The Exit Is a Mirror

Private equity spent last year doing something strange: selling companies to itself. The secondary market ran $240 billion of volume in 2025, a record, and $115 billion of it was GP-led: mostly continuation vehicles, where the same firm sits on both sides of the table, moving an asset out of its old fund into a new vehicle it also controls, setting the price, picking the asset, and collecting fresh fees on the way through. Continuation funds were a niche product in 2019. By 2024 they accounted for about one exit in seven, and 2025 set a record with 147 of them.

The industry files this under liquidity management. I read it as a confession. The sharpest sellers in the world looked at their best assets, ran the auction, and found that the highest bidder was the face in the mirror. When your exit is a mirror, the exit was the problem.

The problem is a clock. A private equity fund is a limited partnership with a contractual life of about ten years, give or take an extension. Money goes in during years one through five, and everything must be sold and returned by the end, which in practice means every company in the portfolio gets re-sold somewhere between year three and year eight regardless of what is happening inside it. Bain’s latest global report puts typical holds around seven years now, up from five or six.

Where I stand, for calibration: I was a founding engineer at Didero, where we built AI agents that ran procurement, about $600K of order volume flowing through daily when I last counted, and these days I deploy AI agents into other companies’ operations at Decagon. For most of my Didero year I treated the capital structure behind our customers as somebody else’s plumbing. I was wrong about that, and the essay where I figured out why is the prequel to this one: institutions grant software authority slowly, one workflow at a time. This essay argues one position without apology. For the businesses AI is about to transform, the holding company — permanent capital, decentralized operations, centralized capital allocation — is the right structure, the ten-year fund is the wrong one, and the gap between them is widening.

The Clock Is the Strategy

Start with the metric. A fund’s headline number is IRR, and IRR rewards speed above everything.IRR (internal rate of return) is time-weighted: turning $1 into $2 in three years scores roughly 26%; the same doubling over nine years scores 8%. MOIC (multiple on invested capital) just counts the dollars that came back. A manager marketed on IRR and an investor who eats MOIC want different things. Return money in year three and a mediocre deal looks brilliant; return the same multiple in year nine and a good deal looks tired. Ludovic Phalippou at Oxford has spent a career documenting how the metric diverges from what investors actually receive. The fund structure doesn’t merely permit selling winners; it is a machine built to sell them on a schedule, and the people operating it are paid to distribute early.

The cost of that machine has been measured, by the people who ran it. Sequoia studied its own distribution history and found that had it held its distributed shares just twelve months longer over the prior fifteen years, its LPs would have collected over $8 billion more. One line item: Square, bought at $0.95 a share in 2011, distributed at $71.95 in December 2019, after which the stock more than doubled within a year. Peter Lynch called this pulling out the flowers and watering the weeds, and Buffett liked the line enough to phone him for permission to reuse it. In October 2021 Sequoia restructured its entire US and European business into a single open-ended fund, with Roelof Botha writing that the change “removes all artificial time horizons on how long we can partner with companies.”

The rest of the asset class is arriving at the same place by uglier routes. Phalippou’s other famous paper found PE funds have returned roughly the same as public equity indices since at least 2006, while collecting an estimated $230 billion in carry along the way. Sponsor-to-sponsor sales, one fund selling to another fund with the same clock, hit 30% of PE exits in early 2025. The CIO of Amundi, Europe’s largest asset manager, described parts of the industry as circular in terms blunt enough to make headlines. And the continuation-vehicle boom is the industry’s own verdict: when a GP moves a company into a vehicle it controls so it can keep owning it, the GP is telling you, with its own money, that the ten-year wrapper destroys value it would rather keep. The conflicts are real; regulators tried to require independent fairness opinions on every such deal for a reason.

But the underlying instinct is correct. They found something that compounds, and the wrapper keeps making them let go of it.

What Never Selling Actually Earned

The counterfactual has a sixty-year track record. The best capital-allocation records of the modern era belong, with strange consistency, to permanent-capital serial acquirers, across decades, countries, and industries.

