Hedge funds hold a record $5.6 trillion, and a Fed note puts the Treasury basis trade at about $830 billion. Every position in it leans on the same overnight repo and futures margin. What platform engineers can take from a crowded exit: name the shared channel, write the eviction order, and rehearse it.
On Sunday, March 15, 2020, the Federal Reserve announced that it would increase its holdings of Treasury securities "by at least $500 billion" to keep the Treasury market working. Eight days later it dropped the number. The Fed would now buy "in the amounts needed to support smooth market functioning."
Treasuries are the asset the rest of finance treats as the floor, and that month the Fed stopped naming a limit. The Financial Stability Board later described one of the reasons. As volatility rose, "margins on US Treasury futures positions increased, as did borrowing rates in US Treasury repo markets." Those rising costs "motivated traders active in the US Treasury cash-futures basis to partially unwind their positions, adding to the significant volumes of sales from other market participants and potentially amplifying market instability."
That trade is the Treasury basis trade. In a note published in June 2026, a Federal Reserve economist estimated it at about $830 billion as of September 2025, roughly double its early-2020 peak. Hedge funds, the machines that run trades like it, have never held more capital, and the regulators who watched March 2020 have been saying so in writing.
By HFR's count, hedge funds held a record $5.6 trillion at the end of June 2026, the 15th consecutive quarter of growth. Most of the recent growth is returns rather than new money: of the record $409.3 billion added in the second quarter, $364 billion was performance gains. But the money is also arriving. Net inflows in 2025 were $115.8 billion, the strongest calendar year since 2007.
Some of that capital sits in multi-strategy platforms, the firms the industry calls pod shops. Two of them publish their own numbers. Millennium, founded in 1989, lists more than $95 billion in assets, more than 7,100 employees and "370+ Investment Teams" on its homepage. Point72 lists "200+ Investing teams" and about $58.5 billion in assets as of July 1, 2026, under the line "Multiple Strategies. Infinite Opportunities. One Firm."
Point72 describes its role as giving its people "the resources and autonomy to turn great ideas into reality." Hundreds of teams inside one firm is division of labor taken about as far as it goes. Adam Smith's pin factory in The Wealth of Nations is the standard picture: making a pin divided into "about eighteen distinct operations," so that ten workers "could make among them upwards of forty-eight thousand pins in a day," where each alone "could not each of them have made twenty, perhaps not one pin in a day." Read as a structure, a pod platform divides the work of taking risk the same way. Each team runs a narrow book it knows well. The platform supplies the capital, the financing, the risk limits and the systems, and it decides how much each team gets.
Neither firm describes its edge in these words, so take this as my reading of the structure rather than theirs. The platform's advantage is not any single trader's insight. It is what happens when you add hundreds of narrow books together.
If most teams' gains and losses do not move together, the total swings far less than the sum of its parts. A platform that trusts that math can give each team more room than it could safely give any one fund standing alone, because the bad days rarely arrive together.
Software engineers know this bet from cluster schedulers. Google's Borg paper states it: "Rather than waste allocated resources that are not currently being consumed, we estimate how many resources a task will use and reclaim the rest for work that can tolerate lower-quality resources, such as batch jobs." In a median Borg cell, about 20% of the workload ran in reclaimed resources. The paper is equally plain about what happens when the estimate is wrong: "If this happens, we kill or throttle non-prod tasks, never prod ones."
Borg's designers wrote down, in advance, who gets evicted when the independence assumption fails. In a market that runs on borrowed money, eviction is a margin call and a forced sale, and the order is set by whoever lends.
The basis trade is the kind of book a netting machine is built to carry. A fund buys a Treasury bond, borrows against it in the repo market to pay for it, and sells the matching futures contract. The profit is the small price gap between the bond and the future, which the OFR writes as "gross basis = Treasury security price - (futures price x conversion factor)." The Office of Financial Research put it in one line: "Because profits for this trade are small, hedge funds scale positions by relying heavily on leverage through repo financing and futures margining." The Fed note says where the borrowing power comes from: "low or zero percent haircuts on the repo borrowing and low margin requirements on the futures."
A caveat belongs here, because the public record is careful about it. Advisers report their funds' positions to the SEC on Form PF, and the published analyses are aggregates, so the data I read do not say which firms hold the trade or how much. I am not claiming the pod shops run it. The claim is narrower. It is a thin spread run on a lot of borrowed money, and that is what a platform built to net risk can carry at scale.
The June 2026 Fed note, by Phillip J. Monin, decomposes large hedge funds' Treasury positions using Form PF data through September 2025. Their gross Treasury exposure doubled from 2023 to $4.0 trillion, of which $2.4 trillion was long and $1.6 trillion short. The basis trade, at about $830 billion, was 35% of the long side. Hedge funds held about 8.5% of all privately held Treasuries by market value, up from 4.5% at the start of 2023, and their holdings "now exceed those of mutual funds and U.S.-chartered depository institutions."
The Office of Financial Research measured it from another direction in August 2026. By its estimate, hedge funds' long cash Treasury holdings reached $2.0 trillion at the end of 2025, nearly three times the level of five years earlier, while marketable Treasury debt grew 29% to $28.9 trillion. That makes a record 7% of marketable Treasury debt. Their short futures positions totaled $1.4 trillion. The two percentages use different denominators, privately held Treasuries for the Fed and all marketable debt for the OFR, which is why they differ.
No filing shows the trade directly. Hedge funds "do not separately report cash Treasury securities and derivatives on SEC Form PF," the OFR notes, and published estimates of the trade's size have ranged "from $350 billion to $1.5 trillion."
