AI Is Wildly Overleveraged. It May Also Be Necessary.

Two documents landed within days of each other this summer, and they cannot both be relaxing. On June 28, 2026, the Bank for International Settlements, the central bank for central banks, used its Annual Economic Report to name three interlocking risks that could crack the global financial system: an AI capex bust, opaque circular financing inside the AI supply chain, and record sovereign debt. Three days earlier, JPMorgan looked at roughly the same numbers and called the multi-year, roughly $5.5 trillion AI capex explosion “profitable, for now.”

Read those two sentences again. The people whose job is to worry about systemic risk are worried. The people whose job is to price the biggest capital cycle in modern tech think the math works. And the strange part is that they are looking at identical data.

That is the whole story, and almost everyone tells only half of it. One camp screams bubble. The other sells the future. Both are describing something real, and both are missing that the other half is also true.

Financed like a bubble, demanded like a utility

Here is the shift, named plainly. The AI buildout is being financed like a bubble and demanded like a utility at the same time. The capital structure has the fingerprints of a mania: debt issued against revenue that does not exist yet, deals that loop money between the same handful of companies, valuations that assume a future none of them can guarantee. The underlying demand has the fingerprints of a new piece of essential infrastructure: capacity absorbed the moment it is switched on, backlogs that keep growing, a workload that behaves less like a fad and more like electricity.

So the useful question is not “bubble or not.” That framing is a trap, because you can prove either side by picking your numbers. The honest question is a timing question.

“

Will the demand arrive before the debt comes due? Everything else is commentary.

”

That is the sentence to carry through the rest of this. Two clocks are running. One counts how fast real, paying AI usage grows into the capacity being built. The other counts how fast the bills on that capacity come due. The outcome depends entirely on which clock is faster, and right now nobody can prove which one wins.

The bubble case, in numbers that do not flatter

Start with the spend, because the spend is genuinely staggering and it is not a projection. It is committed.

In 2026, Amazon is putting roughly $200 billion into infrastructure, Microsoft around $190 billion, Google about $180 billion, and Meta up to $140 billion. Add the rest and the five largest hyperscalers cross a combined $1 trillion in a single year. Zoom out to all of US big tech and the aggregate AI infrastructure commitment rose from roughly $380 billion in 2025 to somewhere around $660 to $690 billion in 2026, with some estimates putting it closer to $725 billion. That is a near-doubling of the largest capital program in the industry, in one year. Nothing at this scale doubles quietly.

Now the part that turns a big number into a nervous one: where the money comes from. Historically, hyperscalers funded capex out of their own enormous free cash flow. That is no longer enough. Commitments are outpacing earnings and cash generation, so the buildout is increasingly funded with debt. Global AI-related debt issuance is forecast at about $570 billion in 2026. By May 31, roughly $236 billion of it had already priced, about four times the pace of the year before. On July 17, 2026, Forbes reported that bond investors had started to push back, demanding more for the risk. When the people lending the money begin to flinch, that is a signal worth respecting.

Then there is the mechanism the BIS singled out as the scariest, and the one hardest to see from the outside: circular financing. Money and commitments loop between the same players. A chip maker invests in a model lab. The lab signs a multi-year contract to buy that maker’s chips. A cloud provider takes an equity stake in the lab and also rents it capacity. Supplier contracts, equity, and debt get braided together until the same underlying asset is effectively pledged more than once, and the revenue on one company’s books is another company’s spending. Layer on that the financing is shifting toward less transparent channels: private credit to AI companies grew from about $3 billion in 2010 to over $40 billion in 2025, and off-balance-sheet structures for data centers are estimated to run around $800 billion. When the funding moves off the public books and into private credit, the risk does not disappear. It just gets harder to price.

And under all of it sits the gap that makes the whole thing vertigo-inducing.

Even the friendly estimates put total AI revenue at roughly $50 to $60 billion a year against $500 billion or more in annual capex. That is on the order of $8 to $10 of investment for every $1 of revenue the industry actually collects today. You do not need a finance degree to feel the tension in that ratio. You are building a ten-lane highway on the strength of a one-lane toll booth, and financing the other nine lanes with borrowed money.

That is the bubble case, and it is not hysteria. It is arithmetic.

The utility case, which is just as real

Now steelman the other side properly, because the bulls are not fools and their evidence is not vibes.

The single strongest argument is the most boring one: the capacity is getting used. All five major hyperscalers report that AI capacity is being absorbed nearly as fast as it is deployed. Cloud backlogs, the contracted demand not yet fulfilled, are large and growing rather than shrinking. That is not the signature of a speculative glut sitting idle. A bubble looks like empty capacity chasing a story. This looks like a queue.

The forward estimates from firms with no incentive to hype say the same thing. Morgan Stanley estimates that global data center capacity needs to grow roughly six-fold by 2035 to meet cloud and AI demand. McKinsey forecasts that data center demand could nearly triple by 2030, and that about 70 percent of that growth comes from AI workloads specifically. If those numbers are even directionally right, then today’s trillion-dollar year is not the top. It is an early installment.

There is also a shift in the kind of demand that matters. Early AI spend was training: the one-time, enormous cost of building a model. The bull case rests on inference, the cost of actually running models to serve users, and inference scales with usage rather than with a single build. As AI moves from experiments into production systems, inference load compounds. Deloitte names agentic AI, software that does not just answer but takes multi-step actions on your behalf, as a top driver of future compute demand, and floats that it could eventually overtake SaaS as a category. Capable agents are token-hungry in a way chatbots never were. A chatbot answers a question. An agent reads, plans, calls tools, checks its work, and loops, burning many times the tokens for a single task. Those agents are only just rolling out. The workload that justifies the buildout has, in the bull’s telling, barely started.

