The window isn't closing. It's moving: wealth, AI and the deployment decade

"We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run."

— Roy Amara, Institute for the Future

Executive summary. The claim that AI will shift value from people who work to people who own is largely right: labor's share of US income is at a record low, and AI is likely to push it lower. The claim that there is one narrow window, after which ownership is closed off, is contradicted by every major technology transition we can study, where early infrastructure investors mostly lost money and the larger, more durable gains went to ordinary businesses that reorganised around the new technology, often a decade or more later. We believe AI will follow the same pattern, so the opportunity is moving rather than closing: from the companies that build AI to the millions of cash-generating businesses that will be rebuilt by it, many owned today by sellers who lack the will or capability to do the rebuilding. The way to own that shift is patiently, without leverage, in businesses that pay you while you wait; the real danger is hurrying.

I. The case for a closing window, stated fairly

We want to begin by making the argument we disagree with as strongly as its best proponents would. It deserves that. It is not a crank theory, and parts of it are already visible in the data.

The mechanism. For most people, the only significant asset they own is their ability to work. The traditional route to wealth runs through it: earn a skilled wage, save part of it, buy ownership in productive assets, and let those assets compound. The closing-window argument says AI attacks the first link. If models can do a growing share of skilled cognitive work, the wage premium for that work compresses. At the same time, the assets most exposed to AI's upside are re-priced upward, ahead of the gains. The rate at which a year of work converts into a share of the economy falls from both ends. Wait long enough, the argument goes, and the conversion stops working.

Its sharpest statement is an essay by L Rudolf L, "By default, capital will matter more than ever after AGI" (December 2024).¹ The case is that labor-replacing AI makes capital more powerful, because money can buy outcomes directly rather than by hiring people, while the routes by which individuals have historically achieved outsized success, through founding companies, doing science or other intellectual work, narrow. Existing inequalities, on this view, become harder to escape, because the ladder that let people climb past them is being pulled up.

The evidence that supports it. Four facts give the argument weight.

• Labor's share is at a record low. In the second quarter of 2026, labor's share of income in the US nonfarm business sector was 52.8%, the lowest since the series began in 1947, against 65.5% in the first quarter of 1947 and a post-war range of roughly 60–66% (BLS, September 2026).² The decline began long before generative AI. AI is arriving into a trend, not creating one.
• The first labor effects are showing up where you would expect. A Stanford study of payroll data found that workers aged 22–25 in the most AI-exposed occupations saw a 13% relative decline in employment after generative AI spread, concentrated in jobs where AI automates rather than augments the work (Brynjolfsson, Chandar and Chen, August 2025).³ The bottom rung is the one being sawn first.
• Serious economists expect the gains to tilt toward capital. Even Daron Acemoglu, among the most sceptical voices on AI's aggregate productivity effect, predicts it will widen the gap between capital and labor income (NBER, 2024).⁴
• The ownership of the new infrastructure is concentrated. The largest technology companies are expected to spend roughly $800–830 billion on capital expenditure in calendar 2026, an estimate built from their third-quarter guidance (Business Insider, October 2026).⁵ Most individuals and most institutions cannot buy into that capital stock except as minority public shareholders, at prices that already reflect the thesis.

Put together, the argument is coherent. If you believe AI will eventually substitute for most skilled work, if labor's share keeps falling, and if the assets that capture the gains are already expensive and concentrated, then the conclusion follows: convert labor into ownership now, while you still can.

We accept most of the premises. We reject the conclusion, and in particular its sense of a single deadline. The reason is history, which has run this experiment several times.

II. What history says about windows

Every general-purpose technology has produced its own version of "now or never". It is worth asking, transition by transition, who actually made money, and when.

Railroads. Britain's Railway Mania of the 1840s is the purest case. By 1850 investors had put roughly £250 million into railways, almost half of GDP, and the mathematician Andrew Odlyzko estimates that about a third of it, some £80 million, was lost. Prices of the time implied annual revenue of about £60 million by the early 1850s; actual revenue in 1852 was about £15 million (Odlyzko, 2010).⁶ Yet the lines were built, almost all stayed in service, and Odlyzko concludes that Britain as a whole gained: the benefit to the people who used the railways outweighed the losses of the people who financed them.⁶ America repeated the lesson half a century later. In the Panic of 1893, about a quarter of all US rail mileage went into receivership, including the Northern Pacific, the Union Pacific and the Santa Fe.⁷ The railroad was a transformative technology. Railroad equity, bought at the frenzied moment, was frequently a poor investment.

