In 1895 you could walk into an American factory and see the future bolted to the ceiling. A steam engine in the basement turned a line shaft that ran the length of the building, and every machine on the floor hung off it by leather belt. When electric motors arrived, owners did the obvious thing: unbolted the steam engine, wired a big motor in its place, and kept the shaft. The shaft was the factory. Nothing improved. Measured productivity from electrification didn't show up for thirty years, because the gain was never in the motor. It was in tearing out the shaft and rebuilding the whole floor around dozens of small motors, one per machine, work flowing in lines instead of around a spinning pole. The technology took a decade to arrive. The payoff waited a generation, and most of it landed in new factories built by new owners, not in old factories that bought motors.

This month Dario Amodei published an essay telling governments to get ready: regulate frontier models like aircraft, build wage insurance and retention credits, prepare for the "decent possibility" that AI causes enduring job loss. The economic logic underneath it is the one everybody now carries around, and it's worth stating in its strongest form because it's correct. AI cuts the cost of cognitive work. A firm that adopts it either produces more with the same people or the same with fewer, and a firm that refuses watches competitors' costs fall until it adopts late from a weaker position or dies. Margin pressure makes adoption compulsory. The only open question is which way each firm harvests the gain, and the scary answer is payroll. Hence the essay, the displacement frameworks, the whole policy apparatus bracing for the layoff wave.

The logic is right and the wave still isn't coming, at least not the way anyone is bracing for it, because it assumes the firm is the unit that adapts. It isn't. Inside the firm, neither branch fires.

Start with the firing branch. No manager wants to cut their own team, and the reason is arithmetic, not kindness. Headcount is budget, budget is status, status is the next promotion. Nobody in the history of corporate America has empire-built downward. Yes, layoffs exist — the tech giants cut tens of thousands when the market demanded a number, and a few CEOs now say the quiet part about agents and support tickets. But look at where those cuts came from: boards and markets, from above, over the org chart's objection, mostly unwinding a hiring binge rather than redesigning anything. The pressure that produces cuts always originates outside the building, which is the tell for everything that follows. Inside it, firms that "embrace AI" don't fire anyone; they stop backfilling. Attrition runs ten to fifteen percent a year and does the work with no layoff event, no headline, no manager's fingerprints on any decision. The displacement is real but it exits through the quietest door in the building: the entry-level job that never gets posted. The wave Dario's framework is built for arrives looking like twenty-four-year-olds who can't get hired anywhere, not like badge boxes in the lobby.

And the firing branch understates its own problem, because the slack it would cut isn't marginal, and it isn't new. There's a confession genre running on Twitter right now: a 38-year-old lists his friends in the $200-400K corporate corridor. One makes $250K as a manager at a biotech and spends the workweek sending pictures of his chicken coop. One plugs a mouse jiggler into his laptop and goes to the gym, homeschools his kids, does yard work, and cannot articulate what his job is when asked directly. One makes $150K at a remote startup and travels the world working maybe an hour a day. The sharpest reply in the thread named the mechanism: a substantial portion of everyone at these companies is playing a cooperative game to grab a portion of the cash flowing by from the core thing the company does, no ill will involved, mostly they don't even know. David Graeber wrote a book saying this and got treated as a provocateur. The confession thread treats it as weather.

The game is stable for a reason worse than politeness: nobody inside the firm can run the audit. Cognitive output has been unmeasurable, so the jiggler is indistinguishable from the producer on every artifact the company collects, and the manager who asks what his report actually produces invites the same question one level up, and one more above that. Mutual assured exposure. Every potential auditor is a beneficiary. That's why the slack survived every prior technology, every consulting engagement, every transformation initiative. And it's the thing AI changes first, before it automates anything: run one function hybrid, agents working beside people, and slack stops being a vibe and becomes arithmetic. AI's first product inside an incumbent isn't a worker. It's a meter. People ask whether AI is a bubble while the capex is being priced against token revenue; price it against the wage pool attached to work that doesn't happen, and the question inverts. But notice what the meter doesn't change: every person who could read it out loud is still a beneficiary of what it would say. The technology to find the slack now exists. The person willing to look at the reading doesn't work there.

