It’s not dot-com. It’s 2008

You know that ridiculously overused, misattributed, kitschy saying of how history doesn’t repeat itself, but it tends to rhyme?

Well… I hear music coming through the walls and it’s the greatest hits album that rocked the charts back in 2008. And not in a good way.

The default comparison pits AI against the dot-comSidenote: Which, arguably, has valid arguments in its favor.Few examples: overvalued equities, tech whose revenue will supposedly arrive years after the capex, and a related, massive infrastructure buildout.However, I think dot-com was mostly an equity valuation story.AI financing is more of a credit story, which hits closer to 2008. End of sidenote. bust of the early ‘00s, but I think that misses the mark by a few years. There are a lot more parallels between current AI-financing and the Global Financial Crisis.

Both on and off the balance sheets.

Now, just to be clear, I’m nowhere near saying that the result will be the same or even similar. Not predicting a crash. It’s just when the same boxes get checked, I think we should pay attention.

Let’s get into it, shall we?

The leverage is real, just not where you’re looking

Institutions borrowing insane amounts of money to increase their exposure to assets is as old as the markets themselves. If you are not leveraged at least two generations going forward, are you even a financially responsible market participant?

Hell, even retail traders use leverage more and more often these days, as painfully evidenced just last month in South KoreaSidenote: Retail investors piled into 2x-3x ETFs and leveraged products to maximise gains on their semiconductor darlings SK Hynix and Samsung.When the unwind came... it was painful. Citi estimates Korean retail losses on leveraged ETFs at whopping $38.7bn.Go watch the reaction videos on YouTube. End of sidenote.. It’s a risky way to play the game, but greed is eternal.

Now, you use leverage when you expect the returns to be fabulous, and you want them to be even more fabulous. Leopold Aschenbrenner can attest to this.

Leading up to 2008, banks, hedge funds, broker-dealers and other institutions borrowed massively to buy, among other things, various subprime mortgage backed securitiesSidenote: I love how underwriters dubbed some of these NINJA: no income, no job, no assets. Gallows humor.Anyway, it would be shortsighted to say that ONLY subprime caused the crisis.The problem was structural: being short commercial paper on long debt. End of sidenote. (MBS) with ratings that were as good as made up.

Asset-to-equity ratios at big investment banks were routinely above 30:1Sidenote: To be exact, biggest 5 investment banks at the time:Bear Stearns 34:1Morgan Stanley 33:1Merrill Lynch 32:1Lehman Brothers 31:1Goldman Sachs 26:1 End of sidenote.. Not exactly pocket change.

We all know how that went: when the subprime part of subprime mortgages reared its ugly head, the dominoes started falling loudly. Cue crisis.

So what’s the hot “asset” that investors/traders can’t get enough of today? The easy answer: anything AI related, basically.

Hyperscaler stocks, hyperscaler bonds, semis, LLM makers, data centers, both public and private companies, the whole AI supply chain.

How much money is riding on all of this?

AI-related debt issuance was, per Goldman Sachs, 322 billion dollars in all of 2025. By mid-July 2026 it had already passed 489 billion. So we’re gaining pace, rapidly.

But the reported numbers aren’t what’s checking the box. Balance sheets are squeaky clean, net debt roughly one times EBITDA for most. In 2008 the thirty-to-one was published and auditable. Today you have to go looking.

Fortunately, Goldman Sachs did our homework for us and in early August 2026 they published an analysis on how much in lease commitments are hyperscalersSidenote: Who are these, anyway?In layman's terms: the handful of companies big enough to build this stuff themselves.Think Microsoft, Alphabet, Amazon, Meta, Oracle. End of sidenote. running OFF balance sheets.

Care to guess?

Upwards of $1 trillion USD. Wall Street Journal said it’s actually at least three times that (they cast a wider net and also counted purchase obligations).

What’s scary is not the amount (which, even by itself, is eye popping), it’s the way HOW this debt is structured.

