Hello! You’re receiving this newsletter as one of 6,559 subscribers (+925 thank you to everyone who singed up, welcome.)
You’ll find the “Share the Newsletter” feature at the end of the current issue. Today’s newsletter contains 933 words, and reading it in full takes about 5 and a half minutes.
Your monthly ChaptGPT subscription costs $20. Maybe another $20 for Copilot. Perhaps a bit more for whatever else has quietly auto-renewed on your credit card. Not bad value, all things considered. But that's not the full bill.
Every time you generate an image, summarise a document, or ask an AI to write your passive-aggressive out-of-office reply, that request gets routed to a data centre, a warehouse full of servers running at intense computational load.
Those servers generate enormous heat. And the cheapest, most practical way to cool them is cycling water through the building to absorb heat from the servers and then venting it into the atmosphere through cooling towers, where most of it evaporates and is gone for good.
AI systems alone consumed somewhere between 300 and 750 billion litres of water last year. To put that in terms that actually mean something, New York City uses roughly 3.8 billion litres a day. So we're talking about AI drinking the equivalent of what a city of 8.5 million people uses over two and a half months, and that's before you count the water embedded in electricity generation itself.
A reader wrote to us after our recent newsletter on the global water crisis asking exactly this question: to what extent is AI making it worse? The answer, it turns out, is quite a lot. And the places feeling it most are the ones where supply was never built to handle this kind of demand.
Northern Virginia is the world's largest concentration of data centres. There, more than 300 facilities process nearly 70% of the world's digital information. And they are thirsty. In 2023, those facilities consumed close to 2 billion gallons of water, a 63% increase from 2019.
The demand has grown so fast that reclaimed water can no longer keep up, forcing the local water authority to increasingly redirect potable drinking water to cool servers instead. That's the same water that comes out of taps, flows into hospitals, and, until recently, nobody thought needed to compete with a server farm.
In summer, when droughts hit and rivers run low, data centres account for around 8% of total water consumption in the Washington metro area. By 2035, that figure is projected to reach 25%.
A similar story is playing out in Newton County, a modest rural county southeast of Atlanta, which is now on track to face a water deficit by 2030, in significant part because a single data centre moved in and the local water supply simply wasn't sized for that kind of industrial demand.
Across the world, of the nearly 9,000 data centres operating globally, around 600 are located in regions where average temperatures already exceed the optimal range for server operation. In 21 countries — including Singapore, Nigeria, Thailand, and the UAE — every single data centre sits in a climate that is technically too hot.
Singapore alone has 72 of them. The country is one of the most humid and densely populated places on Earth, essentially the opposite of what you'd design a data centre around, and yet it's the fifth biggest data centre market in the Asia-Pacific region and growing.
The more heat outside, the harder the cooling systems have to work, which means they use even more water, leaving less for the rest. As with most things in economics, when there is less of something to go around, it tends to become more expensive.
The electricity story follows the same pattern. US data centres consumed more than 4% of the country's total electricity last year, roughly equivalent to the entire annual demand of Pakistan. By 2030, that figure is projected to more than double. And someone has to pay for the grid upgrades required to support that kind of growth. Spoiler: it's not going to be Sam Altman.
The average US electricity bill could rise by around 8% by 2030 just from data centre demand, with some high-density markets looking at increases well beyond that. In parts of Virginia, electricity costs near data centres have already more than doubled over the last five years. These are households that never signed up for an AI subscription. But they're paying for it anyway.
So the costs of production get distributed outward — onto local infrastructure, drying water supplies, and household utility bills — while the value gets captured upstream. The people living near a data centre get higher bills and less water, while the people running the data centre get trillion-dollar market caps.
None of this means AI isn't valuable or that the trade-offs aren't sometimes worth making. But the $20 monthly subscription was never the real price. It was just the only part of the invoice you were shown.
The data centres behind the AI boom need chips. Billions of them. And roughly 90% of the world's most advanced semiconductors come from a single island of 23 million people sitting in one of the most geopolitically contested stretches of water on Earth.
Taiwan's semiconductor industry has grown so dominant, so fast, that the rest of the economy is starting to feel it. It's a pattern economists have a name for, Dutch disease — when one booming industry distorts an entire economy around it.
Taiwan's version comes with a twist: the product driving the boom is one the whole world depends on, which means the consequences of getting this wrong don't stop at the Taiwan Strait.
In our latest video, we get into how that dependence is reshaping the island from the inside out, and what it means for the people living inside the boom without benefiting from it.
