One Fifth of the Money Earns Nearly All the Profit: What 176 Billion Dollars of Float Really Cost

Berkshire Hathaway ended 2025 carrying roughly 176 billion dollars of insurance float — money policyholders have paid in that belongs to the company until a claim eventually comes due. It is routinely described as the cheapest capital in the world, the engine behind six decades of compounding. But work out where the year’s underwriting profit actually came from and very little of that story survives. Almost all of it came from the smallest pool of the three.

Of 7.26 billion dollars in pre-tax underwriting earnings, GEICO alone produced 6.8 billion — the auto insurer whose claims are essentially settled inside a year, and which therefore generates less float per premium dollar than anything else Berkshire owns. Every other insurance operation combined, holding the overwhelming majority of that 176 billion, contributed under half a billion. This is not the quirk of one weak year. It is the governing arithmetic of the business, and it runs opposite to the popular version: the float you most want to own is precisely the float that is hardest to earn a profit on.

Float Is a Loan, Not a Gift

The word invites sloppiness. Float is not a reported balance sheet line; it is a derived figure — loss reserves plus unearned premiums and similar obligations, less the receivables that offset them. Put plainly: money the insurer holds today and must hand to somebody else later. Economically that is borrowing. What separates it from a bond is not that it costs nothing, but that its price is only set in arrears — and that it is not quoted in percent, it is paid in claims.

Warren Buffett laid this out himself long before the concept became a slogan. His 1996 letter tabulated the cost of float year by year back to 1967. In years with an underwriting profit the cost was negative — Berkshire was paid to hold other people’s money. In loss years the float cost roughly seven and a half percent on average. And there were fourteen loss years out of twenty-nine, close to half the sample. In those years float was more expensive than a bank line.

That is the half of the quotation that gets dropped. Being paid to hold money is not a property of the business model; it is a condition. It holds while the combined ratio stays below 100. It does not hold automatically, it does not hold for everyone, and it certainly does not hold in every line of business.

The Combined Ratio Is the Price — Per Premium Dollar

The combined ratio is the price of float expressed as a percentage of premium. Berkshire’s property and casualty operations came in at 87.1 percent for 2025, against a five-year average of 90.7. So 12.9 cents of every premium dollar dropped through as profit before a single dollar of investment income. That is an excellent number, and anyone who stops there will conclude the business is far better than it is.

Because the combined ratio is measured against premium, not against float. What matters to an investor is the cost of the float, because the float is the invested capital. The conversion is trivial: underwriting result divided by average float. For Berkshire in 2025 that is 7.26 billion dollars against average float of roughly 173.5 billion — minus 4.2 percent. Berkshire was paid a little over four percent a year for the privilege of holding 173 billion dollars of other people’s money. Invest that at the three-month Treasury bill yield of 3.69 percent and the arithmetic return on capital the company does not own and pays no interest on approaches eight percent.

That is the calculation that made the model famous. It is correct. It is simply not evenly distributed.

The Second Number Almost Nobody Runs

Alongside price per premium dollar sits a second variable that decides what an insurer is actually worth: how much float each premium dollar creates. That depends entirely on the payout tail — the lag between collecting the premium and paying the claim. Personal auto claims are mostly closed within twelve to eighteen months. Professional liability, asbestos and retroactive covers run for decades.

Line	Typical tail	Float per premium dollar	How knowable is the price?
Personal auto	about 1 year	roughly 0.6 to 0.8	high — mistakes surface in months
Property and catastrophe	months to 2 years	roughly 0.5 to 1.0	high, but violently volatile
General liability	5 to 10 years	roughly 2.5 to 4	low — the price is set later
Workers compensation	10 years and beyond	roughly 3 to 5	low, highly inflation sensitive
Retroactive reinsurance	decades	a multiple	the price is deliberately paid up front

Multiply the two and the ranking inverts. An auto writer at an 87 combined ratio with float turns of 0.7 earns something like eighteen percent a year on its float — a spectacular figure, on a narrow base. A liability reinsurer at the same 87 with float turns of 3.5 earns only around four percent on its float, but has five times the capital in motion. And because long-tail risk is far harder to price, it rarely gets to 87 in the first place.

Berkshire 2025, Taken Apart

The 2025 segment numbers make the point about as cleanly as an analyst could hope for.

