Not Quite a Bubble, but a Reckoning

AI data centers are being built faster than the need for them. The chips age fast. The debt does not.

Jeff Gray · 10 November 2026 · Control Systems Security

Rows of racks that age in five or six years, inside a building financed for decades. The clock is the story.

Any figure for 2027 or later is MODELED by us from public data. It is not a forecast from any firm.

Executive summary

Everyone is building AI data centers as if demand will never stop. But the technology keeps getting cheaper, faster than anyone is building. Every year the same AI work takes less hardware, less storage and less power. At some point, the buildings outrun the need.

Our models say that point arrives around 2028. By the mid 2030s, much of what's being built today could sit half empty.

That alone wouldn't be a disaster. The trouble is the clock. The chips wear out in five or six years. The buildings and the loans that paid for them run for decades. When the chips go stale, the bills keep coming.

This isn't the end of AI. It's what happened to fiber optics in 2001: too much built too fast, painful write-downs, and cheap capacity for whoever is left standing. Not quite a bubble. A reckoning.

1. Everyone wants to be in the room

My father used to say, "If you're not in the room, you're not in the deal."

Right now the room is an AI data center. Picture a windowless building the size of several football fields, humming with racks of chips, fed by its own substation. Tech giants, chip makers, utilities, lenders and bond buyers are all crowding in. Nobody wants to be left in the hallway.

Everyone in that room is making the same bet: that demand for computing will keep rising faster than technology makes it cheaper. Not for a year or two, but long enough to pay off buildings and loans that run for decades.

I wanted to know if that bet holds. So we built two simple models from public data, one for storage and one for AI computing, and followed the money.

What we found is not a bubble that pops overnight. It is a slower collision between chips that age fast and debt that does not.

2. The bet

Start with the size of it. Amazon, Microsoft, Alphabet and Meta alone expect to spend more than $700 billion on buildings and equipment in 2026, by their own guidance.

Google's parent company shows how fast this moved. This year it expects to spend about six times what it spent in 2022.

The forecasters keep raising their numbers. Dell'Oro now expects data center spending to pass $3 trillion by 2030, nearly double its view in January.

Where does it go? McKinsey says most goes to chips and hardware, and the rest to power, cooling, land and concrete. Remember that split. The chips wear out fastest. The power and buildings get financed the longest.

3. Storage, the canary

Storage is where the cracks show first, because the waste is already measured.

You have heard the world is drowning in data. Mostly, it is not. Nearly everything created is streamed, viewed and gone. IDC found that less than 2% of the data created in 2020 was still kept a year later.

What does get kept is mostly copies. One customer file becomes a backup, a disaster recovery copy, a test copy and an analytics copy. IDC found in 2013 that more than 60% of business disk space was copies of something stored somewhere else.

Much of the rest is clutter. Veritas found that only about one stored byte in seven actually mattered to the business.

Meanwhile flash chips pack in about a third more data every year, and each new video format squeezes the same picture into fewer bits.

So we asked: if you subtract what compression, deduplication and fewer copies save each year, when does real need start growing slower than capacity?

Storage need growth versus installed capacity growth, three scenarios. All years after 2026 are MODELED.

In our early scenario, that turn comes next year, in 2027 (MODELED). In our middle scenario, it comes in 2034 (MODELED). If stored data grows at the top of IDC's range, it never comes, and the world stays short of storage.

Two warnings keep this honest. Today's market is short: TrendForce expects flash shortages in 2026, and Seagate says demand could outrun what the industry can make. And storage is small money next to AI spending.

Storage is the canary. It is not the mine.

4. AI compute, the main event

AI computing is where the money is, and where things get cheaper fastest.

Two forces work at once. The software gets smarter about doing the same job with less work, and the chips do more work for the same electricity. Epoch AI finds a given level of AI gets about 13 times cheaper every year.

Picture it this way: a question that cost a dollar to answer last year costs less than a dime this year.

The building has kept pace. The amount of NVIDIA AI computing installed in the world has more than doubled every year since 2019.

So here is the race. On one side is the buildout, adding capacity at a planned pace. On the other is real need: how much useful AI work people want, divided by how efficient the software has become. While need grows faster, chips stay scarce. When the buildout grows faster, the buildings start emptying.

Nobody can measure yet how fast demand will cool from today's pace, so we ran three versions.

AI compute need growth versus installed capacity growth, and implied utilization. All years after 2026 are MODELED.

In our middle scenario, the buildout overtakes the need in 2028 (MODELED). From there, the share of capacity actually needed slides downhill, and by 2034 less than 60% of it is in use (MODELED).

In the fast version the turn comes sooner. In the slow version demand keeps compounding and there is no bubble at all.

The exact year is not the point. The point is that demand does not have to collapse for the buildings to outrun it. It only has to slow from astonishing to merely very fast.

5. The math that breaks

Two pieces of arithmetic carry this story.

