A dollar of artificial-intelligence spending is the most confident money in the world. It is committed years ahead of the revenue meant to justify it, against models of demand that do not yet exist at scale. To know whether that confidence is warranted, you do not need a forecast. You need only to follow the dollar around the loop and see how far it gets.
1 · It enters as a commitment
The dollar begins as capital expenditure. Nvidia alone cites more than $500 billion in booked orders for its Blackwell and Rubin generations; the major hyperscalers are spending at a combined pace measured in the hundreds of billions a year. Crucially, much of this is financed rather than retained — raised as debt and equity against a return that is still a promise. The instant the dollar is committed, a second clock starts: the one on the financing behind it.
2 · It splits across the supply chain
From there the dollar fans outward. It becomes a slice of a TSMC wafer, an ASML lithography tool, a stack of SK Hynix high-bandwidth memory, a Vertiv cooling unit, a Constellation power contract, an Equinix lease. This is where the build-out turns physical — and where the first uncomfortable fact surfaces.
A large share of what looks like demand for AI compute is recycled supply-side capital. By our reading of the funded-cash flows, for every dollar of outside equity that actually reaches the model labs, on the order of twenty-six dollars of compute commitments flow back to the same chip-and-cloud vendors who, in several cases, helped finance those labs to begin with. The dollar can circle inside this ring for a long time — booking revenue at every pass — without ever touching a customer outside it.
3 · It has to reach demand
Eventually the capacity must be sold to someone outside the loop: the thirty-one industries that have to turn AI into real productivity and revenue. Here the honest answer is mixed. Adoption is genuinely broad — across those industries our value-chain read sits near 69 of 100. And in places the payoff clock is audibly ticking: Booking Holdings pulled $250 million of cost out in a single year, Airbnb cut its cost per booking around 10%, Expedia now handles a third of its support with AI.
But those are efficiency gains concentrated in a handful of digitally-native sectors. Across most of the thirty-one, AI still shows up as spend and pilots, not yet as proportionate profit. Adoption is not the same as return — and the distance between the two is the entire argument.
4 · The market prices it as already returned
Before the dollar has demonstrably come home, the market re-rates it. In the second quarter of 2026 our market-versus-ground-truth divergence blew out to +4.06 — the widest reading in the series — as the price signal spiked and the fundamental signal fell. The market is, in effect, booking the first dollar as profit while the filings still carry it as cost.
Where the dollar is now
Trace it honestly and the dollar is stretched across the loop — committed at the top, embedded in capacity, partly recycled, broadly adopted but only narrowly returned, and already re-priced as if the circuit had closed. It has not. The fragility core reads 49: moderate, not severe. So this is stretching, not breaking. But every quarter the dollar sits in the loop without completing it, the financing clock that funded it grows louder.
None of this requires a crash. It requires only that you watch one thing — whether the return at stage three begins to justify the capital at stage one, before the market at stage four or the financing behind stage one forces the question.
Will the return arrive before the financing reprices? Follow the dollar, and you will know before the headlines do.
Sources & method. Forward-order scale from company disclosure (Nvidia); the recycling ratio from our capex model on a funded-cash basis ($21B funded equity against ~$540B of compute commitments); the divergence reading from the model's latest quarter; the 69/100 adoption figure is an aggregate of our thirty-one per-industry value-chain assessments; the industry ROI examples are from company earnings (Booking, Airbnb, Expedia). The loop framing and the regime call are Divergent Compute's editorial reading of those measures. See the open-source method.