John Maynard Keynes

The Lens

Keynes’s operational claim about investment is a single equality. New building continues until the expected return on the marginal asset has fallen to the rate of interest. Two numbers, and everything worth arguing about is which one moves.

His answer is that the rate of interest is the steadier. It is a market price; it can be looked up. The expected return cannot. The General Theory defines it on the yield an asset is expected to produce over its working life against what it costs to build one today — a forecast divided by a fact. A track record is evidence for the forecast, never a substitute for one.

Two things push that expected return down. Build more of an asset and its yield falls, because the new output competes with the old; build faster than the supply chain can absorb and the cost of building rises.

Then the argument that earns the lens its place here. Keynes points out that expecting rates to fall later lowers the expected return on equipment built now — because that equipment will spend part of its working life competing against equipment built later, financed more cheaply, and therefore content with a lower return. Cheap money in the future is not a gift to the asset you are building today. It is competition for it.

Why It Matters Now

The AI buildout is a very large bet on the expected return side of that equality, and the coverage record shows the assumption underneath it being adjusted in public.

The adjustment is depreciation, and the useful move is to notice that the word covers two different objects.

Book life is an accounting policy: how many years a company spreads an asset’s cost across, which decides how much cost lands in each year’s accounts. Economic life is how long the thing actually earns before something better makes it uncompetitive. Keynes’s chapter is entirely about the second. Nothing requires the two to move together.

The record shows the first one moving, and shows what moving it is worth. In August 2022 Microsoft extended the useful life of its cloud servers from four years to six, citing improved efficiency and technology, and expected to save $3.7 billion the following fiscal year. By February 2024 the Financial Times counted the group effect: Alphabet, Amazon, Microsoft and Meta had added about $10 billion to their collective profits over two years by shifting those estimates.

That $10 billion is not evidence of anything improper. It is evidence that the two objects are separable — a revision large enough to move reported profit by that much is being made in the accounting frame, and a company’s book life is not a measurement of how long its hardware stays competitive.

Which is exactly why the distinction matters here. The expected return that justifies borrowing against a GPU runs on economic life: what the chip earns before a faster one takes the work. The profit that reassures an equity holder runs on book life. The same asset supports two arithmetics, and only one of them is the number a lender is exposed to.

Keynes’s chapter points at the second. If a faster part ships every cycle, the thing competing with today’s GPU is next year’s GPU, bought more cheaply and content with a lower return — which pushes economic life down while nothing obliges book life to follow.

The record carries a named version of the same worry. In November 2025 Bloomberg reported Michael Burry’s warning that hyperscalers are underestimating depreciation on AI chips. Read carefully, that is a claim that book life has drifted too far from economic life — a coherent claim precisely because the two are separate things, and one this page does not adjudicate.

Applied to Tech Markets

The reason this stops being an accounting curiosity is that the two halves of Keynes’s equality are now moving in opposite directions at once.

The rate rose through 2026 — the ten-year Treasury from 4.19% in January to 4.75% at the end of August. The expected return is being estimated, in volume, against a yield nobody can observe yet. Keynes’s equality says building stops when the second falls to the first; it does not say when, and this page makes no claim about that.

What the record does show is the squeeze arriving through cash. In February 2026 The Information reported that the year’s projected capex ramp would wipe out free cash flow at Amazon, Google and Meta, leaving buyback cuts or more borrowing as the available responses. That is the moment a capex question becomes a credit question, and it is where this lens hands off to The Credit Wall.

Chapter 12 supplies the second half. Keynes distinguishes enterprise — forecasting an asset’s yield across its whole life — from speculation, forecasting what other people will think. His claim is that the better organised the market, the more reliably the second crowds out the first — not because anyone is foolish, but because forecasting the convention is what pays. A market that funds investment-grade AI paper at three to five times covered while declining project debt on the same buildout is a market doing both jobs at once, on different instruments.

Nvidia, Oracle, CoreWeave, Meta and Alphabet all sit somewhere in that arithmetic. None of this says what any of them is worth. It says which number the case rests on, and which of the two is the forecast.

The Library

The lens comes from The General Theory (1936) — Chapter 11 on the expected return, Chapter 12 on the state of long-term expectation, Chapter 13 on why the rate of interest is the price of parting with liquidity rather than the reward for saving.

It sits beside two others here. Big Debt Crises tracks what borrowing does to a balance sheet; A Monetary History tracks the lenders and their plumbing. Keynes supplies the middle term — the expected return that justified the borrowing, and why it is the least stable number in the calculation.

Sources

John Maynard Keynes, The General Theory of Employment, Interest and Money (1936) — Chapters 11–14 and 22. Reading notes.