What the Fed can see is concentration. The 50 largest funds by gross Treasury exposure account for about 90% of the total, up from 84% at the start of 2023.
None of this makes the trade a villain. The Fed note says it "plays an important role in price discovery and liquidity provision." The OFR adds a reason the public might want it. Primary dealers have traditionally backstopped new Treasury issuance, but capital rules written after the 2007-09 crisis, in part, left them less room to hold bonds, and hedge funds are not subject to those constraints. That borrowing, in the OFR's words, "allows hedge funds to absorb more Treasury issuance at a time when primary dealers face balance sheet constraints." The government sells a great deal of debt, and this machine helps it get bought.
The trouble is that every position in the trade leans on the same two supports. The OFR spells them out: "Hedge funds typically rely on overnight repo because it is cheaper than term financing. If repo rates spike or futures margin requirements increase abruptly, hedge funds may be forced to unwind positions rapidly."
Networking has a name for what happens next. In October 1986, as Van Jacobson and Michael Karels recounted in 1988, the data throughput between Lawrence Berkeley Laboratory and UC Berkeley, "sites separated by 400 yards," "dropped from 32 Kbps to 40 bps," a "factor-of-thousand drop in bandwidth." It was the first of a series of congestion collapses. Each sender followed its own rules, resending whatever had not arrived, and together the resends kept the shared link overloaded. When they asked whether Berkeley's TCP was misbehaving, their answer was "yes." The fix was not better judgment in each sender. It was a rule built into TCP itself: when the network signals congestion, back off. Their paper put the ambition modestly. If the principle were obeyed, "congestion collapse would become the exception rather than the rule."
A trade built on borrowed money has the same shape. A margin increase or a repo spike is the congestion signal. Each fund's sensible response, cutting the position, means selling Treasuries into the same dealers at the same moment, and the selling is what makes the next signal worse. The FSB's account of March 2020 adds the detail that makes this hard to see coming. Funds could borrow "with near-zero haircuts," and the lenders "may have lacked sufficient visibility into clients' overall positioning." Each lender saw its own slice.
The record also shows how fast a crowded exit can move, in a neighboring trade. After the April 2025 tariff announcements, about $60 billion of hedge funds' swap-spread arbitrage "unwound rapidly before recovering within months." That was not the basis trade, and the Fed note does not say the basis trade broke that spring. It shows how quickly a relative-value position built on borrowing can empty out when the signal arrives.
The Fed note ends on the combination rather than any one number: "the combination of large scale, high concentration, and elevated leverage creates the potential for systemic stress if multiple strategies face simultaneous pressure or if severe shocks affect the largest participants."
The incentive problem is my framing, not the regulators'. The spread belongs to the funds. Nothing in one fund's own numbers charges it for the crowding, because its exit alone looks small against the market. And the worst case lands on the public balance sheet, in a press release promising purchases "in the amounts needed." The FSB states its own goal in those terms: to "reduce the need for extraordinary central bank interventions."
The responses the record describes look more like Jacobson's fix than like a lecture to traders. The FSB ties the near-zero haircuts to the lack of "supervisory guidance on maximum leverage" in repo that is not centrally cleared, and it lists mandatory central clearing of Treasury and repo trades among the measures adopted since. In the United States, the SEC's clearing requirement takes effect for eligible cash Treasury trades at the end of 2026 and for repo by June 30, 2027. Clearing puts the trades through a common counterparty with common rules, a protocol in place of bilateral agreements. The Form PF decompositions, the Fed note says, give regulators "a framework for ongoing monitoring," which is the aggregate view no single lender had in 2020.
If you run a shared platform, you are running a small version of this machine. A Kubernetes cluster that overcommits, a multi-tenant database, an API gateway fronting one rate-limited provider, a CI fleet sized for the average day: each makes money, or saves it, by betting that its tenants will not all peak at once. The bet is usually right, and that is the danger, because it trains you to forget it is a bet.
The practical move is an exit drill, and it has three parts.
Name the channel every tenant shares. The basis trade's positions look independent until you notice that they all roll overnight repo and post futures margin. Yours probably share a queue, a region, an identity provider, a database connection pool or a vendor's rate limit. Put that channel on the dashboard, not just the tenants. The FSB's complaint about 2020 was that each lender saw only its own slice.
Write the eviction order before the crowd arrives. Borg's rule was one sentence: non-prod goes first, prod never. Decide what gets throttled, in what order and on what signal, and make it automatic, the way TCP made backoff automatic. Otherwise the order will be set in the moment by whoever controls the shared channel, and it may be your largest customer. Then rehearse it: once a quarter, push every tenant against the shared channel at once in a staging copy, and check that the throttling fires in the order you wrote down.
Watch concentration as it drifts. The top 50 funds went from 84% to 90% of large hedge funds' gross Treasury exposure in under three years, and nothing broke. If your largest tenants' share of load is climbing, your independence assumption is getting weaker even while every graph stays green.
The machine will keep earning its spread, and most years that is good for everyone, the Treasury included. The question the regulators keep putting in writing is who plans for the day everyone leaves at once. In March 2020 the answer was the Fed, on a Sunday.
Sources:
Agent Trust Stack bundles Chain of Consciousness, a tamper-evident, dated record in hash-linked entries, with the Agent Rating Protocol, a portable measure of how agents performed. Record the order you wrote and each quarterly rehearsal against it, and “did the throttling fire in the order we wrote down?” has an answer you can check.
pip install agent-trust-stack npm install agent-trust-stack
Or start without installing anything: Hosted Chain of Consciousness.