Henry Singleton compounded Teledyne at 20.4% a year from 1963 to 1990 against roughly 8% for the S&P 500, then, when his own stock got cheap, repurchased about 90% of it across eight tender offers. Buffett, quoted in 1980, called it “the best operating and capital deployment record in American business.” Buffett’s own version just ran longer: Berkshire compounded at 19.9% annually from 1965 through 2024 versus 10.4% for the S&P 500, an overall gain of 5,502,284%. The engine was insurance float, which Buffett describes as “money we hold but don’t own,” plus a stated refusal to sell: “when we own portions of outstanding businesses with outstanding managements, our favorite holding period is forever.”

The software version is Constellation, public in 2006 at C$17 a share, around C$2,700 twenty years and a thousand-plus acquisitions later, roughly 29% a year in price terms before counting the Topicus and Lumine spin-outs shareholders received along the way. Its stated objective is to be “a perpetual owner of inherently attractive software businesses”; it has reportedly divested exactly one business in its history, a sale Mark Leonard said he regretted; and its Volaris operating group literally markets “buy and hold forever” to prospective sellers. The industrial version is Danaher, built by the Rales brothers from a 1984 shell company into a run so strong that by 2004 Bloomberg was writing headlines about them eclipsing Buffett, 31% a year over the trailing twenty.The pattern survives large-sample scrutiny: McKinsey’s study of the world’s 2,000 largest companies finds programmatic M&A — many small deals, executed continuously — is the only acquisition strategy that systematically outperforms; about two in three programmatic acquirers beat their peers.

Four mechanical advantages produce this, and none of them is mystical.

Taxes. US tax law lets an affiliated group move cash between subsidiaries without a taxable event; a consolidated return eliminates intercompany dividends entirely. A fund’s exit-and-distribute cycle hands the government a slice at every hop. Buffett described Berkshire’s deferred tax liability as resembling “an interest-free loan from the U.S. Treasury that comes due only at our election”.

Reinvestment optionality. Cash from a mature subsidiary can fund the best marginal opportunity anywhere in the portfolio, indefinitely, with no re-underwriting, no fundraising cycle, and no fees taken twice. Koch Industries, private for its entire history, says it reinvests about 90% of its earnings, and no outside shareholder can demand otherwise. Leonard’s final president’s letter, in February 2021, was entirely about this: a Constellation director had argued for years that the company invests capital better than most of its shareholders could and should never return a dollar of it. “I have stopped arguing,” Leonard wrote. “I have converted, and with the fervour of the newly converted, I am busy demonstrating my new-found faith.”

Incentive horizon. Constellation’s senior executives must put 75% of their after-tax bonus into stock bought on the open market and held for years. Compare 2-and-20 on a fund clock, where the rational move is to exit the winner while the IRR still sparkles and raise the next fund on the print.

Seller selection. This one is underrated. A founder who cares what happens to her company and her people will sell to a credible permanent owner at a price a flipper cannot match, because part of what she is buying is the promise. A fund on a clock cannot make the promise, and every seller knows it.

Why AI Raises the Stakes

Everything above was true before language models. Here is what changes.

First: the value curve is back-loaded. This is the part I’ve seen with my own eyes. Model capability never held us up at Didero; getting a controller to let the system touch her ledger did, one workflow at a time, and I’ve written about that trust curve before. It is measured in quarters and years. Elad Gil, making the bull case for AI rollups, talks about taking gross margins from 10% to 40%. Whatever you think of the number, nobody gets it in year one; the early years of a services transformation go to integration, exception discovery, and the slow reassignment of decision rights, and the margins arrive after the trust does. An owner on a fund clock eats those years and then must begin staging the exit just as the curve inflects. The permanent owner keeps the back half.

Second: the operating system finally travels at machine speed, but only inside a trust boundary. Danaher published its edge decades ago: DBS, a kaizen-derived operating system installed into every acquired company through training and repetition, at the speed humans teach humans. The modern equivalent is the exception library, the accumulated memory of how a thousand edge cases got resolved. At Didero I watched a vendor quirk fixed for one customer on Monday become the default for every similar customer by Tuesday morning, propagation DBS could never achieve with any number of kaizen events. I’ve seen that within procurement, across customers sharing an exception taxonomy; whether it transfers across industries is still a hypothesis. The structural point survives the weak version. Shared operational memory is a data asset with legal edges. It propagates freely under common ownership, common liability, and common incentives. Divest a subsidiary and you amputate its contribution to the memory, and the lawyers will spend a year arguing about who owns what it already taught the rest of the portfolio.