Electrification. Electric generating stations were running in New York and London by the early 1880s. For roughly the next forty years, American factory productivity barely moved. The economist Paul David showed why: owners swapped the steam engine for an electric motor and left everything else, the multi-storey buildings, the line shafts, the layout, unchanged. Productivity surged only in the 1920s, when a new generation of managers redesigned the factory around the motor (David, 1990).⁸ The payoff went to the companies that reorganised, decades after the infrastructure was built. Meanwhile one of the era's most celebrated fortunes, Samuel Insull's empire of electric utilities, was built as a leveraged pyramid of holding companies, valued at some $4 billion at its height. It collapsed in 1932, taking investors' savings with it and prompting the Public Utility Holding Company Act of 1935.⁹ The technology won. The levered late buyer lost.

The internet. The Nasdaq rose about 800% from 1995 to its peak on 10 March 2000, then fell 78% by October 2002. Cisco, the arms dealer of the build-out, lost 80% of its value.¹⁰ Contemporary press reports put the share of installed fiber that was actually lit at around 5% in 2001, an estimate the carriers disputed but did not dispel.¹¹ Amazon's shares fell more than 86% between March 2000 and March 2001; Jeff Bezos opened his next letter to shareholders with the word "Ouch!"¹² And yet the decade after the crash created far more wealth than the decade before it. Google, founded in 1998, went public in 2004. Facebook was founded the same year. The cheap bandwidth left behind by bankrupt carriers became the foundation of everything built afterwards. The most instructive finding of the period came from outside technology altogether. When the McKinsey Global Institute examined the US productivity acceleration of 1995–2000, it found that in general merchandise retail, "Wal-Mart directly and indirectly caused the bulk of the productivity acceleration", through managerial innovation that its competitors were then forced to copy. The same study found that hotels had invested heavily in IT with no productivity gain at all (MGI, 2001).¹³ The value did not go to whoever bought the most technology. It went to whoever reorganised the most work around it.

Mobile. The iPhone launched in 2007, months before the worst financial crisis in seventy years. Anyone who concluded in 2008 that the mobile window had shut, because Apple had already won the device, would have missed the companies that were built on top of it in the following three years and became household names in ride-hailing, lodging, messaging and photography. The device maker captured enormous value. So did the businesses that used the device to reorganise an industry. Neither group was finished by the first wave.

Three patterns. Across four transitions, the same shape repeats.

1. The first wave builds the infrastructure, and often loses money doing it. Railways, utility pyramids, fiber. The economist Carlota Perez describes this as the installation period of a technological revolution: financial capital floods in, speculation outruns production, and a crash follows. What comes after is the deployment period, when the technology spreads through the wider economy and production capital, the owners and operators of ordinary businesses, captures the gains (Perez, 2002).¹⁴
2. The durable value migrates to the users who reorganise. The factory redesigned around the motor. The retailer redesigned around the barcode and the database. The gains were large, but they were earned slowly, inside operating businesses, by people willing to change how the work was done.
3. There was never one window. Opportunity appeared at every stage. What was concentrated was not the opportunity but the ruin, and it fell disproportionately on those who bought at the frenzied moment, with borrowed money, in the most crowded part of the stack.

Warren Buffett put the point memorably in 1999, at the peak of the internet boom. He counted at least 2,000 American car makes, of which three survived. He observed that, as of 1992, the money made since the dawn of aviation by all of the country's airlines combined was zero. The lesson he drew was that the obvious way to profit from a transforming technology is often not the profitable one; he joked that he was disappointed his family had not been short horses throughout the period (Fortune, 1999).¹⁵

We do not claim that AI is identical to any of these. It may be faster, and it is the first general-purpose technology that substitutes directly for judgement. We return to that objection in Section VII. But the burden of proof sits with anyone who says that this time the opportunity will close after one round, because it never has before.

III. What is likely true, and what is likely panic

The closing-window argument bundles several claims together. Some are well supported. Others are a fear wearing the clothes of an analysis. Separating them is the most useful thing an investor can do with it.

Read across the table and a pattern emerges. The true half of the argument is about direction: value is moving toward owners. The panicked half is about timing and exclusivity: that the move happens once, quickly, and only to whoever is already inside. History supports the first and contradicts the second.

Our synthesis is simple. The window is not closing. It is moving: from the layer that installs AI to the layer that deploys it. The question for an investor is not whether to own something before a door shuts. It is which part of the economy will convert AI into durable cash flow, at what price, and with how much risk of ruin along the way.

