Thus our first lesson: businesses logically are worth far more than net tangible assets when they can be expected to produce earnings on such assets considerably in excess of market rates of return. The capitalized value of this excess return is economic Goodwill.
— Warren Buffett, Berkshire Hathaway shareholder letter, 1983
Download the Slides Deck: The_Architecture_of_Capital
Listen to the deep-dive discussion – We Went Looking for Hidden Capital — and Found Something Else.
I. When High Returns on Net Tangible Assets Mean Returns on Intangibles
Buffett’s 1983 appendix contains one of the most useful moves in investing, and it is easy to read past. He is not saying that businesses with few tangible assets are better. He is saying that when a business earns far more on its tangible assets than those assets could normally produce, the excess is evidence of something the balance sheet does not contain.
The example he uses is See’s Candies. Blue Chip Stamps bought it in 1972 for twenty-five million dollars, at which point See’s held about eight million of net tangible assets and earned about two million after tax. Inventory and fixed assets do not produce that. Something else did — in See’s case a reputation built over decades, which allowed the value of the product to the buyer rather than its production cost to set the price.
Notice what Buffett does not do. He does not attempt to put a number on the reputation. He reads it off the return. The ratio is not a measurement of the intangible; it is a signal that the intangible is there and worth going to find.
The same move appears elsewhere. A persistently high gross margin says a business can charge more than its costs justify, which means some advantage exists and the analyst’s job is to identify it. A high return on capital says the same thing about the capital side: this business produces a great deal with very little. The two are not the same signal — one is about pricing power, the other about capital efficiency, and a business can have either without the other — but the method is identical. The number is a trace, not a valuation.
II. Accounting Capital Versus Economic Capital
There is a serious problem with reading returns this way, and it has occupied academic accounting for thirty years.
If the productive asset is a brand, a network, a standard or a body of code, then the spending that built it may never have appeared on the balance sheet at all. Research is expensed. Sales and marketing are expensed. Software development is sometimes capitalized and sometimes not. So the denominator of return on invested capital could be missing a large part of the capital actually deployed — and the extraordinary returns of capital-light businesses would be, in part, an artifact of accounting rather than a property of the business.
This is not a fringe position. Baruch Lev has argued it since the late nineties and titled a book The End of Accounting. Peters and Taylor published the standard academic method in the Journal of Financial Economics: capitalize a fraction of selling and administrative expense, amortize it, and rebuild the invested capital that accounting discarded. Practitioners have followed, measuring brand equity, patents and network effects directly.
The objection deserves a test rather than an argument. If it holds, the framework this collection is built on — which distinguishes businesses by how much capital they can absorb — is measuring the wrong thing.
III. What Six Companies Capitalize and What They Expense
The test was run on six businesses chosen because their structures are as different as possible: two light tolls, two heavy tolls, and two former light tolls in the middle of becoming something else.
MSCI turned out to capitalize roughly a third of its development spending as internally developed software, and to separate maintenance from creation in its own expense categories: research that supports existing products sits in cost of revenues, while research and development is defined as the cost of building new or enhanced ones. The company had already made the distinction we were looking for.
Visa was more surprising. Its client incentive payments are large — close to sixteen billion dollars a year — and it would be easy to treat them as invisible capital. But Visa’s policy is explicit: upfront and fixed payments made under multi-year contracts are capitalized as assets and amortized over the contract term, while incentives earned on performance are recognized as they are earned. The capitalized portion sits on the balance sheet at over seven billion dollars, larger than the company’s property, equipment and technology. What is expensed is the recurring part, and it is expensed because it recurs.
Worth noting in passing, because it complicates the picture in a useful way: even a nearly weightless toll can carry a substantial recurring burden. Visa returns roughly twenty-eight percent of gross revenue through client incentives to maintain and expand its claim on the network. Capital burden is not necessarily capital on the balance sheet.
Linde does not disclose research and development as a dollar figure at all — it does not reach the threshold that would require it. Union Pacific publishes its capital spending broken into replacement, capacity, equipment and technology, which is more disclosure than most companies offer. And Microsoft, the case where hidden intangible capital should be largest, spent thirty-six billion dollars on research and development against one hundred sixteen billion on property and equipment. The invisible portion had become the smaller one.
IV. The Accounting Distortion Was Smaller Than Expected
Across six very different architectures, the accounting distortion was smaller than expected — and not large enough to explain away the economics we were observing.
That is a narrower claim than it may appear. Six companies do not settle thirty years of debate about intangible accounting, and there is no reason to think the result generalizes to every industry. Economically durable is not the same as accounting-capitalizable, and it never will be. What the test shows is that in these cases, the spending that produced lasting capacity had largely been captured — and where it had not, the amounts were too small to change any conclusion.
So the returns are real. Which leaves the more interesting question untouched.
V. Asset Life Determines How Much of the Spending Accumulates
If the problem is not that we cannot see the capital, then the problem is what the capital is.
