When Citadel took Situational Awareness's entire public book in a single block before the open on 30 July (WSJ), four of the AI-infrastructure names at the centre of it rallied 21–28% in a session. The consensus read that as confidence: a credible buyer had stepped in, so the selloff must have been overdone. I don't think that reading survives the cross-section. A stock the fund never owned rallied just as hard as the ones it did. That's the whole story.
The discount a forced seller pays is not a verdict on the asset. It's the fee for compressing time. What cleared on 30 July was not a portfolio. It was the market's belief that a levered book still had to come out.
I have named the observation that would prove me wrong, and it resolves within weeks.
The consensus, stated at its strongest
A large, solvent buyer publicly underwriting a battered asset class is real information. Citadel has the balance sheet, the risk systems and the reputation to make that signal credible, and it committed capital rather than commentary. If you believed the AI-infrastructure selloff was a solvency scare rather than a demand story, Citadel's bid was evidence for you.
I'm not going to argue that this is stupid. My claim is narrower: it's incomplete, and I think the incompleteness is measurable.
What the tape actually did
By mid-afternoon on 30 July, the cohort was up hard: Nebius +28.5%, Cipher Mining +28.3%, IREN +27.4%, Core Scientific +21.0%.
Cipher Mining does not appear in any reported holdings list for the fund, and SpotGamma's position mapping states it was never in the book.
Hut 8, which the fund had already exited, rallied with the group. A story about renewed confidence in AI predicts a broad rally roughly proportional to AI beta. It doesn't predict that a name with no connection to the seller trades in line with the very best of the names actually being sold.
That is what an overhang predicts, and it is why I read the session the way I do.
The objection worth taking seriously
These names move together every day. They're the same trade. So the fair challenge is: you have shown correlation that exists on any big up-day in AI infrastructure, and called it a mechanism.
Two things separate this from an ordinary co-movement day, to my eye.
The timing tracks a supply announcement. No fundamental changed. The cohort based through the evening of 29 July as the rumour circulated, then gapped at the 30 July open when the block trade hit the news. Nothing about the demand for AI compute changed between those two prints.
And the cross-section runs the wrong way for the alternative. If the rally were about those specific shares finding a home, the names actually being absorbed should have moved more than the ones that weren't. They didn't: a name with no reported connection to the seller moved with the best of them. Under the overhang read that's the expected result, because the discount was never on the shares.
To me that's a discriminator, not decoration, and it is the same logic the falsifier at the end proposes to test out of sample.
Liquidity is a service someone sells you
Grossman and Miller set this out in 1988 ("Liquidity and Market Structure", *Journal of Finance* 43(3), 617–633; the working paper is free to read at NBER w2641). Market liquidity is the demand and supply of immediacy. A liquidity event creates demand. So does the risk of waiting. Market makers supply it by being present and bearing risk during the interval between the arrival of final buyers and sellers.
I would read that last clause twice, because it is the entire mechanism. The market maker isn't paid for being right about the asset. He is paid for standing between a seller who must transact now and a buyer who has not arrived yet.
Worth sitting with what that interval actually is. A seller who must be flat by the close and a buyer who would happily own the asset next Tuesday are not in the market at the same moment. Somebody has to carry the position across the gap between them, wearing the price risk for as long as the gap lasts. That carry is the service, and the discount is its fee. Widen the gap and the fee rises, whatever the asset is.
Which means the fee is a function of two things that have nothing to do with the company: how fast the seller must be done, and how many balance sheets are willing to stand in the interval. A forced seller has set the first to zero. The second is where this particular week gets interesting.
So the discount on a forced sale is a price for time. What the securities are worth is a separate question. It tells you what the interval was worth. It tells you nothing about fair value.
Why the bid disappeared everywhere at once
Here is the part the confidence story cannot reach.
Shleifer and Vishny showed in 1992 ("Liquidation Values and Debt Capacity: A Market Equilibrium Approach", *Journal of Finance* 47(4), 1343–1366; free copy) that in industry distress, the seller and the next-best user of the asset are encumbered at the same moment, for the same reason. The people who would ordinarily bid are impaired by the identical shock. Assets then clear to buyers who value them less. The discount widens past anything the fundamentals justify.
Now apply it, which is where I think this gets interesting. Who are the natural best-use buyers of levered AI-infrastructure equity? Other funds running levered AI-infrastructure books. In July, they were nursing the same drawdown as the seller.
So the bid did not thin in the names being sold. It thinned across everything those buyers would otherwise have bought. That's the step I think the confidence reading misses. The cohort was marked down for a supply event that hadn't happened yet. Names the seller never touched were marked down alongside it, because the discount was never attached to the shares. It was attached to the absence of anyone able to buy them.
The paper's own conclusion is worth stating in full, because it is sharper than the way it usually gets quoted. When the natural owners cannot bid, the asset does not simply trade lower. It trades to somebody who values it less, which is a different and more permanent kind of loss. Price is the visible part; misallocation is the part that persists after the price recovers.
Carry that into a crowded equity book and the shape changes. The buyer who would pay the most for levered AI-infrastructure exposure is someone who already understands it, already models it, and already has the mandate to hold it. That describes the cohort. It also describes exactly who was nursing a drawdown in July. So the bid that disappeared was not the marginal bid. It was the best bid, and the best bid is the one that sets the price you get.
That is why a never held stock rallied like a held one. Both carried the same discount. The same news removed it.
I ran the full version of this for paying members. It carries the evidence table with the in book column, the sizing view, and why the standard overhang hedge is the wrong instrument for: The Overhang Trade: What the Citadel Block Actually Repriced.
