On Friday, March 10, 2023, the California Department of Financial Protection and Innovation closed Silicon Valley Bank and appointed the Federal Deposit Insurance Corporation as receiver. The announcement arrived at 11:30 a.m. Pacific time. By Monday morning, the FDIC had established a bridge bank, transferred all deposits — including those above the $250,000 insurance limit — and guaranteed that every depositor would be made whole. Shareholders received nothing. Bondholders were wiped out. The 8,500 employees learned their status over the following weeks, in phases, as the FDIC and eventual acquirers sorted through what remained.

Within fourteen days, Signature Bank in New York followed. Within two months, First Republic Bank in San Francisco was seized and sold to JPMorgan Chase at a discount. The three failures represented approximately $548 billion in combined assets — more than every bank failure in 2008 combined. The system held. The economy did not collapse. The contagion did not spread. The response was called a success.

The question is: a success for whom?

Share

The architecture of absorption

Every financial system has a hierarchy of loss absorption. This is not hidden. It is codified in law, in contracts, in the order of priority that determines who gets paid when there is not enough money to pay everyone. Equity holders absorb losses first. Then subordinated debt holders. Then senior creditors. Then depositors. Government insurance — the FDIC — backstops depositors up to $250,000. The structure is designed so that those who took the most risk, the shareholders who bought stock anticipating profit, bear the first losses, and those who took the least risk, the retiree with a checking account, are protected.

The design is clean.

The execution is where things get interesting. When SVB collapsed, the hierarchy operated as scripted at the top: shareholders were wiped out. But the next moves were less predictable. The FDIC invoked a “systemic risk exception,” a provision allowing it to guarantee all deposits, not just those below the insurance cap. This meant that venture capital firms and technology companies with $10 million, $50 million, $200 million in uninsured deposits at SVB were made whole. The stated reason was preventing contagion — if large depositors at other regional banks panicked and withdrew their funds simultaneously, the cascading failures could bring down the banking system.

The logic was sound. Every economist and regulator who supported the decision had good reason to do so. Nassim Taleb has argued for years that systems designed around normal conditions are destroyed by tail events, the improbable shocks that probability models assign near-zero likelihood but that occur with devastating regularity. The SVB failure was this kind of event: a bank whose balance sheet was technically sound under normal interest rate conditions became insolvent when the Federal Reserve raised rates at historical speed. The failure was not caused by fraud or incompetence in the traditional sense. It was caused by a mismatch between the bank’s asset duration and its liability structure, a mismatch that existed at dozens of other banks simultaneously.

So the decision to guarantee all deposits was rational. It prevented a worse outcome.

But notice what happened to the hierarchy. The people who were supposed to absorb losses did absorb them. The people who were not supposed to absorb losses were protected. And the middle layer — the large uninsured depositors who had chosen to concentrate tens of millions at a single institution without diversifying — were also protected. They had taken a risk. They were not penalized for it. The formal hierarchy said they should bear losses. The emergency decision said they would not.

The cost of that protection was borne by the FDIC’s Deposit Insurance Fund, which is replenished through assessments on all FDIC-insured banks. Those assessments are passed through to customers in the form of slightly lower interest rates and slightly higher fees. The cost was socialized — distributed across every depositor at every bank in America, in amounts too small for any individual to notice.

This is how shock absorption works in practice. Not through the clean hierarchy on paper, but through a series of rational decisions that shift losses from those with the capacity to organize toward those without the capacity to object.

Who was in the room

There is a moment in any institutional crisis when the question shifts from “what happened” to “who decides what happens next.” In the SVB collapse, that shift occurred over a single weekend. Treasury Secretary Janet Yellen, Federal Reserve Chair Jerome Powell, and FDIC Chair Martin Gruenberg convened emergency consultations. The people at the table were regulators, banking executives, and representatives of the largest depositors. The depositors who showed up, or whose representatives called, were venture capital firms. They argued, correctly, that if their deposits vanished, they would be unable to make payroll for the startups they funded, which would cascade into layoffs for tens of thousands of workers.

This argument was effective. It was also self-interested in a way that happened to align with systemic stability. The venture capitalists were protecting their own capital. They were also, incidentally, protecting the paychecks of employees who had no seat at the table and no vote on the outcome.

Mariana Mazzucato has documented this pattern across industries in The Entrepreneurial State and The Value of Everything: public institutions absorb the downside risk of private ventures while private actors capture the upside. The pharmaceutical industry develops drugs using publicly funded basic research; the profits accrue to private shareholders. The technology sector builds on publicly funded infrastructure — the internet, GPS, touchscreen technology — and privatizes the returns. Mazzucato’s framework describes what happened at SVB with precision. The banking system’s stability was maintained at public cost. The depositors who benefited most from that stability — those with the largest uninsured balances — contributed least to the insurance fund relative to the protection they received.

The FDIC later estimated the cost of the SVB resolution at approximately $16.1 billion. The special assessment to replenish the fund was levied on banks with more than $5 billion in uninsured deposits, which sounds like the cost was borne by the institutions that created the risk. In practice, those banks passed the cost to their customers. The final bearers of the loss, the people at the bottom of the chain, were individual depositors and borrowers at large banks paying in increments too small to perceive but too consistent to avoid.

I keep returning to a detail from the days after the closure. The notice taped to the glass door of a SVB branch on that following Monday — a single sheet of white paper, the FDIC seal printed slightly off-center, the text in a serif font that conveyed formality without clarity. The glass doors locked behind it. The notice addressed itself to “depositors and customers,” a category that included both the retired schoolteacher with $47,000 in savings and the venture fund with $180 million in operating capital. The same notice. The same language. The same institution standing behind them both. But the schoolteacher’s $47,000 was always protected. The venture fund’s $180 million was protected only because the people in the room that weekend decided it should be.

Subscribe now

The pattern beneath the event

What SVB revealed is not specific to banking. The same structure operates in every complex system that encounters shocks: the formal hierarchy of loss absorption exists on paper, but in practice, losses flow toward those with the least capacity to redirect them.

Consider the 2008 financial crisis. Homeowners absorbed the initial shock: lost homes, destroyed credit, erased savings. Banks received $700 billion in TARP funds. The largest banks repaid those funds, with interest, and within five years were reporting record profits. Homeowners who lost their properties between 2008 and 2010 — approximately 3.8 million foreclosures in that period alone — did not receive comparable restoration. Their credit records carried the foreclosure for seven years. Their lost equity did not return. The shock was absorbed permanently by those at the bottom and temporarily by those at the top.

(There is something about this asymmetry of duration that I have not yet been able to state with enough precision — the same dollar amount of loss means fundamentally different things depending on how long it persists on the balance sheet. A bank writes down a quarter and moves on. A household carries a foreclosure for a decade. The temporal dimension of shock absorption matters at least as much as the dollar dimension, and I am not sure the existing economic frameworks account for it adequately. I want to note this gap rather than pretend I have resolved it.)

Or consider pandemic-era layoffs. In March 2020, 22 million American workers lost their jobs in four weeks. The stock market dropped 34% and recovered within five months. By January 2021, the S&P 500 had reached new highs. Median household income did not return to 2019 levels until 2023, and that figure conceals enormous variation: workers in the bottom wage quintile were six times more likely to be laid off than those in the top quintile, according to Brookings Institution data. The recovery operated identically to the crisis — capital recovered first, labor recovered last, and no one designed it that way.

Every system has a failure mode. What matters is which tier of people absorbs the failure.

The hierarchy sounds orderly on paper. Equity first, then debt, then depositors, then the public. In practice, the hierarchy reshuffles in real time based on who can organize, who can advocate, who can present their losses as threats to the system itself. The venture capitalists at SVB could argue — and did argue — that protecting their deposits was protecting the economy. The homeowner facing foreclosure in 2009 could not make a comparable argument, because individual foreclosure does not threaten systemic stability.

