Data Center Electricity Costs & Power Management
Data centers consume 100–200 kWh per square foot per year — the highest energy use intensity of any commercial building type by a wide margin. A 1 MW facility pays $60,000–$100,000 per month in electricity alone. But unlike any other building, the electricity bill isn’t an operating expense to minimize — it’s directly tied to revenue-generating capacity. Every watt saved on cooling is a watt available for compute. Electricity isn’t overhead. It’s cost of goods sold. This page covers electricity cost management for data center operators. For industry data on U.S. data center growth, facilities, and market impact, see our Data Center Tracker.
PUE: The Number That Runs Everything
Power Usage Effectiveness is the defining metric for data center efficiency. Every operational decision — site selection, cooling architecture, rack density, even what time of year you do maintenance — ultimately shows up in PUE.
PUE = Total Facility Power ÷ IT Equipment Power. A PUE of 1.0 means every watt entering the building reaches a server — impossible in practice, because cooling, lighting, and power conversion losses always exist. A PUE of 2.0 means you’re spending as much on overhead as you are on compute. The industry average sits at 1.55–1.60 (Uptime Institute), while hyperscale leaders report 1.10–1.12.
| PUE Range | Rating | Typical Facility | What It Means |
|---|---|---|---|
| 1.10–1.20 | Best-in-class | Hyperscale (Google, Microsoft, Meta) | Custom-built facilities with free cooling, liquid cooling, and purpose-designed power distribution. 10–20% overhead. |
| 1.20–1.40 | Excellent | Modern colocation, well-designed enterprise | Hot/cold aisle containment, economizers, efficient UPS. 20–40% overhead. |
| 1.40–1.60 | Average | Typical enterprise, older colocation | Industry average range. Significant efficiency opportunities likely exist in cooling and airflow management. |
| 1.60–2.00+ | Below average | Legacy facilities, converted office space | 40–100%+ of IT power wasted on overhead. Retrofit or replacement typically justified. |
Where the Power Goes
A data center’s power budget breaks down into four categories. The ratio between them defines your PUE and determines where efficiency investments produce the largest return.
50–65%
IT Load
Servers, storage arrays, and networking equipment. This is the revenue-generating load — every other system in the building exists to keep this running. The goal is to maximize this percentage relative to total facility power. AI/GPU racks now draw 30–50 kW each, compared to 7–10 kW for traditional compute.
25–35%
Cooling
CRAC/CRAH units, chillers, cooling towers, and pumps. The largest non-IT load and the primary target for PUE improvement. The industry is moving from air cooling to liquid cooling to immersion cooling — each step compresses this percentage. AI workloads are accelerating the transition because they generate 3–5× more heat per rack than traditional compute.
8–12%
Power Distribution
UPS systems, power distribution units (PDUs), switchgear, and transformers. Conversion losses are inherent but can be minimized with high-efficiency UPS topologies, right-sized transformers, and higher-voltage distribution (415V vs. 208V). N+1 and 2N redundancy architectures add capacity you hope never operates — but the idle losses are real.
2–5%
Lighting & Misc
Lighting, security systems, fire suppression monitoring, office space for on-site staff. Small in absolute terms but often overlooked in efficiency audits. LED lighting with occupancy sensors in cold aisles and storage areas eliminates continuous draw in spaces that are occupied only during maintenance windows.
Cooling Economics
Cooling is where the money is. It’s the largest controllable cost in the facility and the primary lever for PUE improvement. The right cooling strategy depends on climate, rack density, and whether you’re building new or retrofitting existing infrastructure.
-
Free cooling / economizer hours Location-dependent
Uses outside air when ambient temperature is below the supply air setpoint, eliminating mechanical refrigeration during those hours. In the Pacific Northwest or upper Midwest, economizer mode can provide 4,000–6,000 hours per year of free cooling. In Texas or Arizona, you’re looking at 1,000–2,000 hours. The difference in annual cooling cost between a facility in Oregon and one in Dallas can exceed $1 million at scale. The site selection decision is an energy cost decision. -
Hot aisle / cold aisle containment 10–15% improvement
Prevents mixing of supply and return air by physically separating hot exhaust from cold supply aisles. Basic but still not universal — Uptime Institute surveys show roughly 20% of facilities still lack full containment. The investment is modest (curtains, panels, or hard enclosures) and the efficiency improvement is immediate: 10–15% reduction in cooling energy from eliminating recirculation. -
Raising inlet temperatures 4–5% per degree
ASHRAE A1 class allows server inlets up to 113°F. Most operators run at 75–80°F. Every degree Fahrenheit you raise the inlet temperature reduces cooling energy by 4–5%. Moving from 72°F to 80°F saves 32–40% on cooling energy alone. The limiting factor is usually comfort (for staff working in the space) and equipment warranty terms, not the hardware itself. -
Liquid cooling Essential for AI racks
Direct-to-chip cooling or rear-door heat exchangers eliminate the need to cool room air. Water’s heat capacity is 3,500× that of air, making it dramatically more efficient at removing heat from dense compute. At 30–50 kW per rack (typical for GPU-dense AI workloads), air cooling simply cannot keep up. Liquid cooling is no longer optional for high-density deployments — it’s a prerequisite. -
Evaporative cooling Best in dry climates
Uses water evaporation instead of mechanical refrigeration. Much more efficient in arid regions — Google’s data centers in The Dalles, Oregon use this approach. The trade-off is water consumption: a large evaporative cooling system can use 1–3 million gallons per year. In water-stressed regions, this creates a cost-vs-sustainability tension that’s becoming increasingly contentious in local permitting.
