Skip to content
Slow Hiss or big bang? About the AI Bubble
ArticlesAI translation· Slow Hiss or big bang? About the AI Bubble · Teil 4· 24 July 2026

The 2026 Bubble Update

Anything that can go wrong will go wrong. Almost always.

by 9 min read

From Hiss to Crack: An Update on the AI Bubble

In December 2025, I wrote that the AI bubble would rather hiss slowly than burst with a bang. My argument back then: the canaries in the AI coal mine are singing so loudly that nobody will walk blindly off the cliff, and the physical limits of power, cooling and permits act as a natural brake on the inflow of capital.

At the last “AI@Southcross - Episode II” on 13 July, I briefly picked up the topic again as the “Trillion Dollar Bet”, describing what I consider a very unhealthy financial situation as worrying, and arguing that any AI strategy has to take it into account.

Now I'm sitting in front of the latest analyses and have to be honest: part of my last thesis needs an update. Not because the canaries have gone quiet. Because I underestimated how creatively you can build cages in which nobody hears the birds.

The Analysis: Everyone Sees the Bubble, Nobody Sees the Debt

Here's the elephant in the room, and by now it weighs 1.65 trillion US dollars: according to reports by Nikkei Asia, the “hidden” liabilities of the big US tech companies, packaged in leases, special purpose vehicles and the footnotes of SEC filings, now exceed their officially reported debt of around 1.35 trillion.

Meta has reportedly parked around 420 billion outside its actual balance sheet, almost three times the official figure. Oracle's lease obligations for data centres have grown thirtyfold in four years. The Hyperion data centre in Louisiana? Meta is the sole user but holds only 20% of the shares. That's how 27 billion in investment elegantly vanishes from the earnings report.

All of this is legal. All of it can be read in the footnotes. And that's exactly the problem: you can't see the trees for the forest.

The Paradox: Balance-Sheet Feng Shui

Let's call this phenomenon balance-sheet feng shui: the debt doesn't disappear, it just gets moved around until the room looks harmonious. The energy of the liability keeps flowing, only now through Blue Owl Capital, Fluidstack, Cipher Mining and whatever else they're called, instead of through your own balance sheet.

And this is where the real paradox of this cycle lies:

We are witnessing the most transparent bubble in history, financed by the most opaque structures since 2008.

My December thesis treated loud scepticism as a protective mechanism. But scepticism only works where it can look. The public debate rages over OpenAI's valuation and NVIDIA's share price, in other words over the visible part of the system. Meanwhile, the credit leverage has been shifted to where no earnings call ever takes place: into private credit, infrastructure funds and SPVs.

2008 taught us: the bubble itself isn't what's systemically dangerous. The danger is when nobody knows any more who actually owes what to whom.

The Mechanism: Deterministic Debt, Probabilistic Revenue

At the heart of the construct is a timing mismatch worth understanding:

A lease for a data centre is deterministic. The payment is due every month, for 15 to 20 years, whether the GPUs inside are glowing or gathering dust.

The revenue meant to cover those payments, on the other hand, is probabilistic, and in a remarkable cascade at that: the SPVs rely on the hyperscalers. The hyperscalers rely on demand from essentially two major customers, OpenAI and Anthropic. And those two rely on a continuous inflow of venture capital, because they themselves are unprofitable.

Add to that the circular financing: NVIDIA invests in neocloud providers and guarantees to take unused capacity off their hands. These guarantees serve as collateral for billion-dollar loans, which are then used to buy NVIDIA chips. If “revenue flowing back from your own investment” gives you a slight sense of déjà vu, you watched The Big Short closely.

The data also suggest the numbers don't add up: around 190 GW of capacity is planned, with estimates putting that at 15 times current demand. At the macroeconomic level, Goldman Sachs has so far found no significant link between AI adoption and productivity growth. So the bridge to the AI revolution keeps getting longer, without anyone seeing the other bank, let alone knowing where it is.

The Real-World Test: Who Actually Holds the Risk?

Let's make it concrete. A teacher in California pays into her pension fund, CalSTRS. CalSTRS holds stakes in Blue Owl and CoreWeave. Blue Owl finances Meta's data centre. CoreWeave's creditworthiness depends on NVIDIA guarantees and OpenAI contracts. OpenAI's ability to pay depends on the next funding round.

The teacher never bought a single NVIDIA share. She carries the risk anyway, four steps removed from the source, without knowing it. This is exactly the constellation the Bank for International Settlements (which, incidentally, was already warning back in 2003 about what would hit in 2008) has in mind when it warns of “growing financial vulnerabilities” and collateral pledged multiple times over. (BTW: Elon Musk has jumped on the pension fund train with SpaceX too, to cash in.)

And the crucial difference from 2008: back then, publicly held US government debt stood at 35% of GDP. Today it's around 100%. The fire extinguisher that prevented the worst in 2008 is pretty much empty this time.