It only threatens the individual.

What the structure reveals

Joseph Stiglitz has written about what he calls “ersatz capitalism,” a system that socializes losses while privatizing gains. The phrase captures something real, but I find it incomplete. It implies a deviation from some truer form of capitalism in which both gains and losses would accrue to those who earned them. What the SVB episode and the 2008 crisis and the pandemic recovery collectively suggest is something different: the asymmetric distribution of shock is not a corruption of the system. It is an output of the system functioning under stress. The capacity to redirect losses is itself a form of capital — perhaps the most consequential form. The advantage of scale is not only higher returns in good times but insulation from losses in bad ones.

This does not require conspiracy. It does not require coordination. It requires only that the people making emergency decisions are responsive to the arguments of those who are present, and that the people who are present are, by definition, those with the resources to show up.

Subscribe now

I do not know whether the pattern I am describing constitutes a design flaw or an inherent property of systems that distribute resources unequally and then encounter shocks. The distinction may not matter in practice. What I observe is that the pattern repeats — in banking, in housing, in healthcare, in labor markets — and each repetition follows the same structure. Losses settle. They move from balance sheets that can absorb a write-down to households that cannot absorb anything further. Not because anyone decided they should, but because the architecture of response makes no other distribution possible.

The system held. The economy recovered. The losses were absorbed. It is worth asking, with some precision, what it means that the answer to “who absorbs the shock” is always legible after the fact and never negotiated before it.

-Aimé

Aimé Halden writes Uninsurable, a newsletter about the systems that shape who is protected and who is not. Subscribe for weekly analysis.

In October 2021, the Federal Emergency Management Agency did something it had never done in the fifty-three-year history of the National Flood Insurance Program. It changed how it priced flood risk. The old system — in use since the program’s creation in 1968 — assigned premiums based on flood zone maps, broad geographic designations that sorted properties into categories: in a floodplain or not, in a special hazard area or not. The maps were crude. They were updated slowly, sometimes decades apart. They priced entire neighborhoods identically. A house three feet above the flood line and a house three feet below it paid the same rate if both sat inside the same zone boundary. The system was inaccurate. It was also functional, in the specific sense that it kept the pool together.

Risk Rating 2.0, as FEMA named the replacement, introduced property-level pricing. Each structure is now assessed individually based on distance to a water source, elevation relative to flood levels, the cost to rebuild, the type of flood risk (river, storm surge, coastal erosion, heavy rainfall), and historical loss data. The methodology is, by actuarial standards, a significant improvement. The Government Accountability Office confirmed as much in a 2023 review: the new system is more actuarially sound. Premiums now correspond more closely to the actual risk each property faces. Seventy-seven percent of policyholders saw their premiums increase. Nine percent will eventually require increases exceeding 300 percent. The national average annual premium sits at roughly $926 — but this is a number in transit. The GAO estimates that it will take until 2037 for 95 percent of policies to reach their actuarially correct price, and the cumulative shortfall between what policyholders are currently paying and what the risk actually costs is approximately $27 billion.

The program itself owes $22.5 billion to the U.S. Treasury, borrowed against a $30.4 billion ceiling that Congress periodically raises. It borrowed another $2 billion in February 2025 to cover claims. Eight hyperclustered storm events — all within the last twenty-one years — account for more than half of all claims in the program’s history. The NFIP is not bankrupt in the legal sense, because a federal program cannot be bankrupt. It can only be perpetually insolvent and perpetually re-authorized, which is a different condition with the same arithmetic.

Share

The paradox of accuracy

The instinct is to treat Risk Rating 2.0 as a correction. The old system underpriced risk. The new system prices it accurately. Accurate pricing is better than inaccurate pricing. This seems obvious.

It is not obvious. It depends entirely on what you think insurance is for.

Ian Hacking, in The Taming of Chance, traced how the rise of statistical thinking in the nineteenth century transformed institutions that had previously operated on judgment, custom, and rough approximation into institutions that operated on probability. Insurance was one of the first. The fundamental insight was that individual events are unpredictable but aggregate behavior is not. One house might flood or might not. Ten thousand houses distributed across a geography will flood at a knowable rate. The insurer does not need to know which house. It needs to know the rate. The pool exists to absorb the individual uncertainty that the mathematics have rendered collectively predictable. The price of admission — the premium — is set not at the individual’s actual risk but at the average risk of the pool. This is the mechanism. It works because it is imprecise. The low-risk subsidize the high-risk. The young subsidize the old. The lucky subsidize the unlucky. The cross-subsidy is not a flaw in the system. It is the system.

What happens when you make the pricing precise?

The cross-subsidy disappears. Each person pays according to their own risk profile, and the pool fragments into a collection of individual transactions. The person whose house sits fourteen feet above the flood level pays almost nothing. The person whose house sits two feet above it pays a premium that approaches the expected cost of the flood itself. At the mathematical limit — where pricing is perfectly accurate — the premium equals the expected loss, which means the insurance provides no economic benefit to the policyholder. You are paying a dollar to receive a dollar. The risk transfer has collapsed into an accounting exercise. The pool has not been improved. It has been dissolved.

This is not a theoretical concern. Since FEMA implemented Risk Rating 2.0, new NFIP policy uptake has declined by roughly 11 to 39 percent depending on the magnitude of the premium increase, and renewals of existing policies have dropped by 5 to 13 percent. The declines are largest in lower-income communities. The people leaving the pool are not the people whose risk decreased. They are the people whose risk was most accurately priced — and who discovered they could not afford the accuracy.

The pool is shrinking from the bottom.

What the pool was holding

A flood insurance pool that includes both high-risk and low-risk properties functions. A flood insurance pool that includes only high-risk properties — because the low-risk have been priced into a cheaper tier or have dropped coverage entirely — is not a pool. It is a collection of expected losses, priced at cost, with no surplus to absorb the unexpected. The actuarial term is adverse selection, and it has been understood since at least the 1970s: when pricing becomes granular enough to separate the healthy from the sick, the safe from the dangerous, the cheap from the expensive, the people who remain are disproportionately the people who will file claims. The pool becomes sicker as it becomes more accurately priced. The sicker it becomes, the more expensive it becomes. The more expensive it becomes, the more people leave. The spiral does not reverse.

(I keep circling a formulation here that I cannot land precisely. The NFIP was designed as a public program because the private market would not insure flood risk — the losses were too concentrated, too correlated, too catastrophic. The private market looked at flood risk and said: this cannot be pooled profitably. The government stepped in and said: we will pool it anyway, because the social cost of not pooling it is worse. That was a policy decision, not an actuarial one. Risk Rating 2.0 applies actuarial logic to a program that was created specifically because actuarial logic could not solve the problem. I am not sure whether this means the reform is contradictory or whether it means the original premise was always unstable. Both possibilities lead somewhere I have not yet worked through.)

Consider the specific geography. Along the Gulf Coast, from Galveston to the Florida Panhandle, NFIP premiums for properties in high-risk zones are reaching 20 to 30 percent of the home’s assessed value. A house worth $150,000 carrying $30,000 in annual flood insurance premiums is a house that cannot be sold, because no buyer can finance the carrying cost. It cannot be mortgaged, because lenders require flood insurance in special flood hazard areas and the insurance exceeds what the property can service. It cannot be abandoned without consequences, because the mortgage still exists. The homeowner is locked into a property whose market value the insurance cost has destroyed, unable to sell, unable to refinance, unable to leave, unable to stay without paying a premium that the property itself does not justify. The insurance has not protected the asset. It has priced the asset out of the category of things worth protecting.

There is a particular quality to the FEMA flood map when you look at it on a screen — the color gradations from blue to orange to red, the way a parcel boundary can straddle two zones, the thin black line of your property sitting half in one color and half in another, the precision of it suggesting a certainty that the underlying hydrology does not actually possess. The map is a model. The model is good. The colors are vivid and clean. The experience of locating your own address on it and watching the color tell you what your future costs is not something the model accounts for.