Segment	Premiums earned 2025	Pre-tax underwriting result	Combined ratio
GEICO (auto)	$44.5bn	+$6.80bn	84.7% (prior year 81.5%)
Reinsurance, property and casualty	$20.4bn	+$1.90bn (down 32%)	84.5%
Berkshire Hathaway Primary Group	—	+$0.79bn	95.8%
Life and health, retroactive, periodic payment annuities	—	roughly −$2.2bn	structurally above 100
Group total	about $68.4bn	+$7.26bn (prior year 9.02)	87.1%

Berkshire does not disclose float by segment, so the split that follows is an estimate built from payout tails rather than a reported figure. It is robust nonetheless. GEICO’s float consists essentially of unearned premium on six-month policies plus loss reserves worth less than one year of claims. Against 44.5 billion dollars of earned premium that lands somewhere around 30 to 35 billion — roughly a fifth of group float. The remaining 140 billion odd sits in reinsurance, retroactive covers and annuity obligations with tails measured in many years.

So the ledger reads: about one fifth of the money produced 94 percent of the underwriting profit. Measured against GEICO’s own float the cost is somewhere near minus twenty percent a year. Measured against all the rest — the capital that actually funds the investment portfolio — it is close to zero. Not negative enough to carry anything, not positive enough to hurt. It is interest-free money, and not much more than that. And 2025 was a good year.

Retroactive Reinsurance: Where the Price Is Printed

There is one place in Berkshire’s accounts where the cost of float does not have to be estimated, because it is disclosed. In a retroactive cover the reinsurer accepts old claims that have already occurred but not yet been paid, in exchange for a single up-front premium. That premium is by construction lower than the expected ultimate losses — otherwise nobody would write the contract. The gap is capitalized as a deferred charge asset and amortized over decades, and the amortization shows up permanently as an underwriting loss.

In the first quarter of 2025 that loss was 209 million dollars, and the deferred charge stood at 8.6 billion at the end of March. Those 8.6 billion are nothing other than the unrecognized interest expense on purchased float. Berkshire consciously paid money for capital, because the capital stays a very long time and because management believed it could invest it for more than it costs. That is a perfectly sound calculation. What it does is retire the idea that float is free: in the one line where Berkshire buys float most deliberately, the price sits on the balance sheet as an asset and is paid down quarter after quarter.

Interest Is the Second Price, and Somebody Else Sets It

2025 delivered that lesson involuntarily. Float grew, from 171 to 176 billion dollars. Insurance investment income fell over the same year, from 13.67 to 12.51 billion. More money, less income. That is not an operating failure; it is what happens when short rates come down. Float parked mainly in short government paper — of Berkshire’s 373.3 billion dollars of year-end liquidity, roughly 305 billion sat in Treasury bills — is a floating-rate instrument. It earns whatever the central bank says, and nobody inside the company has a vote.

As of late August 2026 the effective federal funds rate stood at 3.63 percent, three-month bills at 3.69, the two-year at 4.20 and the ten-year at 4.67. Every point off the front end costs a company holding 300 billion in bills roughly three billion dollars of pre-tax income a year, without a single contract getting worse. Buying an insurance stock always means buying a position on short rates. It is simply a position you cannot close without selling the business.

The reverse deserves the same skepticism. A large share of the dazzling insurer results from 2023 through 2025 came not from better underwriting but from float that earned essentially nothing between 2010 and 2021 suddenly earning four to five percent. Same balance sheets, same contracts, different rate.

For Liability Float, the Price Is Set by a Jury

In short-tail lines the insurer learns the price of its float almost immediately. In liability nobody knows it at first, and it gets set later, in court. That is happening in the United States right now on a scale that is retroactively invalidating whole underwriting years.

A tally published on 25 August 2026 counted close to 200 verdicts of at least ten million dollars for 2025 — up 40.7 percent on the prior year and the highest count since 2009. Those verdicts summed to about 25.6 billion dollars; forty exceeded 100 million, four cleared a billion, and the average ran above 51 million. California led with 24, ahead of Florida at 16, Texas at 14, Georgia at 12 and New York at 11. Behind it sits a litigation funding industry now estimated at 18 to 20 billion dollars and a plaintiffs’ bar that has systematically re-anchored what a claim is worth.

The consequence lands in the reserves. For 2025 the US property and casualty industry booked roughly two billion dollars of deficiency in commercial auto liability alone, mostly against accident years 2023 and 2024 — contracts that were considered adequately priced the day they were signed. General liability added 7.3 billion dollars of adverse development. None of that is catastrophe loss. It is the retroactive repricing of float that was collected years ago and assumed to be cheap.

