The Demand Side Is a Presumption

Every number in the artificial intelligence build-out is documented except the one that pays for all of them.

The supply side is a matter of public record. Anthropic’s amended filing lists roughly $895 billion in aggregate commitments — purchase obligations and leases, stretching to 2045. The hyperscalers’ leases, backstops, and construction obligations are in their filings too. Add them up and you get a web of promises, each one signed, dated, and enforceable. I know. I mapped it.

The demand side is a matter of faith.

For those commitments to be serviced, AI has to be adopted at the rate search was adopted. Not tried — adopted, the way search was: default, universal, unremarkable, a verb. Search earned that curve because it asked almost nothing of anyone. You typed, it answered, it left. What is being sold now asks for considerably more: agents inside your accounts, your files, your workflows, your money — and it is being financed on the old curve anyway, as if the trust required were the same trust a search box required.

Ask the public what it thinks of the technology. The most common honest answer is not excitement. It is that the stuff is creepy. That word does not appear in any prospectus. It should, because it is the demand curve, stated in plain language by the people who are supposed to generate it. The entire capital structure of the largest technology build-out in history rests on a presumption about the future behavior of people who were never consulted and are, on the available evidence, uneasy.

Meanwhile, the price is being quietly renegotiated by the people who were consulted. Institutional investors have told Anthropic in recent days that its offering should price nearer $1.5 trillion than the $2 trillion figure carried all summer. This is not happening in public. It is happening the way repricing always happens — privately, in models, desk by desk — and it will become visible only when the roadshow range prints and makes it official. The information asymmetry is not a bug in this process. It is the process. The institutions get the map early and quietly. The public gets the opening print.

There is a second asymmetry underneath the first. A map of obligations can be read in two directions. Forward, it is a financing chart: who promised what to whom. Backward, it is a blast-radius chart: when a node fails, the loss travels the same lines in reverse until it stops at whoever is holding the end of the chain — and the end of the chain, as the filings show, is the public: their index funds, their pensions, their 401(k)s, holding a record share of household wealth in equities. But reading the map backward is only actionable if you have the capital to act on it — the borrow, the margin, the risk desk. The desks get a weapon from the map. The public gets a warning. Same paper.

And even the warning has a limit, which is the subject of my previous paper. If a node in this web fails, there is no counterparty to transfer the loss to. The obligations point at entities — special-purpose vehicles, uncommenced leases, commitments stacked on commitments — with no independent balance sheet standing behind them. You cannot sue an empty structure and collect. You cannot collect from a presumption. The loss does not get transferred. It gets absorbed, by whoever is last in line, and the line has already been drawn.

One fact in the ledger deserves plain statement, because it decides how any failure actually unfolds. Amazon holds every one of Anthropic’s convertible notes — all $6.2 billion of principal. Whatever else that means, it means there is exactly one standing buyer with a structural reason to catch falling pieces, and it is not the public.

None of this requires a villain. It requires only a sequence: commitments made in public filings, a price set by presumption, a quiet repricing for those in the room, and an opening print for those outside it. The remedy is not complicated either. The map is public. It was assembled entirely from documents any investor can read. Read it before the range exists — not after the loss does.

The demand side is a presumption. Presumptions are not counterparties.