The direction of fragmentation

The NFIP is one program in one country covering one type of risk. The pattern it illustrates is not limited to flood insurance.

Health insurance underwent a version of this fragmentation when the Affordable Care Act’s individual mandate was effectively eliminated in 2019. The mandate existed to keep the pool intact — to ensure that healthy people subsidized sick people, which is the mechanism by which insurance functions. Without it, the pool began to segment. Healthy enrollees left. Premiums rose for those who remained. The coverage became more expensive precisely because it became more concentrated among those who needed it. The same mathematics applied: accurate sorting of who needs insurance most produces a pool composed of the most expensive members, which produces a pool that cannot sustain itself.

Private auto insurance has been granulating for two decades. Telematics devices — small units mounted in vehicles that track speed, braking, cornering, time of day, miles driven — allow insurers to price policies based on individual driving behavior rather than demographic category. Safe drivers pay less. Dangerous drivers pay more. Each driver is priced according to their own data. The cross-subsidy between safe and dangerous drivers shrinks. The dangerous drivers who remain in the pool face premiums that reflect their concentrated risk, and the concentrated risk is expensive enough that some of them stop carrying insurance, which transfers the cost of their accidents to the people they hit, who now bear a risk that the insurance pool was supposed to absorb.

Genetic testing presents the terminal case. If an insurer can assess an applicant’s genome and price the policy according to their heritable disease risk — and in most U.S. states, life insurers and disability insurers face no prohibition against using genetic information — the pool fragments along the most granular possible axis. The person with a BRCA1 mutation pays for the cancer they have not yet developed. The person without it pays almost nothing. The mathematics are impeccable. The pool no longer exists in any meaningful sense, because there is no uncertainty left to distribute. Risk pricing is a form of selection. What it selects, at sufficient resolution, is who deserves to be pooled and who does not — and the category of people who do not deserve pooling grows in exact proportion to the accuracy of the instrument doing the selecting.

What accuracy costs

The pattern is consistent across domains and it moves in one direction. As risk assessment improves — as models become more granular, as data becomes more abundant, as pricing becomes more precise — the cross-subsidy that makes insurance functional erodes. The people identified as high-risk are priced out. The people identified as low-risk see no reason to subsidize anyone. The pool fragments into individual risk profiles, each accurately priced, none of them insured in the sense that the word originally meant.

This creates a problem that better modeling cannot solve, because better modeling is the cause. The accuracy of risk models is their undoing. Perfect information makes insurance impossible — not because the information is wrong but because insurance depends on a specific kind of ignorance: the inability to distinguish, at the point of underwriting, between the person who will need the pool and the person who will not. That inability is what justifies charging both of them the same price. Remove it, and you have not improved insurance. You have replaced it with something else — a system that charges each person the precise cost of their own misfortune, which is not risk transfer. It is prepayment.

Whether a system that converts insurance into prepayment is better or worse than a system that pools risk inaccurately is not a question the actuarial science can answer, because the actuarial science is optimized for accuracy, and accuracy is what produces the fragmentation. The question belongs to a different discipline — one that asks not “what does this person’s risk cost?” but “what happens to a society in which everyone knows exactly what their risk costs and some of them cannot pay it?” The data shows the cost. The data does not show what to do when the cost exceeds what the person can bear. That calculation — the one that weighs the accuracy of the model against the durability of the commons — is not a calculation the model was designed to perform. It may not be a calculation at all. It may be a decision, and decisions require something the model does not have, which is a preference for one outcome over another that cannot be justified by the mathematics alone.

The mathematics are getting better every year. The pool is getting smaller.

-Aimé

Aimé Halden writes Uninsurable, a newsletter about the systems that shape who is protected and who is not. Subscribe for weekly analysis.

In March 2025, Bloomberg’s Billionaires Index crossed a threshold that received less attention than it should have. The combined wealth of the world’s five hundred richest people exceeded $10 trillion for the first time, a figure roughly equal to the combined GDP of Germany, Japan, and India. The top ten alone held more than $1.8 trillion. In the same twelve months, median household wealth in the United States grew by approximately 2.3 percent in nominal terms — roughly tracking inflation, meaning no real gain. Billionaire wealth grew by 64 percent. The disparity is not new. The rate of divergence is.

The instinct is to explain this through narrative. Billionaires are smarter, or luckier, or more ruthless. They built companies, took risks, innovated. Or: they exploited workers, captured regulators, gamed the tax code. Both narratives have evidence. Neither explains the mathematics. Because the mathematics do not require narrative. They do not require genius, corruption, or any particular human quality at all. They require only a differential in return rates — sustained over time, compounded without interruption — and the arithmetic does the rest.

This is worth understanding precisely, because the mechanism is less dramatic and more durable than any story about individual ambition or individual villainy.

Share

The differential

Thomas Piketty’s central observation in Capital in the Twenty-First Century was deceptively simple: when the rate of return on capital exceeds the rate of economic growth — when r > g — wealth concentrates. The formulation is algebraic. It does not depend on policy. It does not depend on character. It depends on a mathematical relationship between two rates, and that relationship has held, with brief interruptions, for most of recorded economic history.

The mechanism works like this. A person who earns a wage can save a portion of it. If median household income is $75,000 and the savings rate is 5 percent, the household accumulates $3,750 per year. At a 4 percent annual return — generous for a savings account, conservative for a diversified portfolio — those savings grow slowly. After thirty years of consistent saving and compounding, the household has approximately $215,000. This is meaningful. It is not transformative. The accumulation is linear in character even if technically exponential, because the base is small and the compounding has limited material to work with.

A person who begins with $10 million in investable capital — not earned income, but capital — and achieves the same 4 percent return adds $400,000 in the first year alone. After thirty years, the $10 million has become roughly $32 million. At 7 percent — a reasonable long-run average for equities — it becomes $76 million. The person contributed no additional labor, took no additional risk beyond what the portfolio already represented, and added no new productive capacity to the economy. The money grew because money, at sufficient scale, grows. The rate of return applies to the base, and the base is what determines the outcome. A 7 percent return on $75,000 in savings is $5,250. A 7 percent return on $10 million is $700,000. Same percentage. Same market. Same economy. Radically different trajectory. The divergence is not a distortion of the system. It is the system, operating as compound interest operates:

(There is something in this that I keep trying to state more precisely and have not managed yet. The mathematics are impersonal. A compound function does not know or care whether its input is a retirement account or a hedge fund. But the impersonality of the function does not mean the outcomes are impersonal. A system that treats all inputs identically will amplify whatever differences exist in those inputs. Identical treatment of unequal starting positions is not neutrality. I am not sure whether that constitutes a design flaw or a design feature, and I suspect the answer depends on whether you think mathematics can have intentions.)

What capital can purchase

The differential in return rates is only the first mechanism. The second is that capital, at sufficient scale, purchases access to higher returns — returns that are unavailable to those below the threshold.

A household with $50,000 in retirement savings invests through a 401(k), typically in index funds with annual fees of 0.03 to 0.5 percent. The returns track the broad market. A household with $10 million invests through a private wealth management firm that provides access to private equity, venture capital, real estate syndications, and tax-advantaged structures that are either explicitly restricted to accredited investors (net worth above $1 million, excluding primary residence) or practically restricted by minimum investment thresholds of $250,000 to $5 million. The returns on these instruments have historically exceeded public market returns by 2 to 4 percentage points annually. The differential is not speculative. Cambridge Associates’ data on private equity performance over the past twenty-five years shows a median net internal rate of return exceeding public equity benchmarks by roughly 3 percent per year.

Three percentage points sounds modest. Over thirty years of compounding, it is the difference between $76 million and $174 million on the same $10 million base.

