The media has found its favorite new climate villain: the artificial intelligence data center.
Every time the summer temperature spikes, the headlines write themselves. Journalists look at a sprawling concrete warehouse in Virginia or Ireland, look at the local power grid, and sound the alarm. They claim that AI workloads are pushing local infrastructure to a breaking point, evaporating millions of gallons of pristine drinking water, and cooking the planet. They want you to believe that training a single large language model is equivalent to running a fleet of diesel trucks through a suburban neighborhood.
It is a neat, emotionally compelling narrative. It is also completely wrong.
The current panic over data center energy and cooling consumption misses the fundamental mechanics of industrial thermodynamics. The mainstream critique treats data centers like static, open-loop black holes. In reality, modern hyperscale facilities are among the most intensely optimized thermodynamic systems ever engineered. The real strain on our infrastructure isn’t coming from the servers; it is coming from an obsolete, inflexible utility sector that is hiding its own decades-long failure to modernize behind the specter of big tech.
Stop blaming AI for an infrastructure crisis that we built ourselves.
The Evaporation Fallacy: Where the Water Actually Goes
Let’s dismantle the biggest lie first: the idea that data centers are "consuming" and destroying local water supplies.
When a critic states that a facility uses millions of gallons of water a day, they are counting on your ignorance of the difference between withdrawal and consumption. They want you to picture a giant pipe sucking water out of a river and spitting out toxic sludge, or worse, making that water vanish from the earth entirely.
That is not how industrial cooling works.
Modern hyperscale data centers primarily utilize evaporative cooling systems or closed-loop chilled water systems. In a closed-loop system, the water stays inside the pipes forever. It absorbs heat from the servers, travels to a heat exchanger, rejects that heat into the atmosphere, and cycles back. The net consumption of water is effectively zero.
Even in traditional evaporative cooling towers—where water is evaporated to cool the air—the water is not lost to the universe. It enters the troposphere as water vapor and returns to the hydrological cycle. More importantly, an increasing percentage of hyperscale facilities do not use potable drinking water at all. They run on industrial gray water, treated sewage effluent, or brackish water that is entirely unfit for human consumption or agriculture.
I have spent years auditing enterprise infrastructure deployments, and I have seen companies throw millions of dollars at PR campaigns trying to prove they are "water positive." They are playing a losing game because they are accepting the flawed premise of the argument. If a data center uses 100,000 gallons of recycled municipal wastewater to keep its GPUs at 20°C, it has deprived exactly zero citizens of drinking water. It has actually monetized a waste product for the local municipality.
The Power Grid Inversion: Data Centers Are the Solution, Not the Problem
The second pillar of the data center panic is the electrical grid. "AI is going to break the grid," the pundits cry. They point to projections from the International Energy Agency (IEA) or local grid operators like PJM Interconnection showing skyrocketing load growth.
Here is what they are missing: data centers are the most predictable, manageable, and credit-worthy electricity consumers on the planet.
A traditional manufacturing plant or a residential suburb has a chaotic load profile. People turn on their air conditioners at 5:00 PM when they get home from work, creating massive, unpredictable spikes known as the "duck curve." Grids struggle to handle these rapid ramps, forcing them to fire up dirty, expensive "peaker" plants running on natural gas or fuel oil.
Data centers, conversely, represent a flat, baseload demand. A cluster of training servers runs at a sustained, predictable wattage 24 hours a day, seven days a week. Furthermore, because AI training workloads are asynchronous—meaning the training of a model does not require instantaneous real-time human interaction—these workloads can be shifted.
Imagine a scenario where a hyperscale operator detects that the local solar grid is oversaturated at 1:00 PM. The operator can ramp up non-urgent AI training jobs to soak up that excess renewable energy that would otherwise be curtailed (wasted). When the sun goes down and the grid tightens, they can throttle those same workloads back down.
Data centers do not destabilize grids; they act as economic anchors for new energy infrastructure.