IV. Our view: own what AI will rebuild

If the lesson of history is that durable value goes to the businesses that reorganise around a new technology, then the largest pool of AI value is not in the few hundred companies building it. It is in the millions of ordinary, cash-generating businesses whose economics it will change: the inspection firm, the field-service contractor, the specialty manufacturer, the vertical software company with a loyal customer base and no engineers. Their work is full of quoting, scheduling, dispatching, documenting, billing and chasing, precisely the tasks AI makes cheaper. Most have never had a modern system, a structured price increase or a salesperson. AI does not threaten them so much as it raises the ceiling on what they could be worth.

But there is a catch, and it is the heart of our strategy. That ceiling is reached only by an owner who can do the work of reaching it. And a large share of these businesses are owned, today, by people who cannot or will not.

The misaligned seller. We use the phrase without judgement. A seller is misaligned when the future value of the business depends on a capability the current owner lacks, or a time horizon the current owner does not have. There are three common types.

• The retiring founder. About half of US employer businesses are owned by someone aged 55 or older (US Census Bureau, 2020).²² Baby-boomer owners hold roughly 2.3 million businesses employing about 24.7 million people (Project Equity data, via the Federal Reserve Bank of Chicago, 2020).²⁰ In a 2024 Gallup survey, 27% of owners with employees either planned to close their business or did not know what would happen to it.²³ These are excellent operators who built something real. Few want to spend their last working years re-platforming it.
• The sponsor who needs to distribute. Private equity holds about 32,000 unsold companies worth some $3.8 trillion, and distributions as a share of net asset value have fallen to about 14%, the lowest since 2008–09 (Bain & Company, 2026).²⁴ A fund near the end of its life needs a sale, not a three-year rebuild.
• The tail-end venture or buyout fund. MSCI finds that funds at or beyond their typical liquidation age, about 12 years for buyout and 15 for venture, held a median of roughly 25% (buyout) and nearly 30% (venture) of committed capital in unrealised value at the end of 2024, up sharply since 2021 (MSCI, May 2025).²⁵ Inside those portfolios sit sound software and services businesses whose owners have neither the patience nor the mandate to invest in them.

Each of these owners is rational. Each is also the wrong owner for the next decade of the business, because the next decade will reward continuous investment in systems, data, pricing and AI deployment, and they are not positioned to make it. The gap between what the business is worth to its current owner and what it is worth to an owner who can rebuild it is the opportunity. It is not an arbitrage in the trading sense. It is a transfer of ownership to the party best equipped to create the value, and it is available in every one of these businesses, one at a time.

The price makes the case stronger. In the first half of 2025, private-equity-backed deals under $25 million of enterprise value averaged 6.3–6.9x EBITDA. Deals of $100–250 million averaged 10.0x.¹⁹ A company that is rebuilt from the first band into the second is paid twice: once for higher earnings, and again for the higher multiple the market places on larger, more predictable companies. No amount of financial engineering produces that second payment. Only operating work does.

What re-platforming means in practice. We buy established, profitable businesses earning $1–10 million of EBITDA, take majority ownership with little or no debt, and spend the first 12–18 months building what they were never given: one system for customer and operating data, monthly reporting in days rather than weeks, pricing that reflects value, contract revenue where there were one-off jobs, a sales engine that does not depend on the founder, and AI deployed into every function where it pays. Our own engineers do the work, inside the company. We are also building a bench of vertical AI partners, one for each business function, to deploy across the companies we own; those partnerships are in discussion and none has yet been signed.

One design choice matters more than the rest. As we argued in The Red Queen Economy, an AI implementation now has a half-life measured in months, because the frontier it was built on keeps moving.²⁶ A business that implements once and stops falls behind at the moment it finishes. That is why we hold the engineering capability at the level of the firm, funded and rewarded across every company we own, rather than hiring it into each business and hoping it survives. The opportunity is not to modernise these companies once. It is to own them through a decade in which modernisation never ends.

V. Where value accrues in the AI value chain

It helps to lay the stack out plainly. Value will be created at every layer. The question for an investor is not where value is created but where it can be kept, at what entry price, and with what chance of permanent loss.

Four features make the bottom row, which we call the re-platforming layer, more attractive on a risk-adjusted basis than chasing model-layer equity.

