A note before we start.
I work with AI tools every single day. I thought I knew this space pretty well. Then I went down a rabbit hole.. research papers, water utility reports, chip engineering docs, environmental filings and I came out the other side genuinely unsettled. Not in a dramatic way. In that quiet, uncomfortable way where you realise something big is happening and nobody in your circle is really talking about it. So I’m writing this. Not to scare anyone. Just to lay it out plainly, the way I’d explain it to a friend over chai at MTR.
The numbers here are real. The chemistry is real. The damage is real. And the hard truth is: we are already late to this conversation.
Okay, picture this.
It’s 2 AM. Somewhere outside E-city (for an example), past the toll booth, past the last petrol pump, there’s a building the size of three cricket grounds sitting behind a fence. No name on the gate. A lot of cameras. It hums constantly, this low, steady industrial sound that the nearby village/households has just accepted as part of life now. Inside that building, thousands of chips are running at full speed. Generating images. Finishing sentences. Summarising documents. Answering questions. Every second, without a break.
That building is a data center. And every time you open ChatGPT, or ask Gemini something, or use any AI tool, one of those buildings wakes up on your behalf. It draws power. It generates enormous heat. And it uses water. A lot of water.
We’ve all been so excited about what AI can do that we forgot to ask what it costs to keep it running. That’s what this is about.
Let’s start with the chips, because the physics here is actually fascinating.
The GPUs running AI.. chips like NVIDIA’s H100, which powers most of the big AI systems you use are not like the chip in your phone. Each one draws between 300 and 700 watts of power. One chip. A single server holds maybe eight of these chips. A rack holds several servers. A data center holds thousands of racks. Now add all that up.
What happens when you push that much electricity through silicon? Heat. Enormous, relentless heat.
At 70 degrees Celsius, the chip is doing its job fine. Uncomfortable to touch, but happy.
At 85 degrees, it starts slowing itself down on purpose. Engineers call this thermal throttling. The chip is basically saying “bhai, I need a minute.”
At 110 degrees and above, things start breaking. Permanently. The insulating layers inside the chip degrade. The tiny solder joints holding everything together start warping. In bad cases, the chip just dies. This is what people in the industry quietly call silicon melt or junction burnout and it’s not a metaphor, it’s an actual failure mode with an actual replacement cost.
But here’s the really nasty bit. Silicon gets less efficient as it heats up. Less efficient means it produces more heat. More heat means less efficient. It’s a loop.. engineers call it thermal runaway and the only thing stopping it is the cooling system working perfectly, all the time, forever.
And unlike your laptop which gets warm and then you close it and it cools down, these data center chips never get a break. They run continuously, under some of the most demanding workloads ever designed. Training a large AI model can run for weeks without stopping.
The chips aren’t the only things suffering either. The circuit boards around them, the power regulators, the capacitors all of it gets cooked slowly. Data center operators have documented swollen capacitors, boards that literally peel apart from heat stress, connectors degrading. The infrastructure was designed for regular cloud computing, which runs maybe 3 to 5 times lighter than AI workloads. Nobody fully updated the plan when AI took over.
The scariest part? These servers don’t dramatically fail when they overheat. They quietly throttle. They accumulate tiny invisible cracks in solder joints that no monitoring tool catches. They slowly degrade until one day, mid-conversation, mid-task, the chip just stops responding. The server was reporting everything as fine, right until it wasn’t.
So how do you cool a building full of machines running like this?
Water. Shocking amounts of water.
The most common method is called evaporative cooling. Cold water absorbs the server heat, gets warm, turns into vapour, and floats away into the sky. Fresh water refills what was lost. About 80% of the water that enters this system never comes back.. it evaporates and is gone.
Now here are the numbers that stopped me cold.
One Google data center in Iowa, USA consumed roughly one billion gallons of water in 2024. That is one building, one year, one billion gallons. Researchers calculated that this is enough to cover all residential water needs in the entire state of Iowa for five days. Just the people. Not the farms, not the industries.. just people drinking, cooking, bathing. Five full days. One building.
Now think about Bengaluru for a second. We know what a water crisis looks like. We lived through 2023 when parts of the city were getting water tankers every three days. The Cauvery has been a political flashpoint for decades. We understand scarcity in a way that, say, someone in Seattle maybe doesn’t.