But the access premium extends beyond investment returns. Capital at scale purchases lower borrowing costs — a billionaire borrows against assets at rates below 2 percent, while a median household carries credit card debt at 22 percent. It purchases optionality — the ability to wait, to absorb losses, to hold illiquid assets until conditions improve, none of which is available to someone who needs next month’s paycheck. It purchases geographic arbitrage, legal arbitrage, regulatory arbitrage. It purchases the ability to structure income as capital gains taxed at 20 percent rather than wages taxed at 37 percent. Each of these is individually rational, individually legal, individually available to anyone who meets the threshold. The thresholds are what make them mechanisms of concentration rather than mechanisms of opportunity.

The investor who can wait is not smarter than the investor who cannot. They are more capitalized.

Subscribe now

The feedback structure

A system that compounds returns at rates proportional to the existing base, in which larger bases access higher return rates, and in which higher returns enable access to further return-enhancing instruments, is not a system that tends toward equilibrium. It is a system that tends toward concentration. The question is not whether concentration occurs but at what rate and whether anything interrupts the compounding.

Historically, the interruptions have been specific and violent: world wars, revolutions, pandemics, hyperinflation. Piketty documented that the period of relatively low wealth inequality in the mid-twentieth century — the anomaly, not the norm — was produced by the destruction of capital in two world wars and the subsequent policy environment (high marginal tax rates, strong labor unions, Bretton Woods capital controls) that restrained the return differential. That environment has been systematically dismantled since the 1980s. Top marginal tax rates fell from 70 percent to 37 percent. Capital gains rates fell from 28 percent to 20 percent. Union membership declined from 20 percent to 10 percent of the workforce. Capital controls were lifted. Financial deregulation expanded the instruments available to large capital holders. Each policy change was individually justified on efficiency grounds. Collectively, they restored the conditions under which r > g operates without interruption.

The result is visible in the data. In 1980, the top 1 percent of U.S. households held approximately 22 percent of total wealth. By 2000, this had risen to 28 percent. By 2024, it reached roughly 33 percent. The trajectory is not linear — it accelerates, because concentration feeds further concentration, because the return differential widens as the base grows, because access to higher returns compounds the advantage, because the tax treatment of capital income diverges further from the tax treatment of labor income at every level where the distinction matters.

None of this requires conspiracy. None of it requires exceptional talent. None of it requires a single decision by a single person to concentrate wealth at the expense of others. It requires only that the rules of arithmetic continue to apply to a system in which starting positions are unequal and return rates favor larger positions. The spreadsheet runs. The math accumulates. The gap, year by year, becomes visible — not as a decision but as a gradient, the way a river does not decide to erode its banks but does so because water moves downhill and the bank was already there.

What concentration produces

The consequences of concentration are structural, not personal. They alter the operating environment for everyone inside the system, including those who benefit from it.

Consider housing. An open house in a mid-range neighborhood: the listing sheet printed on card stock, the agent’s name in small serif type at the bottom, the square footage rounded to the nearest ten. Twenty people walk through the same three-bedroom in a single Saturday afternoon, opening closet doors, running fingertips along countertops, checking the water pressure — and none of them know that an institutional offer arrived by email before the house was listed. When institutional capital enters residential real estate markets — as it did aggressively from 2020 onward, with firms like Blackstone, Invitation Homes, and American Homes 4 Rent acquiring hundreds of thousands of single-family properties — the effect is not that individual home buyers compete against individual home buyers. It is that households bidding with mortgage-constrained budgets compete against entities bidding with fund-level capital that has no borrowing constraint, no contingency requirement, and no emotional attachment to any particular property. The institutional buyer can offer cash at above asking price and absorb the loss if the market dips, because the loss on a single property is noise in a portfolio of 80,000 units. The individual buyer cannot. The market clears at a price set by the better-capitalized participant. The individual buyer is not outbid because they are less capable. They are outbid because capital, at scale, operates in a different market than labor, and the two happen to be competing for the same asset.

The same dynamic replicates in healthcare (private equity acquisition of physician practices and hospitals has restructured pricing and care delivery around return targets), in education (endowment size increasingly determines institutional viability and the resources available to students), in media (ownership concentration has collapsed local news infrastructure in regions where advertising revenue cannot sustain private equity return expectations), and in politics, where the cost of influence scales with the capital available to deploy it.

Inequality is rarely the result of malice. It is the result of systems working exactly as designed for people with capital.

There is an asymmetry embedded in the mechanism that is worth naming. The person who benefits most from compound returns did not design compound interest. The person harmed most by the access premium did not design accredited investor thresholds. The policy environment that restored r > g was not architected by a single actor. It was the accumulated output of thousands of rational decisions — each defensible, each incremental, each responsive to the incentives facing the person who made it. A legislator who votes to lower capital gains taxes is responding to donor incentives and constituent preferences and economic arguments that are, taken individually, coherent. A fund manager who structures returns to minimize tax liability is fulfilling a fiduciary obligation. A wealth advisor who moves a client into private equity is optimizing the portfolio.

What I do not yet know how to think about clearly is whether a system that produces concentration through arithmetic — rather than through intention — is more or less susceptible to intervention than a system that produces it through identifiable decisions. A decision can be reversed. A policy can be reformed. An individual can be held accountable. But a mathematical tendency — a property of how compound functions operate on unequal inputs — is not a decision anyone made. It is a feature of the structure. Reforming the structure requires changing the relationship between r and g across an entire economy, which requires either reducing the returns available to capital (which capital will resist and route around) or increasing the growth rate available to labor (which decades of policy have moved in the opposite direction). Whether either is achievable within the existing institutional framework is a question the data does not answer. What the data does show is the curve, and the curve does not flatten on its own.

The math does not need anyone to believe in it. It runs regardless.

-Aimé

Aimé Halden writes Uninsurable, a newsletter about the systems that shape who is protected and who is not. Subscribe for weekly analysis.

In 2014, Amazon began building a hiring tool. The system was designed to review résumés and score applicants on a scale of one to five, the way the company rated products. Engineers trained it on ten years of hiring data — patterns extracted from the résumés of people the company had previously selected. The model learned to identify what a successful Amazon applicant looked like. By 2015, the team noticed the model had taught itself to penalize résumés containing the word “women’s,” as in “women’s chess club” or “women’s studies.” It downgraded graduates of two all-women’s colleges. The engineers attempted to neutralize the variable. The model found proxies. They corrected the proxies. It found others. In 2017, Reuters reported that Amazon abandoned the project. The tool was never used as the sole determinant in hiring. But the pattern it revealed was not a malfunction. The model had done what it was told: identify what a successful Amazon hire looks like. What a successful Amazon hire looked like, based on ten years of data, was a man.

The instinct is to treat this as a cautionary tale about flawed data. Fix the data, fix the outcome. But the data was not flawed. The data was accurate. Amazon had, in fact, hired overwhelmingly male candidates for technical positions over the previous decade. The historical record was correct. The model ingested it faithfully. The bias was not introduced by the algorithm. It was already in the institution, encoded in a decade of human decisions, and the algorithm made those decisions operational at scale. The difference between a biased hiring manager and a biased algorithm is not the bias. It is the speed, the consistency, and the invisibility.

A human interviewer who systematically penalized candidates from women’s colleges would eventually be noticed. A model that does it is noticed only when someone audits the model — and most models are not audited.

Share

The inheritance problem

A credit score is a three-digit number, typically ranging from 300 to 850. It is calculated by three private companies — Equifax, Experian, TransUnion — using proprietary formulas that are not publicly disclosed in full. FICO, the most widely used scoring model, weighs five categories: payment history, amounts owed, length of credit history, new credit, and credit mix. These categories sound neutral. They are not. Each one encodes assumptions about what kind of financial behavior constitutes reliability, and those assumptions were calibrated on the behavior of populations that had access to credit in the first place.

