The Machine That Eats Money, Power And Water (AI Bubble)
Category: Investing & Markets
Nothing here is financial advice - I'm not an adviser, and I don't know your circumstances. This is an attempt to connect three stories that are usually told separately, because I think they're the same story.
Three stories, one machine
For months now I've written about three things as if they were separate.
I wrote about water, how the UN declared "global water bankruptcy," how the planet is drawing down its freshwater savings faster than they refill.
I wrote about energy, how artificial intelligence is straining electricity grids so hard that power, not chips, has become the thing limiting how fast it can grow.
And I've circled, repeatedly, around markets, how concentration, mispricing, and crowd euphoria create the conditions for spectacular mistakes.
Today I want to do something different. I want to show you that these three stories are not separate at all. They meet in one place, an enormous, windowless building in a field somewhere, and understanding how they connect is, I think, one of the most important pieces of thinking an ordinary person can do about the economy right now.
Because there is a single machine at the centre of the modern economy that eats three things at once: staggering amounts of money, staggering amounts of power, and, the part almost nobody talks about, staggering amounts of water.
Let's follow all three, and then ask the uncomfortable question at the end: what happens if the money runs out first?
The bubble question, asked properly
Let's start with the word everyone's dancing around. Is AI a bubble?
First, what a bubble actually is, because the word gets thrown around lazily. A bubble isn't just "prices going up a lot." A bubble is when the price of something detaches from the underlying reality that's supposed to justify it, when people are paying not for what a thing is worth, but for what they're convinced someone else will pay for it later. Tulips in 1637. Dot-com stocks in 1999.
Houses in 2007.
The honest answer to "is AI a bubble" is: not obviously, not yet, but it's developing the exact features that have preceded every historic bubble, and one of them should genuinely worry you.
Let me give you the bull case first, because intellectual honesty demands it. Unlike the dot-com companies of 1999, many of which had no revenue and no path to any, the big AI players are generating real money. The chair of the US Federal Reserve has pointed out that these are real firms with real output, not vapour. Corporate cash flows are vastly healthier than they were in 1999.
There is genuine demand: businesses and consumers are paying real money for these tools. This is not nothing. Anyone telling you it's definitely a bubble and definitely about to pop is overclaiming.
Now the part that should worry you.
The money that goes round in a circle
There's a phrase you need to know, because it's the single most important concept for understanding the risk in AI right now: circular financing.
Here's how it works, stripped to its bones.
A chip-maker (let's call it what it is, Nvidia) sells the chips that AI runs on. Its customers are companies like OpenAI, which need to buy tens of billions of dollars of those chips. But those customers are losing enormous amounts of money and can't easily afford the chips.
So the chip-maker does something clever, and slightly alarming: it invests in its own customer. It gives OpenAI money, reportedly up to $100 billion, which OpenAI then uses to... buy chips from the chip-maker.
Read that again. The seller is funding the buyer to buy from the seller.
Money leaves Nvidia, travels to OpenAI, and comes straight back to Nvidia as revenue, revenue that makes Nvidia look like it's selling enormous quantities of chips into roaring demand, which pushes its share price up, which gives it more firepower to fund more customers to buy more chips.
And it's not a two-way loop, it's a web. Microsoft owns a large chunk of OpenAI and is one of Nvidia's biggest customers. OpenAI partners with a cloud company, CoreWeave, that Nvidia also owns a stake in, and Nvidia has agreed to buy whatever capacity CoreWeave can't sell.
Wall Street analysts have estimated that for every $10 billion Nvidia "invests" in OpenAI, it sees around $35 billion in chip purchases come back. By 2026, the total value of these interlocking deals ran past $800 billion.
Here's why this matters, and why it's not just clever business. Circular financing makes demand look bigger than it really is. When companies buy each other's products with each other's money, you can no longer tell how much of the "demand" is real customers wanting a real product, and how much is just money spinning around a closed loop, creating the appearance of a booming market.