Third: the conglomerate discount was an attention tax. The canonical finding, Berger and Ofek’s 13-to-15% average discount on diversified firms, dates from 1995.Berger & Ofek (1995) imputed stand-alone values for each segment of US multi-segment firms, 1986–1991, and found the whole traded 13–15% below the sum of the parts. Later work (Campa & Kedia 2002; Villalonga 2004) argues much of the discount is a selection effect: firms that diversify were already discounted before they did. Take the discount at face value and ask why it exists: a small headquarters cannot genuinely understand capital allocation across unrelated businesses. It is the same bounded-rationality problem I saw in procurement, one level up; the map doesn’t fit in the allocator’s head, so capital gets peanut-buttered, cross-subsidized, and mispriced. The best holdcos beat the tax with radical decentralization. Leonard: “my personal preference is to instead focus on keeping our business units small, and the majority of the decision making down at the business unit level.” Orchestration attacks the same problem from the other side, by making the portfolio legible: uniform decision traces, comparable cost per successful outcome, exception rates you can read across fifty subsidiaries the way Singleton read cash statements. Nobody has run a holdco on this stack long enough to prove that. File it under argument.

There is a counter-current, and it deserves its own paragraph. AI raises the span of control and, at the same time, the disruption rate of the assets themselves. Constellation, the arch-holdco of software, has spent the past year in a drawdown fed partly by fears that AI erodes vertical software moats, and when Mark Leonard resigned for health reasons in September, the stock fell roughly 17% in a week. Permanent capital is a license to hold, not a license to stop underwriting change, and the market repriced both risks at once that week: the allocator was gone, and the things he had allocated into might not be permanent after all.

The Graveyard

The graveyard is bigger than the pantheon, so walk it first.

The 1960s conglomerates were the first mass experiment in permanent diversified ownership, and most of them were accounting tricks wearing a philosophy. The game was earnings-per-share arbitrage: use your high-multiple stock to buy low-multiple companies, report the combined earnings at your multiple, and call the arithmetic growth. Litton Industries ran 57 consecutive quarters of growth until January 1968, when earnings came in at a third of the prior year and a stock trading at 40 times earnings fell from around $90 to $53. Jimmy Ling built LTV into the 14th-largest company in the Fortune 500 in fourteen years; by 1970 the stock had gone from $167 to about $11 and his own board removed him.

When the multiple machine ran backward, there was nothing underneath. And the sequels kept coming. GE peaked at $600 billion in August 2000 as the most valuable company in the world, destroyed over half a trillion dollars of value as GE Capital’s leverage and two decades of top-of-market capital allocation came home, and completed its three-way breakup in 2024. Tyco’s Dennis Kozlowski, celebrated on magazine covers and dubbed “Deal-a-Day Dennis” by the business press, announced 88 deals in his first six years and later acknowledged roughly 700 more acquisitions that were never announced at all; he left in handcuffs. Valeant rolled up pharma with debt, gutted R&D, raised prices, briefly passed RBC as Canada’s most valuable company in 2015, and lost more than 90% within two years. The base rate is quantified: Carroll and Mui, working from a study of 750 major corporate failures, report that more than two-thirds of roll-ups fail to create any value for investors. Even the friendly modern cases wobble; Tiny Ltd, pitched for years as the Berkshire of the internet, trades well below its listing price and is on its third CEO in two years. The pantheon has its own memento mori: Teledyne did not outlast Singleton in recognizable form, and whether Constellation outlasts Leonard is being tested in public right now.

So what separates the compounders from the wreckage? The winners bought cash flows and paid for them with cash flows; the losers bought earnings optics and paid with inflated paper. The winners decentralized and left acquired businesses intact; the losers integrated, synergized, and managed by press release. The winners set hurdle rates centrally and let a small office say no; the losers had a charismatic CEO who could not stop. The winners put managers’ bonuses into stock they had to hold; the losers put them into shower curtains. And the winners kept leverage boring. Valeant’s capital was permanent right up until its creditors decided otherwise.