Length of credit history, for instance, rewards duration. A person whose parents added them to a credit card at eighteen has a longer history than a first-generation immigrant who opened their first account at thirty. The immigrant’s financial behavior may be identical in every respect — same payment record, same debt ratios, same income — and they will score lower because the metric measures time, and time is not distributed equally. Amounts owed, weighted at 30 percent, penalizes those who use a higher proportion of available credit — but available credit is itself determined by previous credit scores, creating a recursion in which past disadvantage compounds into present penalty. A person with a $2,000 limit who carries $1,000 in debt is measured differently from a person with a $20,000 limit who carries $1,000 in debt, even though the dollar amount is identical. The metric does not measure financial responsibility. It measures financial headroom. And headroom is a proxy for prior advantage.

Cathy O’Neil documented this architecture in Weapons of Math Destruction, tracing how models trained on historical data do not merely reflect historical patterns — they operationalize them. A model that scores creditworthiness based on past credit access is not predicting future behavior in a vacuum. It is predicting future behavior as a function of prior access, which is itself shaped by redlining, by discriminatory lending, by decades of policy that determined who received credit and on what terms. The algorithm inherits the history. It does not interrogate it. O’Neil called these models “opinions embedded in mathematics.” The phrase is precise. The score presents itself as a measurement — as neutral as a thermometer reading — but what it measures is a combination of behavior and circumstance, and the circumstance carries the weight of decisions made long before the person being scored was born.

(Between the discriminatory lending practices of the 1960s and a credit score generated in 2025, there are six decades, multiple policy reforms, and several generations of algorithmic iteration. The bank that redlined a neighborhood in 1965 no longer exists. The FICO model that penalizes the grandchild of someone denied a mortgage in that neighborhood is a different instrument built by different people. The effect persists. The responsibility has evaporated somewhere in the iterations. I do not yet know how to think clearly about accountability across that kind of distance — whether the absence of a responsible party is a genuine philosophical problem or merely a convenient one.)

What the score touches

The credit score was designed for lending. It is no longer used only for lending. Landlords use credit scores to screen tenants. Employers in most U.S. states can check credit history as part of hiring decisions. Auto insurance companies in forty-seven states use credit-based insurance scores to set premiums. Utility companies use them to determine deposit requirements. The three-digit number that was built to assess the likelihood that a borrower would repay a loan now determines whether a person can rent an apartment, get a job, insure a car, or turn on their electricity.

Each of these expansions follows its own logic. Landlords use credit scores because they predict, statistically, the likelihood of rent default. Insurers use them because credit history correlates, actuarially, with claims frequency. Employers use them because — the reasoning gets thinner here — financial instability is assumed to indicate unreliability or vulnerability to fraud. Each user of the score can point to a statistical correlation that justifies the practice. The correlations are real. The question is what a correlation means when the variable being measured is itself shaped by the conditions being predicted.

A person denied a job because of a low credit score may have a low credit score because they were previously denied employment. A person charged higher insurance premiums because of poor credit may have poor credit because a medical emergency forced them into debt, and the medical debt exists because their previous insurer denied a claim. A person denied housing because of a thin credit file may have a thin file because previous landlords required the credit history they could not yet build. Each system treats the score as an input. None of them treat it as an output of the other systems — which is what it also is.

The circularity is not a bug. It is the architecture.

Consider the texture of the interaction. A person applies for an apartment. They have toured the unit, measured the bedroom doorframe with a tape measure to confirm a bed frame will fit through it, checked the water pressure in the kitchen sink. The property management company runs a credit check through an automated screening service. The applicant never meets the person making the decision because there is no person making the decision. The screening service returns a recommendation — accept, deny, or require additional deposit — based on a threshold the property manager set months ago and has not revisited. The denial arrives by email, a two-paragraph template in the system’s default sans-serif. At the bottom, in smaller type than the denial itself, a notice required by the Fair Credit Reporting Act informs the applicant of their right to request a copy of the report that was used. The right is real. The process for exercising it is bureaucratic enough that most people do not.

The question of neutrality

The architecture of these systems distributes a specific kind of power. Not the power to decide — no individual decides — but the power to determine the framework within which outcomes become inevitable. The person who designed FICO’s weighting system made a choice about what matters: that payment history should count for 35 percent and credit mix for 10 percent. This is a judgment, not a discovery. A different weighting would produce different scores, different populations flagged as risky, different distributions of access and denial. The choice was made. It was not made democratically, or publicly, or with input from the populations it would sort. It was made by a private company, in the course of building a product, and it became infrastructure.

This same structure replicates in every domain where algorithms make or shape decisions. Predictive policing systems like PredPol — now rebranded as Geolitica, in a nominal distancing from its own reputation — deploy patrol officers to neighborhoods flagged as high-risk by models trained on historical arrest data. Neighborhoods with more prior arrests get more patrols, more patrols produce more arrests, more arrests confirm the model’s prediction. The feedback loop is not hidden. Multiple studies have demonstrated that the predictions track policing patterns more closely than they track crime patterns. The algorithm is not predicting where crime will occur. It is predicting where police will be sent, which is a function of where police have already been, which is a function of enforcement priorities that predate the algorithm by decades. The model inherited the pattern. It did not create it. But it made the pattern faster, more consistent, and harder to see because the output carries the authority of mathematics rather than the visible discretion of a precinct commander.

Optimization is not neutral. The act of optimizing requires a target, and the choice of target determines what is maximized, what is minimized, and what is ignored. A hiring algorithm optimized for “candidates who resemble past successful hires” will reproduce the demographics of past hiring. A credit model optimized for “predicting default among populations with established credit histories” will penalize populations without established histories. A policing model optimized for “deploying resources to areas with high arrest rates” will intensify policing in already-policed areas. Each optimization is technically correct. Each achieves its stated objective. And each encodes a set of values into a system that presents itself as value-free — a system that claims to measure reality while actively constructing it, sorting populations according to patterns it inherited and then treating those patterns as though they were discovered rather than chosen, and the further the algorithm runs, the more the pattern solidifies, because each output becomes the next cycle’s input and the distance between the original human decision and its algorithmic descendant grows until the decision’s fingerprints are no longer visible on the surface of the data, which is precisely where the accountability evaporates, not in any single moment of error or malice but in the accumulation of iterations, each one correct, each one justified, each one moving the pattern one step further from the point where someone could have chosen differently.

Subscribe now

What replaces discretion

Algorithmic systems were adopted, in most domains, for a defensible reason: human judgment is inconsistent. Two loan officers reviewing the same application will reach different conclusions. A hiring manager’s assessment shifts depending on whether the interview is before or after lunch. An algorithm removes that variation. It applies the same rules to every applicant. It produces the same output for the same inputs. Consistency, at scale, looks like fairness.

But what the algorithm actually does is replace individual discretion with structural discretion — the discretion embedded in the training data, the weighting of variables, the choice of optimization target. Individual discretion is visible, local, and contestable. Structural discretion is none of these things. It does not reside in any person. It resides in the system. And the system, when questioned, can produce a mathematically precise explanation for every output, which looks like transparency but functions as a wall.

The question is what happens to accountability when a consequential decision has no identifiable decision-maker. When a human denies your loan application, there is a person. You can appeal. You can argue. The person can be wrong, and their wrongness can be named. When an algorithm denies it, there is a score. The score can be explained. The explanation will be technically accurate. And the process of challenging it requires you to argue not with a judgment but with a formula — which requires expertise most people do not possess, resources most people cannot afford, and access to proprietary systems most people are not granted.

The burden of proof shifts from the decision-maker to the person affected by the decision.

The system does not need to be right about everyone. It needs to be right often enough that the exceptions — the people wrongly scored, wrongly denied, wrongly sorted — are too few to organize, too expensive to litigate, too invisible to matter to the metrics by which the system evaluates itself. Whether it works for the people inside it is a different question, measured by a different instrument, and that instrument does not yet exist.