The obligations are held by various off-book “vehicles” and investment structures and not by the hyperscalers directlySidenote: I'm simplifying here purposefully, but the basic description and functional mechanism hold.Just to create these financing schemes takes an army of five-figures-per-month accountants, bankers and lawyers.Methinks that's also a risk. It obfuscates cash flows, which hinders potential damage control.Hopefully we never find out how much. End of sidenote.. Don’t be fooled, though, they ARE on the hook for eventually honoring them, they just don’t show up on their balance sheets.

The eerie part is that this was exactly the modus operandi pre-crisis, where massive debt was “hidden” by offloading it on these so-called structured investment vehicles. I mean, Big Money Inc. does it all the time, but it becomes more prevalent especially during financing sprees.

This creates an unnecessary fog of war, because if AI assets, god-forbid, turn sour, it’s hard to gauge who holds what exactly.

Same as in 2008.

Massive FOMO

I’ve read somewhere that you know a theme is hot when your barber starts talking about it, or something like that.

Now my barber doesn’t talk about AI at all, for which I am eternally grateful, but you don’t need to get a haircut to know that it’s been the thang of the past few years.

You may be aware that trends have this nasty side-effect of creating FOMO, or fear of missing out. Meaning the infectious belief that if you don’t jump on the bandwagon right now, you might miss out on an opportunity of a lifetime.

So a lot of people do jump. Often blindly and with leverage, because in the markets, once in a lifetime opportunities come roughly twice a year.

Let’s travel back 18 years and see if that was the case also back then.

By post-crisis analyses, the 25 biggest subprime lenders wrote about 72% of all subprime loans between 2005 and 2007. The list runs through Citigroup, Wells Fargo, HSBC and Countrywide, not fringe mortgage shops.

No one wanted to miss out on easy money, or, as Chuck Prince, then CEO of Citigroup famously put itSidenote: He later testified to the Financial Crisis Inquiry Commission that it was not about subprime FOMO, rather about leveraged buyout lending. End of sidenote.: “As long as the music is playing, you’ve got to get up and dance. We’re still dancing.”

It’s no secret that it rolls this way today, also. Everyone wants a piece of the AI cake: retail traders, banks, quants, pension funds, hell, even cybercriminalsSidenote: Per Interpol: AI-enhanced fraud is 4.5 times more profitable than traditional methods, and agentic systems can now run a full fraud campaign from reconnaissance to ransom demand.Talk about artificial intelligence enhancing productivity. End of sidenote..

FOMO and leverage walk hand in hand and together they form a self-reinforcing spiral.

A lot of players piling in on the same thing, all borrowing heavily to gain as much exposure as possible, thus acting as a force multiplier on… whatever comes out of this.

I honestly don’t know if the AI spending supercycle ends in a boom or a bust, but it’s gonna be big.

Securitize everything, until…

Let’s have a little lecture on creating assets out of thin air.

In 2008 it worked something like this:

1. take thousands of American mortgages,
2. dump them together in a single security/asset,
3. chop it into multiple slices or, if you will, tranches,
4. get a rating agency to slap an AAA rating on the top one.

Boom, your mortgage backed Frankenstein is complete and you name him… collateralized debt obligation, or CDO.

Then simply sell these en masse to pension funds, municipalities and insurers, who are allowed to only buy safe assets.

Why is it considered safe, beside that dubious AAA rating, you ask?

Because the assumption is that US house prices don’t fall. But, shockingly, they then do and along with them the whole house of cards.

What does the 2026 remaster of this evergreen jam sound like?

Well, data centre securitizations have been around since at least 2018, their issuance in the first half of 2026 alone is around $17 billion USD, up 29% year on year and on pace for a record.

But on August 10 the level got pushed up a notch or twenty.

