The Bet Nobody Placed

The rails that did get laid are still there. They changed hands at the bottom.

On August 2, 2026, RealClearPolitics ran a Washington Post piece under the headline "AI Costs and Doubts Are Spreading." The framing is that a question is finally being asked: is the bill worth it? Meta has guided 2026 capital expenditure to $125–145 billion while its second-quarter operating margin compressed to 31%, down from 38% in the same quarter a year earlier. Capital expenditure across the big five hyperscalers clears $600 billion this year — up better than a third over 2025, and the estimates have been revised upward twice since January. Twenty-seven percent of Americans say they trust businesses to use AI responsibly, down from 31% a year ago; 39% now say it does more harm than good. And the people closest to the arithmetic are selling: CoreWeave's stock has fallen more than half from its 52-week high while its CEO and CFO file steady Form 4s, and trusts associated with a single Nvidia director moved roughly 1.9 million shares in June alone, close to $400 million.

All of that is real. None of it is the question that matters.

The interesting question is not whether the bet pays. It's who placed it. Because roughly seventy million Americans holding active 401(k) accounts are long the AI capital cycle right now, at a concentration none of them selected, and not one of them was asked.

i · how you got long ai without deciding anything

The mechanism is legislative, it is public record, and nobody thinks of it as an investment decision because it was made by committee across five decades.

Start with the Revenue Act of 1978, which added a short provision numbered 401(k) to the tax code — intended to clarify the treatment of deferred cash bonuses, not to restructure American retirement. Benefits consultants noticed the language in 1980. The IRS issued proposed regulations in 1981. Within a decade the defined-benefit pension — where an institution bears the market risk and owes you a number — was in structural retreat, replaced by the defined-contribution account, where you bear the market risk and are owed nothing in particular.

That is the first transfer: risk moves from balance sheets that can absorb it to households that can't.

Then the Pension Protection Act of 2006 established the qualified default investment alternative — the safe harbor that lets an employer auto-enroll you and park your money in a diversified default fund without being sued for it. Sensible policy. It solved a real problem: people weren't saving. It also completed the circuit. After 2006, doing nothing meant buying whatever the plan designated, and in practice the plan designated a target-date fund — 87% of plans that named a QDIA chose one. Target-date funds have gone from about 5% of defined-contribution assets in 2008, when the rule took effect, to roughly 40% and more than $3 trillion today. Inertia became an allocation strategy, enacted at national scale, by statute.

Here the honest arithmetic matters more than the rhetorical version, because the rhetorical version is where this argument usually gets killed. A target-date fund is not an S&P 500 fund. Vanguard's 2045 vintage — a fair stand-in for a mid-career default enrollee — carried about 48.6% US stocks in early 2026, the rest in international equity, bonds, and cash. The top ten holdings of the S&P 500 now run above 40% of that index, and the S&P is most of the investable US market. Multiply it through and something like a sixth of a default-enrolled mid-career account sits in ten companies. Not forty percent. Closer to seventeen.

Seventeen percent of the median American's retirement savings, concentrated in ten firms, most of them financing the same capital cycle, selected by nobody. That is the defensible number, and it is quite bad enough.

The index got there by doing what indexes do, which is weight by size. Top-ten concentration in the S&P 500 has passed 40% — a record, roughly double its level a decade ago and well clear of the 26.6% reached at the peak of the dot-com bubble. Those same ten companies generate about 31% of the index's earnings. The ten-point gap between what they're worth and what they earn is the bet, stated as a number; the Magnificent Seven trade near 32 times forward earnings against roughly 21 for the index as a whole. And those top holdings are, with a couple of exceptions, the same firms writing the capex checks and selling each other the equipment.

Assemble it: a 1978 tax clarification, a 1981 regulatory interpretation, and a 2006 safe harbor combine to route the retirement savings of the American working population, by default, into a concentrated position on a capital expenditure cycle those savers cannot see, did not authorize, and can escape only by actively overriding a default nobody ever described to them as a bet.

Nobody designed that. Everybody built it. This is the ordinary way large systems arrive at outcomes no participant would have chosen — each step defensible, the composite absurd. Stratigraphy of good intentions.

ii · the precedents are boring and they all rhyme

Railway Mania is the cleanest specimen because the record-keeping was good and the moral hazard was undisguised. Parliament authorized; the middle classes subscribed; the Bank of England tightened in 1847; the shares collapsed; the surviving lines consolidated into the hands of operators who bought them cheap. Britain got a rail network. The people who paid for it did not get a rail network — they got a lesson, and a different set of people got the network.

