Everyone keeps watching Nvidia.
That’s probably not where I’d look first.
The more interesting question is where the money financing the AI buildout comes from.
Because the AI boom has become much more than companies spending their own cash on servers.
It’s turning into a financing chain.
Hyperscalers borrow.
AI companies raise money.
Private credit funds provide financing.
Data-center developers sign enormous leases.
Chip companies sell the equipment.
Then the infrastructure and contracts help support higher valuations, which make it easier to raise more money.
And that money goes right back into more AI infrastructure.
That’s the part that should make people uncomfortable.
Oracle is probably the clearest example.
Oracle’s cloud infrastructure revenue just grew 93%.
Sounds fantastic.
But Oracle also spent roughly $55.7 billion on capex in fiscal 2026, up from $21.2 billion the year before.
Free cash flow was negative $23.7 billion.
And Oracle is planning roughly $70 billion of capex next year.
At the same time, Oracle’s remaining performance obligations exploded to $638 billion, up 363%.
That sounds like an enormous demand signal.
But look at what has to happen before that backlog becomes cash.
Oracle has to build the data centers.
Buy the chips.
Secure power.
Finance the construction.
Operate the infrastructure.
And collect the revenue over years.
The market is treating that $638 billion as evidence of future earnings.
The credit market is asking a different question:
How are you going to finance all of this?
Oracle’s credit default swaps have climbed to multi-year highs, according to recent market analysis. Reuters has also reported that Oracle is approaching the lower edge of investment-grade credit territory while making one of the most aggressive AI infrastructure bets in the industry.
That’s a much more interesting signal than another AI company announcing record demand.
Because if the financing gets more expensive, the economics change.
And this is where the Reddit discussion goes deeper.
There is a theory circulating that part of the AI boom has developed a circular financing structure.
Private capital finances AI companies and infrastructure.
Those companies spend the money on Nvidia hardware.
Nvidia’s extraordinary revenue growth reinforces the valuation of the AI ecosystem.
Those valuations support more financing.
That financing creates more demand for Nvidia.
And the cycle repeats.
I’m not saying this proves some giant illegal scheme.
It doesn’t.
But you don’t need fraud for a circular financing structure to become dangerous.
You just need the money to slow down.
And there are already signs that credit investors are paying closer attention.
CoreWeave is heavily dependent on a small number of customers, while its credit spreads have reflected growing concern.
Oracle’s credit risk has risen.
Banks are becoming more selective around some AI-related financing.
And private credit is now a huge part of the financial system.
That’s the part I think the market is underestimating.
The AI bubble doesn’t necessarily need AI demand to disappear.
It only needs the cost and availability of financing to change.
Imagine AI revenue keeps growing.
But not fast enough to support the infrastructure being built.
Then companies have to raise more debt.
Debt costs more.
Private lenders demand higher returns.
Banks become more cautious.
Data-center projects get delayed.
Chip orders get pushed out.
Hyperscalers reduce capex.
Then Nvidia’s growth slows.
Then valuations fall.
Then raising money gets harder.
And suddenly the thing that looked like a demand problem becomes a credit problem.
That’s why the $5 trillion calculation in the original post is so interesting.
If the existing AI capital base really requires something like $2.4 trillion of annual revenue to justify itself, then the industry needs roughly six times today’s hyperscaler cloud revenue.
That requires extraordinary growth.
But there’s another problem hiding underneath that calculation.
The price of AI is falling.
If token prices fall roughly 25%, the industry needs about 33% more volume just to keep revenue flat.
So the system needs explosive volume growth at exactly the moment competition is making each unit of AI cheaper.
That’s not impossible.
But it creates an enormous burden on demand.
And demand isn’t the only constraint.
Power is becoming one too.
Some data-center projects are waiting years for grid connections.
So you can have a bizarre situation where companies have already committed billions to infrastructure that may not be fully usable when they expected.
Meanwhile the financing costs continue.
The depreciation continues.
The leases continue.
The debt continues.
This is where the 2027 problem becomes important.
Capex boosts economic activity while the infrastructure is being built.
But once the spending rate stops accelerating, that boost disappears.
The depreciation from yesterday’s spending doesn’t disappear with it.
So if 2026 really is the peak of the AI construction boom, you could enter 2027 with:
less incremental capex contribution
while simultaneously getting
more depreciation
more interest expense
more lease obligations
and potentially
slower AI revenue growth.
That is a very different problem from “AI stocks are expensive.”
It’s a cash-flow problem.
And it can spread.
Maybe the first losses sit with AI companies.
Then lenders.
Then private credit.
Then insurers and pension funds holding those assets.
Then the companies supplying the infrastructure.
Then the broader market.
One Reddit discussion made an especially interesting observation: unlike 2000 and 2008, much of this risk is visible before the collapse.
We can see the spending.
We can see the debt.
We can see the leases.
We can see the private credit exposure.
We can see Oracle’s negative free cash flow.
We can see the enormous commitments.
The strange part is that everyone can see it and the machine keeps accelerating anyway.
That’s why I don’t think the most useful question is:
“When will the AI bubble burst?”
Nobody knows.
The better question is:
“What breaks first if the financing loop slows down?”
Because if AI demand merely grows slower than expected, that’s manageable.
But if slower demand causes financing to tighten, and tighter financing causes capex to fall, and falling capex hits chip sales, data centers, private credit and valuations at the same time…
Then you’re no longer talking about an overpriced technology sector.
You’re talking about a financial system built around an investment boom beginning to unwind.
And that could look very different from 2000.
It might not be one spectacular crash.
It could be years of projects getting cancelled.
Data centers sitting unfinished.
Private credit funds restricting withdrawals.
AI companies refinancing at much higher rates.
Hyperscalers cutting capex.
Nvidia orders slowing.
And trillions of dollars of promised future spending quietly disappearing.
The scary part isn’t that the AI boom needs to be fake.
It only needs the money behind it to stop flowing fast enough.