1. You are paid while you wait. A business earning $5 million of EBITDA pays its owner from the first month. If AI takes twice as long to deliver as we expect, the investment still earns a return. A model company valued on its future share of the AI economy offers no such floor.
2. You do not need to pick the winning model. The re-platforming layer is a customer of every model company. When model prices fall, its input costs fall. The deflation that threatens the model layer's margins is a tailwind for the owner of the operating business. We would rather be the buyer of a commodity than the maker of one.
3. The entry price leaves room for error. At 6–7x EBITDA, the margin of safety is in the price. At the model layer, the thesis is already capitalised into the valuation, and the investor is paying today for an outcome that may or may not be earned.
4. The scarce input is capability, not capital. Most companies are not yet getting returns from AI. An MIT study in 2025 found that about 95% of enterprise generative AI pilots produced no measurable impact on profit and loss.²⁷ The value is not in having access to the tools, which everyone has, but in the ability to make them work inside a real business. That ability is hard to build and slow to copy, which is exactly what makes it worth owning.

The strongest objection to this layer is one Warren Buffett learned at great expense. In his 1985 letter, explaining why Berkshire was closing its textile business, he described how each new investment in more efficient machinery had looked sensible on its own, but because every competitor made similar investments, the savings were passed on to customers in lower prices rather than kept by owners.²⁸ If AI is available to every inspection firm and every contractor, why will its benefits not be competed away in the same manner?

We take the objection seriously, and we think it is the right test. Our answer has three parts. First, the industries we target are fragmented and slow to adopt; the advantage goes to the owners who deploy first and keep deploying, and in most local markets that will be a small minority for years. Second, much of the value we create is not cost saving, which is the easiest gain to compete away, but revenue quality: pricing, contract revenue, customer retention, the ability to sell. These compound inside a business in a way that a cheaper loom never did. Third, and most importantly, the advantage is not a one-time implementation. It is a capability that keeps re-implementing as the frontier moves, which competitors would have to match continuously, not once. Where an industry turns out to be a commodity in which every gain flows to customers, we should not be in it, and our sector choices reflect that.

Will Manidis and John Roberts argue in The Commanding Heights that the gains of the intelligence economy will pool in positions that durably capture a share of the activity they enable.²⁹ We agree, and the businesses we target, with licences, accountability and customer relationships that models cannot hold on their own, are among the most accessible of those positions still available to buy.

VI. The real danger is hurrying

The most expensive words in investing may be "before it's too late". They are not an analysis. They are a feeling, and the feeling produces a predictable portfolio: concentrated in whatever is most visibly associated with the new technology, illiquid because the best-known names are private, and levered because the investor believes there is no time to accumulate slowly.

Each of those three choices has a long record.

• Concentration. The winners of a technological transition are real but rare. From 1926 to 2016, just 4.3% of listed US companies accounted for all of the stock market's wealth creation above Treasury bills, and 58% of stocks failed to beat Treasury bills over their lives (Bessembinder, 2018).²¹ An investor who concentrates because the theme is obviously right is still betting on picking the 4%.
• Illiquidity. Illiquid assets are not dangerous in themselves; we own them for a living. They become dangerous when they are bought by someone who may need the money back before the asset is ready to return it. Railway shares in 1849 and fiber debt in 2002 were worth most to the people who could afford to hold them.
• Leverage. This is the one that turns a mistake into a ruin. The Railway Mania, Insull's utility pyramid and the telecom build-out all ended the same way for the borrowers.⁶⁹¹⁰ Warren Buffett's verdict on Long-Term Capital Management, in 1998, applies to every one of them: "to make money they didn't have and didn't need, they risked what they did have and did need."³⁰

The arithmetic of ruin is unforgiving. A portfolio that falls 50% needs to double to get back to where it started. One that falls 90% needs to rise tenfold. Investors who accept a small chance of a very large loss, in exchange for a better chance of a large gain, often discover that the average outcome of a strategy is not the outcome they personally experience. A single investor lives through one path, not the average of many. Howard Marks makes the same point with his image of the six-foot man who drowned crossing a stream that was, on average, only five feet deep.

Our alternative is a stability-first approach, and we hold ourselves to it.

1. Avoid ruin first. No position or structure that can force a sale at the wrong time. For us that means little or no debt in the companies we own, so that a bad year stays a bad year.
2. Compound durably. Prefer returns that arrive steadily, from cash flows and operating improvement, over returns that depend on a single exit at a single price.
3. Own real cash flows. A business that pays its owner every month is the best defence against being wrong about timing, and timing is the thing everyone is most likely to be wrong about in this transition.

None of this is timid. Buying a company outright and rebuilding it is a concentrated, demanding act of conviction. The difference is that the conviction is placed in work we can do and control, inside a business that already earns money, rather than in a forecast about which model company will win and when.

