Skip to content

Own the Vacuum

The Melting Ice Cube#

An AI-era arbitrage that isn't — and the question underneath it I haven't been able to put down.

About a year ago I was sitting in a wine bar with an investor, having just finished a project for a client in a week that should have taken four months. I was marveling at myself — putting on a show, walking him through every detail of this beautiful solution to a genuinely hard problem, and how I'd built it with no employees at all. Just my AI agents: multiple Claude instances I'd configured through prompts to work alongside me like colleagues.

And somewhere in the performance I stopped and asked the question out loud: who captures this value?

Because here was the thing. A year earlier, someone would have had to pay my heavy salary for four months to get that outcome. They'd just gotten it for one week of my time. Call it $5,000 of cost for $80,000 of value. So who pockets the $75,000 of surplus?

You'd say "the client," and you'd be close. But here's the wrinkle: my client was itself providing professional services to its client. The savings didn't stop with the company that hired me. They flowed past it, downstream, to the customer it served.

I haven't been able to stop asking the question since. Who captures the surplus?

Fast-forward a year. I'm having coffee with a couple of investors — good coffee, and one of them lays a thesis on me. It goes like this. There's a window right now — three years, five, maybe seven — where you can still buy a small, profitable, unglamorous business at the multiple those businesses have always sold for. Five times earnings. A company doing one, three, five million in EBITDA, run by someone who's been at it thirty years and is ready to get out. Blue-collar, mostly: HVAC, industrial equipment, the services that keep the physical world running.

And he's far from the only one saying it — some version of this thesis is animating a whole wave of search funds right now, sharp young MBAs fanning out to buy these businesses before the window closes.

Here's what makes it a thesis and not just private equity. These businesses are prime for AI intervention. You buy one at five times — the price the market sets because it's the price it's always set — then you bring tools the seller has never heard of and squeeze out margin nobody's squeezed before. Two hundred basis points. Three hundred. Maybe five hundred, over time. And none of that upside is in the purchase price, because the person selling is in his sixties and has never typed a sentence into a chatbot. He doesn't know it's coming. There's a capture moment, my friend says — and if I were in your shoes, that's what I'd be salivating at.

It's a good pitch. It's specific, the math is clean, and the asymmetry is real: one side of the table knows something the other side doesn't, and the not-knowing is worth millions.

And here's the embarrassing part — I didn't see the flaw. I'd spent a year asking who captures the surplus, and here was a smart, confident person handing me an answer: the buyer does. You do. Go get it. I wanted it to be true.

I didn't see the flaw until the caffeine wore off.

The trade, on a whiteboard#

Let me step out of the room and put the trade on a whiteboard, because the arithmetic is the seduction.

You find a business throwing off $1 million a year in EBITDA — earnings, roughly; the cash the thing produces. Businesses like this sell for a multiple of that number, and for small, unglamorous companies that multiple has sat around five for as long as anyone can remember. So you buy it for $5 million.

Now you go to work. You bring in the AI tools the previous owner never touched and you automate away $500,000 of annual cost — a couple of roles that used to be people, a workflow that used to eat a week and now takes an hour. The business does exactly what it did before. It just costs less to run. EBITDA goes from $1 million to $1.5 million.

Then you sell. Same kind of buyer, same convention, same multiple: five times $1.5 million. $7.5 million.

You paid five. You sold for seven and a half. You pocket $2.5 million — for half a million dollars of cost-cutting you did with tools that cost you almost nothing.

And here's the part that makes people lean across the table: the magic isn't the $500,000. It's the multiple. Every dollar of EBITDA you add isn't worth a dollar — it's worth five, because that's what the next buyer pays for it. The multiple is a lever. You're not saving half a million dollars; you're manufacturing two and a half million in enterprise value, nearly out of air. Do it across a portfolio of these and you don't have a job — you have a machine.

On a whiteboard, it's the best trade in the world: buy a dollar of earnings for five, conjure a new dollar for pennies, sell it for five.

On a whiteboard.

But a business is not a whiteboard#

Here's what the arithmetic leaves out: the tools you used to cut that $500,000 aren't yours. You rented them by the month, the same as anyone can. Which means the company across town — your competitor, running the same HVAC installs or the same back office — can rent the exact same tools and cut the exact same $500,000. And he will, because you just proved it's possible.

