6 hours ago
74
Every time you ask ChatGPT to write a cover letter or summarize a PDF, a data center somewhere gets a little thirstier. OpenAI is now pushing back on the narrative that AI’s water consumption is spiraling out of control, pointing to closed-loop cooling systems as the industry’s answer to sustainability concerns.
The clarification comes at a moment when the gap between AI’s environmental promises and its environmental reality has never been more visible. And the details matter more than the talking points.
The water problem, in plain terms
Data centers generate enormous amounts of heat. Thousands of GPUs running inference and training workloads produce thermal output that would make a commercial kitchen jealous. That heat needs to go somewhere, and for decades, the answer has been water.
Traditional evaporative cooling systems work like a giant sweat gland. Water absorbs heat and evaporates into the atmosphere, taking the thermal energy with it. It’s effective, but it’s also a one-way trip for that water. In English: the water doesn’t come back.
Closed-loop cooling systems take a fundamentally different approach. Water (or another coolant) circulates through sealed piping, absorbs heat from servers, dumps that heat through a heat exchanger, and loops back around. The water is recirculated rather than evaporated, which dramatically cuts the amount of fresh water a facility needs to pull from the local supply.
OpenAI is highlighting this distinction as the preferred path forward. Most data centers, the company notes, use closed-loop systems that recirculate water, reducing water withdrawal.
Here’s the thing, though: reducing withdrawal is not the same as reducing consumption. Closed-loop systems still require an initial fill, periodic top-offs, and the energy used to run them has its own water footprint upstream at power plants. The framing is accurate but incomplete, which is precisely the kind of nuance that tends to evaporate (pun intended) in public discourse.
What the numbers actually say
The water cost of AI is not theoretical. Training OpenAI’s GPT-3 was estimated to consume around 700,000 liters of fresh water. That’s roughly enough to fill a standard Olympic swimming pool, and it was just one training run for a model that’s now two generations old.
On the usage side, ChatGPT is estimated to require over 2 liters of water per 50 queries when you account for both cooling and the water consumed by power generation upstream. Scale that across hundreds of millions of users, and you start to see why environmental groups have taken notice.
Looking ahead, future AI workloads are projected to cause up to 6.6 billion cubic meters in annual water withdrawals by 2027. For context, that figure approaches the total annual water consumption of entire small nations. The trajectory is steep, and it’s accelerating alongside the AI arms race.
Microsoft, one of OpenAI’s closest partners and largest cloud infrastructure providers, reports a water usage effectiveness of approximately 0.30 liters per kilowatt-hour across its data centers. The company says closed-loop cooling is expected to save more than 125 million liters per site annually compared to evaporative systems. That’s a meaningful reduction per facility, but Microsoft is also building facilities at a pace that could offset those per-site gains in aggregate.
Oracle has gone even further with its newest AI data centers, deploying direct-to-chip, closed-loop, non-evaporative cooling systems that the company says result in effectively zero ongoing community water usage. The coolant runs through pipes directly attached to the chip, pulling heat at the source before it can radiate into the broader server environment.
Why closed-loop cooling is more than a PR move
Direct-to-chip liquid cooling isn’t just about water savings. It can reduce server-fan power consumption by up to 80% compared to traditional air cooling. Fans are one of the biggest parasitic power draws in a data center, so eliminating most of that load has real energy implications too.
The technology is projected to account for roughly one-third of the data-center cooling market by 2030, up from a small fraction today. That growth curve tells you something about where the economics are heading: companies aren’t adopting this out of altruism. It’s cheaper to run, easier to site (you don’t need to be near a massive water source), and it makes the ESG narrative cleaner.
Site selection is actually one of the underappreciated implications here. Traditional evaporative cooling systems anchored data centers to regions with abundant water access. Closed-loop and direct-to-chip systems loosen that constraint, opening up arid or water-stressed regions as viable locations. That’s a big deal for companies racing to build capacity in markets where land and power are available but water is not.
What investors and the industry should watch
The shift toward closed-loop cooling is real, but the industry’s transparency problem hasn’t been solved. There’s a growing call for data center operators to clearly distinguish between initial water fills and ongoing consumption, between water withdrawal and water that’s permanently removed from the local cycle. Those are very different numbers, and blending them together, whether intentionally or not, makes it harder to assess actual environmental impact.
For investors evaluating AI infrastructure plays, water efficiency is becoming a material factor. Facilities running advanced cooling systems face lower operational costs, fewer regulatory risks in water-stressed jurisdictions, and stronger positioning for government contracts that increasingly include sustainability requirements. Companies like Oracle that can credibly claim near-zero water usage have a competitive edge that extends beyond the environmental story.
The risk is that the industry’s rapid buildout outpaces its sustainability improvements. Training runs are getting larger. Inference demand is exploding. Even if each individual data center uses less water per kilowatt-hour, the total number of kilowatt-hours is growing faster than efficiency gains can offset. OpenAI’s clarification is a step toward better public understanding, but the math still has to work at scale, and the 2027 projections suggest it might not without a much broader adoption of next-generation cooling across the entire industry, not just flagship facilities.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·