What makes it genuinely uncomfortable is the contrast with the headline. The same industry posted its strongest underwriting year in a decade in 2025: over 45 billion dollars of profit in personal lines and over 19 billion in commercial, some 84 billion across 2024 and 2025 combined. For 2026 AM Best expects the combined ratio to rise 1.9 points to 96.9 as rates soften while claim severity does not. Record profit and reserve deficiency coexist comfortably, because they belong to different accident years.

Reserves Are an Opinion

Which brings us to the least comfortable truth in the business. The combined ratio is not a measurement, it is an estimate with a wide band of legitimate discretion. Reserve conservatively and you report a worse profit while quietly building a cushion. Reserve tightly and you report a better profit while moving the problem into the future. Both are compliant.

Progressive shows what the disciplined short-tail version looks like from the outside. Net premiums written reached 83.2 billion dollars in 2025, up 12 percent, at a combined ratio of 87.4 against 88.8 the prior year, with the property book at 75. Net income came in at 11.3 billion dollars, or 19.23 dollars a share. That is a formidable underwriting machine — and note what it is not: it is not a float machine. Personal auto pays out too fast to accumulate much. Progressive earns its money by pricing risk better than the next carrier, roughly twelve times a year, not by investing a mountain of other people’s cash.

The long-tail version is visible in Europe, where reserving culture is measurably more conservative. Hannover Re has been building an explicitly labelled resilience buffer in its loss reserves: from 2.1 to 2.5 billion euros in 2024, and to roughly 3.2 billion in 2025. In the same year it reported 2.6 billion euros of group net income and grew its property and casualty reinsurance book by 12.3 percent — while burying another 700 million euros of potential profit. To an investor watching the price-to-earnings ratio that looks like a weaker year. It is in fact the purchase of an option on its own credibility. Munich Re showed the same handwriting from the other side: over 6.1 billion euros of net result in 2025, a fifth consecutive year above guidance, a 74 percent combined ratio in property and casualty reinsurance, and major losses down from 2.807 to 1.627 billion euros.

The practical instruction for a US investor is specific and free: read the loss development triangles in the statutory filings, not just the headline combined ratio. If a company has released prior-year reserves every year for five years and those releases carry most of the reported profit, the profit is being withdrawn, not earned.

Athene: Renting Float Instead of Earning It

An entire industry has grown up around acquiring float rather than underwriting it. Athene, the annuity arm of Apollo, does not collect float through property and casualty policies; it buys it, through retail annuities and institutional funding agreements that pay a contractual rate. Funding agreement inflows alone came to 35.4 billion dollars in 2025. Cost of funds rose over the same year from 7.7 to 10.1 billion dollars. Net investment spread — the gap between what the assets earn and what the liabilities cost — fell from 1.78 to 1.61 percent, producing 3.36 billion dollars of spread related earnings, up 4.2 percent.

Set the two side by side. Berkshire was paid roughly four percent in 2025 to hold its float. Athene pays for its float and keeps 1.61 points of spread. Both models can work, but they are not the same business and they fail in opposite directions. Compress the spread and Athene earns less while the obligation stays exactly where it was. Let claim inflation run and Berkshire earns less, but at least the obligation was never contractually fixed. Before comparing any two insurance stocks, establish which of the two machines you are looking at.

Model	Where the float comes from	Annual price	Principal risk
Berkshire 2025	underwriting, earned	about −4.2% (income)	claim inflation, adverse development
GEICO alone (estimate)	auto, short tail	about −20% (income)	price competition, small base
Athene 2025	annuities, purchased	positive, 1.61 points of spread	spread compression, credit
German life insurers 2011–2021	guaranteed savings premiums	above the achievable yield	guarantees without matching assets
Amazon, Starbucks	customers and suppliers	zero	shrinks when sales shrink

The Cautionary Tale Europe Already Lived Through

Anyone who believes negative float costs are a law of nature should look at German life insurance between 2011 and 2021. There was float in abundance there — premiums that sit for thirty years, theoretically the ideal case. Except the contracts carried guaranteed crediting rates of up to four percent while government bonds yielded zero or less. The cost of that float sat permanently above the achievable return on the assets funding it.

