non-intermittent) power source — and SMRs require far less land than large conventional reactors.^[1]

However, every announced SMR co-location deal in the United States as of 2026 is pre-NRC application. No commercial SMR has completed NRC licensing and operated. The NRC's combined construction and operating license (COLA) process, even on an expedited timeline, typically runs 3–7 years from pre-application engagement to operating license.^[2] The only advanced reactor that has completed NRC design certification to date is the NuScale SMR (DC approved 2022, later delayed).^[3]

SMR co-location announcements are therefore better understood as long-term hedges — potentially replacing or supplementing gas generation in the 2032–2038 timeframe — rather than as solutions to near-term water and emissions impacts. A facility permitted today with an on-site gas plant will run on gas for the foreseeable future, regardless of what SMR contracts are announced alongside it.

Renewable claims: what counts and what doesn't

Many hyperscalers claim to operate on "100% renewable energy." This claim almost always rests on the purchase of Renewable Energy Certificates (RECs), which represent one megawatt-hour of electricity generated by a renewable source — but not necessarily the electricity physically delivered to the data center.

Unbundled RECs do not remove electrons from the grid or reduce the carbon intensity of the power actually consumed. They are an accounting mechanism. Google, Microsoft, and Amazon have each acknowledged this limitation and have shifted to more demanding 24/7 carbon-free energy (CFE) commitments — matching hourly consumption to hourly carbon-free generation within the same grid region.^[4] Most smaller operators have not made the same shift.

For a facility in Texas, the relevant question is not "does the developer buy RECs?" but rather:

1. Is there an additionality-bearing PPA — a contract that funds the construction of new renewable capacity that would not otherwise exist?
2. Is the PPA within ERCOT, so the electrons are physically correlated with the load?
3. Does the PPA provide hourly matching or only annual matching?

Annual, system-wide RECs are the weakest form. Additionality-bearing, ERCOT-based, hourly-matched PPAs are the strongest. The Central Texas Data Center Tracker records whether disclosed renewable procurement is additionality-bearing and distinguishes it from unbundled RECs.

Cooling technology choices

After power source, cooling technology has the largest effect on a data center's local environmental footprint. There are seven main categories:

• Evaporative (wet) cooling: cooling towers evaporate water to atmosphere. The dominant technology in hot climates. Water Usage Effectiveness (WUE) of 1.5–2.5 liters per kWh of IT load in Central Texas summer heat. The most water-intensive option.
• Air-cooled (dry) cooling: ambient air removes heat; no evaporative water loss. Requires more electricity (higher PUE) and is less efficient above 35°C / 95°F. WUE ≈ 0 L/kWh, but PUE is typically 1.4–1.7 versus 1.2–1.4 for evaporative facilities.
• Hybrid (dry/wet): primarily air-cooled with evaporative assist during peak heat. WUE ~0.3–0.7 L/kWh depending on design and climate.
• Closed-loop liquid cooling: water or glycol circulated in a sealed loop between server heat exchangers and a dry cooler. Minimal evaporative losses; WUE ~0.1–0.3 L/kWh.
• Two-phase immersion: servers submerged in a dielectric fluid that boils at low temperatures and re-condenses in a heat exchanger. Near-zero water; WUE ~0.05 L/kWh.
• Single-phase immersion: servers in a non-evaporating dielectric fluid, with dry cooler. Near-zero water consumption.
• Direct-to-chip liquid: cold plates on CPUs and GPUs carry coolant directly from server to building heat exchangers. Can be paired with a small cooling tower for final heat rejection; WUE ~0.2–0.4 L/kWh.

AI/ML workloads using dense GPU clusters (NVIDIA H100, H200, B200) generate 2–5× more heat per rack than traditional server workloads, accelerating industry adoption of direct-to-chip and immersion cooling. Several hyperscale facilities now target WUE below 0.3 L/kWh even in hot climates.

PUE and WUE: the two efficiency metrics

Power Usage Effectiveness (PUE) measures total facility power divided by IT equipment power. A perfect PUE is 1.0; the industry average is ~1.5. PUE improvements reduce both electricity consumption and the water needed to generate that electricity (indirect water), but do not necessarily reduce direct cooling water.

