The Capital Sink

The Thesis

A capital glut is supposed to end with the money coming back out. This one is built so it doesn’t.

The worry is old. Adam Smith wrote it down in 1776: as capital piles up, the next unit gets harder to employ, so the return falls. Every version since ends the same way. Too much capital kills the return on capital, and capital that cannot earn goes home — as a dividend, a buyback, or an idle balance.

That ending hides an assumption. Somebody has to decide, every quarter, whether the marginal dollar is worth more inside the business or outside it. And when the answer is outside, the dollar has to actually leave.

The AI buildout offers a third door. Not distribution. Not idleness. Feedstock.

The productive use of the next dollar of AI profit is more AI. Chips earn the money that buys chips. Models earn the money that trains models. The return never has to fall to release the capital, because the capital is never released — nothing inside the loop is set up to ask whether the marginal dollar earns its keep outside it.

Strip it to a stock and two flows and there is nothing exotic here. Capital grows through a reinforcing loop, output reinvested. It shrinks through a balancing loop, depreciation. That is the ordinary shape of every business that owns anything. What makes a normal firm normal is that it throttles the reinforcing loop on purpose. It pays a dividend. It retires stock. It declines to build the plant.

Since 2024 the record shows that throttle coming off. In public. On a published schedule. By the largest cash generators in the world, at the same time.

This page does not argue the loop is a mistake. It sets out where the profit actually goes, how far the throttle has opened, and the one number that decides whether a self-feeding capital stock compounds or just runs in place.

The Evidence

Start with the vocabulary, because the words moved before the cash did.

Count archive records whose headline or summary carries a given word, over matched Jan 1–Aug 28 windows each year. “Capex”: 2 records in 2023, 11 in 2024, 23 in 2025, 59 in 2026. “Buyback”: 19, 21, 17, 30.

In 2023 the archive held nine and a half buyback stories for every capex story. In 2026 it holds two capex stories for every buyback story.

Normalising matters more than usual here, because the archive nearly doubled in volume and raw counts flatter everything. Done properly, the two words move in opposite directions. Capex mentions up sixteenfold per thousand records. Buyback mentions slightly down. That is not the record getting more interested in corporate finance. It is the record following the money to a different exit.

Then the cash flows caught up.

In February, The Information read the year ahead and said the capex ramp would wipe out free cash flow at Amazon, Google and Meta, forcing either buyback cuts or more borrowing. By May the Financial Times had the arithmetic. Combined free cash flow at Amazon, Alphabet, Microsoft and Meta was headed for about $4B in the third quarter, and the lowest full-year level since 2014.

Four of the most profitable enterprises ever built. A rounding error between them.

Meta got there first and alone: free cash flow down 91% to $784M in the second quarter, with capex guidance narrowed upward at the bottom end. The aggregate tells the same story with more names on it. Big-four capex hit a record $130B in Q1 and was tracking to $725B for the year, up 77% from $410B in 2025 — already above the ~$650B those same four forecast in February. The plan gets revised up faster than the year runs.

The tell is not the size of any number. It is what the companies say comes next. Alphabet raised 2026 guidance to $180B–$190B in April and said capex would “significantly increase” again in 2027. That is the throttle setting, said out loud. It describes a schedule, not a peak.

And the shortfall runs the wrong way for the classic ending. Hyperscalers are cutting buybacks and boosting capex. Alphabet — which spent a decade handing cash back — planned roughly $85B of equity offerings, nearly 40% of it covering tax obligations on employee equity awards.

A capital glut does not issue equity. This is the opposite condition wearing the same clothes: not more capital than uses, but more uses than capital.

The overflow is parked off the balance sheet.

Reported capex is only the visible part. In August the Wall Street Journal counted about $3T of AI-related off-balance-sheet commitments across nine tech companies, against roughly $600B of reported capex. Five dollars committed for every dollar shown. A separate study put off-balance-sheet debt at Alphabet, Microsoft, Amazon, Meta and Oracle at ~$1.65T, up roughly eightfold since 2022 — now bigger than their $1.35T of on-balance-sheet debt, which itself grew by ~$350B in five years. The plumbing is SPVs and GPU-backed debt pioneered by CoreWeave.

When the reinvestment rate outruns what earnings can fund, the excess does not stop. It moves somewhere harder to see.

The loop closes on itself.

The clearest evidence that this is a circuit and not a spending spree is Nvidia, which now stands on both sides of its own demand. It invests in OpenAI for non-voting shares, and OpenAI buys Nvidia chips with the cash. It rents its own chips back from Lambda — 10,000 of them, $1.3B over four years. It backstops young cloud providers by renting their unused GPUs back for a cut of revenue. It committed $40B+ of equity in 2026 alone, disclosed $18B more committed and $47.9B held in private companies as of late July, doubled its own server rentals to $26B, and announced a $500B financing package days after the SEC loosened rules on data-centre securitisations.