Now imagine that kind of consumption a billion gallons a year, sitting on top of whatever groundwater exists in the area. And then imagine dozens of buildings like that in the same region.
Across the US alone, data centers consumed 17 billion gallons of water directly in 2023, according to Lawrence Berkeley National Laboratory. And that’s just the direct figure water used in cooling. The indirect figure: water used to generate the electricity that powers the buildings, is estimated to be 12 times higher.
That means when you ask an AI to write you a 100-word message, you’ve consumed roughly 519 millilitres of water. One water bottle, per small paragraph. That stat comes from researchers at UC Riverside, and I sat with it for a while before it really landed.
By 2028, US data center water use is expected to be 4 times what it is today. The Great Lakes, which supply drinking water to over 40 million people are being actively targeted for new data center campuses because of their sheer scale. The same freshwater that a continent depends on is now in competition with server cooling.
This is not a future problem. This is a now problem.
Here’s the part that genuinely surprised me, and I think it’ll surprise you too.
That water that goes through the cooling system, it isn’t just water by the time it comes out.
Plain water would destroy the pipes within weeks. It would grow Legionella bacteria inside cooling towers, same bacteria that causes Legionnaires’ disease, a serious lung infection. It would leave mineral deposits on every surface until nothing moved properly. So operators treat the water with a long list of chemicals. (Sounds like science class, but it gotta be serious to know)
Biocides like chlorine and glutaraldehyde to kill biological growth. Corrosion inhibitors to stop the pipes rusting. Scale inhibitors to prevent mineral buildup. pH adjusting chemicals to keep the water chemistry stable. Antifreeze compounds in some systems. And in the newer immersion cooling setups where entire servers are literally submerged in liquid dielectric fluids, some of which contain what are called PFAS compounds.
PFAS. The nickname is “forever chemicals.” They don’t break down in the environment. They accumulate in soil, in water sources, in the bodies of living things, and they just stay there. Essentially forever. Some are linked to cancer, hormone disruption, immune system effects. The science on this is still developing but the direction is not reassuring.
The water that gets periodically drained out of these cooling systems called blowdown, carries all of this out into the world. Most operators treat it before releasing it. Most. Not all. And what “treated” means varies by country, by state, by municipality. There’s no global standard. There’s no strong enforcement in many places. There’s a lot of goodwill being relied upon from companies whose main job is keeping servers online, not protecting local groundwater.
In countries like India, where environmental monitoring infrastructure is still being built, and where regulations are unevenly applied across states, this is a particularly sharp concern. A data center going up near a water-stressed area in AP (iykyk) or other states operates in a very different accountability environment than one in Germany or Japan.
I want to give credit
Where it’s due, because the engineering world is genuinely trying to fix this.
Singapore made a sharp move here. The city-state ran a moratorium on new data centers from 2019 to 2022 specifically to figure out the sustainability problem before letting the industry grow further. When they reopened approvals, they came with strict energy and water efficiency requirements. A small country making a firm call that’s worth noting.
Iceland has built a legitimate data center industry almost entirely around free cooling. It’s cold there most of the year, geothermal energy is abundant, and renewable electricity is close to 100%. Companies like Verne Global run GPU clusters in Iceland that use almost no water for cooling and run on clean energy. The geography does the work.
In the Netherlands, Microsoft has a data center that returns waste heat to a local greenhouse complex, which uses it to grow vegetables. The heat that would have been dumped into the atmosphere is instead growing food. This is the kind of thinking that scales.
DeepMind built an AI system to manage cooling at Google’s data centers and cut cooling energy consumption by 40%. The machines are optimising the problem they helped create. There’s something poetic about that.
Direct-to-chip liquid cooling, where cold plates sit directly on the GPU surface reduces water usage by up to 95% compared to traditional evaporative cooling. NVIDIA and HPE are shipping hardware with this built in now. This is not experimental, it’s production-ready.
Immersion cooling, where entire servers are submerged in non-conductive fluid, eliminates evaporative water loss entirely. No fans, no cooling towers, 30 to 40% lower energy use. Companies like GRC and Submer are doing this at real scale today.
