In the first quarter of 2026, AI capex reportedly contributed more than a third of US GDP (per Chartbook/Tan Kai Xian). That is not a metaphor. It is a national accounts line item, and it means the American economy has quietly rewired itself so that high equity prices are an input to growth, not just an outcome of it.
The buildout is expensive, front-loaded, and financed against future returns. It only pencils out if the companies doing the spending — and the investors funding them — believe valuations will hold. The stock price is the enabling condition. Take it away and the capex thesis stops closing.
So consider the timing of a post the New York Fed published in August, which — read naively — says the opposite of alarming. It says the economy is now safer from rate hikes than it has been in decades.
Here is the finding. In the mid-1980s, fewer than 30 percent of US households held equity. By the early 2000s, more than half did — through 401(k)s, IRAs, mutual funds. As participation widened, the way stock swings pass through to spending changed.
The Fed's author builds a model where participants hold leveraged, procyclical equity and nonparticipants hold only bonds. When rates rise, equity prices fall and financing costs climb, so participants cut consumption hard. That is the transmission channel: rate up, stocks down, wealthy leveraged holders retrench, investment and output follow.
The counterintuitive part is what happens as more households join. Each participant ends up holding a smaller, less-leveraged slice. The wealth effect of any given price move gets diluted; the financing pressure eases. Comparing a 25-percent-participation economy with a 55-percent one, the model reportedly finds the output response to an unexpected rate increase is 20 percent smaller.
The conventional read is reassurance. Broad ownership spreads risk, smooths the cycle, makes the Fed's hikes less brutal. Fine.
But run the mechanism the other direction and it is not reassuring at all. The same dampening that shields output from rate hikes dampens the entire equity-to-real-economy channel — in both directions.
That channel is precisely what normally throttles a capex boom. Rates rise, valuations compress, the market value of installed capital falls relative to the cost of building new capital, and firms pull back on investment. The paper says it plainly: smaller stock-price movements translate into smaller adjustments in investment spending.
In an ordinary cycle that is the automatic stabilizer. It keeps the buildout from running past where policy would brake it. When the Fed leans on rates, the wealth-and-financing channel does the enforcement.
But if that channel has gone quiet, the brake is disconnected. Valuations and capex can run past the point where monetary policy would have stopped them — because the mechanism that would have stopped them barely registers anymore.
If that channel has gone quiet, the brake is disconnected. Valuations and capex can run past the point where monetary policy would have stopped them.
That is arguably why AI-linked activity could contribute nearly 30% of real US GDP growth over the last three years (per BlackRock) without the Fed's tightening cleanly cooling it. The blunting the Fed documented is arguably the same blunting that let the boom get this large. The lever slipped, and the buildout kept accelerating.
Now watch the plumbing that has grown up on top of that assumption, because it is reflexive all the way down.
Start with the fuel. Money-market funds, per FT Alphaville, are reportedly providing the stock market "rocket fuel" — parked cash flowing into risk. The "AI 33," the companies most directly tied to the buildout, reportedly delivered roughly 78 percent of the S&P 500's year-to-date return, per BlackRock. The rally is narrow, and it is the same narrow set of names whose spending is a third of GDP growth.
Then the collateral. This is where it gets genuinely new. GPU rental prices, which everyone in 2024 and 2025 assumed would decline, are instead — in tech investor Gavin Baker's words — "going vertical."
Silicon Data reportedly has the H100 renting at $2.71 an hour, up from $1.96 at the end of November. Baseten disclosed its provider will lift B200 rents from $2.63 to $5.10 at renewal in October. Even 2020-vintage A100 chips are being leased out to 2029 at attractive prices.
Rising, durable rents make a stream of cash flows. And cash flows can be financed. Jensen Huang's framing is explicit: CUDA makes the chips fungible, fungibility drives utilization, and that makes NVIDIA compute "rentable, durable and financeable."
On the back of that, a consortium — Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR — lined up to mobilize over $500 billion of third-party capital for the buildout. Blackstone's Jon Gray: "People are going to begin to recognize that this is a financeable asset class." Larry Fink compared it to the birth of mortgage-backed securities in the 1970s.
Read that comparison carefully. Fink means it as promise. It is also a warning, because MBS was an asset class whose collateral value was assumed stable until it wasn't.
Here is the load-bearing assumption underneath all of it. GPU rents stay high because AI demand stays high — and AI demand stays high because the companies buying compute can keep spending, which they can do because their equity valuations support the buildout. The collateral is only worth $500 billion of financeable value if the stock prices hold.
So you have three things chained together: capex that is a third of GDP growth, a rally where 78 percent of the return comes from the buildout's own names, and a new half-trillion-dollar asset class collateralized by rents that depend on the buildout continuing. Every link assumes equity prices stay elevated.
Now reverse the loop.
The stress test is not hypothetical. A disappointing Nvidia print — the company reportedly guided to roughly $91 billion in revenue, per intellectia.ai, an extraordinary bar — or any valuation air-pocket in the AI 33 is enough to start it. The Magnificent Seven ETF was reportedly already down over 8 percent from its June peak; the semiconductor index down more than 19 percent, per ainvest.com (as of mid-August 2026).
When stocks fall in a normal economy, the wealth-and-financing channel cushions the descent as the automatic stabilizer kicks in the other way. But the Fed's paper just told us that channel is 20 percent quieter than it used to be. The shock absorber has been drained.
So a stock-price shock does not get dampened on the way down. It hits at close to full force — the same un-throttled directness that let the boom overshoot on the way up.
And it hits three things at once. Consumption, through the wealth effect, however muted. GDP, because a third of the growth was the capex itself, and capex is the first thing cut when the valuation thesis breaks. And the collateral, because falling equity prices undercut the demand story that made GPU rents "vertical" — which reprices the $500 billion of financing built on those rents.
That is the fragility. The transmission-blunting that made the AI boom possible, and that reads as safety in a Fed blog post, is the exact same feature that removes the cushion when the loop turns.
An engine that needs high equity prices as oil runs beautifully until the pressure drops. Then there is arguably nothing between the moving parts and the fall — because the mechanism that would have caught it is arguably the one the economy spent four decades quietly disconnecting.