This is not a new trick. It was a hallmark of the telecoms bubble and the dot-com crash, companies financing their own customers to inflate perceived growth, and when those bubbles burst, the circular deals didn't cushion the fall. They accelerated it. Even Sam Altman, the head of OpenAI, has said out loud that "someone is going to lose a phenomenal amount of money." CNBC's Jim Cramer, no permabear, said the arrangements revived his memories of 2000: "I don't want the sequel."
And underneath the financial engineering, the raw economics are sobering. OpenAI is reportedly on track to lose around $14 billion in 2026, nearly triple its losses the year before, while projecting profitability years away and hoping the demand curve keeps climbing steeply enough to justify it all.
That's the money story. Now here's where it stops being abstract, and starts touching the physical world, and the two earlier stories I promised to connect.
The power bill of a thought
Everything I've just described, every chip, every model, every chatbot reply, has to run somewhere physical. And "somewhere physical" means a data centre: a vast building full of computers that
consume electricity and, crucially, turn nearly all of it into heat.
I've written about the energy side before, so I'll be brief on the headline: AI's electricity demand has grown so fast that access to power has become the binding constraint on the whole industry. New AI data centre campuses are being designed to draw as much electricity as a large nuclear power station produces, for a single site. Forecasts for how much of America's electricity data centres will consume keep getting revised upward, and the strain is already turning up on ordinary people's power bills.
But electricity is only half of the physical bill. The other half is the one almost nobody has connected to any of this, and it's the one that brings us all the way back to where this series started.
The thirst, where all three stories meet
Here is the fact that reframed this entire subject for me.
Those computers get hot. And the cheapest, most common way to cool them down is to evaporate water. Vast quantities of it.
Not metaphorical water. Actual, clean, drinkable freshwater, poured through cooling systems and boiled off into the sky.
Let me give you the numbers, because they are genuinely staggering:
A single large data centre can consume up to 5 million gallons of water a day, the daily water use of a town of up to 50,000 people.
Google's data centres alone used around 7.8 billion gallons of water in a recent year, up sharply from a couple of years before.
In Loudoun County, Virginia - which hosts more data centres than anywhere on Earth, data-centre water use grew more than 250% in four years, reaching 1.6 billion gallons and approaching 10% of the entire county's water consumption.
Roughly 80% of the water a data centre withdraws simply evaporates, it doesn't come back.
And industry projections suggest hyperscale data centres could be drinking up to 33 billion gallons a year by 2028.
Now connect that to the story I opened this whole series with. The UN has declared global water bankruptcy, the planet is already drawing down its freshwater faster than it refills. And into that shortage, we are now building, at frantic speed, a fleet of machines whose cooling systems evaporate drinking water by the billions of gallons.
Worse: because of where the land, power and tax incentives are cheapest, an enormous number of these data centres are being built in exactly the wrong places, hot, dry regions that are already water-stressed. More than 160 new AI data centres have gone up in the US in recent years in areas with high competition for scarce water. In Arizona, they compete with farmers for water during an active Colorado River shortage. In rural Georgia, residents living near a data centre have reported their own water running contaminated or dry. One resident's summary of the situation was five words long: "I can't drink the water."
Sit with the full shape of this for a moment, because it's the whole point of the piece.
We have a machine that:
Runs on money that is increasingly spinning in a circle, making demand look bigger than it may be.
Consumes electricity on a scale that's straining grids and raising ordinary people's bills.
Evaporates drinking water by the billions of gallons, often in places that have already declared water emergencies.
Three of my three stories, the bubble, the power, the bankruptcy, are not three stories. They are three appetites of the same machine.
Why this is the connection that matters
Here's the insight I most want you to take away, because it's the kind of pattern-recognition that this whole publication exists to build.
Most people analyse AI as a technology question: is it clever, will it get cleverer, which company has the best model? That's the glamorous, visible layer, and it's the layer where almost everyone stops.