Who do you think is financing the construction of utility-scale solar, wind, and next-generation geothermal plants across the globe? It isn't the utility companies, and it certainly isn't the state governments. It is tech companies signing Power Purchase Agreements (PPAs). According to BloombergNEF, tech companies accounted for over half of all corporate clean energy PPAs globally. They are single-handedly funding the decarbonization of the very grids they are accused of destroying.
The Flawed Metrics: Why PUE is Outdated
Part of the reason the public conversation is so warped is that the industry relies on a deeply flawed, outdated metric: Power Usage Effectiveness (PUE).
$$PUE = \frac{\text{Total Facility Energy}}{\text{IT Equipment Energy}}$$
A PUE of 1.0 is the theoretical perfect score, meaning every watt of electricity entering the building goes directly to the compute hardware, and zero watts are spent on cooling, lighting, or distribution losses.
Hyperscalers have driven PUE down from an industry average of 2.0 a decade ago to around 1.1 or 1.2 today. But PUE is a purely operational metric. It measures the efficiency of the building, not the efficiency of the work.
You could run an incredibly efficient building (PUE of 1.05) filled with obsolete 14nm CPUs that deliver miserable computational throughput per watt. Conversely, you could have a slightly less efficient building (PUE of 1.4) packed with the latest 3nm custom ASICs or advanced GPUs that deliver ten times the computational output for the same energy footprint.
By focusing entirely on the physical footprint of the data center building—the heat it rejects, the water it cycles—critics miss the massive, compounding efficiency gains happening at the silicon and software layers. We are getting exponentially more intelligence out of every joule of energy. The carbon intensity of a single AI inference query has plummeted by orders of magnitude over the last three years, yet the headlines only look at the aggregate power draw of the facility.
The Real Crisis: Incumbent Utility Incompetence
If the data centers aren't the problem, why are we seeing localized strain during heatwaves?
Look at the regulated monopolies running our electrical grids. For decades, traditional utilities have underinvested in transmission lines, distribution hardware, and high-voltage substations. They operate on guaranteed rates of return, disincentivizing them from taking risks or deploying agile infrastructure.
Now, suddenly, an agile, hyper-capitalized industry turns up wanting to buy gigawatts of power, and the utilities are caught flat-footed. It takes up to seven years to build a new high-voltage transmission line in the United States, mostly due to bureaucratic red tape and NIMBY lawsuit loops. A data center can be erected in eighteen months.
The strain we are experiencing is a regulatory and bureaucratic bottleneck, not a physical shortage of energy or water.
There is a downside to my contrarian view, and we must be honest about it: in the short term, the friction between fast-moving tech capital and slow-moving utility bureaucracy will cause local flashpoints. In regions like Northern Virginia or Dublin, Ireland, the sheer concentration of facilities has outpaced the local utility's ability to upgrade substations. This can lead to localized price inflation for electricity if regulators allow utilities to pass the costs of rapid upgrades onto residential consumers.
But the solution is not to ban or restrict data centers. The solution is to force the unbundling of the utility sector, fast-track transmission permits, and allow data center operators to build their own behind-the-meter, next-generation nuclear or geothermal power generation.
Stop Demanding Less Compute; Demand Better Infrastructure
The narrative that we must limit AI development to save our resources is a form of techno-pessimistic degrowth. It assumes that the resource pie is fixed, that our infrastructure is static, and that human ingenuity cannot engineer its way out of thermal constraints.
Every major industrial leap forward—from the steam engine to the electrification of factories to the birth of the internet—was accompanied by a massive, terrifying spike in energy demand. And in every instance, that energy consumption unlocked efficiencies across society that far outweighed the initial input cost.
AI is already being used to optimize the logistics routes of global shipping fleets, reducing global fuel consumption. It is being used to discover new materials for high-efficiency solar cells and to manage smart grids with microsecond precision. The compute we deploy today is the very tool that will automate and optimize the resource distribution of tomorrow.
The heatwaves aren't the problem. The data centers aren't the problem. The problem is a collective lack of ambition that prefers rationing existing scarcity to engineering future abundance.
Turn the servers up. Build the reactors. Let it run.
