It is worth starting with what a data center actually is, because the term covers far more variety than the headlines suggest. At its simplest, a data center is a building full of computers plus the power and cooling to keep them running. For decades, most organizations ran their own small server rooms on-site. That model was inefficient, with those server rooms mostly sitting idle, built for peak demand but rarely used to capacity. Over the past twenty years, computing has migrated out of thousands of these small rooms and into a smaller number of large, specialized facilities run by cloud providers. The data room in the basement became a data center in the cloud.
It is also worth remembering that a data center is downstream of demand, not a source of it. These buildings exist because of what households and businesses already do every day—stream, search, shop, bank, and now prompt AI. That does not make their footprint imaginary. But it does mean the building is an effect of consumption we have collectively chosen, which is worth keeping in mind before treating the facility itself as the problem to be solved.
Another point worth understanding, as it is often misunderstood: “data center” describes very different business models. At one end sits the pure landlord who builds the shell and brings power to the site, then leases it to a large technology company that installs its own computers and contracts for the electricity and water. At the other end, a company builds and operates the entire facility itself. In between, there is a range of hybrids.
Ownership of the buildings is dispersed across specialist developers, publicly traded REITs such as Digital Realty and Equinix, private investment funds, and the technology giants themselves. The tenants are far more concentrated. A handful of “hyperscalers”—principally Amazon, Microsoft, Google and Meta—account for an outsized share of demand, alongside a newer group of AI-focused “neocloud” providers. Importantly, even these giants do not own everything they use; a large share of their capacity is leased rather than self-built.
The consolidation of many users’ moving on-premise computing into a smaller number of large facilities had a consequence that runs counter to the headlines. The common assumption is that all this construction simply means more energy use but for most of the last decade, the opposite was true. A large, shared facility pools the demand of many users across the same hardware, so the machines do useful work rather than sit idle. The result was a striking decoupling: between 2010 and 2018, the computing done in the world’s data centers rose by roughly 550 percent while the electricity they consumed rose just 6 percent. Consolidating computing is, in effect, the equivalent of commuters moving from individual cars to mass transit. The footprint became more visible because it was concentrated in large buildings, but the energy required to perform a given task kept falling.
The industry measures facility efficiency with a simple ratio called Power Usage Effectiveness, or PUE—the total energy a data center draws divided by the energy its computers actually use. A value of 1.0 would mean every watt reaches the servers; 1.5 means the building consumes half as much as the IT load itself, the rest going to cooling and power conversion. The typical on-premise server room runs around 1.6, while the hyperscalers run near 1.1. A hyperscale facility wastes less than a tenth of its power on overhead; a typical enterprise site wastes more than a third.
And PUE, if anything, understates the advantage. It captures only facility overhead - cooling and power loss - not how fully the servers themselves are used. The larger efficiency gain comes from utilization: pooled hyperscale hardware runs far busier than the idle servers of a company server room, so the true gap in useful computing per unit of energy is even wider.
The trend, however, is now reversing. AI is far more power-hungry than ordinary web tasks, and it is adding a heavy new layer of demand on top of the steady growth in everyday cloud use. The efficiency dividend that kept energy use flat for a decade has largely been spent.
Much of the concern around data centers blends water and energy into a single worry. It is more useful to separate them, because they have very different answers.
On water, the concern is more tractable than it appears. The traditional way to cool a data center is to evaporate water, running it past the hot equipment and letting evaporation carry the heat away, which consumes the water and requires continual replacement. A facility cooled this way can draw large volumes from the local supply, which is what is fueling most community opposition. The newer approach is “closed-loop” cooling, in which the cooling fluid is sealed inside and sheds its heat to the outside air, much like a radiator in a car. The same water circulates again and again rather than evaporating away. A building designed this way consumes almost no water for cooling on an ongoing basis, although of course, it must be filled upon its commissioning.
