Here is the contradiction most investors are holding in their heads right now without noticing it.
On one screen, Goldman Sachs CEO David Solomon is calling this an "AI capex super cycle" — a phrase that lands like a promise of years of upside. SpaceX just closed an $86 billion IPO. SK Hynix sold $26.5 billion in ADRs. ASML raised guidance again on what the Wall Street Journal called "unrelenting AI demand."
On the other screen, Morgan Stanley's accounting team is quietly cataloguing roughly $1.8 trillion in debt that isn't on anyone's balance sheet.
Both are true. The banks minting fees from the buildout and the analysts flagging its hidden liabilities are describing the same machine — just from opposite ends. The super cycle and the time bomb are the same object.
The banks minting fees from the buildout and the analysts flagging its hidden liabilities are describing the same machine. The super cycle and the time bomb are the same object.
Last week we laid out that $1.8 trillion off-balance-sheet figure and asked what it was made of. This week the picture sharpened in two specific, uncomfortable ways: the leverage behind it doubled, and the bond issuance funding it got quantified.
Start with leverage, because it's the number that moved.
Gross leverage across the hyperscale cloud companies went from 0.9x in Q3 2025 to 1.8x now — a full doubling in two quarters, or roughly 0.3x added per quarter. That pace is the story as much as the level.
It also crossed a line that should make bond investors pause: these tech balance sheets are now more levered than the entire energy sector.
The market noticed. Credit spreads on these names have drifted from the AA range down into the A range, and Morgan Stanley expects them to widen further under supply pressure.
Meta is the tell. Its credit spread is now wider than the CDX IG benchmark itself — meaning the market is pricing one of the most profitable companies on earth as a below-average investment-grade credit.
Now the issuance, which is where the doubling gets funded.
Global AI-related bond issuance hit $236 billion by the end of May 2026, up 357% year over year. Morgan Stanley expects the full-year figure to clear $570 billion.
April alone saw more than $74 billion issued, a new high for the year. And the composition matters: project-finance structures for data center construction made up 85% of high-yield supply and 40% of investment-grade supply.
Amazon, Meta, Google, Microsoft and Oracle now account for 4% of the entire investment-grade bond index. Five companies. That's concentration that reshapes the market they're borrowing from.
Here's why the debt has to keep coming. Morgan Stanley projects that Amazon's and Meta's free cash flow in 2026 will approach zero or turn negative.
When your free cash flow is gone, incremental spending gets funded almost entirely by new debt. That's not a choice; it's the arithmetic of capex growth outrunning revenue growth.
Then there's the $1.8 trillion itself, now broken out cleanly.
It's three pieces: roughly $982 billion in purchase commitments, over $822 billion in non-cancelable lease commitments, and about $110 billion in accounts payable and supplier financing.
The purchase commitments are the sneakiest. Under accounting rules, a long-term contract to buy chips — from Nvidia, say — isn't recorded as a liability until the goods are delivered, unless the company expects a loss on it.
So nearly a trillion dollars in contractually locked-in future cash outflows sits on no balance sheet anywhere. The obligation is real; the disclosure is not.
Layered on top is a financing structure that pushes the risk further out of view.
Private credit shops like Apollo and Blackstone are using special purpose vehicles to move leverage off corporate balance sheets and down into the supply chain — the report cites a $35 billion "chip-backed" SPV deal for Anthropic.
Morgan Stanley's word for the result is "circular." Leverage gets shifted from the companies everyone watches to entities almost no one can see through. If AI commercialization disappoints, or enterprise customers migrate to cheaper alternatives, that's where the chain snaps first.
And the depreciation hasn't even arrived yet.
Construction-in-progress balances are ballooning — Oracle's rose 200% year over year — which means the assets aren't yet running through the income statement. Once they do, Morgan Stanley estimates cumulative depreciation for the Big Four could top $520 billion over four years.
That's a $520 billion drag on margins that is currently deferred, not avoided.
The structural mismatch underneath all of it: capex forecasts are being revised up far faster than revenue forecasts. Oracle's capex forecast was lifted 175%. Revenue upgrades haven't come close to matching that.
You are watching spending get committed against demand that is projected but not yet proven.
Which brings us back to the two screens. Goldman's $1.1 trillion-to-$1.4 trillion 2027 capex projection — the number underpinning the "super cycle" — comes with a condition attached that rarely gets quoted: it holds provided large language models keep their pricing power and enterprise customers stay sticky.
That "provided" is doing an enormous amount of work.
For a smaller, cleaner read on how the market is actually feeling, look at Credo Technology, a SerDes connectivity supplier riding the same wave. Its stock is up 121% over the year even after falling more than 30% from its high in a few weeks.
Insiders have been trimming on schedule — the COO sold about $23.9 million across two same-day sales, the co-founder CTO another $6.6 million — all under preset 10b5-1 plans, all tiny slices of nine- and ten-figure stakes. That's diversification, not a fire alarm.
But note the fragility in the business even where the story is working: about 90% of Credo's revenue comes from its top 10 customers, two of them topping 10% each. Concentration on the supplier side mirrors the concentration on the borrower side.
Here's what to hold onto. The AI buildout is genuinely producing record revenue, record fees, and real earnings — Credo tripled revenue past $1.3 billion and grew non-GAAP net income fivefold. None of that is fake.
But the financing of it is being pushed steadily off the balance sheet, down the supply chain, and into the future — into commitments that don't count as debt, into SPVs that resist inspection, into depreciation that hasn't hit.
The bull case and the bear case aren't competing narratives about different things. They're the same $1.8 trillion, read forward versus read backward.
What changed this week is the margin of safety. Leverage doubled, spreads widened, and free cash flow is heading to zero at the two companies leaning hardest into the trade. The super cycle needs everything to keep working. The off-balance-sheet math is what happens if it doesn't.
Watch the spreads, not the IPO headlines. When Meta already trades wider than the index, the bond market has started answering the question the equity market is still cheering through.