Regulators responded by mandating an additional interest reserve, a supplementary buffer that pre-funds the gap between the guaranteed rate and the market rate. It grew to roughly 96 billion euros by the end of 2021 — 96 billion that never became profit because it had to be tied up simply to make the float survivable. Rising rates have reversed it: the balance stood at 84 billion at the end of 2024, below 80 billion at the end of 2025, with releases of around five billion expected in 2026 and low double-digit billions across 2027 to 2029. The catch for shareholders is that those funds belong predominantly to policyholders rather than owners, and are first in line to absorb unrealized losses on low-coupon bonds bought during the zero-rate era.

The transferable lesson is that duration cuts both ways. A life insurer’s float looks vastly larger on paper than a property casualty writer’s, and it is structurally worse: long, contractually rate-fixed, regulator-constrained, and largely credited to customers. Length is only an advantage when the price is below the yield.

Float Without Insurance: The Clean Version

The last row of the table deserves its own attention, because it shows the principle in its purest form. Amazon runs a negative cash conversion cycle: it collects from customers before it pays suppliers. Depending on the calculation the cycle ran somewhere around minus 55 to minus 65 days in 2025, driven mainly by days payable stretching toward 125. That is roughly two months of revenue held as an interest-free loan that grows with every increment of growth.

Starbucks carries a stored value card liability of around 1.84 billion dollars — money customers paid in advance, some of it for coffee that will never be collected, with the unredeemed portion recognized as breakage revenue in proportion to actual redemptions. Revenue with no cost of goods attached. Airlines show the seasonal variant: unflown ticket liabilities build through the booking season and unwind as passengers travel.

The decisive difference from insurance float is the absence of a loss reserve. Amazon’s suppliers cannot develop adversely. A Starbucks card cannot be repriced by a jury. This kind of float has a known, fixed price of zero and only shrinks when the business does. That is why a dollar of customer float is worth more than a dollar of liability float — and why it is a mistake to hunt for the model only among insurers.

What It Means for Valuation

All of which reduces to a short, hard checklist. First: how much float per premium dollar, and from which lines? A high-turns insurer is a leveraged bond portfolio with underwriting risk attached; a low-turns insurer is an operating business with some interest income. Either can be excellent, but they deserve different multiples.

Second: what do the last five years of prior-year development look like? Consistent favorable development is a quality signal on its own — but when it carries most of the reported earnings, substance is being consumed. Third: how much of the return is simply the front end of the yield curve? A carrier that shone in 2024 and 2025 because bills paid four to five percent will shine less at three, without having made a single mistake.

Fourth, and this is where price-to-book comparisons mislead: float is debt, but debt at a negative interest rate behaves economically like equity. That is precisely why first-rate underwriters trade above book and weak ones below. The market is not pricing the capital, it is pricing the probability that the combined ratio stays under 100. An insurer’s price-to-book ratio is a statement about future underwriting discipline, not about the balance sheet.

Three Scenarios Through 2028

Scenario	Trigger	Who wins	Who loses
Softening market, rates hold	prices give way, front end near 3.5%	short-tail writers with a cost advantage	liability writers without a reserve cushion
Full easing cycle	front end toward 2%	carriers with real underwriting margin	float-heavy models, annuity writers
Claim inflation accelerates	verdict severity keeps compounding	conservative reservers, buffered reinsurers	US commercial liability, funding-exposed books

The likeliest outcome blends the first and third, which is exactly what Berkshire chief executive Greg Abel signalled when he said the company expects to write less property and casualty business for a period of time, citing increased competition and lower rates. It is the most striking sentence of the reporting year: a company whose entire model rests on float deliberately choosing to collect less of it, because it considers the price too high.

The Bottom Line

Float is not a property you own. It is a price you renegotiate continuously — with customers, competitors, courts and the central bank. Calling it free capital really means the price has been low lately.

The practical conclusion cuts against the usual heuristic. The most impressive float — long, large, locked up for decades — is the most expensive and the hardest to price. The unglamorous float — auto policies, customer prepayments, supplier credit — is reliably free and grows with the business. Berkshire’s own 2025 numbers say precisely that: 6.8 of 7.26 billion dollars from the line with the shortest breath.

So do not screen for the biggest float number. Screen for the most reliable price. Three questions fit on an index card: does the profit come from current underwriting or from releasing past reserves? Does the return come from the business or from the front end of the curve? And would this management write less business if the price were wrong? Berkshire answered the third one this year. Most carriers never do.