-Aimé

Aimé Halden writes Uninsurable, a newsletter about the systems that shape who is protected and who is not. Subscribe for weekly analysis.

In January 2025, the Bureau of Labor Statistics reported that approximately 59 million Americans — roughly 36 percent of the workforce — earned income through independent contracting, freelancing, or gig platform work in the previous twelve months. This number has been growing steadily since 2015 and accelerating since the pandemic. The IRS tracks these workers through 1099 forms. The Department of Labor does not count them as employed. Unemployment insurance does not cover them. Workers’ compensation does not apply to them. Employer-sponsored health insurance, which covers 155 million Americans, excludes them by definition. Social Security contributions are their own responsibility, at a higher effective rate than salaried employees pay, because they fund both the employer and employee portions. Disability coverage, paid family leave, pension contributions — none of these exist unless the worker purchases them individually, at retail prices, without employer subsidy.

Fifty-nine million people. This is not a fringe category. This is larger than the population of any single U.S. state. It exceeds the combined enrollment of Medicare and Medicaid. And yet the regulatory architecture that determines who receives protection in this country — unemployment benefits, workplace safety enforcement, retirement security, health coverage — was built on a model that assumes these people do not exist.

The assumption is not accidental. It is structural. The entire system of worker protection in the United States was designed around a specific arrangement: one employer, one employee, continuous tenure, defined benefits. This arrangement, which peaked in the mid-twentieth century, produced most of the safety infrastructure Americans still rely on. It assumed stability. It assumed legibility. And legibility, in this context, means something specific: the ability of a system to read, classify, and act upon a person’s circumstances using the categories the system has built for itself.

Share

The readable and the invisible

James C. Scott, in Seeing Like a State, traced how large institutions simplify human complexity into categories they can manage. Forests become timber yields. Villages become cadastral maps. Citizens become tax registries. The simplification is not optional — it is the mechanism by which the institution operates at scale. A state that cannot categorize its population cannot tax, conscript, inoculate, or insure that population. Legibility is the precondition of governance. Scott was writing about states, but the framework applies with equal precision to every large system that processes people: insurance underwriting, credit scoring, benefits administration, healthcare billing. Each system constructs categories. Each system processes individuals through those categories. Each system produces coherent, defensible outputs for the people who fit the categories and silence about the people who do not.

The American safety net is a legibility machine. It reads W-2 forms. It reads employer-sponsored group plans. It reads payroll deduction records. It reads continuous employment histories. These are the inputs it was built to process. A person with a W-2, a single employer, and twelve months of continuous employment is legible to every system simultaneously: the employer’s insurance carrier, the state unemployment office, the Social Security Administration, the IRS. That person exists in the system. Their risks are pooled, their benefits calculated, their protections activated. The architecture works for them because it can see them.

A person with fourteen 1099 forms from fourteen different platforms, no single client relationship exceeding three months, income that fluctuates by 40 percent quarter to quarter, and no employer-sponsored coverage of any kind is not illegible in the sense of being invisible. The IRS can see them — their tax obligations are clear. The system can see them when it wants to collect. It cannot see them when they need protection. The asymmetry is not a failure of vision. It is a design specification. The system was built to process a different shape of person, and when a person arrives in the wrong shape, the system does not reject them. It does not even acknowledge the mismatch. It processes what it can read and produces silence about the rest.

(I am not certain this framing is complete. Scott’s argument is that legibility serves the institution, not the individual — that simplification benefits the entity doing the simplifying. But what I keep circling is whether modern systems have moved beyond even that. The institution benefits from legibility, yes. But the illegible person does not merely fail to benefit. They actively bear costs that the legible person does not. The gig worker is not just unprotected. They subsidize the system that excludes them — through higher self-employment taxes, through retail-price insurance, through the absence of bargaining power that comes from being uncategorizable. The illegibility is not neutral. I do not yet have a precise term for what it is instead.)

The shape of exclusion

Consider the specific mechanics. A driver for a rideshare platform works forty hours per week. The platform classifies the driver as an independent contractor, which is a classification decision — not a description of the work arrangement, but a determination about which regulatory category the worker occupies. The classification has consequences that cascade. As a contractor, the driver is ineligible for the platform’s group health insurance. The driver must purchase individual coverage, which costs more per unit of coverage because individual policies cannot achieve the risk pooling that group plans provide. The driver is ineligible for unemployment insurance if the platform reduces available rides. The driver is ineligible for workers’ compensation if injured while driving. The driver bears the full cost of vehicle maintenance, fuel, and commercial insurance, none of which is deductible at rates that offset the actual expense.

Each of these consequences follows from the classification. The classification follows from a legal determination that the platform made and lobbied to preserve. The legal determination follows from a regulatory framework that defines “employee” using tests designed in the 1930s for industrial labor relationships. The tests measure supervision, location, scheduling, and tool provision. A rideshare driver, supervised by algorithm rather than foreman, working from a personal vehicle rather than a factory floor, scheduled by demand prediction rather than a shift roster, fails these tests. Not because the tests are wrong about what they measure, but because what they measure no longer describes the relationship they were designed to identify.

The driver works. The driver earns. The driver pays taxes. The driver bears risk. The driver is, by every functional measure, performing labor that creates value for the platform. But the system that determines whether this person receives protection — health coverage, injury compensation, income security — cannot read the arrangement. It looks at the driver and sees a 1099. It processes the 1099 correctly. It produces the correct tax obligation. And it produces nothing else, because nothing else was requested by the form.

The form is a three-by-eleven-inch piece of paper. The typeface is Helvetica. The boxes are small and evenly spaced, designed for a printer that deposits ink in a single pass. There is a box for nonemployee compensation. There is no box for hours worked, for injuries sustained, for dependents covered, for risk absorbed. The form measures what it was designed to measure. Everything it was not designed to measure is, from the system’s perspective, nonexistent.

The cost of not counting

The rationality of each actor in this arrangement is not in question. The platform classifies workers as contractors because the classification reduces labor costs by an estimated 20 to 30 percent. This is a defensible business decision. The regulatory agencies apply the tests they have because the tests are established law. This is correct procedure. The insurance market prices individual policies higher than group policies because individual risk pools are smaller and more volatile. This is sound actuarial practice. The IRS collects self-employment tax at the combined rate because the law requires it. This is the statute operating as written.

Every actor is behaving rationally. Every institution is functioning correctly. Every rule is being applied as designed.

Fifty-nine million people with no unemployment protection in an economy where platform algorithms can reduce their income overnight. Fifty-nine million people purchasing health coverage at retail prices in a market designed around group negotiation. Fifty-nine million people with no employer-funded retirement contribution in a system where retirement security was built on the assumption of employer funding. Fifty-nine million people bearing the full cost of workplace injury — medical bills, lost income, rehabilitation — in a system where that cost was supposed to be socialized through workers’ compensation pools that they are excluded from by classification.

The numbers are large enough to be abstract. They should not be. Each of those 59 million people encounters the system’s categories at specific moments: when they are injured and discover no workers’ compensation exists for them. When a pandemic closes their income source and the unemployment office has no category for their claim. When they turn sixty-five and discover their Social Security benefit is lower than expected because their self-employment income was structured in ways the formula does not reward. These encounters are not abstract. They are the moments where the gap between what the system can see and what the person needs becomes a lived experience. The system processes the encounter correctly. The person absorbs the gap.

Modern suffering is rarely caused by villains. It is caused by spreadsheets.

Subscribe now

What legibility costs

This is the pattern I keep finding across domains, and it is becoming more precise as the evidence accumulates: a system that works correctly for its stated purpose may work catastrophically for the people inside it — not because the system fails, but because the system succeeds at something narrower than what the people inside it need. Insurance succeeds at pricing risk and fails at distributing protection. Employment law succeeds at classifying relationships and fails at recognizing labor. Tax collection succeeds at revenue extraction and fails at distinguishing between a sole proprietor with negotiating power and a gig worker without any. The system sees what it measures. Everything else is noise.