Hiss or Bang? Four Possible Scenarios

My December forecast of a “slow hiss” hasn't turned out wrong (yet). It's still one of several plausible scenarios. Here's my current read:

Scenario 1: The controlled hiss (still possible, but more fragile).
The physical brakes hold, capital flows in stages, weak players die quietly, the big ones digest their overcapacity over years. This works as long as no single credit event triggers the SPV cascade. The catch: a hiss assumes the air escapes evenly. Experience shows that off-balance-sheet structures don't discharge evenly, though. They go all at once.

Scenario 2: The credit event.
A small gust of wind can set the whole house of cards wobbling: a funding round that fails to materialise for one of the two big buyers, a default at a neocloud provider, an SPV that can't make its payments. Rising prices for credit default swaps suggest that institutional players are already weighting this scenario more heavily than the stock market. If private credit is hit, pension funds and insurers are the next stop. At that point we're no longer talking about a “tech correction” but a “financial crisis”.

Scenario 3: Too strategic to fail.
Governments declare AI infrastructure a matter of national security and prop up the sector, empty coffers notwithstanding, via the printing press if need be. That would avert the crash and rebook it as inflation and public debt. Losses would be socialised, profits would stay privatised. Sound familiar? Exactly.

Scenario 4: The late payoff.
The productivity gains do arrive, just later and less evenly distributed than promised. Historically, not far-fetched: the internet eventually delivered on its dot-com promises too, just ten years and one crash later. In this scenario, the infrastructure outlives its first owners. Whoever bought fibre out of bankruptcy estates in 2003 did very well. Whoever laid it in 1999 did not.

What Does This Mean for Us?

For executives and technology leaders, this adds up to some uncomfortably concrete homework:

Take counterparty risk seriously. Anyone signing long-term contracts with AI vendors today should scrutinise their financing structure at least as closely as their benchmarks. A model that hallucinates brilliantly is annoying. A vendor that goes bust in the middle of your project is seriously expensive.

Build in the ability to exit. Multi-provider strategies and portable architectures cost efficiency today and buy you room to manoeuvre tomorrow. The first realistic step: an honest assessment of how long your own organisation could keep working without its primary AI vendor.

Decouple your business case from the hype. Whether the bubble bursts or hisses changes nothing about whether your specific AI use case creates value. If you can prove it does, you'll survive any market phase. If you're only investing out of “strategic FOMO”, you're helping to fund the overcapacity.

And Europe? Just Spectators with the Illusion of Full Coverage?

In Berlin, Paris or Brussels, it might be tempting to lean back: after all, this is an American bubble. The masked debt sits on US balance sheets, the SPVs in Delaware, the data centres in Louisiana. That, however, would be the same mindset that in 2007 held American mortgages to be an American problem, until IKB and the German Landesbanken suddenly had to explain why they were stuffed with subprime paper.

A burst would hit Europe through at least three channels:

The financial channel. European insurers, pension funds and asset managers have long been invested via private credit and infrastructure funds, often in the very structures the BIS is warning about. How deeply, nobody knows exactly. That's not reassuring. That's the problem.

The real-economy channel. European industry depends on the US tech stack. If vendors consolidate, prices explode or services get shut down, it hits every European organisation that built its AI strategy on two or three American providers. One possible consequence would be an involuntary brake on digitalisation, right in the middle of the transformation process.

The economic cycle channel. A US recession following an AI crash would hit an export-dependent European economy that's already skating on thin ice.

And that's exactly why it would be negligent for policymakers not to prepare for this scenario. Not because the crash is certain to come. Nobody knows that. But because preparation is cheap and surprise is expensive: stress tests for European institutions' exposure to AI private credit. An honest look at which critical infrastructure and administrative processes now depend on US AI vendors. And yes, also the uncomfortable question of whether European sovereignty in compute is a nice-to-have or an insurance policy. Whoever only starts thinking about resilience after the bang is buying fire insurance while the roof is already in flames.

The Uncomfortable Bottom Line

Which scenario will play out is still open. But one thing has become clear, and with it I'm correcting myself: the loud scepticism I counted as a safety net in December only protects against risks you can see. Balance-sheet feng shui has moved the relevant risks precisely to where the canaries don't sing.

And in systems like this, Murphy's Law almost always applies: anything that can go wrong will go wrong, just not where everyone is looking, but in the footnote on page 247 of the SEC filing nobody has read. Murphy isn't a pessimist, he's a statistician with enough patience (haha). With 1.65 trillion in hidden liabilities spread across hundreds of opaque structures, the question isn't whether a link in the chain will snap somewhere. It's only which one goes first, and who's standing behind it in the cascade.

So the decisive question is no longer whether the AI bubble implodes or goes psssssht. It's this:

If the profits of this boom go to the top ten per cent of shareholders, the risks sit in the pension funds of ordinary people, and Europe, four steps removed, still has a seat at the table: who is actually preparing while preparation is still cheap?

As long as the answer is “nobody”, Murphy has an easy job.


Live long and prosper 😉🖖


Soundtrack 😉

▶Listen