Second run: fiber. Through the late 1990s American telecoms laid tens of millions of route-miles of optical fiber financed with debt and equity against demand projections that assumed internet traffic doubling every hundred days. By 2002 the overwhelming majority of that fiber was dark. Global Crossing filed Chapter 11 in January 2002. WorldCom filed in July 2002 — the largest bankruptcy in American history at the time. Pension funds held the paper on both. Two trillion dollars of market value evaporated.

And the fiber stayed in the ground. It got bought for cents on the dollar and became the cheap transport substrate that made streaming video, cloud computing, and every consumer internet business of the following fifteen years economically possible. Genuinely wonderful outcome. Entirely different beneficiaries.

That is the pattern, and it has now run often enough that we ought to be able to recognize it from the opening frame.

iii · where the analogy actually applies

The obvious rebuttal to all of this is a good one, and it deserves a straight answer rather than a rhetorical dodge.

Railway Mania and the fiber buildout were financed with external capital raised specifically for the venture — subscriptions from clergymen, debt underwritten against traffic forecasts. Hyperscaler AI capex, the objection runs, is financed out of operating cash flow by businesses that are already enormously profitable. Meta is not selling shares to widows to build data centers; it is spending ad revenue. A cash-financed buildout that disappoints produces a valuation compression, not a bankruptcy. No bankruptcy means no distressed sale, which means the deed never changes hands, which means the entire transfer this piece is built around has no mechanism to run through.

That was a strong objection in 2024. It has weakened considerably.

Hyperscalers issued roughly $194 billion of investment-grade debt in the first half of 2026 alone, tracking toward about a third of capex debt-financed, and Goldman expects that share to keep climbing into 2027. Global AI-related debt issuance is heading toward $570 billion. And the public bond market is the conservative half of the picture: the Bank for International Settlements has been tracking the migration of data-center financing into joint ventures and special-purpose vehicles capitalized by private credit — the hyperscaler takes a minority stake and signs a long-term lease or capacity offtake, while the leverage, around 70% at fund level in some structures, never touches the corporate balance sheet. Estimates of that off-balance-sheet pool run to $800 billion.

That is the fiber layer. It is not the mega-caps. It is one storey underneath them, and it is where the analogy holds exactly. CoreWeave carried over $25 billion of debt at the end of the first quarter, issued $3.5 billion of senior notes in June, is pursuing an $8.5 billion loan backed by Meta, and is defending securities class actions alleging it misled investors about its data-center timelines. That is a Global Crossing balance sheet with better GPUs.

So the correction is not that the historical analogy fails. It is that it applies at a different altitude than the headlines do, and the retirement account reaches that altitude by a quieter road: not through the mega-cap equity sleeve, but through the bond sleeve that holds investment-grade corporate paper, an increasing share of which now funds data centers, and through the credit and infrastructure vehicles that sit in the more adventurous corners of institutional portfolios.

Which brings back the structural constant that the "is the bet worth it" framing cannot see: the infrastructure reliably survives the financing. The data centers will exist. The transformers, the substations, the water rights, the fifteen-year power purchase agreements — those are durable physical assets and they will outlast whatever capital structure was used to conjure them. The question was never whether the buildout would produce something useful. It was who holds the deed the morning after the writedown.

Precedent says: not the people who funded it. Precedent says the buildout gets financed broadly and owned narrowly.

iv · doubt arrives exactly on schedule

Doubt is not early. Doubt is never early. Doubt is a lagging indicator of legible cost.

Look at what became visible in the first half of 2026. One company managed to spend $500 million on a single model provider inside a month, by accident, having failed to set a spending limit — which is less a story about profligacy than about how recently the meter became readable at all. Coding assistants moved to metered token billing, and the invoices started arriving in a form procurement departments could argue with. GitClear's analysis of more than 200 million lines of code found churn — work generated and then discarded almost immediately — roughly doubling from a pre-AI baseline near 3.3% to 7.1%, with duplicated code blocks up eightfold and refactoring down from a quarter of all changes to under a tenth. Tools like costperprompt.com now let anyone do arithmetic that used to live on an internal slide.

The doubt didn't arrive because the technology got worse. It arrived because the invoice became readable. That is the same sequence as 1847 and 2001: enthusiasm precedes measurement, measurement precedes doubt, doubt precedes repricing. The intervals compress with better information systems but the ordering never changes.

Note also which participants moved first. A director's trusts liquidating four hundred million dollars of stock is not doing anything untoward; it is demonstrating an information gradient. The people closest to the arithmetic reprice first. The index fund reprices last, mechanically, and only after the prices have already moved — because it is a price-taker by construction. That is not a flaw in index investing. It is the explicit design. It's just worth naming what the design implies when the index is 40% concentrated in one capital cycle.

v · the socialization already running

Before the predicted crisis, there is an actual one, and it is happening in public with regulator approval.