Now watch what happens to your incremental half million gain in EBITDA.

When two competitors both have a lower cost floor, that floor becomes a weapon. One of them — maybe the hungry one, maybe just the nervous one — decides to win some business by charging less, because now he can and still make money. To hold your customers, you match him. He cuts again. You match again. Price competition reopens, and it reopens at the new, lower cost base. The 500,000 you saved doesn't sit in your pocket as margin; it walks out the door as lower prices. Your costs fell by half a million — but so did your revenue, and the 1.5 million settles back toward 1 million.

The money didn't vanish. It flowed downstream, to your customers, who now pay less for the same work — which, if you were paying attention in that wine bar a year earlier, is exactly where it always goes.

Here's the principle the whiteboard hides: a cost advantage is only profit if your competitors can't copy it. If they can, competition turns your saving into the customer's discount. And AI is, by its very nature, the most copyable advantage in the history of business — a general-purpose capability that anyone, including the man you're trying to beat, can switch on for a few hundred dollars a month. An AI cost-out isn't an advantage. It's the new table stakes. You don't get to keep the margin. You get to keep existing.

We've seen this movie. When the spreadsheet arrived, every accountant in America got faster overnight. Did fees hold and margins double? Of course not — fees fell, because every accountant got faster, and the savings went to the clients. Email, the internet, the CNC machine: each handed early adopters a brief, beautiful margin, then handed the same gift to everyone else, and the gift became the price of admission. AI is that same story, running faster and reaching wider than any of them.

So this is the melting ice cube. You bought $1.5 million of earnings — real, verified, sitting right there on the day of the sale. But some of that ice was always going to melt in your hands on the walk to the exit. And what melted didn't disappear. It ran downhill, to the customer, exactly as it has every other time.

A window you can't fit through#

The strongest defense of the trade is that I've described the melt as if it's instant, and it isn't. The ice doesn't vanish the moment you close; reversion takes time — months, while competitors notice, adopt, and start cutting. So there's a real window: buy early, harvest the margin before it melts, get out. The window is genuine.

It's just not one you can fit through.

Look at the actual clock. The window — the months before competition catches up — is what you're racing. But the transaction is slower than the race. Sourcing a business, signing an LOI, running diligence, and closing is six months to a year by itself. Then you have to implement the AI, run the business long enough to show the lift on a trailing twelve months — no buyer pays for margin you booked last quarter; they need to see it hold — then re-market and sell. Another year, easily. Your harvest window is measured in months. Your round trip is measured in years. By the time you can hand the business to the next person, the ice you bought it for is a puddle.

That's the quiet contradiction in the pitch. The same people describing a three-to-seven-year window are running a three-to-seven-year hold — which means they aren't flipping inside the window at all. They're buying the business and owning it straight through the melt. And even if you could somehow buy, juice, and flip inside a single year, you'd still need a buyer at the far end who hasn't done this arithmetic. The first wave finds those buyers. The fifth wave is selling to people who price AI-juiced earnings for exactly what they are. The window shuts on the exit even faster than it shuts on the entrance.

Timing, then, won't save the trade. But the buyers are right about one thing: not every business melts. Some keep the gain entirely. The question is which ones — and why.

The multiple is a diagnosis#

The businesses that keep the gain are the ones with a moat — something the competitor across town can't copy, no matter how many AI subscriptions he buys. If your customers can't easily leave — because switching is painful, because you hold data no one else has, because regulation favors the incumbent, because the network grows more valuable the more people are on it, because you're the only game in your town — then the competitor can't force your price down, and the 1.5 million stays 1.5 million. The ice doesn't melt, because the moat keeps the sun off it.

A few months ago I wrote an essay called The Cost of Software Is Now Zero. In it was a rubric: score a business across a handful of dimensions — switching costs, regulatory position, proprietary data, network effects, problem complexity — and you get a read on whether AI commoditizes it or can't touch it. I built it to tell a software founder whether to panic. It turns out it answers the acquirer's question just as well, because it's the same question wearing a different hat: is the advantage defensible, or merely replicable?