The revenue gap. In 2024 David Cahn of Sequoia asked how much AI customers must earn to pay for all these chips. His method starts with NVIDIA's sales, adds the building and power around the chips, then adds a profit for whoever rents them out. On NVIDIA's latest quarter, the answer is about $1.4 trillion a year (MODELED, our arithmetic).

The most famous AI company, OpenAI, reports about $50 billion a year. That is a big number that looks small next to the bill.

The fair caveat: Google, Amazon, Microsoft and Meta do not report how much of their cloud and ad money comes from AI, so the real gap is smaller than it looks, by an amount nobody outside knows.

The clock. This is the part that turns a slowdown into a bust. Think of a fleet of delivery vans parked inside a warehouse with a 30 year mortgage. The vans wear out in six years. The mortgage does not.

That is the data center. The chips are written off in five or six years. The buildings run 15 to 40 years, and the bonds and leases 10 to 40. Microsoft announced on its July 2026 earnings call that it will stretch its own buildings to 25 years, a change not yet in its annual filing.

When the chips go stale, the mortgage is still there.

Filed useful lives, and last full year cash versus capex. Source: company 10-Ks.

6. The debt layer

So is this a debt bubble? Not exactly, and the difference matters.

The giants still mostly pay cash. Amazon, Google, Meta and Microsoft each earned enough last year to cover their building budgets. But the cushion is thin. Amazon had about $11 billion left over.

But they are borrowing like never before. This year these four plus Oracle sold more than $370 billion in high grade corporate bonds. Not all of it is AI debt, but it is a step change.

Oracle is the stressed edge. It spent far more on buildings and chips last year than its business brought in, ending about $24 billion short. Its credit rating sits one notch above junk, and the cost of insuring its debt hit a record in October.

Some of it is off the books. Meta's Hyperion campus in Louisiana sits in a joint venture it owns only a fifth of, so about $27 billion of building cost stays off its balance sheet. Meta still guarantees much of the value if things go wrong.

Some of it is backed by the chips themselves. CoreWeave, which rents out GPUs, borrowed $8.5 billion against its computing gear and a customer contract. That ties repayment to the asset with the shortest life.

Michael Burry has estimated big tech will understate depreciation by $176 billion from 2026 to 2028 by stretching how long servers are said to last. That is his estimate, not independently confirmed. The filings cut both ways: Meta stretched its servers and booked a profit boost, while Amazon shortened some and took a hit.

So the thesis is not "it's all borrowed." It is that a growing layer of bonds, leases, side deals and GPU loans sits on top of assets with a short clock. If the buildings empty before that layer is paid down, lenders feel it, not just shareholders.

7. The case against

Here is the strongest argument that I am wrong, and it is happening right now. 2026 is a shortage, not a glut.

Renting a top NVIDIA chip for a year got about 40% more expensive between October 2025 and March 2026, according to SemiAnalysis. Old chips held their value because demand surged.

Google says it is short of capacity. The amount of AI text it processes each month grew about seven times from May 2025 to May 2026.

Then there is the Jevons effect: when something gets cheaper, people use more of it. Researchers find the cost of running the newest, biggest models is still rising every year, because each new model thinks harder and longer.

If AI agents and reasoning keep multiplying demand, our slow scenario wins. No bubble. Just higher prices. That is a real possibility, and it is why I call this a reckoning, not a prediction of collapse.

8. What the reckoning looks like

If it comes, it probably looks like 2001, not like AI failing.

In the late 1990s, companies laid fiber optic cable as fast as they could dig. They were right that the internet would change everything, and wrong about how fast and what people would pay. The builders took the losses. The survivors bought the capacity cheap and built the next decade on it.

The AI version would rhyme. People keep using more AI. Chips get written off faster than planned. Some buildings and power contracts sit half used. The most stretched players, spending more than they earn and borrowing against chips, feel it first.

And whoever still has cash when it settles gets the cheapest computing in history.

9. What OT and grid owners should do now

If you run a utility or a plant, this is your story too. You are being asked to commit power lines, substations and grid connections for these buildings. Three things to do now:

1. Be skeptical of load forecasts. Ask what happens to a new connection if the customer's building is half empty by the mid 2030s. Write contracts so the stranded cost does not land on your ratepayers or your capital plan.
2. Build security into every new connection from day one. A rushed hookup means new control points, new remote access paths and new vendors. Security bolted on later costs more and covers less.
3. Cut duplicate data. If most stored bytes are copies, most stored bytes are extra places to break in. Fewer copies means less exposure and a smaller storage bill at the same time.

A footnote on quantum, since people ask: it does not change this decade's math. IBM is aiming for its first large error-corrected machine in 2029, and Google has not put a date on its own. Neither is built for the kind of math AI runs on.

At Cyborama we help OT teams turn questions like these into practical risk decisions.

← Back to all articles