Now the productivity branch, which is where the dynamo comes back. AI compresses the parallel part of work, the drafting, the code, the analysis, the part you can hand to something tireless and check later. What it doesn't compress is the serial part: deciding what to build and who to sell it to. That runs through management and product, through meetings and alignment and taste and somebody willing to own the call, and it's the bottleneck the moment execution gets fast. Speed up the parallel fraction tenfold and the serial fraction becomes the whole schedule. Worse, the serial part is made of exactly the coordination you'd be cutting if you took the firing branch. Shrink the org and you either ship fewer things or you silo into one person per feature and lose the coherence that made the product worth buying. The companies posting absurd revenue per employee are not proof that adoption is easy. They were born as one pod with full context. There is no shaft to tear out because they never built one. An incumbent cannot get there by subtraction.

Meanwhile the tokens themselves are the opposite of removable. Once people have the tool they do not give it back; pulling it is a morale event, so the spend converts from experiment to benefit, sticky as health insurance. Which produces the actual near-term P&L of AI adoption at an incumbent: payroll unchanged, token bill on top, productivity gains gated behind an org redesign nobody in the building is incentivized to perform. Costs go up. So the rational incumbent move, and I'd argue the modal one right now, is the slow roll: pilot programs, an AI governance committee, a responsible-deployment review board, every one of them a brake dressed as diligence. Ride the existing P&L as long as it lasts.

As long as it lasts. The compulsion doesn't dissolve, it relocates. If firms won't adapt internally, the economy adopts AI by replacing firms, which is how every general-purpose technology has actually diffused — the dynamo's gains landed in new factories built by new owners, and the AI-native pod with absurd revenue per employee is the new factory. Selection, not adaptation. That channel is the big one and it grinds slow: corporate death by competition takes a decade or more even when deserved, moats and contracts and balance sheets all buying time, and the giants at the top of the market will get their decade. But there's a predator in the middle of the timeline and the middle of the market, and it doesn't wait for bankruptcy.

Name the predator. It's private equity, and the weapon is the leveraged buyout, which is worth explaining plainly, because the engineers reading this have mostly never had it explained and the buyout guys reading this will enjoy hearing it said out loud. A sponsor raises a fund, borrows two or three dollars more against every dollar in it, and buys the company. Not a position in the company — the company, board and all, at a price the years of drift have already marked down. The debt lands on the target's own balance sheet, and the target's own cash flow pays the interest. Then comes the part management could never do to itself: an operating team that never sat in the meetings and owes nothing to anyone in the building reads the meter and cuts to the reading. The leverage is what turns a margin story into a fortune. Buy at eight times depressed earnings with sixty percent debt, remove the third of the payroll the model priced before anyone made a phone call, and margins double; sell five years later at a healthier multiple, and the equity check returns three to five times its money on a company that got smaller. The cut is not a cost of the deal. The cut is the deal.

Say the objection out loud, because it's the right one: didn't I just argue you can't subtract your way to AI-native, that cutting the org kills the coherence that makes the product worth buying? You can't, and the sponsor doesn't need to. The incumbent can't cut because it's still competing: it has to ship the roadmap, defend the logo, justify a growth multiple to public shareholders every ninety days. The sponsor bought at the depressed multiple, and its exit story is margin, not features. It cuts to the installed base, reprices customers who would have to rip out ten years of plumbing to leave, and lets the roadmap die quietly. Broadcom closed VMware in late 2023, took the workforce down by more than half, multiplied the license bills, and the customers screamed in every trade publication while renewing anyway, because the switching cost was higher than the insult. Margins went vertical on a product that stopped getting better. The play has a perimeter, and half of LBO history is the graveyard of sponsors who missed it: it needs switching costs, an installed base that can't leave without surgery. Where customers can walk, the gutted company decays faster than the margin model, and some of what looks like slack turns out to have been the immune system. That is not a comfort. It's targeting criteria — the predator hunts lock-in first, and the stickier your revenue, the better you look in the model. A company that intends to keep winning can't subtract. A company that has been bought to yield doesn't have to win.

