Two numbers tell you more about a data center's operating discipline than almost anything else in the data room. The first is PUE, power usage effectiveness. The second is WUE, water usage effectiveness. One prices the power overhead. The other prices the water overhead. When I am reviewing a facility for a lender or a buyer, these are among the first figures I ask for, and the answers shape how much of the rent I believe.

Here is where both numbers stand in 2026, and how I read them as an appraiser.

PUE: the average barely moved, again

PUE is total facility power divided by IT equipment power. A PUE of 1.0 would mean every watt reaches the servers, which no real building achieves. The overhead above 1.0 is cooling, power distribution losses, lighting, and everything else it takes to keep the IT load alive.

Uptime Institute's 2026 global survey puts the industry average PUE at 1.52. That is a small improvement from the 1.54 the same survey found in 2025, and Uptime is blunt about the pattern: the 2026 figure falls right in line with a seven year trend of relative stasis. The easy gains were taken a decade ago. What remains is a slow grind, held back by the legacy buildings that stay in service because demand keeps them full. Uptime notes that 27% of respondents work in data centers over 15 years old.

The average hides a wide spread, and the spread is the story. Respondents' largest data centers average 1.45, recent builds routinely produce 1.3 or better, and 23% of respondents work in a data center averaging below 1.3. At the front of the pack, Google reports a fleet wide average PUE of 1.09 for 2025, which it calculates as 83% less overhead energy than the industry average.

What a tenth of a point is worth

Run the arithmetic the way an underwriter does. Take a 100 megawatt IT load. At the industry average PUE of 1.52, that building draws 152 megawatts in total, so 52 megawatts of the power bill buys overhead instead of compute. At Google's 1.09, the same IT load draws 109 megawatts, with only 9 megawatts of overhead. The difference, 43 megawatts, is capacity a grid constrained developer does not have to buy, interconnect, or wait in a queue for. It is also operating expense the tenant or the owner does not pay every hour of every year.

That is why I do not accept a PUE at face value in a data room. I ask which PUE it is. Design PUE is what the engineers modeled. Annualized PUE is what the meters recorded across all seasons and all load levels. A facility that runs at partial load, which is most facilities in their early years, usually posts a worse annualized figure than its design figure. Uptime itself cautions that PUE is best used to track one site over time, not to rank two different buildings against each other, because climate, redundancy, and load factor all move the number. A 1.4 in Phoenix and a 1.4 in Oregon are not the same achievement.

For valuation, the practical read is this. A below average PUE on a modern building is evidence of durable operating advantage, and it belongs in the expense analysis, not just the marketing deck. A high PUE on an older building is not automatically a defect to be cured. It may be a retrofit bill, and as I noted in my piece on construction cost per megawatt, bringing older stock up to modern standards costs close to new construction money.

WUE: the number half the industry still does not track

WUE is liters of water used per kilowatt hour of IT energy. It captures the cooling trade that PUE cannot see. Evaporative cooling lowers the power bill and raises the water bill. Air cooling does the reverse. A facility can post an excellent PUE by spending water, which is why reading PUE alone in a dry market is a mistake.

Start with how thin the disclosure still is. Uptime's 2026 survey found that only 53% of operators collect water usage metrics for sustainability purposes, up from 47% in 2025, while 79% collect PUE and 87% track power consumption. In other words, nearly half the industry cannot tell you its WUE because it does not measure it. When a seller cannot produce a WUE, I treat that as a diligence finding in itself.

The best national benchmark comes from Lawrence Berkeley National Laboratory. Its average site WUE for U.S. data centers was just over 0.36 liters per kWh through 2023, and LBNL projects that figure rising to 0.45 to 0.48 liters per kWh by 2028 as hotter, denser AI capacity comes online. Total on site water consumption for U.S. data centers reached 66 billion liters in 2023. Scale that with my earlier example: a 100 megawatt IT load running a full year at the 0.36 average uses roughly 315 million liters of water, and at the projected 2028 intensity it would use roughly 400 million. Water supply is becoming a site selection input on the same level as power, which is exactly the shift I track in my work on powered land and how to value data center land.

The hyperscalers report far below the national average. AWS reports a WUE of 0.12 liters per kWh for 2025. Microsoft reports a fleet WUE of 0.27 liters per kWh for fiscal 2025. Meta reports 0.19 liters per kWh for 2024, on a withdrawal basis. Read those three with care: the bases differ. Some operators report water withdrawn, others water consumed, and those are different quantities, so the figures show the distance between hyperscale practice and the national fleet better than they rank the three companies against each other.

The honest gaps

Two cautions. First, there is no audited, facility level public database for either metric. PUE has a long survey history behind it, which makes its trend credible, but it is still self reported survey data. WUE is worse: disclosure is voluntary, definitions vary, and the most recent national benchmark is modeled from 2023 data. Anyone quoting you a precise industry WUE for 2026 is ahead of the evidence.

Second, neither metric measures useful work. A building can improve its PUE while its servers sit idle, and a low WUE site can still sit in a stressed watershed where every liter is contested. That is why the underwriting checklist I use for data centers puts PUE and WUE alongside power sourcing, tenant credit, density, and residual value, instead of letting either ratio carry the conclusion on its own.

The bottom line

In 2026, underwrite the average data center at a PUE of about 1.52, expect modern large facilities nearer 1.3 to 1.45, and treat anything near 1.1 as hyperscale best practice rather than a market standard. On water, use just over 0.36 liters per kWh as the U.S. fleet benchmark, expect it to rise toward 0.45 to 0.48 by 2028, and ask every operator for its own measured figure. If it cannot produce one, that silence tells you something about the asset, and about the market it will sell into when disclosure is no longer optional.

Sources: Uptime Institute Global Data Center Survey 2026 (July 2026) and Global Data Center Survey 2025; Google 2026 Environmental Report; Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report; AWS Sustainability Report 2025; Microsoft environmental reporting, fiscal 2025; Meta 2025 Environmental Data Index.