Nvidia announced it was teaming up with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to round up over $500 billion USD (in signed intent) over time and turn computeSidenote: Compute is, simply put, processing power.Chips running calculations and carrying out tasks, measured and sold by the hour, kinda like electricity.Compute is "produced" by graphics processing units (GPUs) in data centres.Data centres are essentially sheds or warehouses full of racks of GPUs. End of sidenote. into an investable asset.

The craziest thing was Blackstone’s president Jon Gray then going on CNBC and literally saying out loud that it is similar to how mortgage lenders underwrite homes. You cannot make this up, I mean, I almost applauded.

So what is the assumption this time around?

That a GPU is still worth something in a few years, thus worth investing into AND lending against.

How many “few” years?

Amazon says five years. Meta five and half. Microsoft says six years.

A one year difference is nothing, you might say. But a longer GPU shelf life means a smaller yearly write-down, that means fatter reported profit, which means the debt looks more affordable. And, what do you know, they choose the number themselves.

Remember, these are people talking up their own book.

But the prom king crown goes to the Securities and Exchange Commission (SEC), the markets watchdog.

Late July 2026 it issued staff guidance that data centre securitizations aren’t really asset-backed securities at all. They are actually operating assets, not a loan that pays itself down, such as the infamous mortgage backed securities.

The consequence of this being that the issuers of this data centre paper do not have to keep a slice of them. They could create them and sell them.

It bypasses the exact goddamn rule that was created because of 2008, written specifically to prevent institutions from packing subprime garbage and walking away clean.

Let me repeat that for emphasis: one of the most important guardrails just got powersawed offSidenote: I need to be fair here and say this: data center operators already keep around 30% of these deals, because otherwise they wouldn't get the top ratings on these securities.And, data centers produce genuine revenue, unlike a foreclosed house somewhere in Chippewa Falls, Wisconsin. End of sidenote. the fastest growing theme on Wall Street.

But I’m sure this time it’s different.

…everyone’s connected

So the “crisis” part of the Global Financial Crisis spread like a wildfire because of how interconnected the world’s financial institutions were (and still are, because we humans hilariously never learn).

They borrow money from each other, they buy, lend and sell each other securities, they fund and co-finance a lot of what keeps the global economic gears spinning. When a banker in San Francisco sneezes, one in Tokyo needs to wipe his nose.

So you can imagine that when one bank’s balance sheet got tainted by toxic assets, it didn’t stay contained too long. No, brother, suddenly everyone was wary of what nasty surprises the others were hiding and international capital flowsSidenote: There is something called interbank repo.Banks lend each other huge sums overnight, secured against bonds. It's the plumbing under everything and literally how banks finance their operations. And its clients operations and yours and mine, too.In 2008 nobody could tell whose bonds were good because of the toxic assets.Banks genuinely stopped lending each other money for days. End of sidenote. almost ground to a halt. Chaos and madness ensued and the rest is history.

Beam us back to the present, Scotty.

Since banks lend heavily into AI, their inter-connectedness is also a big part of the potential risk down the line. But it’s not just the banks.

Financial Times published an interesting piece on how hyperscalers themselves borrowing massive amounts of money through (but not limited to) bond issuance actually drives up borrowing costs all over the place.

For example: Amazon and Alphabet have both issued bonds in the Swiss franc and British pound, among other currencies. Due to their immense size and financial backstops they are able to “out-issue” other companies in these foreign credit markets, who therefore have trouble finding buyers, have to offer cheaper prices/higher yields, thus ending up with worse borrowing terms.

An American behemoth needs dough and thus a Swiss window maker gets a crappy loan. Capitalism in its purest form.

Anyway, this is called crowding out and it is not only company-to-company, it’s also happening, by Bank of America’s economists, company-to-government.

Yes, you guessed right, company bonds are also able to potentially crowd out government bonds, including US treasuries.

Shoot!