In Liu Cixin's Three-Body books the galaxy is a dark forest. Every civilization is an armed hunter moving between the trees, and the one law is that revealing your location gets you killed, so everyone hides and the forest stays silent. The economy now runs the same game with one difference: the lights are on. This forest has a filing requirement. The gap between your cost structure as it stands and your cost structure rebuilt AI-native is computable from public data. Filings show what you spend to produce a dollar of revenue. LinkedIn shows your headcount by function. Your own job postings list the things you still do by hand. The diligence that used to take a deal team a quarter runs across the whole Russell 3000 in a weekend, and every quarter you slow-roll, the gap widens, which means your signal gets louder. Be precise about what the outside model sees, because it isn't the chicken coop — no filing shows which employee is slack. It sees the distance between your cost per dollar of revenue and what the same revenue costs to run AI-native. Buyout shops screened against peer margins for forty years; the new thing isn't the screening, it's the benchmark — a comparison company that didn't exist until now, and one that gets leaner every time the models improve. The screen finds the gap from outside. The meter finds the bodies after close. Electrification's selection took a generation partly because nobody could compute which factories were obsolete; the market found out one bankruptcy at a time. This gap publishes quarterly. There are a couple trillion dollars of private-equity dry powder pointed at that signal, software multiples sitting at decade lows, and the buyer is the only actor at the table who was never in the cooperative game. No political attachment to the org chart, no exposure to the mutual audit, no chicken-coop photos in the group chat. The cut is the product they sell.

Everyone in the forest understands this, which is why the visible data is junk. Firms run secret pilots while publicly being thoughtful about AI, because announcing adoption invites employees, unions, and journalists to ask who's getting replaced, while announcing nothing invites the question of whether you're dying. One level down, the surveys keep finding workers who use AI heavily and hide it from their bosses, and of course they do: reveal it and you either get handed more work or you've documented your own redundancy, so the gain gets captured as slack. Every layer of the economy is concealing something from the layer above. Adoption statistics understate adoption. Productivity statistics understate productivity. But all of that hiding conceals the remedy, not the condition. You can hide what you're doing about the gap. You cannot hide the gap, because the gap is the payroll, and the payroll is in the filings.

Which leaves the income statement as the only honest witness in the building. Every AI program reports up as a success: the pilots are promising, the committee is convening, transformation is on track. The P&L can't play along, because the token bill is already in it. So there are exactly three readings, for any company, including yours. Margins expanded: the gains are real and you kept them. Cost lines fell: the gains are real and you spent them on price. Neither moved: the program is lip service, and the proof is published quarterly, machine-readable, signed by your own CFO. There is no fourth reading where the transformation is working and the P&L hasn't noticed. Either the AI made you leaner, or your filings are announcing, to anyone with a model and a weekend, that someone leaner can run your revenue without you.

So picture the meeting where this resolves, because it isn't a layoff announcement. It's a board room in 2029. A CEO who spent four years being thoughtful about AI sits across from a partner whose model says the business runs on a third of the people at twice the margin, and the bid is a premium to a beaten-down stock and a fraction of the old high. Management takes it. They were never going to make the cut themselves, that was the whole problem, and here is someone offering to pay them to leave so a stranger can do it. Nobody at that table gets fired. Everybody under them does, eighteen months later, by a sponsor whose name they'd never heard the day the journalists wrote that AI still hadn't shown up in the productivity numbers.

There are three exits from the forest, and only two of them are yours. If your margins expand without the headcount games — real output per payroll dollar rising, not attrition dressed as transformation — the redesign is happening inside and the dynamo is wrong about you. If you make the work measurable before an outsider does — what coding already proved, where the test suite is the manager and nobody can hide — the slack converts into capacity instead of into someone else's model. The third exit is the credit markets refusing to fund the buyouts, and that one is weather: the predators starve for a while and the old rules hold for years longer than the gap deserves. Two decisions and a prayer. Every incentive in the building runs against the decisions.

Dario's essay prepares for the loud version: the layoff wave, the displacement stats, the policy response. The quiet version is the likely one. Nobody fires anybody. The company just gets old around its payroll, the margin gap compounds in public filings like a beacon, and one day the offer arrives. The slow-rollers think the silence is protecting them. In this forest, silence is the signal. They're not hiding.

They're glowing.