Water Usage Effectiveness (WUE) measures liters of water consumed per kWh of IT equipment energy. Best-in-class air-cooled facilities achieve WUE near 0; industry average evaporative facilities are ~1.8 L/kWh; Texas summer conditions can push evaporative WUE to 2.5 L/kWh or higher.

PUE and WUE are in tension for evaporative facilities: improving PUE by running servers harder at high utilization reduces total power draw but concentrates heat, potentially increasing WUE. The combination — total water per unit of useful IT work — matters most for local resource impact.

Waste heat: an underused resource

Data centers generate enormous quantities of waste heat at low temperatures (40–70°C for air-cooled, up to 80°C for immersion). This heat can in principle be captured and used for district heating, greenhouse operations, aquaculture, or industrial processes.

In Northern Europe — particularly Denmark, Sweden, and Finland — data center waste heat recovery is legally mandated above certain facility sizes, and several campuses supply heat to thousands of households.^[5] In the United States, heat reuse remains rare. The economics are challenging in warm climates where district heating demand is low and the temperature differential between waste heat and ambient air is smaller.

For Central Texas, waste heat reuse at scale is unlikely given the climate and lack of district heating infrastructure. However, greenhouse or aquaculture applications at smaller scales are feasible and worth tracking as a disclosure item.

Backup generators and air quality

All data centers maintain diesel or natural gas emergency generators for power continuity. A large campus may have 50–500 individual generators, each 2–3 MW. These are tested regularly (typically 1–4 hours per month per generator) and run during grid outages.

In Loudoun County, Virginia, data centers collectively hold permits for over 1,000 diesel generators — aggregate NOₓ and PM₂.₅ emissions from testing alone rival those of a medium-sized power plant.^[6] TCEQ air permits cap annual runtime for each generator (typically 500 hours/year for emergency use), but enforcement of testing restrictions is often light.

Natural gas turbines used for backup power emit lower particulates than diesel but can have significant NOₓ emissions depending on combustion technology. Battery energy storage systems (BESS) are increasingly proposed as an alternative or supplement, eliminating combustion emissions from backup operations entirely — at the cost of lithium supply chain and end-of-life impacts not captured in Scope 1 emissions.

The Texas tradeoff

Every power and cooling choice in Central Texas involves a tradeoff between local impacts:

• Evaporative cooling minimizes electricity demand but maximizes local water withdrawals — the worst possible choice during Stage 3 drought on the Edwards Aquifer.
• Dry cooling eliminates water demand but increases electricity draw by 5–15%, adding load to an ERCOT grid already at elevated risk of energy emergencies (NERC 2024).
• On-site gas provides firm, grid-independent power but adds local Scope 1 emissions, local water for power, and air quality impacts — concentrated in a county where residents already face drought restrictions.
• Grid-only with renewables PPA spreads emissions and water impacts across the ERCOT region, but adds large-load interconnection demand that may require expensive transmission upgrades billed to ratepayers.
• SMR co-location offers the best long-term carbon and water profile, but no commercially operating U.S. SMR exists yet and licensing timelines are measured in years to decades.

There is no single "right" answer — but there is a spectrum of choices, and those choices can be made transparent. The Central Texas Data Center Tracker's environmental grade synthesizes these tradeoffs into a single per-project score so residents and decision-makers can compare facilities on a common basis.

Sources

1. ^ IPCC Sixth Assessment Report (AR6), Working Group III, "Mitigation of Climate Change," Annex III, Table A.III.2 — Lifecycle emission factors by technology, 2022. Nuclear: median 12 gCO₂e/kWh.
2. ^ U.S. Nuclear Regulatory Commission, "Licensing Process for New Reactors — Combined License Process," 2024. Available at nrc.gov.
3. ^ NRC, "Design Certification — NuScale Power Module," 2022; NuScale announcement of project cancellation (UAMPS Carbon Free Power Project), November 2023.
4. ^ Google DeepMind, "24/7 Carbon-Free Energy: Cleanly Powering Our Data Centers," 2020. See also Microsoft Sustainability Report 2023 and Amazon Sustainability Report 2023.
5. ^ Danish Energy Agency, "Data Centers and District Heating in Denmark," 2022; Meta Odense data center waste heat case study, 2021.
6. ^ Piedmont Environmental Council, "Data Centers and Air Quality in Loudoun County," 2023. See also Virginia Department of Environmental Quality permit inventory.