The graveyard is about to get a new wing, and Benaich and Mrkšić’s critique of AI rollups is the right admissions test: buying services EBITDA and bolting on a chatbot is operational improvement, not business-model transformation, and public markets already price the difference. Their exhibit is Concentrix, which deployed generative AI with over a thousand clients and still trades at low-single-digit EV/EBITDA on 10% margins; the market’s exhibit is Capgemini buying WNS at roughly 2.5 times revenue, a services price for an “agentic AI” story. The holdco structure protects you from forced selling. It does not protect you from buying labor and calling it software.

The Bar

Put the history and the technology together and the specification writes itself. Five requirements, none optional.

Permanent capital, honestly structured. A C-corp, an evergreen vehicle, or LPs who signed up for forever in writing. Not a ten-year fund with a marketing deck about patience. General Catalyst funded its hospital acquisition through a permanent-capital vehicle precisely because the work would not fit a standard fund timeline. The continuation-vehicle version, permanence rebuilt deal by deal with the sponsor setting its own price, is the same instinct with worse governance.

Decentralized operations, centralized only where compounding lives. Capital allocation and the shared exception memory sit at the center; everything else stays at the subsidiary. GE centralized everything else, and its dashboards were excellent right up to the end.

Underwrite operations, not re-rating. McKinsey attributes about two-thirds of buyout returns in the last cycle to leverage and multiple expansion. With $1.3 trillion of dry powder chasing deals, that trade is crowded and rate-dependent. The AI holdco’s model has to clear its hurdle with exit multiples frozen; if the thesis is right, there is no exit to re-rate.

Owner incentives all the way down. Constellation’s 75% rule, extended to the people who actually build the systems. An engineer with equity in the holdco has a reason to promote her exception fix to the global library; a contractor billing hours does not have one.

The orchestration test, applied before every acquisition. Who will own the exception path after close? What is the cost per successful outcome today, and what is the credible path down? If the answer involves the word “copilot” and a headcount target, walk away. That deal is Valeant with GPUs.

The current field is at least attempting the shape. QXO is the purest capital-structure play and tellingly the least AI-branded: about $5 billion of equity raised in 2024, roughly $30 billion now deployed into building-products distribution, TopBuild closed two days ago, and Brad Jacobs, on at least his fifth act, states his integration method as asking the frontline what’s broken and listening. Thrive Holdings launched last April as an explicit permanent-capital vehicle and is putting about a billion dollars into consolidating accounting firms; in December it gave OpenAI an equity stake in exchange for embedded engineering teams, which buys real technical depth at the price of everyone involved grading each other’s homework. Metropolis is the one that has already run the full loop, sell the software, stall, buy the market instead, and its November raise at a $5 billion valuation is the market agreeing the loop worked. Behind them: Long Lake in HOA management, General Catalyst’s hospital, Alpine sourcing 18,000 deals to close 190 by its own scorecard. Most will fail; the base rate says so. The structure removes exactly one failure mode, the forced sale of a winner, and leaves every other one intact.

What the Clock Was Hiding

Leonard called his own conversion a religious experience, and from a man that dry, that was the loudest available register. The best software capital allocator of his generation ended twenty years of letters by arguing his shareholders should never see their cash again.

Two years of deploying agents into companies, procurement then broader operations, taught me a smaller version of the same lesson. Trust is the slowest asset I have ever watched accumulate. It arrives in quarters, workflow by workflow, signature by signature, and it is specific: to this system, this team, this owner, these receipts. I have never watched a deployment survive a change of control, and that is partly the point; I have never seen one try. What I have seen is what the trust is physically made of: a named controller who owns the kill switch, thresholds with her initials on them, a risk budget the CFO signed after two months of arguing. None of that is in the data room. All of it reports to whoever owns the company next, starting from zero, on a clock that is already running.

The ten-year fund asks how fast you can sell what you’ve built. The holdco asks what you would build if you never had to answer that question. AI is turning service businesses into accumulated trust and accumulated memory, and for assets made of those two things, only the second question is worth asking.