Latest / The Cloud Business Podcast with Fexingo: AWS, Azure, GCP, and Enterprise Infrastructure / The Carbon Cost of Cloud: Microsoft and Google's Cooling Water Problem
Transcript
- Lucas: So Microsoft published its 2025 sustainability report two weeks ago, and one number jumped out at me — the company consumed 2.4 billion liters of water across its global data center fleet in fiscal 2025. That's up 34 percent year-over-year. Luna: Thirty-four percent — that's massive. Is that all from cooling? Lucas: Almost entirely. Data centers generate a huge amount of heat, and the most efficient way to remove it, especially in warmer climates, is evaporative cooling. You run water over cooling towers, the heat evaporates some of it, and the rest is cycled back. But a portion is lost to evaporation and blowdown — that's the water that's discharged and replaced. Luna: So the 2.4 billion liters is mostly water that evaporates or gets treated and sent down the drain. It's not recycled back. Lucas: Right. And Microsoft's own target is to be 'water positive' by 2030 — meaning they replenish more water than they consume in stressed regions. But the consumption keeps climbing because they're building new data centers faster than they can retrofit older ones with more efficient cooling or alternative technologies. Luna: How do the other hyperscalers compare? Google is usually the most transparent on environmental metrics. Lucas: Google reports a different number — they focus on 'total water consumption' across their data centers and offices. In 2024, they reported about 5.6 billion liters consumed globally. But they also claim to have replenished 120 percent of the water they used in water-stressed regions through watershed restoration projects. Luna: So Google is actually net-positive in stressed basins, while Microsoft is still climbing toward that goal. What about AWS? Lucas: AWS is the least transparent of the three. Amazon doesn't break out water consumption for AWS separately from the rest of the company. They publish a 'water utilization effectiveness' metric — WUE — which measures water used per kilowatt-hour of IT load. But they don't give you an absolute gallon number for the cloud business. Luna: Interesting. So if you're an enterprise customer trying to pick a cloud provider based on water impact, you can't really compare them directly. Lucas: Exactly. And that's becoming a real issue because data center water use is drawing regulatory attention. Northern Virginia — which is the largest data center market in the world — has seen local utilities impose moratoriums on new connections due to water supply limits. Prince William County put a pause on new data center permits last year partly because of water concerns. Luna: Wait — so the water constraint is actually affecting where hyperscalers can build now? Lucas: It is. Singapore had a moratorium on new data center builds from 2019 to 2022 largely because of energy and water constraints. When they lifted it, they imposed strict efficiency requirements — including a maximum water usage effectiveness of 2.0 cubic meters per megawatt-hour of IT load. That's roughly the industry average, but any new facility has to prove it can meet that. Luna: So water is becoming a site-selection criterion on par with energy cost and latency. What are the practical alternatives to evaporative cooling? Lucas: There are two main alternatives. One is air-based cooling — using outside air when the climate is cool enough, which works well in places like the Pacific Northwest or Northern Europe. The other is liquid immersion cooling, where servers are submerged in a non-conductive dielectric fluid that absorbs heat directly. That fluid is then cooled through a closed-loop system with minimal water loss. Luna: Immersion cooling — that sounds like it could be a game-changer for water consumption. Is it being deployed at scale? Lucas: It's still early. Microsoft tested immersion cooling in a production environment in 2021 and said it reduced power consumption by up to 15 percent, but they haven't rolled it out broadly. The challenge is retrofitting existing data centers — immersion tanks require different rack configurations, different maintenance procedures, and the fluids themselves are expensive. New builds can incorporate it, but the hyperscalers are building so fast that most new capacity still uses conventional cooling. Luna: So the water problem is going to get worse before it gets better. But I'm curious — is water the new carbon? Are we going to see water offsets like we have carbon offsets? Lucas: We're starting to. Google's water replenishment program is essentially a form of offset — they fund projects that restore watersheds or improve water access in the same basins where they consume water. Microsoft does something similar with their 'water replenishment' projects. But offsets for water are trickier than carbon because water is local. Carbon is global — a ton saved anywhere helps. But a liter saved in Northern Virginia doesn't help a stressed basin in Arizona. So you have to match the offset geographically. Luna: That makes sense. So the real solution has to be reducing consumption per workload, not just offsetting. What kind of improvement in water efficiency have the hyperscalers shown? Lucas: They've gotten much better. Microsoft's water utilization effectiveness — WUE — has improved from about 1.2 liters per kilowatt-hour in 2019 to under 0.6 liters per kilowatt-hour today. Google claims an average WUE of 0.4 globally. So the efficiency per unit of compute is improving. But the total compute demand is growing so fast — AI workloads especially — that absolute water consumption is still going up. Luna: Right — because AI training runs generate a lot of heat. A single GPT-4 class training run could consume as much energy as a small town for a month. That heat has to go somewhere. Lucas: Exactly. And inference — the actual use of AI models — is even more distributed and can be harder to cool efficiently because it's spread across many smaller servers rather than concentrated in one giant cluster. Luna: So what's your takeaway for the enterprise customer listening? How should they think about this when choosing a cloud provider? Lucas: I'd say two things. First, if your company has its own sustainability targets, ask your cloud provider for their water efficiency data by region. AWS won't give you an absolute number, but they will provide WUE data for each availability zone. Second, consider the location of your workloads — if you're running batch AI training, you might choose a region with naturally cool air or access to non-potable water, like Microsoft's data centers in Sweden that use treated wastewater for cooling. Luna: That's a great point — using non-potable water eliminates the competition with local drinking water supplies. Are any providers doing that at scale? Lucas: Google is piloting a project in the Netherlands that uses reclaimed water from a nearby treatment plant. Microsoft's Swedish facility uses treated wastewater. But it's still niche. Most data centers still draw from municipal water supplies because it's cheaper and more reliable than building a parallel pipeline for reclaimed water. Luna: So we're at an inflection point. Water availability is beginning to constrain data center growth, and the hyperscalers are scrambling to find alternatives. But the demand from AI is relentless. Lucas: That's the tension. I think the next five years will see a real shift toward closed-loop cooling systems — either air-based or liquid immersion — that drastically cut water use. But the economics have to work, and right now, water is still so cheap that there's little financial incentive to innovate. That's starting to change as communities push back. Luna: So water may end up being the thing that forces the industry to adopt those technologies faster than carbon ever did. Lucas: I think that's right. Carbon is a global externality — hard to price locally. Water is a local externality that shows up in your utility bill and your neighbor's well. The pressure is going to come from communities, not just shareholders. Luna: Okay, so we've covered the problem, the numbers, and the potential solutions. One last thing — what's the one data point every cloud buyer should know about their provider's water use? Lucas: Ask for the water utilization effectiveness by region. If they won't share it, that's a red flag. And for the regions where you're deploying the most compute, ask whether that region is classified as water-stressed by the World Resources Institute. If it is, and your provider isn't actively replenishing water there, you might want to reconsider. Luna: Great advice. Water in the cloud isn't going away as a topic. Thanks, Lucas. Lucas: Thanks, Luna. Next time, we'll look at how the big three are handling the nuclear option — literally — with small modular reactors for data center power.