Here is the honest core of the bull case, stated without spin. The gap between infrastructure spend and enterprise return on investment is the central question for 2026 and 2027. Mass productive deployment, the point where ordinary companies run AI in ways that clearly pay for themselves, has not fully materialized yet. The bulls are not claiming it has. They are claiming it is coming, and that building ahead of it is the correct move even though the receipts are not all in.

Why both things are true at once

Most takes force a choice: bubble or revolution, doom or hype. That framing is wrong, and seeing why is the point of this whole piece.

Both are true because they describe different layers of the same thing. The demand can be a genuine, durable, utility-grade shift and the financing of the rush to meet it can be reckless. The dot-com era is the cleanest analogy, and people usually draw the wrong lesson from it. The internet was not a hoax. It was the real thing, and it did reshape the economy exactly as promised. And the financing around it in 1999 was still a bubble that vaporized a lot of capital when it popped. The technology was right. The timing of the money was wrong. Both, at once.

“

A real revolution and a real bubble are not opposites. History’s biggest infrastructure booms were usually both, and which word you use depends only on where you are standing in time.

”

So the fight between the BIS and JPMorgan is not really a disagreement about facts. It is a disagreement about which clock wins. The BIS is watching the debt clock: leverage rising, financing turning opaque, buyers of that debt starting to balk, all against a backdrop of record sovereign debt that leaves less room to cushion a shock. JPMorgan is watching the demand clock: usage compounding, backlogs growing, inference and agents just beginning to scale. “Profitable, for now” is not a contradiction of the risk. It is a bet on the timing. The “for now” is doing enormous work in that sentence, and JPMorgan knows it.

The uncomfortable truth is that nobody can currently prove which clock is faster. Not the central bank, not the biggest lender on Wall Street, not the hyperscalers writing the checks. That is not a failure of analysis. It is the actual state of the world. The buildout is a leveraged bet that demand shows up on schedule, and the schedule is unknown.

What would actually break it, and what would not

If you want to move past the shouting, stop asking “is this a bubble” and start asking “what specifically breaks, and what happens if it does.”

The fragile part is the financing, not the technology. The demand could keep growing and the system could still take a hard hit, if the circular financing unwinds faster than the revenue fills in. That is the BIS’s real fear, and it is worth stating precisely. When assets are pledged more than once and money loops between the same firms, a stumble at one node propagates. If a major model lab misses badly, the chip contracts, the equity stakes, and the debt written against all of it reprice together. Opacity makes that worse, because when risk sits in private credit and off-balance-sheet vehicles, the market cannot see the exposure until it is already moving. The sovereign-debt backdrop matters here only as capacity to absorb a shock: with public balance sheets already stretched, there is less slack if something large breaks. That is a market-structure observation, not a verdict on anyone’s policy.

The resilient part is the usage. Even in a sharp financing correction, the capacity built does not evaporate. The chips still compute. The data centers still run. A correction would wash out the most leveraged players and the silliest valuations, and it would be brutal for whoever is holding the riskiest paper. It would not un-invent inference demand or send agents back into the lab. This is exactly the dot-com shape again: the pets.com layer gets wiped out, the fiber in the ground gets bought for cents and powers the next decade.

What to actually watch (and do)

You do not get to know the ending in advance. But you are not helpless either. The whole debate collapses into one measurable thing, and once you can watch that thing, the noise stops controlling you. Watch whether revenue is catching capex. That is the signal that settles the timing question, and it shows up before the headlines do.

If you build with AI, the leverage question is not abstract, it is your dependency risk. The infrastructure you rely on is being financed on the assumption that demand shows up on time. So watch the ratio, not the announcements. Is real, paying AI revenue growing toward the capex, or is the gap widening while the debt column grows? Widening gap plus rising leverage plus lenders demanding more is the pattern that says the debt clock is winning. Backlogs converting into recognized, paid usage is the pattern that says the demand clock is. Design so a repricing of your suppliers is survivable: know who is funding the capacity under you, and do not assume today’s cheap tokens are permanent.

If you invest or allocate, separate the two layers the crowd keeps merging. The technology being real tells you nothing about whether a given security is priced sanely, and a bubble in the financing tells you nothing about whether the demand is fake. The dangerous position is the one that bundles them, that treats “AI is important” as if it settled “this AI debt is safe.” It does not. This is not financial advice, and I am not telling you which way it breaks. I am telling you which number to keep your eyes on while everyone else argues about the label.

The reframe is the thing to leave with. You have been handed a binary, bubble or revolution, and asked to pick a team. Put the binary down. The real position is more useful and more honest: this is a genuine utility-scale shift being funded on a bubble’s terms, and your job is not to guess the label but to watch the one ratio that decides it. That is a calmer, sharper place to stand than either camp, and it is the only place from which the next two years actually make sense.

Sources

• How the AI bubble could pop and take down the global economy, according to the BIS (The Register, on the BIS 2026 Annual Economic Report)
• What bubble? JPMorgan says the $5.5 trillion AI capex explosion is ‘profitable, for now’ (Fortune)
• Bond investors push back as AI debt heads toward $570 billion (Forbes)
• The State of AI: Infrastructure, Demand, Costs, and Custom Silicon (ARK Invest)