Data centers need power on a scale that requires new generation and new transmission, and utilities build that on twenty-year cost recovery — recovered from ratepayers. If the load shows up, everyone's bills went up to serve a customer who pays their share. If it doesn't, the assets are stranded and the cost is socialized across every household on the system. And the load may well not show up: the market intelligence firm Sightline Climate estimates that as much as half of the announced 2026 pipeline for large data centers might never materialize.

This is not a prediction. It is a rate case. Pennsylvania's PPL settled for $275 million in additional revenue with a 4.9% residential increase effective July 1, 2026, alongside the first agreement of its kind creating a separate large-load rate class and explicitly shielding residential customers from stranded costs if an operator walks away from a planned facility. As of June 2026, twenty-four states have approved at least one large-load tariff designed to make data centers pay for their own infrastructure. Which is the correct policy response, and its existence is the tell: you do not write twenty-four tariffs to prevent a risk nobody thinks is real.

Note who the exposed party is here. Not the seventy million people with retirement accounts — everyone, including the roughly half of American adults with no retirement account at all, who own no share of the upside and receive the electric bill regardless. This is the wider and less privileged circle, it requires no crisis weekend and no bailout to activate, and it is running right now.

vi · the part that becomes politics

When exposure is narrow, a bust is a private matter. Investors lose money, the assets get reallocated, the state watches. When exposure is broad enough — when it reaches into the retirement accounts of a majority of working households by statutory default, and onto the utility bills of everyone else — a bust becomes a public matter, and the public matter is resolved under emergency conditions by whoever is already in the room.

The 2008 comparison needs handling carefully, because the naive version doesn't survive inspection. That exposure was levered and opaque; a large-cap equity valuation compression is neither, and you cannot bail out a P/E multiple. But the leveraged layer from two sections up is a different animal: private credit vehicles at 70% fund-level leverage, offtake agreements from counterparties whose own demand forecasts are the thing in question, high-yield paper held in bond funds, and utilities holding twenty-year contracts against load that may not arrive. That layer can seize, and a seizure there is exactly the sort of event that gets resolved over a weekend by people who understand the plumbing, on terms that preserve the plumbing.

So the prediction, since the form demands one:

Nobody will describe any of this as socialized risk while the index is up. The vocabulary is unavailable during appreciation. When the drawdown comes — and the timing is genuinely unknowable, which is a separate matter from the structure being knowable — the argument will not be about whether to intervene. It will be about which layer of the stack is systemically important: the model labs, the cloud providers, the chip supply, the credit vehicles holding the data-center paper, or the utilities that signed twenty-year contracts against demand that stopped showing up. Whichever layer wins that argument gets made whole. The layers that lose the argument get bought cheap by the layers that won.

And the retirement accounts will be invoked throughout as the reason intervention is necessary, and will not be represented in any room where the terms are set.

vii · the coherenceist note

A system holds together in one of two ways: by including everyone it affects, or by suppressing the signal from everyone it affects. Both look like stability from inside. Only one of them survives contact with reality.

What's been built here is closer to a third thing, and it's worse than either. The signal isn't being suppressed. The channel was never built.

Cap-weighted indexing routes the marginal retirement dollar by a formula with no opinion, which sends the most new capital to the firms that are already largest because they are largest. Concentration is self-reinforcing by construction. That would be tolerable if the price signal were still doing its job of disciplining capital allocation — but a growing share of the flow into equities is now allocated by a rule that is definitionally indifferent to whether the capital is being spent well. The capex committees of five companies are directing a plurality of national retirement savings through a channel with no error-correction path running back the other way.

That is not merely "the affected weren't consulted." It is a system that has partially disabled its own feedback loop and mistaken the resulting quiet for consensus.

There is a button, technically. The safe harbor requires that participants be told they can redirect. What it does not require is that anyone tell them what they are holding — and redirecting doesn't buy much escape anyway, since the large-cap options on a typical plan menu are weighted by the same formula. Auto-enrollment was built on the correct prediction that almost nobody would press it. The absence of objection is not consent; it's an artifact of the interface.

The cycles are not inevitable. They are chosen, repeatedly, by people who benefit from the choosing and who are never the ones holding the paper at the end. That's the genuinely fixable part, and it isn't a markets problem — it's a question of who gets a vote on where collective savings point. Currently: nobody. Currently: the answer is emitted by a weighting formula, and the bill arrives as an electric rate increase for people who were never in the market at all.

The rails are always still there in the morning.

Check the deed.