I’m not sure letters on a computer or phone screen can convey how significant this is. Massive interest in AI is actually making it more expensiveSidenote: Bank of America: by roughly 0.3 percentage points added to the 10-year bond rates.However, this also includes general corporate bond sales and mortgage-backed issuance, not AI borrowing alone. End of sidenote. for the US government, the biggest damn economy in the world, to borrow money.

Were all that invested capital to prove un-returnable, the reverberations would cause a chain reaction from which no one would be really protected.

Have your napkins ready.

Wait… what about the differences?

Sure, let’s balance this out.

The whole point of stuffing all this money into AI is that one day the returnsSidenote: Real, tangible ones, such as productivity boost and economy growth, not just return on equity invested. End of sidenote. will make it all worthwhile. Now, I need to be honest here, the scale of total investmentSidenote: Roughly $3 trillion of global data centre investment through 2028 per Morgan Stanley. End of sidenote. makes me a bit sceptical if that is achievable even in the next decade or so… but, okay, I’ll bite my tongue here.

Because there’s no denying AI already is generating returns.

Perhaps not at the scale imagined, but some companies are reporting decreased costs and/or increased productivity through the use of artificial intelligence tools. And we are, reportedly, still in the early phases of integrating these into our economies.

The adoption is widespread with both consumers and companiesSidenote: A NBER survey of corporate executives (link here) shows that AI adoption is widespread but uneven.Larger firms are generally further along, and productivity gains are measurable but modest, expected to accelerate. End of sidenote., so a very important factor - demand - is there and is, admittedly, strong and growing.

This differentiates AI from MBS, as these were only good as long as property prices were rising, people were paying their mortgages and even then there were early defaulters.

With AI, the collateral actually makes money day-to-day. Data centres generate revenue, GPUs do too. A foreclosed house earns nothing, instead it costs you money every month it sits empty.

The financial backing of it all is also, admittedly, healthier this time around. Subprime borrowers often had no tangible income, while today hyperscalers still cough up roughly two thirds of the buildout expenses without borrowingSidenote: The problem lies further down the chain with the neoclouds and labs burning cash like they are on a mission. End of sidenote.. And they all have non-AI related pipelines too, anyway.

Banks themselves, while still exposed, are not neck deep like 18 years ago. Back then they held the toxic paper on and off balance sheets, often with crazy leverage above 30-to-1Sidenote: Rest in pieces, Bear Stearns. End of sidenote., and when it inevitably went bad, the whole financial plumbing broke.

Today their exposure to AI is reportedly around 0.8%Sidenote: Caveat here, that's direct exposure.Chicago explicitly flags additional indirect exposure through bank-to-non-bank lending. End of sidenote. of their total assets, at least per the Chicago Fed.

AI debt, while huge, mostly consists of long-dated fixed-rate bonds and locked-up private credit. Unlike short-term commercial paper in 2008, it cannot really be withdrawnSidenote: This was one of the reasons why 2008 was as bad as it was.Lenders simply stopped rolling the debt used to finance purchase of CDOs and this initiated a fire-sale of toxic assets into a market... which had no buyers, basically.It's like you bought a stock worth $100 on Monday and at Tuesday open it had a $0 value.But multiply it by billions. End of sidenote. overnight, although it can lose value.

Aaand, that about wraps it up.

Those are the most significant differences I can think of, though there are more if you care to look for them.

So what to take away from all this?

Nothing dramatic. Anyone who tells you they know how this whole shebang ends is guessing, however confidently they smash those keyboard letters.

I think it’s worth holding onto both the similarities and differences. The first ones are not necessarily damning and the second aren’t a saving grace.

Vagueness aside, if you are invested in today’s markets, chances are you’ve got a stake in AI too. Directly or indirectly.

You simply can’t avoid the biggest trade of the last 3 years (and running).

I mean, you can hold all cash, sure.

But you wouldn’t want to miss out, would ya?

And now, I hope you’ll excuse me. The music is still playing and I’m aching to dance.