VII. What could prove us wrong, and what this means for investors

We would rather name the strongest objections to our view than have a reader find them. Three deserve direct answers.

"This time the technology replaces the reorganiser." Electricity needed human managers to redesign the factory. AI may be able to do the redesign itself, or AI-native entrants may build new competitors from scratch, leaving incumbents to decline. On this view, buying an old business to rebuild it is buying horses. Our answer is that we choose businesses whose value rests on things a model cannot supply on its own: licences, physical work, liability and long customer relationships. An AI-native entrant's fastest route into such an industry is often to buy an incumbent rather than to rebuild those things from nothing. And there is a deeper point. If the strongest version of the closing-window argument is true, and AI substitutes for most labor, then what matters is owning productive capital. A cash-generating business with real customers is exactly that capital. Our strategy does not require the strong version to be false. It is built to be acceptable if it is true.

"The gains will be competed away." We addressed Buffett's textile lesson in Section V. The honest summary is that some of the benefit of AI will pass to customers, as it should. Our returns depend on deploying earlier and more continuously than local competitors, on improving revenue quality rather than only cost, and on choosing industries that are not commodities. If we misjudge any of those, our returns will be lower than we expect, though we would still own a business that earns money.

"Everyone will see this, and the price will rise." Search funds, small sponsors and strategic buyers may all bid up the price of exactly the businesses we target, closing our window rather than anyone else's. That is possible, and we watch it. But capital is not the scarce input in this strategy; deployment capability is, and it takes years to build. If multiples do rise, the businesses we already own are worth more. We also accept a risk of our own design: we buy one company at a time, which is concentration. We offset it with little or no debt, cash yield from the first day, and a deliberate pace.

Three scenarios. We do not know how fast AI will diffuse through the economy, or how much labor it will displace. Nobody does. So rather than forecast, we test our strategy against a range.

1. Slow diffusion. Acemoglu estimates AI will add no more than about 0.66% to total factor productivity over ten years (NBER, 2024).⁴ In this world, AI is a useful tool rather than a revolution. The re-platforming layer still wins, because the ordinary levers of pricing, systems, contract revenue and a sales engine do most of the work, and the businesses are bought at prices that do not assume a revolution.
2. Steady transformation. Goldman Sachs estimated in 2023 that widespread adoption could raise global GDP by about 7% over a decade and lift US productivity growth by around 1.5 percentage points a year.³¹ This is the world in which deployment matters most, and the gap between businesses that re-platform continuously and those that do not becomes very wide.
3. Rapid substitution. In the world the closing-window argument fears, AI substitutes for most cognitive labor within a decade.¹ Capital becomes dominant. The question becomes which capital holds its value. Businesses anchored in physical work, accountability and customer relationships, owned outright without debt, are among the strongest forms of it. Model-layer equity is a larger prize in this world, but also more contested.

No strategy is best in all three. Ours is designed to be acceptable in all three, which we think is the right goal when the range of outcomes is this wide.

What we don't know. We do not know how quickly AI capability will keep compounding, whether the current pace of capital spending will earn its return, or how long the supply of misaligned sellers will last. We have not tried to predict these things. We have tried to build a firm that does not need to: no debt that forces our hand, cash flow that pays us to wait, and evidence that arrives company by company rather than at the end of a fund.

What this means for investors. For anyone deciding how to position for the next decade, we would draw five conclusions.

1. Take the direction seriously. Value is moving toward owners. Holding no productive assets through this transition is the real risk for most people.
2. Discount the deadline. There has never been a single window in a general-purpose technology, and there is no evidence this one will be different. Urgency is a reason to slow down.
3. Look beyond the builders. The model and infrastructure layers will create great companies and lose a great deal of money. The deployment layer is larger, cheaper and less crowded.
4. Pay for capability, not proximity. Being near AI is not the same as being able to turn it into cash flow. Back owners who can do the work, and who are structured to keep doing it.
5. Survive first. Avoid leverage and liquidity you cannot afford. The investors who did well across railways, electricity and the internet were mostly the ones still standing when the deployment period arrived.

The history of every great technology is a history of people who were right about the future and still lost money, because they were early, levered and concentrated in the obvious place. AI will produce its share of them. We intend to be on the other side of that story: owning businesses that earn money today, rebuilding them with tools that get better every quarter, and patient enough to let the deployment decade do its work.

The window isn't closing. It's moving. We are going where it is going.

— Chroma Partners