Now run the search-fund thesis through it. The cheap, unglamorous, five-million-dollar businesses — the HVAC shops, the back-office services, the ones whose aging owners "don't know any of this" — where do they score? Low. They score low because their advantage was never deeply defensible to begin with; they were coasting on a sleepy local market and a competitor who hadn't bothered. Which is precisely why they're available at five times earnings, from an owner happy to walk away. The multiple isn't a mispricing waiting to be exploited. It's a diagnosis. A business that changes hands at five times its earnings is very often telling you, in the only language balance sheets speak, exactly how much moat it has: not much.

And the businesses that would hold the AI gain — the ones with real switching costs, real data, real regulatory walls? Their owners know exactly what they're sitting on. They are not selling at five times to a kid out of business school. The sophisticated end of the M&A market sorted this out a while ago; it's been quietly paying up for defensibility and discounting raw margin for years now. So the window everyone's salivating over is, mostly, a window onto the meltable businesses — because the screen the thesis uses, "cheap and easy to improve with AI," is very nearly a perfect filter for ice.

Who captures the surplus?#

A year of asking, and the answer turns out to be the oldest one in economics — the one nobody wants to hear. Mostly, the customer. The economist William Nordhaus once estimated that the people who produce innovations capture only about two percent of the value those innovations create. The other ninety-eight flows past them — past the inventor, past the company, past the acquirer — downstream, until it hits something that can stop it. Usually nothing does, and it lands in the lap of whoever buys the finished thing a little cheaper than they did last year.

That's what I'd seen in the wine bar without naming it. The surplus from my impossible week didn't stop at me, and it didn't stop at the company that hired me. It ran downhill to their customer. And it's what the search funds are about to learn at a much larger scale: the EBITDA they conjure with AI won't stop at them either. It'll run downhill to their customers — because the thing they used to conjure it is the same thing their competitors are using, and the same thing dissolving the walls that used to make the running stop.

That last part is what should give the salivating buyers pause. AI doesn't only hand you a new way to create surplus. It is, at the same time, a solvent for the moats that used to trap it — the proprietary process, the specialized expertise, the software that used to be hard to build. It pours value into the stream and washes out the dams in the same motion. So in the AI era, more value flows further downstream, faster, than ever before. Wonderful, if you're the customer. Treacherous, if you just paid five times earnings betting the stream would pool in your reservoir.

None of which makes the buyers villains. They're not preying on the old men so much as about to join them on the wrong side of the same misunderstanding. The seller is mispricing the near-term gain he could have captured and didn't. The buyer is mispricing how long that gain will last once he has it. Two smart parties, both wrong about AI in opposite directions, shaking hands over a melting ice cube.

If you're going to play this game anyway — and people will — there are only two questions that matter, and you should ask them out loud, of every deal: What stops everyone else from doing exactly what I'm about to do? And how long do I have before they do it? If you can't answer both, you are not buying earnings. You are renting them, at a purchase price, with the meter running.

Because here's the thing the whiteboard never shows you. In a world where intelligence is cheap and getting cheaper, the scarce thing isn't the ability to improve a business — everyone will have that. The scarce thing is a business the improvement can't leak out of. The durable money was never in capturing the surplus. It's in owning the one rare thing AI can't melt: a reason the customer can't leave.

[SCAFFOLD — not yet drafted. These are the remaining Part I beats per architecture v0.3. Draft in later passes; decide whether they fit this piece or split into their own articles — see open question to Ben, 2026-06-29.]

One rung too high (Nadella) — [SCAFFOLD]#

  • Steelman the smartest optimist; he confirms the machinery (models commoditize expertise = the diamond→water solvent; "own the learning loop" = build a moat the solvent can't dissolve) but argues one altitude too high: his "broad distribution" is across firms, not people. The worker's version of his own advice — be an owner of a loop, not a contributor to someone else's — is the hinge into the next beat.

Own the vacuum — [SCAFFOLD]#

  • If the only rung above labor is ownership, "moving up" means becoming capital — owning the machine replacing you. The ticket is prior capital, so it concentrates. The honest shape of "it's political": redistribute equity, not income, because the wage channel is closing. The ramps that once let labor reach capital are being removed one by one.