The solutions are real and they work. The barrier is not engineering. It’s the pace of adoption, and adoption pace is shaped by regulation and public awareness.
The Local-First Approach
Now here’s the argument that I think is especially relevant for us in India.
We don’t actually need all this compute concentrated in giant facilities thousands of kilometres from where the work is happening. That model massive centralised data campuses in specific regions was an accident of early internet economics, not a permanent truth about how computation has to work.
A local-first model puts smaller compute capacity closer to where it’s actually being used. Edge nodes at the city level cut data transmission distance by 80 to 95%. Local renewable energy, rooftop solar, battery storage can power these smaller nodes without touching a coal grid at all. Cooler ambient temperatures in cities like Pune, Shillong, or Dharamshala could eliminate most cooling requirements for significant parts of the year. Rainwater harvesting can handle residual cooling needs. And if the facility is owned or governed locally, think a municipal technology cooperative then the entity using your city’s water supply is actually accountable to your city.
India’s data center market is one of the fastest growing in Asia. Mumbai, Chennai, Hyderabad, and Bengaluru are already major hubs. The decisions being made right now about where to build, how to cool, and what standards to require will shape the water and energy reality of these cities for the next 20 to 30 years. We are not watching this from the outside. We are in it.
The EU passed regulation in 2024 requiring data centers to publicly report their Water Usage Effectiveness making water consumption visible and comparable for the first time. India’s Data Center Policy is still evolving. This is exactly the window where pushing for similar requirements actually changes something.
The carbon picture makes this worse
Yes! you heard it right. It deserves its own moment.
Global data center electricity demand is heading toward 945 terawatt hours by 2030. To put that in perspective that is roughly equal to Japan’s entire annual electricity consumption. Most of this power comes from grids that still run on fossil fuels.
AI systems alone could be responsible for 79.7 million tonnes of CO2 in 2025, according to analysis published in ScienceDirect. And that figure doesn’t include manufacturing building a single NVIDIA H100 chip generates around 143 kilograms of CO2 before it even ships. It doesn’t include the chemicals, the facility construction, the shipping of equipment.
India’s electricity grid is still heavily coal-dependent, though that’s changing quickly. A data center running on India’s current grid mix has a significantly higher carbon footprint per unit of compute than one running on Scandinavian renewable energy. This is not a criticism.. it’s a fact about where we are in our energy transition. But it does mean that the localisation argument and the renewable energy argument are the same argument. Local solar-powered compute nodes are better on every axis: water, carbon, accountability, and data sovereignty.
What We Can Actually Do? And I Mean Really Do
Let me be straight with you. I’m not here to guilt trip anyone. We’re all using these tools, including me, and that’s not going to stop. The goal isn’t to quit AI cold turkey and go live in a forest. The goal is to be smarter about how we use it, and to push for better options to exist in the first place.
So here’s what actually moves the needle.
Use AI for things that are genuinely worth the compute.
This sounds obvious but honestly we’ve all done it asked an AI to write a two-line reply to a WhatsApp message, or generated five variations of something we already knew we were going to write ourselves. Every one of those requests hit a real server, drew real power, and used real water. That doesn’t mean stop using AI. It means stop using it the way we mindlessly scroll — reflexively, without thinking about whether it’s actually the right tool for the job. Use it when it saves you serious time. Use it when it genuinely does something you can’t. Don’t use it to avoid thinking for thirty seconds.
Try local models for everyday tasks.
This is the one I’d really push. Tools like Ollama let you run genuinely capable AI models directly on your own laptop or desktop.. no internet connection needed, no data going to a server farm in Virginia, no water being evaporated in Iowa on your behalf. It’s free, it runs on a mid-range laptop, and for everyday tasks like summarising a document, drafting a message, explaining a concept, or writing basic code.. it’s more than capable. LM Studio is another option with a clean interface if you prefer something more visual. If you’re on a Mac with Apple Silicon, MLX models run surprisingly fast and Apple has been quietly building out local AI infrastructure that is genuinely efficient.
Are local models as powerful as GPT-4 or Gemini Ultra? No. But for maybe 60 to 70 percent of what most of us actually use AI for day to day, they’re completely fine. And the gap is closing fast.