Money leaves as revenue and comes back as demand. Each of those deals is ordinary on its own. Together they describe a set of processes that each supply something another one needs — a system that rebuilds itself given nothing but outside energy and materials.

And the self-assembly is getting literal.

Foxconn is putting Nvidia’s humanoid robots into the Houston plant that builds AI servers for Nvidia. Meta is testing robots that swap cables and reset servers in its data centres. Cadence sells an AI agent for chip design that runs some tasks up to tenfold faster, to chipmakers including Nvidia. Musk announced Terafab, a Tesla and SpaceX project making chips for Tesla, xAI and SpaceX.

The machine designs its chips, builds its servers, and maintains its own halls. None of that needs a singularity. It needs only that each step is cheaper inside the loop than bought outside it.

The number that decides all of it is depreciation.

A stock fed by a reinforcing loop and drained by a balancing one compounds only while the first outruns the second. AI capital wears out fast. That is why the accounting is contested, and why Microsoft’s change to its schedule cut a 2026 capex forecast from $190B to $175B while the spending plans stayed put.

It is also why the most-quoted good news of the year was so thin. Global AI sales excluding China hit $25B in the first quarter, against an estimated $21B of data-centre and chip depreciation in the same quarter. Revenue cleared decay by four billion dollars.

On the record as it stands, the loop is running at about the rate that replaces itself.

The Companies

Nvidia is where the circuit shows up in one set of filings — supplier, lender, landlord and shareholder to its own customers. Read those disclosures as a balance sheet and it is a chip company. Read them as a flow diagram and it is the clearing house for the loop. Different risk, different failure mode.

Alphabet is the cleanest measurement, because it has the longest record of doing the opposite. A company that returned cash for a decade now guides capex up two quarters running, tells the market to expect another increase in 2027, and issues equity. None of that is distress. It is a stated preference for scale over distribution — the whole thesis in one issuer.

Meta shows what the preference costs when it lands all at once: free cash flow down 91% in a single quarter. Microsoft and Amazon are the same trade with more cushion. SpaceX runs it with no cushion and on purpose — $15.8B of AI capex in a quarter against $2.56B of AI revenue. Oracle and the neoclouds, CoreWeave and Nebius, run it on borrowed money, which turns a preference into an obligation with a date attached.

The instructive contrast sits one layer down. In August SK Hynix announced a ~$29B buyback, explicitly to calm worries about how long AI spending lasts. The chokepoint owners are distributing while the platform layer reinvests. That split is the sharpest read available on who currently believes their own returns will last.

The chokepoint rotation tracks who keeps the dollar. The capex supercycle tracks who funds the build. The token ceiling asks whether demand arrives in time. This page asks the question underneath all three: was that money ever coming back out?

The Lenses

Hamilton Helmer supplies the test, and it cuts against the thesis. In 7 Powers, the Benefit half of Power is anything that materially lifts cash flow, and he lists three routes to it: higher prices, lower costs, or lessened investment needs. A permanent capital sink is that third route running backwards.

Worse for the argument, his fundamental equation prices a business as market size × growth × share × m — long-run margin above the cost of capital. Scale is the first two terms. Power is the last one. An industry that optimises for scale and leaves m unstated has not found a new form of value. It has found a very large market with the value term missing.

Helmer’s warning is that Benefits without Barriers get arbitraged away in full. The arbitrage here would look exactly like this: everyone reinvesting everything, nobody earning above their cost of capital, all of them enormous.

Benjamin Graham made the same objection sixty years earlier, in ownership terms. The principle in Security Analysis is that stockholders are entitled to the earnings on their capital except to the extent they decide to reinvest them in the business. The load-bearing word is decide.

Graham and Dodd’s category was earnings power — the cash a business throws off after whatever spending it genuinely takes to stay in business. It is the well every return is drawn from, and a valuation is only an estimate of it. A self-replicating capital regime does not steal that cash. It removes the decision. If the reinvestment rate is set by what the loop needs instead of what the owners choose, the entire value of these businesses moves into a terminal year nobody will name — which, as the $/GW arithmetic this site keeps as its source of record shows again and again, is what every argument about this buildout turns out to be about.

Neither lens says the loop fails. Both say the same thing about what would make it work. At some point the machine has to hand something back. The date it does is the whole investment case.

What Moved

The entries split by layer, and the split is the finding. Everything at the platform and model layer is money going in. Everything at the component layer — the memory buyback, the supplier commitments taken in advance — is money coming out.

The record does not yet contain one quarter where the reinvestment rate came down by choice rather than by an accounting change. That is the entry to watch for. Its absence is the most informative thing on this page.

Sources

Word counts are drawn from TEXXR’s archive over matched Jan 1–Aug 28 windows in each year, counting records whose headline or summary contains the word; the archive held 7,919 records in the 2023 window and 14,563 in 2026, which is why normalised rates rather than raw counts carry the argument. Both terms were sampled for false positives and returned none. Capex, free-cash-flow and debt figures are as reported on the dates cited, and are not restated to a common basis across companies. Article IDs resolve at texxr.com/<id>.