Stop Asking Whether AI Is a Bubble. Start Reading the Lease Agreements.

Questo articolo non è ancora disponibile in italiano. Viene mostrata la versione inglese.

On 6 April 2026, Jamie Dimon told JPMorgan shareholders that AI spending “is not a speculative bubble; rather, it will deliver significant benefits.” Speaking on a podcast on 30 July, the same man said the payoff would “definitely not” arrive on the timetable people expect, and that the risks “are probably bigger than other people think.”

Nothing broke in between. The numbers just got bigger.

That is the trouble with the bubble question: two defensible answers, both available on the same morning, neither useful. The popular fix is to split the market into layers, applications, model makers, data centres, then ask which gives way first. Better. Still not enough.

The three layers will not fail on their own schedules, because they are stitched together by lease contracts and residual value guarantees. Watch the stitching rather than the layers and you get a date to watch, which is more than the bubble debate has ever offered.

Three layers, three completely different balance sheets

The layers really are different. The application layer, every product built on somebody else’s model, is capital starved. In the first quarter of 2026, three deals took 172 billion dollars of global AI funding, 67.3 percent of the total, according to PitchBook. The other 1,543 deals shared 83.5 billion.

The model layer has the opposite problem. Anthropic raised 65 billion dollars in May at a 965 billion dollar valuation, with run rate revenue that had “crossed 47 billion dollars”, meaning current revenue stretched across a year rather than money already banked. That kind of company does not fail on a repayment date. It fails, if it fails, at a funding round.

The infrastructure layer is the only one running on borrowed money at scale. From the Bank of England’s July 2026 Financial Stability Report: the five AI hyperscalers were 3 percent of outstanding US investment grade corporate debt at the end of 2025, but “over 15% of year-to-date issuance” by early May. Private credit financing AI investment went from 9 percent in 2024 to 34 percent in 2025, on an OECD estimate the same report cites.

Three funding models, three clocks. So far the layer view holds.

Then you read the contracts

On 28 July, Meta announced a data centre venture in El Paso with BlackRock: roughly 14 billion dollars, one gigawatt, online in 2028. BlackRock holds 80 percent and brings about 4.9 billion in equity plus 12.5 billion in debt.

Then the details that matter. The lease has a four year base term with four optional four year extensions. And Meta provides a residual value guarantee with a threshold of around 13 billion dollars.

Read that slowly. A 14 billion dollar building, a four year lease, and a promise from the tenant that it will still be worth roughly 13 billion when the lease ends. That is not renting. It sits much closer to a bond that never appears on Meta’s balance sheet.

Meta is the template, not the outlier. Moody’s counted 662 billion dollars of data centre lease obligations across Amazon, Meta, Alphabet, Microsoft and Oracle in February 2026 that had not yet commenced and so appear on nobody’s books. Initial lease terms have fallen from a historical 10 to 15 years to 4 to 6. The Bank for International Settlements named the pattern in March: “shadow borrowing”, meaning “obligations that are economically akin to debt but largely reside outside corporate balance sheets.”

Why the stitching beats the layers

Because it comes with dates and numbers. A layer never fails on a Tuesday. But a guarantee has a threshold, a lease has an end date, and a depreciation schedule has an assumed useful life, the accounting guess about how long a thing stays valuable. All three are published, and all three are moving the same way.

Microsoft told investors on 29 July that from financial year 2027 it will depreciate data centres and buildings over 25 years instead of 15. Fed governor Michael Barr noted in February that chips “have historically been depreciated over three years” but that “some firms have stretched the depreciation of AI chips to five years or more.” Each is a bet that expensive equipment holds its value longer than people used to assume. That can be entirely justified. It is still a lever, and the direction it moves tells you what is under strain.

The live test is CoreWeave, the largest company that buys GPUs and rents them out by the hour. For the quarter ending June 2026 it reported 2.575 billion dollars of revenue, a 626 million dollar net loss, and 640 million dollars of interest expense, double the year before. Interest now costs it more than it loses, and its debt maturities cluster in 2026 to 2028. If the stitching frays, it frays there first, and it will show up in a refinancing announcement.

The strongest case against this

Two careful teams read this market in 2026 and reached opposite conclusions. Man Group’s February analysis puts the fragility in infrastructure, because “capex is accelerating while revenue growth is stalling.” A May paper by Wang and Chen finds the reverse: hyperscalers and chipmakers show “weaker bubble signals after controlling for realized revenue growth”, while applications have the shakiest foundation.

The doom case for applications is weaker than the mood suggests. ICONIQ’s 2026 survey of around 300 software leaders found AI gross margins rising, 41 percent in 2024 to 52 percent in 2026, and companies now use 3.1 model providers on average.

But the strongest objection is the simplest. None of this matters if demand keeps outrunning supply. Andy Jassy said Amazon will spend about 220 billion dollars in cash capex in 2026 and will still “not have enough capacity to meet all the demand.” If that holds through 2027, no guarantee is ever tested. I cannot rule that out, and I am not predicting a crash. Lease terms are a stress gauge, not a prophecy. But a gauge you can read beats an argument nobody can settle.

What this looks like from Europe

Europe’s constraint is physical rather than financial, which changes the risk you carry.

Grid connection queues in the main data centre markets, Frankfurt, London, Amsterdam, Paris and Dublin, run seven to ten years, up to thirteen in the worst spots. Dublin has an effective moratorium until 2028, the Netherlands and Frankfurt until 2030. Electricity for mid sized business customers varies by more than a factor of three inside the EU, from about 0.075 euro per kilowatt hour in Finland to 0.255 in Ireland (Eurostat, second half of 2025).

Put the EU’s InvestAI plan, 20 billion euros over several years, next to Amazon’s 220 billion dollars in a single year, and the shape is clear. Europe will not be where the leverage sits: a real disadvantage for building things, and an accidental shelter if the stitching tears. We argued something related about where European AI money goes in Europe’s AI problem is not the AI Act, it’s the second cheque.

Europeans feel American financing decisions in price. OpenAI charges a “10% uplift” for regional data residency endpoints on models released on or after 5 March 2026, so keeping data in the EU carries its own surcharge.

What this means for you

For most people, nothing changes this week.

If you are curious: the price per token, the chunk of text a model reads or writes, has fallen a long way. GPT-4 launched at 30 dollars per million input tokens, the current mid tier costs 1 dollar. The sticker price is not where the pressure shows. Usage limits are.

If you build on models: the average company already uses three providers. Copy them. Portable prompts and data cost a little effort now and buy the option to move later. If you need EU data residency, budget that 10 percent on purpose rather than finding it on an invoice.

What would change my mind

1. A hyperscaler renegotiating, writing down or paying out on a residual value guarantee, or a data centre venture failing to place its debt. That would be the thesis working.
2. Lease base terms lengthening back toward 10 to 15 years. If landlords and tenants both agree the assets hold value that long, my hinge stops being a hinge.
3. CoreWeave refinancing its 2026 to 2028 maturities at flat or tighter spreads, or real failure data in the application layer. Either one would mean I am watching the wrong floor.

Sources

• Bank of England, Financial Stability Report, July 2026
• BIS Quarterly Review, “Financing the AI infrastructure boom: on- and off-balance sheet borrowing”, 16 March 2026
• Meta, “Meta Announces New Strategic Venture with BlackRock to Develop Data Center in El Paso”, 28 July 2026
• Amazon, Second Quarter 2026 Results, SEC Exhibit 99.1, 30 July 2026
• PitchBook, “Q1 2026 AI funding blows past 2025 total”, 12 May 2026
• Eurostat, Electricity price statistics, retrieved 12 August 2026