The question this raises is whether better measurement solves the problem. The instinct — and it is a reasonable instinct — is that if the categories are wrong, we should update the categories. Make the system see gig workers. Expand the definition of employment. Redesign the form. But Scott’s argument suggests a more uncomfortable possibility: that the act of making something legible to a large system necessarily simplifies it, and that simplification necessarily excludes something. You can change what gets excluded. You cannot eliminate exclusion from a system that operates through categories. A broader category captures more people and loses more nuance. A narrower category captures nuance and loses more people. The tradeoff is structural, not political. It belongs to the act of measurement itself.

Whether this means the current arrangement is inevitable or merely the product of choices that could be made differently is a question I cannot answer from the analysis alone. What I can observe is this: the number of people who do not fit the system’s categories is growing. The categories are not expanding to match. And the gap between the system’s legibility and the population’s reality is producing a class of people who are visible to the system only when they owe something and invisible when they are owed something in return. That asymmetry — legible as taxpayers, illegible as claimants — is not a temporary condition. It is the system’s current operating specification. Whether it remains the specification depends on whether the system measures the cost of what it cannot see. The system, so far, does not.

-Aimé

A handheld inventory scanner weighs about twelve ounces. It fits in the palm. Each time a warehouse worker logs an item — picked from a shelf, placed in a bin, moved to a conveyor — the scanner emits a short, high-pitched beep. Across an eight- or ten-hour shift, the sound becomes ambient. Workers stop noticing it after the first hour. What they do not stop noticing is the absence of it — the silence that means they have fallen behind rate, that the algorithm monitoring their throughput has flagged a gap, that a supervisor may appear. In Amazon’s fulfillment centers, the target pick rate is calibrated to the capacity of the system, not the capacity of the person holding the scanner. The beep is not the problem. The silence is.

Amazon operates more than 1,000 of these facilities globally, employing roughly 1.5 million people. The architecture of each building is designed around a single proposition: minimize the time between a customer clicking “Buy Now” and a package leaving the dock. Every process is measured. Every measurement has a target. Every target is calibrated to the system’s capacity, and the humans inside the system are expected to match it.

In 2023, the company’s net income increased roughly 190 percent. In the same twelve months, it eliminated approximately 27,000 positions. Meta posted a 201 percent increase in quarterly profit and cut 10,000 jobs. Alphabet reported record revenue and laid off 12,000 employees. These companies were not cutting costs to survive. They were cutting costs to optimize. The distinction — between managing crisis and engineering margin — changes what the numbers mean.

Share

The logic of removal

A company has revenue and expenses. Profit is the gap. Widening the gap is the job. One can increase revenue or decrease expense, but decreasing expense is faster, more predictable, and more legible to investors. A cost line eliminated is immediately visible on a balance sheet; a revenue opportunity foregone is invisible. The asymmetry is structural: cutting is concrete, building is speculative. Financial markets reward certainty, and cost reduction is the most certain move available. Meta’s Mark Zuckerberg declared 2023 the “Year of Efficiency.” Amazon’s annual letter to shareholders referenced “operational efficiency” fourteen times. The term operates as a directive: applied to any cost structure, it targets whatever the spreadsheet identifies as excess. Headcount, benefits, safety margins, maintenance budgets — all are legible as cost lines, and cost lines are what efficiency eliminates.

The Senate HELP Committee’s 2024 investigation into Amazon’s warehouse operations found that the company’s internal total injury rate — including injuries not required to be disclosed to OSHA — was just under 45 per 100 workers. Nearly half the workforce, in a given year, was injured. The publicly reported recordable rate ran 6.5 per 100 employees, 71 percent higher than the rate at comparable non-Amazon warehouses. Sixty-nine percent of workers surveyed had taken unpaid time off due to pain or exhaustion. Fifty-three percent attributed their injuries directly to productivity pressure. Musculoskeletal disorders — the specific cost of performing the same reaching, bending, and lifting motions at speed for hours — account for a disproportionate share. These injuries are not accidents. They are the predictable output of a system that measures throughput and does not measure what maintaining that throughput does to the body over time. The injury rate is not a failure of the efficiency model. It is the efficiency model’s externality — the cost that appears on no spreadsheet, in no quarterly report, in no earnings call where the word “efficiency” was used fourteen times.

The metrics work. Amazon ships billions of packages per year at speeds that would have been logistically impossible two decades ago. The metrics are not miscalibrated. The targets are not arbitrary. The system is, by every measure designed to assess it, performing.

The boundary

Efficiency, as a practice, requires a boundary. A line is drawn around the things being optimized, and everything inside that line improves. Everything outside it is, by definition, someone else’s problem. The smaller the boundary, the more efficient the interior becomes — and the more cost accumulates on the exterior.

The people on Meta’s cost line had names, mortgages, health insurance that ended the day they were let go. None of these facts are legible to the efficiency metric. The metric sees full-time equivalents and total compensation burden. It sees what it was designed to see. In healthcare, hospitals that spent two decades optimizing nurse-to-patient ratios — the leanest, the most efficient by every measure their boards reviewed — were the ones that broke first when a respiratory virus arrived, because surge capacity was the cost that had been cut. In retail, companies that shifted to part-time workers scheduled by algorithm, cutting hours to just below the threshold that triggers benefit obligations, saved money that appeared on the balance sheet and created costs — turnover, lost institutional knowledge, workers holding three jobs and performing well at none — that appeared nowhere.

An efficiency metric incentivizes the removal of cost. If cost includes human labor, the incentive is to remove labor or reduce its price — lower wages, fewer benefits, faster rates, automation. Each choice is rational within the boundary the metric draws. Each transfers cost from the entity being measured to entities outside the measurement: workers, families, public health systems. The transfer is not a malfunction. It is the mechanism. Optimization is always optimization of something, which means it is always optimization away from something else. The spreadsheet does not know what it is optimizing away. It knows only that whatever it is costs more than the alternative.

What gets optimized out

This produces brittleness. Systems stripped of redundancy perform well under normal conditions and fail under abnormal ones, because the buffer was the cost that got cut. Southwest Airlines spent years optimizing its crew scheduling for maximum aircraft utilization — planes in the air, not on the ground; crews rotating through tight connections with minimal downtime. The system was efficient. Then, in December 2022, Winter Storm Elliott disrupted enough flights that the scheduling software could not reassign crews. Not because the software crashed, but because it had been built with no tolerance for deviation. The airline canceled more than 16,700 flights in ten days. Two million passengers were stranded over the holidays. The eventual cost exceeded $1.2 billion. The airline had known the software was outdated. It had been warned. But modernizing the scheduling system was a cost, and the existing system was working — by every metric used to evaluate it, it was working well — right up to the point where it encountered a situation it had no capacity to absorb. The scheduling system that had saved the airline money every quarter for years could not handle a single week of weather. The margin was the cost that got cut.

Geoffrey West, in his work on organizational scaling, identified a version of this pattern in biology: living systems maintain inefficiency as a survival strategy. An organism that operated at peak efficiency — no fat reserves, no redundant neural pathways, no excess lung capacity — would be optimized for exactly one set of conditions and would die the moment those conditions changed. Living systems carry slack. They are, by the narrow measure of input-to-output ratio, inefficient. They are also durable. Corporate systems face the opposite selective pressure. Markets reward tightness. The quarterly earnings call is a report on how much slack was removed, how lean the operation has become. The incentive runs in one direction: toward the minimum viable configuration, toward the system that performs well today and has nothing left to absorb tomorrow.

