Start with the contradiction, because it's the most honest way into this story.
Two Federal Reserve presidents looked at the same $3 trillion and could not agree on whether it was a problem. New York's John Williams says he doesn't see a bubble. San Francisco's Mary Daly calls the scale "potentially very worrisome."
When the two people whose job is to spot systemic risk publicly disagree about the largest financing structure in the economy, the disagreement is the signal. It means the central bank has no settled way to model what's happening — because what's happening was built specifically to be hard to model.
When the two people whose job is to spot systemic risk disagree about the largest financing structure in the economy, the disagreement is the signal.
Last week we walked through the $1.8 trillion in hidden leverage that Morgan Stanley reportedly traced through purchase commitments, uncommenced leases, and vendor financing — the SPV chains that Apollo and Blackstone reportedly use to push leverage down into the supply chain. This week the aggregate got bigger and the accounting got clearer: Morgan Stanley now reportedly pegs the total AI support commitments across nine tech giants at over $3.1 trillion, and the market reportedly started pricing the capex, with Alphabet and Tesla both sliding on spending forecasts.
But the number that should actually keep shareholders up isn't $3.1 trillion. It's the speed.
Alphabet's purchase and contractual commitments reportedly rose from $332 billion at the end of March to $811 billion at the end of June. That's a 152% increase in a single quarter — roughly half a trillion dollars in new obligations booked in ninety days.
And it lives in a footnote. Not the balance sheet, not the debt schedule that ratings agencies stress-test — the fine print, the financial equivalent of the small type on a rental car agreement.
That location is the entire story. This is the point most readers get wrong: they treat the $3.1 trillion as a size problem, a "that's a lot of money" problem. It isn't. It's a structure problem.
Here's why structure beats size. On-balance-sheet debt comes with brakes. Covenants that trip if leverage climbs too far. Rating agencies that downgrade and raise your cost of capital. Disclosure that forces the number in front of every analyst who models your free cash flow.
Off-balance-sheet commitments trigger none of that. As Crypto Briefing put it plainly, these obligations don't trip the same covenants or draw the same rating scrutiny as conventional debt — which means they can pile up with almost no external friction.
Frictionless accumulation is exactly what you'd expect when the normal disciplinary machinery has been routed around. Alphabet didn't triple its commitments in a quarter because the discipline failed. It tripled them because, structurally, there was no discipline to fail.
Now the part nobody is stress-testing.
The comforting assumption baked into every bull case is that these commitments only bite if demand collapses. It's the wrong threshold. As one VC underwriting hyperscaler-adjacent exposure framed it, the scenario to model isn't a collapse — it's AI capex growth simply flattening for two quarters.
Because these purchase obligations don't disappear if demand merely softens. Uncommenced leases still have to be honored or renegotiated. Purchase commitments still have to be fulfilled or written off.
That's the trap in one sentence: the obligations are fixed, but the revenue they're underwritten against is not. Flatten the top line while the commitments keep converting into cash outflows, and you get fixed costs grinding against slowing revenue — a margin structure nobody's spreadsheet currently contains.
And some hyperscalers are already there. Morgan Stanley reportedly expects free cash flow at Amazon and Meta to approach zero or turn negative in 2026, with financing increasingly dependent on new debt. The buffer is thinner than the headline profitability suggests.
Now collide that with the demand-side evidence the market is holding in a completely separate hand.
The share of companies abandoning most of their AI initiatives before production reportedly surged from 17% to 42% between 2024 and 2025, per 451 Research at S&P Global. Nearly half of proofs of concept are being discarded.
And MIT's NANDA report reportedly found that roughly 95% of generative AI pilots produced no measurable financial impact.
Sit with the two facts side by side. Contractual commitments to buy chips and data-center capacity are tripling in a quarter — while the enterprise demand meant to fill that capacity is failing to convert to revenue at a rate that just doubled.
The collateral, in other words, is quietly weakening while the leverage against it accelerates. This is the demand-side counterpart to the pre-revenue narrative reset — but here it functions strictly as evidence: the ROI case underwriting these commitments is eroding in real time.
Then there's the deflation that never showed up. The whole thesis — Altman's abundance, Musk's plummeting costs — held that AI would push prices down.
Instead the buildout is arriving as inflation. The scramble for data centers has tangled supply chains and driven up electricity prices before any productivity gain shows in the statistics. OpenAI's own chief economist concedes the immediate costs of AI are more visible than the benefits.
So the commitments are frictionless, the obligations are fixed, the ROI evidence is collapsing, and the promised cost relief has inverted into an inflation problem the Fed is openly split on. Which brings us to the defense you hear most from the companies signing these deals.
We can't stop. It's an arms race. Stopping is riskier than continuing.
Read that carefully, because it's usually delivered as a bull case. It isn't. It's a confession.
"We can't stop spending even though the returns aren't there yet" is not the language of a company exercising discipline. It's the precise internal logic of a fixed-cost trap — the rationalization you construct when the obligations have already grown large enough that the only tolerable story is the one where you keep feeding them.
A rational fixed-cost trap always looks reasonable from the inside. That's what makes it a trap.
The 2008 comparison gets invoked constantly here, and it's usually invoked wrong. The Bank for International Settlements has warned that an AI-spending pullback could hit credit markets as hard as 2008 — but not because the assets are junk.
GPUs and data centers are real, productive, revenue-generating. That's the genuine difference from subprime, and the honest counterweight to the whole bear case.
The parallel isn't asset quality. It's opacity and interconnection.
2008 wasn't fundamentally a story about how bad any single mortgage was. It was about how the instruments stacked on top of those mortgages were so interlocked and so opaque that no one — including the Fed — could map who was exposed to whom.
Swap the mortgages for a $3 trillion web of purchase commitments, uncommenced leases, and private-credit vehicles routing leverage through SPVs, and you have the same failure mode with better collateral. The productive assets don't protect you from a chain no one can trace.
We know it can't be traced, because the people whose job is to trace it have said so. The Fed doesn't fully know who's financing the boom. Its two presidents can't agree on whether to worry.
That's where this stands. The assets are real. The structure is not sound.
And the tell was never the $3.1 trillion. It was $332 billion becoming $811 billion in ninety days, in a footnote, with nothing in the system built to slow it down.