Dead Load

Executive briefing

The disclosure regime around the AI financing stack survived its first ninety days of testing. The assumption underneath it did not. Between mid-April and mid-July 2026, the rules governing how the largest offerings in history are indexed, researched, litigated and licensed were rewritten — each change lawful, each disclosed, each decided by a different institution on its own mandate, and each pointing the same way. The result is a market that is not being manipulated but administered. The disclosures exist; the rulebook they are filed under has been moving; and the second fact is why the first no longer settles the question.

The mechanism is visible in one listing. SpaceX priced the largest IPO ever conducted on 11 June — $75bn, rising towards $86bn with the over-allotment — having converted itself, in the thirty days before the book was built, into the very structure this series has called the neocloud: a loss-making frontier lab wrapped in cash-generative collateral, letting some 325,000 GPUs to Anthropic and Google for $6.5bn a quarter against $818m of quarterly AI revenue of its own. The offering was engineered for scarcity — a float below 5 per cent, a retail tranche held in place by penalty, insiders on a tiered lockup, a directed 5 per cent with no lockup at all — and it delivered the standard result of administered listings: not failure but disappointment. The stock ran from $150 to above $225, then decayed back through its own offer price within a month.

What caught the decay was the scaffold, and the scaffold is the finding. Nasdaq rewrote its index-inclusion rules in a public consultation that concluded six weeks before the listing, replacing a three-month seasoning requirement and a 10 per cent float minimum with a regime that admitted SpaceX on its sixteenth trading day — the fastest entry in the index’s history — and directed a $4.3bn forced bid from index-tracking funds into a 4 per cent float. The SEC consented, six months before the offering, to retire the 2003 Global Research Analyst Settlement, the architecture built after the dotcom crash specifically to police underwriter research in hot offerings; on the day the quiet period expired, the syndicate initiated coverage at targets up to $800 against a $160 market, and the one target below the market came from the one prominent firm outside the syndicate, whose analysts found “no credible financial model” for even a fifth of the top target’s implication. All ten bookrunners were, per the prospectus, creditors of the company through a bridge whose covenant directed where the IPO money would go. And S&P Dow Jones, which ran the same consultation Nasdaq did, kept all three of its requirements — a decision that reads as the control group.

The asset being priced degraded on the same schedule. The US government took Anthropic’s flagship model dark for nineteen days by letter, and restored it on conditions amounting to a co-signature on the release calendar; OpenAI staggered a release of its own after a government request. Four days after listing, SpaceX exercised a pre-disclosed option to pay $60bn in stock for Anysphere — the routing layer that exists precisely because the models beneath it are interchangeable. If the moat were the model, the money would have gone into the model. And the comparison set inverted: Anthropic filed first, reportedly profitable, at a private mark of $965bn — leaving OpenAI’s non-negotiable trillion roughly 4 per cent above a peer with better numbers, while its own leaked audited financials show a $38.5bn net loss on $13.1bn of revenue.

Marked against the Keystone’s six preconditions, the register reads: none cleanly satisfied, one broken. The offering itself is sliding towards 2027 because its sponsor watched the dress rehearsal. Oracle, the anchor’s landlord and this series’ bellwether, was cut to BBB- with S&P conceding it had “underestimated the scale of the investments required” — $638bn of obligations, half owed by OpenAI, against negative $42bn of projected free cash flow. Nvidia’s chief executive described the $100bn commitment as “never a commitment.” Amazon’s $50bn entry carries $35bn contingent on the IPO happening at all: a precondition for the offering now has the offering as its own precondition.

The consequence is carried by the parties who were never asked. The dead load — the structure’s own weight, borne before the load it was built for arrives — rests on a landlord one notch above junk, on hyperscalers whose free cash flow is compressing toward zero, and on the index-tracked savers conscripted on 7 July. They are also the book the OpenAI offering will need when it comes. An arch under dead load does not fall because something pushes it; it falls when the crowd holding it up concludes that holding it up is someone else’s trade. The acute question is how long they stay.