But every AI model rests on a physical foundation of chips, power and water, and that foundation rests, in turn, on a financial foundation of money that must keep flowing for the whole thing to stand up. And those two foundations are connected in a way that creates a genuinely dangerous feedback loop.
Think about it. The enormous spending on power and water infrastructure, the gigawatt data centres, the cooling systems, the grid upgrades - is being justified by the expectation of enormous future AI demand. And a chunk of that "demand" is being manufactured by circular financing. So if the financial bubble deflates, if the money stops spinning round the circle, if lenders demand proof of profit that isn't there - then the demand forecast collapses.
And you are left with a landscape of half-built, power-hungry, water-guzzling data centres that were sized for a future that didn't arrive.
Economists call that a stranded asset, something built at enormous cost that becomes worthless because the world it was built for never showed up. The dot-com bust left behind thousands of miles of unused fibre-optic cable, laid in the confident expectation of demand that took another decade to materialise.
An AI bust could leave behind something with a much heavier physical footprint: enormous facilities that drew down real communities' real water and real power, to serve demand that turned out to be partly an accounting illusion.
That's the risk that ties everything together. Not "will the robot get smart." But "what have we physically consumed, and what have we financially promised, on the strength of a demand number that some very smart people are quietly warning may be inflated?"
What I actually want you to take away
Not a prediction. I genuinely don't know whether AI is a bubble that pops in 2026, deflates slowly, or turns out to be justified by demand that really does arrive. Anyone who tells you they know is guessing with confidence, which is the most dangerous kind of guessing.
What I want to give you is the lens, and it's the same one every time: follow a thing all the way down to the physical world, and all the way back up to who's paying for it.
The people who only look at the top layer, how clever is the AI, will be genuinely surprised if this ever unwinds. The people who understood that the clever AI sits on top of chips, which sit on top of power and water, which sit on top of a tower of circular financing that has to keep spinning, those people won't be surprised. They'll have seen the whole machine, not just the shiny front of it.
And here's the thought I'll leave you with, because it's the one that stays with me. Whatever happens to the share prices, the water is already gone. The gallons evaporated in Arizona and Georgia and Virginia don't come back if the bubble pops. That's the real asymmetry buried in all of this: the financial risk is reversible, markets fall and eventually recover, but the physical resources, drawn down from communities that were already short, are spent for good.
Three appetites. One machine. It's worth understanding all three before you form a view on any one of them.
References
INSEAD Knowledge - "Are We in an AI Bubble?": https://knowledge.insead.edu/economics-finance/are-we-ai-bubble
FXEmpire - "Can the AI Bubble Survive 2026 Credit Tightening and Reality Checks?": https://www.fxempire.com/forecasts/article/ai-market-faces-2026-test-as-credit-stress-and-valuations-peak-1569164
CNBC - "Jim Cramer warns AI's circular financing frenzy echoes the dot-com bubble": https://www.cnbc.com/2026/07/27/jim-cramer-warns-ai-circular-financing-echoes-dot-com-bubble.html
Bloomberg - "AI Circular Deals: How Microsoft, OpenAI and Nvidia Keep Paying Each Other": https://www.bloomberg.com/graphics/2026-ai-circular-deals/
EESI - "Data Centers and Water Consumption": https://www.eesi.org/articles/view/data-centers-and-water-consumption
Bloomberg - "The AI Boom Is Draining Water From the Areas That Need It Most": https://www.bloomberg.com/graphics/2025-ai-impacts-data-centers-water-data/
WaterVerge - "Data Centers Are Drinking Your Water" (Loudoun County figures): https://www.waterverge.com/news/data-centers-ai-water-consumption-2026/
UNU-INWEH - "Global Water Bankruptcy": https://unu.edu/inweh/collection/global-water-bankruptcy
IEA - "Energy demand from AI": https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
Comments