Whether the measurement boundary is a design choice — something that could be drawn to include injury rates and surge capacity alongside throughput and margin — or whether it is an inevitability of measurement itself is not yet clear. Both possibilities lead somewhere difficult. If the boundary is a choice, the failure to redraw it is a decision. If it is inherent to measurement, then every act of optimization necessarily produces an exterior that absorbs what the interior discards.

Subscribe now

What the data does show is where the boundary currently sits. The workers who absorb the injuries are the workers who move the packages. The laid-off engineers were the ones building next year’s product. The communities priced out of insurance are the communities that generate the premiums. The exterior is not separate from the system. It is where the system’s inputs come from. At some point the exterior stops being someone else’s problem — not because the system develops a conscience, but because the system runs out of exterior. The data, by design, only covers the interior. The exterior is where the costs collect, and the costs are not yet anyone’s metric.

Not yet.

-Aimé

In May 2023, State Farm announced it would stop selling new homeowners policies in California. The statement was three paragraphs long and cited “historic increases in construction costs” and “a challenging reinsurance market.” Within eighteen months, seven of the state’s twelve largest home insurers had either paused or restricted new policies. Allstate had already frozen sales in 2022. AIG and Chubb pulled back. By early 2025, the California FAIR Plan — the state’s insurer of last resort, designed as a temporary backstop — held nearly 452,000 policies, more than double its count from 2020. The backstop had become the market.

Share

Florida followed a different timeline but arrived at the same destination. Farmers left. Progressive reduced exposure. AAA retreated. Smaller carriers declared insolvency outright. In Louisiana, homeowners now pay roughly three times the national average for coverage. And these are just the states where the crisis is loudest. Insurify’s analysis flagged fifteen states at risk of the same pattern. The geography is expanding. It is no longer coastal.

The instinct is to read this as failure. Insurance companies fleeing. Homeowners abandoned. A system breaking under pressure. But that reading misses the more interesting question, which is whether the system is breaking at all — or whether it is doing precisely what it was built to do.

Insurance, at its origin, is one of the more elegant financial inventions

The basic mechanics are old: a group of people who face uncertain, potentially catastrophic loss agree to pool their resources. Each contributes a small, predictable amount. When catastrophe strikes one member, the pool pays. No individual can predict their own loss, but the group’s losses are statistically predictable. The law of large numbers converts individual chaos into collective order. This is beautiful mathematics. It transferred catastrophic risk from the person least able to bear it — the individual — to the entity best positioned to absorb it — the pool.

The modern insurance company is still doing this, technically. It still collects premiums. It still pays claims. But somewhere between the actuarial tables and the quarterly earnings call, the emphasis shifted. Kenneth Arrow identified the tension in 1963: insurance markets are inherently unstable because the insurer and the insured have different information about risk, and that asymmetry warps the market. Arrow was writing about healthcare, but the structure applies everywhere. The question stopped being “how do we pool risk?” and became “how do we identify risk?” — which is a different question with different consequences.

Here is how the shift works

An insurer’s profitability depends on the gap between premiums collected and claims paid. The wider the gap, the better. There are two ways to widen it: charge more, or pay less. But there is a third approach, subtler and more powerful than either — select better. If the insurer can identify which policyholders are likely to file claims and exclude them before they do, the pool becomes cheaper to maintain. The remaining customers are lower-risk, claims fall, and profitability rises. The insurance company has not improved outcomes. It has improved its portfolio.

The tools for this selection have become extraordinary. Satellite imagery assesses roof condition. Credit history predicts claim frequency. Zip codes carry actuarial weight that their residents cannot see. A homeowner in a coastal Florida county might have a perfect record — no claims filed, roof replaced last year, hurricane shutters installed — and still receive a nonrenewal notice because the county-level risk model recalculated. The individual record is irrelevant. The geography is the variable. The algorithm prices the portfolio, not the person.

(It is worth pausing on the precision involved. These are not crude instruments. The actuarial models processing risk across millions of households are among the most sophisticated prediction engines in commercial use. They are more accurate than most medical diagnostics. The science is sound. The math works. If you were running an insurance company — if your job were to keep the pool solvent and the shareholders satisfied — you would make the same decisions.

That is worth sitting with for a moment, because the system’s logic is not wrong. Its logic is what makes the consequences so difficult to argue with.)

So the logic holds. Every actor in the chain is behaving rationally. The insurer identifies risk. The reinsurer prices it. The shareholder demands a return. The regulator permits what the law allows. No one is making an error. No one needs to.

Now look at what that rationality produces

The better insurers get at identifying risk, the more finely they can segment their customer base, and the more people fall into categories that are unprofitable to cover.

The insurance pool — which only works when it includes a mix of high and low risk — fragments into ever-smaller subgroups, each priced according to its own statistical destiny. The low-risk pay less. The high-risk pay more, or find themselves priced out entirely. The pool that was supposed to distribute risk has become a sorting mechanism that concentrates it.

Consider what happens at the margins

A nonrenewal notice arrives in a standard envelope, the company logo in the upper left corner, the paper weight identical to the quarterly statement that preceded it. There is no red ink, no bold type, nothing to distinguish it from routine correspondence. The homeowner in a newly reclassified flood zone opens it and learns that premiums will double. The house, which was an asset yesterday, is a liability today — not because anything about the house changed, but because the model updated. The homeowner cannot sell at the previous value because the next buyer faces the same insurance costs. Cannot refinance because the lender requires coverage the homeowner cannot afford. Cannot stay because the carrying costs have exceeded the household budget. The actuarial table has made a determination about this person’s future. No human made a decision. The spreadsheet updated. And this is what the California exodus looks like from the inside — not a sudden catastrophe but a quiet reclassification, household by household, county by county, each one the product of a model recalculating, a threshold breached, a portfolio rebalanced, until the insurer’s map of California no longer matches the state’s map of itself, until the places where people live and the places where coverage exists are two different geographies, and the distance between them is not measured in miles but in basis points on a reinsurance treaty that the homeowner will never see and could not read if they did. State Farm did not leave California because homes were burning. State Farm left California because its models determined that the probability of homes burning, multiplied across its portfolio, exceeded the return threshold its shareholders required. The homes were still standing. The math had moved.

Subscribe now

The math moves in one direction

Insurance companies do not return to markets they have left. Once the models have recalculated, the recalculation is permanent — or at least persistent enough to function as permanent. Louisiana homeowners paying three times the national average are not waiting for the market to correct. The market has corrected. This is the correction.

The state backstops — FAIR Plans, Citizens Insurance, federal flood programs — absorb what the private market discards. These programs were designed as temporary safety nets, last resorts for the small number of people who fell through the cracks. They are now absorbing hundreds of thousands of households in a single state. California’s FAIR Plan exposure has exceeded $650 billion. The backstop was never engineered for this load. (I keep returning to this point and I am not sure I have it right yet — but it seems like the backstop only functions as a backstop if the primary market still exists to define what “normal” coverage looks like. When the primary market leaves, the backstop isn’t catching exceptions anymore. It is the system. And a system designed for exceptions cannot operate as the default. Or maybe it can, and we just don’t have a word for what that becomes.)

Insurance does not distribute risk equally. It reveals which risks the system has decided to bear and which to abandon.

This distinction matters because it changes the nature of the problem

If insurance were failing — if the models were inaccurate, the pricing wrong, the companies mismanaged — there would be a fix. Better models, better regulation, better management. But the models are accurate. The pricing reflects real risk. The companies are well managed, by every metric that matters to their shareholders.

The actuaries’ problem is not that they cannot calculate risk. They can calculate risk with remarkable precision. The problem is that precise calculation, applied at scale, produces a result that nobody designed but everybody enabled: a growing category of people, properties, and places that the mathematics have determined are not worth protecting. The Uninsurables.

Whether those people understand the mathematics is, from the system’s perspective, irrelevant. The model does not require their comprehension. It requires only their data.

-Aimé