Choose Indian and regional products where they exist and are good enough.
This matters more than people realise. Sarvam AI is building genuinely useful models trained on Indian languages like Hindi, Tamil, Telugu, Kannada, Bengali and more. If you’re working on something that involves Indian language content, using Sarvam instead of defaulting to an American model means the compute potentially stays closer to home, the product gets better with use, and you’re supporting an ecosystem that actually understands the Indian context. Krutrim, built by Ola, is another one worth watching early days but the intent is clearly to build infrastructure that serves this market specifically.
For productivity tools, Zoho is the obvious example most people already know but underuse. Built in Chennai, servers in India, serious enterprise-grade software across CRM, email, docs, sheets, projects almost everything the Google Workspace or Microsoft 365 stack does. I’m not saying it’s perfect for every use case. I’m saying it’s worth checking before you default to the American option purely out of habit.
BharatGPT, developed by IIT Bombay in partnership with Persistent Systems, is building language models specifically for Indian languages and Indian contexts. These projects need users and feedback to improve. Using them, even imperfectly, is how they get better.
For voice AI and regional language tools, Vernacular.ai and CoRover are doing interesting work in Indic language AI, the kind of work that a San Francisco lab has no business incentive to prioritise. If your product or your company touches regional Indian markets, these are worth a serious look.
Be deliberate about which cloud you’re using and where it actually is.
Most people pick AWS or Google Cloud or Azure without thinking much about region. But region matters both for latency and for the energy and water profile of the facility serving your workload. AWS Mumbai, Google Cloud Mumbai, and Azure Pune are all running in India, which means your data isn’t travelling halfway around the planet and back. For Indian workloads, defaulting to an Indian region isn’t just better for sovereignty it’s meaningfully better for the energy and emissions picture too.
NxtGen Datacenter and ESDS are Indian data center operators with facilities in multiple cities. For businesses that have flexibility in where they host, these deserve a serious look not out of blind nationalism but out of genuine accountability. An operator whose facility is in Navi Mumbai is answerable to local environmental regulators, subject to Indian law, and at least in theory, easier to hold accountable than one whose primary governance happens in Seattle.
Push the apps and services you use to be transparent.
When was the last time you asked a product you pay for where it runs its AI features and what the energy and water footprint looks like? Most companies have never been asked. Start asking. On social media, in app reviews, in feedback forms. “Where does your AI inference run and do you publish sustainability data for it?” is a totally reasonable question and most product teams have no answer prepared. The fact that they don’t have an answer prepared is itself information worth surfacing publicly.
Support the policy conversation even if you’re not a policy person.
iSPIRT and NASSCOM are both actively engaged in India’s data center and AI policy space. The Ministry of Electronics and IT is developing frameworks right now. State governments are signing land deals with hyperscalers. None of this is being decided in secret.. it’s being decided in rooms where the loudest voices are the industry players who showed up. The more awareness there is among regular technical people about water usage, energy efficiency requirements, and local accountability standards, the more that conversation shifts.
You don’t have to become an activist. You just have to know enough to have the conversation when it comes up.
I want to end with something I genuinely believe.
The chip is not the villain. An H100 GPU is an extraordinary object, billions of transistors doing things that would have looked like science fiction when I was in school. The technology is real and it’s valuable and some of what it enables is genuinely good for the world.
But it runs somewhere. There is always a building, a power line, a cooling system, a water source. The server that answered your last AI query has a physical address. It sits above an actual water table. It draws from an actual grid. It operates inside an actual regulatory environment, or sometimes the absence of one.
We built the internet on the comfortable fiction that it existed “in the cloud”, somewhere weightless and consequence-free. We’re running the same story with AI and it’s not true this time either.
The path forward is not complicated. Distribute the compute. Power it with renewables. Cool it without wasting water. Build it close to the people it serves. Hold the operators accountable to those same people. The technology for all of this already exists.
It just needs more of us paying attention.
So, pay attention.
AI Doesn’t Care About Your Water Crisis was originally published in Age of Awareness on Medium, where people are continuing the conversation by highlighting and responding to this story.