Set the six side by side and something appears that no amount of denominator adjustment would have revealed.
| Capex / revenue | Capex / operating cash flow | Capex / depreciation | Life of the asset bought | |
| Visa | 3.7% | 6.4% | 1.21x | network already built |
| MSCI | 4.1% | 8.2% | n/m | software, index history |
| Linde | 15.5% | 50.8% | 1.40x | plants, 15–20 yrs under contract |
| Union Pacific | 15.5% | 40.8% | 1.54x | corridor laid in 1862 |
| Alphabet | 22.7% | 55.5% | 4.33x | servers, six years |
| Microsoft | 34.9% | 63.4% | 3.01x | servers, two to six years |
Fiscal years ending in 2025, except Microsoft (June 2026). MSCI’s capex-to-depreciation is not meaningful: amortization of its capitalized software is reported together with acquisition intangibles.
Microsoft now spends more than a third of its revenue on property and equipment. Alphabet spends nearly a quarter. Both are above Linde, an industrial gas company that builds air separation plants, and above a railroad. The two businesses most often cited as archetypes of the asset-light model have become the most capital-intensive companies in the sample.
But intensity is not the finding. The finding is in the last column. Microsoft depreciates servers and network equipment over two to six years, and spends three dollars for every dollar of depreciation. Alphabet depreciates them over six, and spends more than four. Linde’s plants run for fifteen to twenty years under contract. Union Pacific’s corridor was laid in 1862 and is still carrying freight.
A dollar of capital expenditure is therefore not one thing. A dollar that buys a rail corridor is bought once. A dollar that buys a server is rented for six years and must be bought again. And because the replacement obligation grows with the base, the maintenance capital expenditure of the future is being created by the growth capital expenditure of the present. Microsoft’s property and equipment went from two hundred five billion dollars to three hundred thirteen billion in a single year. Whatever fraction of that is short-lived will have to be repurchased before the decade is out.
This is also why the historical average has become dangerous on exactly these two companies. Microsoft’s reported return on capital describes a software business with almost no physical base. That business is being replaced, in real time, by one with three hundred thirteen billion dollars of plant. The average is accurate and increasingly beside the point. The historical figure describes the old machine; the buyer owns the new one. The same caution applies to everything written here about Microsoft and Alphabet: it is a photograph of an experiment still running, and the numbers will look different in two years. That is not a weakness of the data. It is the demonstration.
VI. Capital Spending Divided by Depreciation
One number does most of this work and takes ten seconds to compute: capital expenditure divided by depreciation, both from the cash flow statement.
It is not a verdict. A ratio near one can mean a mature business, a mix of assets with different lives, an inflationary replacement cost, or simply the timing of a project cycle. A high ratio can mean genuine expansion or an acquisition binge. The ratio does not answer anything on its own.
What it does is tell you which question to ask. When capital expenditure runs at three times depreciation, as Microsoft’s does, the physical base is growing quickly — and the question becomes how long the assets being added will last, because that determines how much of today’s growth spending becomes tomorrow’s maintenance obligation. When it runs near one, the base is barely growing — and the question becomes how much of the spending is replacement rather than expansion, which is the difference between a business that is compounding and one that is standing still expensively.
Neither answer is in the ratio. Both are in the notes to the financial statements, where the estimated useful lives are disclosed and almost never read.
VII. When New Capital Raises the Return on Old Capital — an Unproven Case
Union Pacific does something in 2025 that this framework does not fully explain.
Volume grew one percent. Revenue grew one percent. Earnings per share grew eight percent, the operating ratio improved to 59.8, and railcar velocity rose eight percent while workforce, locomotive and fuel productivity reached their best recorded levels. The company spent 3.8 billion dollars, of which two billion is explicitly classified as replacement of existing infrastructure and only 258 million as capacity expansion.
One reading is that the capital deployed is not merely adding to the base but improving the return on the base that already exists. A second mainline track, a reconfigured terminal, a better planning system — these do not extend the network. They make the network laid in 1862 carry more at lower cost. If that is what is happening, then new capital is acting on the rate earned by old capital, which is not an interaction this framework has made explicit.
That is a hypothesis, not a finding. Rising earnings on flat volume is consistent with the explanation, and also consistent with pricing, mix, cost control and share repurchases. The causal link has not been demonstrated, and one company is not a pattern. It is recorded here because it is the most interesting thing in the sample that the rest of this essay does not account for.
VIII. What Must Be Spent Again Simply to Stay in Place
The experiment began with an accounting suspicion and ended somewhere more useful.
Gross investment is not the same thing as accumulated productive capacity. A company that spends a hundred billion dollars and must spend forty of it again within six years has not added a hundred billion of anything. This does not require a new variable. It is a property of the one already there: the amount of capital a business can absorb at an attractive return is a net figure, and the netting happens over the life of the asset.
Which returns to Buffett, and to the two faces of the same observation. A business earning enormous returns on very little tangible capital is telling you that something not on its balance sheet is doing the work — and for the purpose of understanding how it compounds, you may never need to price that thing. A business that requires enormous tangible capital is telling you something else: that its returns must be judged against how much of that capital will have to be rebuilt to keep earning them. Neither is good or bad. They are different burdens attached to the same earning power.
So the question is not how much a business invests. It is how much productive capacity remains after the capital it bought has lived, aged, and had to be replaced.
And how much must be spent merely to preserve the earning power already earned.
Watch the video