This is not aspirational; the largest operators are already committing to it. Microsoft announced that every data center it has designed since August 2024 uses a “zero-water” evaporation design, eliminating the more than 125 million liters of water a typical facility previously consumed each year. As the company’s engineering lead put it, “once the system is filled during construction, it will continually circulate water between the servers and chillers to dissipate heat without requiring a fresh water supply.” Water, in short, is largely a design choice, and the operators that consume it are moving quickly to design it out. Our own partners confirm that all future projects are being designed with closed-loop cooling.
Energy is the harder problem. At the level of a single facility the industry has become remarkably efficient, wasting little power, particularly when compared to inefficient on-premise facilities. But efficiency per task is no longer the whole story, because the sheer volume of computing is now growing faster than efficiency can offset.
The numbers are stark. U.S. data centers consumed roughly 192 terawatt-hours of electricity in 2024 - about 4.7 percent of the nation’s total, already more than double the 76 terawatt-hours they used in 2018. The Department of Energy’s national laboratory projects that figure will reach roughly 649 terawatt-hours by 2030, or close to 12 percent of all U.S. electricity, with a scenario range running as high as 843. This is no longer an efficiency story; it is an infrastructure-investment story.
The binding constraint is not only how much electricity is needed, but how quickly it can be delivered. New supply depends on transformers, gas turbines and grid interconnections, all of which carry long lead times and multi-year queues. In many markets, getting power to a site is now the gating item. Meeting this demand will require substantial capital in both generation and transmission, and it has revived interest in firm, around-the-clock power: nuclear restarts and next-generation reactors alongside natural gas and renewables.
All of this—the generation, the transmission, the grid upgrades have to be paid for, which raises a question as politically charged as any in the debate: who pays the bill?
Here the trend is moving in a constructive direction: increasingly, the cost of new infrastructure is being placed on the large users driving the demand rather than on ordinary households. Virginia—home to the world’s densest concentration of data centers—offers the clearest template. In late 2025, its State Corporation Commission approved a new rate class for Dominion Energy, effective January 2027, that places data centers in their own category and requires them to commit to long-term contracts and to pay for the bulk of the transmission and generation capacity they reserve, whether or not they use it. The state has since gone further, enacting the nation’s first statewide consumption tax on data-center energy use and ordering that data centers bear the full cost of transmission lines built specifically to serve them. The large user, not the household, is increasingly being made to pay for the infrastructure its demand requires, and it is a framework now taking shape in other states and municipalities.
The tax incentives themselves are also drawing public outcry. Yet it is worth remembering that those exemptions are what drew these facilities to Virginia rather than Texas or Arizona in the first place; without them, the local tax base simply would not exist. In Loudoun County, which has the densest data-center market in the world, data centers now fund roughly 40 percent of the county’s general fund, which has allowed Loudoun to cut its residential property tax rate nearly every year for the past decade, to the lowest in Northern Virginia. Absent that revenue, the county’s own budget arithmetic implies the residential rate would need to rise by roughly 90 percent to hold services constant.
It is worth noting that these large consumers of energy are not passive consumers of dirty power; they have become the largest buyers of clean energy on the planet. In 2025, Amazon, Microsoft, Google and Meta together accounted for roughly 49 percent of all corporate clean-power purchase agreements signed globally – and nearly 87 percent in the United States. The two single largest corporate clean-energy buyers in the world last year were both data-center operators: Meta and Amazon, each contracting more than 10 gigawatts, with nuclear representing approximately a quarter of their activity as they reach for firm, around-the-clock power. This does not erase the problem that demand is currently outrunning clean supply, and their emissions have been rising as a result; but it is a very different picture from the one the headlines paint.
The grid these tenants draw on is itself getting cleaner: over the past decade U.S. electricity generation has grown even as the mix has shifted toward renewables and away from coal.
AI has genuinely changed the energy equation, and the sector’s rising power demand is a real challenge that governments, utilities and the technology companies will need to manage together.
For investors, the implication follows from these dynamics. As power, permitting, and now the cost of capital increasingly ration the buildout, the advantage accrues to owners of well-located, powered and contracted capacity, particularly assets on the better end of the efficiency and water spectrum, leased to creditworthy tenants, in the markets best equipped to absorb them.