Electricity Procurement at Scale
Data center electricity procurement is nothing like small commercial. At 1 MW and above, you’re not picking from published rate plans — you’re negotiating directly with generators, retailers, and in some cases, the wholesale market itself.
- Power Purchase Agreements (PPAs) — Long-term contracts (10–20 years) directly with renewable energy generators. Provides cost certainty, sustainability reporting, and often below-market rates in exchange for the developer’s financing security. Most hyperscale operators now procure the majority of their power through PPAs. The risk: you’re locked in for a decade or more, and if market prices drop significantly, you’re paying above-market.
- Behind-the-meter generation — On-site solar, fuel cells, or natural gas generation that reduces grid dependence. Increasingly common as a hedge against both price volatility and grid reliability concerns. Some facilities use on-site generation for base load and grid power for peak, inverting the traditional relationship with the utility.
- Wholesale market participation — In ERCOT (Texas), PJM (Mid-Atlantic), CAISO (California), and other ISOs, large data centers can participate directly in wholesale markets or through qualified scheduling entities. This exposes you to real-time pricing — which can mean 2–3¢/kWh at 3 AM and $9/kWh during a grid emergency. Requires sophisticated energy management and risk tolerance.
- Demand response — Data centers with flexible workloads (batch processing, backups, non-real-time analytics) can participate in utility demand response programs, earning $50–$200/kW per year in payments for agreeing to shed load during grid emergencies. Some operators shift non-critical workloads across regions to reduce exposure to peak pricing. The constraint: mission-critical facilities with strict SLAs have limited flexibility.
- Redundancy costs — N+1 or 2N power architecture means you’re paying for UPS, generators, and switchgear that you hope never operates at full capacity. This overhead is baked into the rate structure and the capital budget. A 2N design essentially doubles your power infrastructure investment for the same IT load. The trade-off between uptime guarantees and energy cost is one of the most consequential decisions in facility design.
Compare commercial electricity rates for your facility
Enter your ZIP code to see available rates from multiple providers.
Data Center Electricity FAQ
How much does electricity cost for a data center per month?
A 1 MW data center facility typically pays $60,000–$100,000 per month for electricity at average commercial rates. Hyperscale facilities running 50–100 MW+ pay millions per month but negotiate rates well below the national average of 13.51¢/kWh — often 4–7¢/kWh through long-term power purchase agreements or wholesale market participation. The actual monthly cost depends on PUE (a facility at 1.5 PUE uses 50% more total power than one at 1.0), geographic location, rate structure, and whether the facility operates its own on-site generation.
What is PUE and what’s a good PUE for a data center?
Power Usage Effectiveness (PUE) measures total facility power divided by IT equipment power. A PUE of 1.0 would mean every watt goes to compute — impossible in practice because cooling, lighting, and power conversion losses always exist. The industry average is 1.55–1.60 according to Uptime Institute surveys. Enterprise facilities typically achieve 1.4–1.6, well-designed colocation runs 1.2–1.4, and hyperscale leaders like Google and Microsoft report 1.10–1.12. Every 0.1 PUE improvement on a 5 MW facility saves approximately $350,000 per year.
How do data centers buy electricity differently from other businesses?
At 1 MW and above, data centers negotiate directly with generators and retailers rather than selecting from published rate plans. Common procurement strategies include long-term power purchase agreements (PPAs) with renewable generators for 10–20 year price certainty, behind-the-meter generation using on-site solar or fuel cells, and direct participation in wholesale electricity markets through qualified scheduling entities in ERCOT, PJM, or CAISO. The 85–95% load factor that data centers maintain makes them highly attractive customers to providers, translating to stronger negotiating leverage and lower per-kWh rates.
What is free cooling and where does it work best?
Free cooling uses outside air or water to cool the data center when ambient temperatures are below the supply air setpoint, eliminating the need for mechanical refrigeration during those hours. In northern climates like the Pacific Northwest, Scandinavia, or the upper Midwest, free cooling can provide 4,000–6,000 hours per year of cooling at near-zero marginal energy cost. In hotter climates like Texas or Arizona, economizer hours drop to 1,000–2,000 per year. The annual cooling cost difference between a facility in Oregon and one in Dallas can exceed $1 million at scale, making site selection one of the most consequential energy cost decisions.
How do AI workloads change data center electricity costs?
AI and GPU-intensive workloads generate 3–5 times more heat per rack than traditional compute, pushing rack densities from 7–10 kW to 30–50 kW or higher. This concentration of heat makes conventional air cooling insufficient and accelerates the transition to liquid cooling or immersion cooling. The higher power density also compresses PUE improvements because cooling infrastructure must work harder, and it increases total facility power draw significantly — a single GPU training cluster can consume as much power as an entire traditional data center wing. Operators building for AI workloads should plan for 2–3× the power and cooling infrastructure of equivalent traditional compute space.
Can data centers participate in demand response programs?
Yes. Data centers with flexible or deferrable workloads — batch processing, backups, non-real-time analytics — can participate in utility demand response programs and earn $50–$200 per kW per year in payments for agreeing to reduce load during grid emergencies. Some operators shift non-critical workloads to off-peak hours or across geographic regions to reduce exposure to peak pricing. However, mission-critical facilities with strict uptime SLAs have limited flexibility, and participation requires careful coordination between operations and energy management teams to ensure that shedding load doesn’t violate customer commitments.
