A walk through the whole proof
The machine didn’t take your job. It needed your paycheck.
In 2026, companies cut 157,763 jobs and called it efficiency (the count as of this essay’s date — the live board runs the current number). The filings tell a quieter, stranger story — the money freed by the layoffs is the money buying the compute. This is the walk: from a person, to a payroll line, to a data center, to the strain on a distant power grid, and the receipt at every step.
The Catch · figures as of July 13, 2026 · every figure is either recomputed from a primary filing or attributed to a dated company disclosure — and each one says which
01 · The Handoff
Efficiency is a soft word for a hard trade.
Start with a person. A support rep in Ohio. An engineer in Bengaluru. A program manager who found out by email, at six in the morning, that his badge had already stopped working. In 2026, a hundred and fifty-seven thousand of them heard the same sentence in a dozen corporate dialects: the company is becoming more efficient.
Sometimes they said the quiet part out loud. Salesforce cut four thousand people and its CEO explained he simply needed “less heads.” Amazon trimmed fourteen thousand and called it clearing bureaucracy so AI could innovate faster. The word changes; the trade doesn’t.
When a company cuts ten thousand roles, it doesn’t set the money on fire. That payroll — real dollars, someone’s mortgage — is freed. And in the filings, you can follow exactly where it goes. It doesn’t come back as profit. It walks out the door and lands, nearly to the dollar, on a hyperscaler’s invoice. As compute.
The layoffs aren’t AI replacing workers. They’re the bill for buying it.
The Method
Anyone can shout “bubble.” We prove it.
Here’s the part the other five hundred can’t bring.
Every number on this page comes with its evidence class attached, because not all numbers are the same kind of thing and pretending otherwise is where most reporting goes wrong.
A filed actual is a number a company has put its signature under — in a 10-K or a 10-Q, with an accession number you can look up. Those we recompute ourselves from the filing and cite to the document.
Guidance is a company telling you what it intends to spend. It is not in any filing, because the future isn’t. It lives in a press release or an earnings call, and the most we can do — and we do it — is quote it exactly and date it.
An estimate is somebody’s model. We say whose, and we say it is an estimate.
That distinction is the whole method. It means you can argue with any line on this page on its own terms, and it means we cannot hide a guess inside a paragraph of facts. Where the honest answer is a range, we publish the low end and show you the high one.
So there is no take here you have to trust. Only arithmetic you can redo. If a number on this page is wrong, it is wrong in a way you can demonstrate — and we would rather be corrected than believed.
Built to be broken. It holds.
02 · The Build
The money has to live somewhere.
Compute isn’t an abstraction. It’s concrete and steel and copper, rising in the desert and the exurbs faster than the grid was ever built to carry it.
Start with what is already signed and filed. In their most recent annual reports the four biggest buyers put down a hundred and thirty-two billion dollars, ninety-one billion, seventy billion and sixty-five billion. Amazon, Alphabet, Meta, Microsoft. That is money already spent, not money promised.
Then look at what they say comes next. Meta told investors in April it expects to spend a hundred and twenty-five to a hundred and forty-five billion in 2026 — roughly double what it actually filed for 2025 — in the same season it cut eight thousand people for efficiency. Alphabet raised its own 2026 number to a hundred and eighty to a hundred and ninety billion.
And Oracle’s headcount fell by about twenty-one thousand in a year, from roughly a hundred and sixty-two thousand to a hundred and forty-one thousand. That one you can derive yourself from two consecutive annual reports, which is exactly how we got it.
Record spending. Record cuts. The same filings.
Record results. Record layoffs. The same page.
03 · The Draw
Someone pays the physical bill.
The machines are thirsty and they never sleep.
The data centers filling with the freed payroll draw power a local grid never planned for — an estimated four hundred and thirteen terawatt-hours — and in the densest corridors, like Northern Virginia, that load is already landing on other people’s bills. Capacity prices spiking. Water tables pulled down. The cost of the boom externalized onto towns that never voted for it.
It’s the same move, one layer down. The bill travels from the balance sheet to the municipality.
The cost didn’t disappear. It moved next door.
04 · The Divergence
Then the story and the receipts stopped agreeing.
For three straight quarters the gap ran negative — the filings were actually ahead of the market’s story about them. Minus 1.80, then minus 1.16, then minus 0.32. Narrowing, but pointing one way the whole time.
In the second quarter of 2026 it inverted. The market term jumped to plus 2.83 as the semiconductor index closed at 639.45 on eighty-eight percent three-month momentum. The ground-truth term fell to minus 1.23, dragged down by the AI-layoff share and discretionary insider selling both hitting their window highs. The gap went from minus 1.80 to plus 4.06 — a swing of nearly six points in a single quarter, and a change of sign.
Here is what we will not tell you: that this is statistically rare. It is a four-quarter series. Four points cannot carry that claim — and we logged ourselves saying so in June, after we first published the number with more confidence than it deserved. It is in our falsifiers page under the date, and you can read it.
What we will tell you is the direction, because the direction is not ambiguous. For three quarters the story and the filings moved together. Now they are moving apart, hard, and the machine didn’t do it — AI has no opinion about its own valuation. This gap is a choice, made by the people writing the story and the people buying it, over the numbers those same companies filed.
The filings are on the record. The story is on television. We just measure the distance.
Everything above is the shape of the argument. Everything below is the count that supports it — eleven independent, filing-sourced signals, the composite that aggregates twelve measured ones, and the method you can use to re-derive all of it. The desk offers no view. It offers a sequence.
The Two Cases
The desk draws both, as plainly as it can — because the honesty lives in the distance between them.
The bull case — the bridge. The heavy capex is a strategic land grab. The hyperscalers are betting that once the cost of intelligence falls far enough, the total addressable market expands into every vertical — healthcare, materials science, logistics — and the initial funding gap is rendered irrelevant within five years. On this reading, today's spend is the option premium on owning the next platform.
The bear case — the cliff. The premise holds if the technology hits a diminishing-returns wall. If the scaling laws bend — if the energy and compute to train the next model stop buying a corresponding jump in capability — then the $700B–$1.4T annual funding need becomes un-financeable, and the spend stops suddenly rather than slowly.
Which case is right is not the desk's call. The signals that follow are the count that tells them apart, and the distance between them is the Divergence.
Signal 1 — The Tape
Start with the only thing that is not an estimate: price. The SOXX semiconductor index — the market at the dead center of the AI trade — closed at 566, roughly +50.5% above its own 200-day trend. This is a statement about where price sits relative to its own history. It is not a forecast, and a trend line is not a target.
Signal 2 — The Spend
What the price is paying for. Hyperscaler AI capital expenditure did not drift up; it stepped, every major spender at once, straight from the FY2025 10-Ks: Amazon $78B→$128B, Alphabet $53B→$91B, Meta $37B→$70B, Microsoft $44B→$65B, Oracle $6.9B→$21B — aggregate growth above +67% in a single year. Filed figures, primary source.
Signal 3 — The Demand
Set the spend against the demand it must eventually serve. Sequoia's framing is “AI's $600B question” — roughly the annual end-revenue the build-out implies. Filed 2025 capex for the five largest spenders is already $375B. Against it, arm's-length revenue is thin: OpenAI's reported annualised revenue is near $12B (private, not filed), and MIT NANDA reports 95% of enterprise generative-AI pilots show no measurable P&L impact. The bar is drawn; most of it is empty. What fills it is a question, not yet an answer.
Signal 4 — The Financing
Stated in its own numbers, the bull case is denominated in future funding, not current revenue. JPMorgan's figures put the annual funding need at ~$700B in 2026, rising past $1.4T by 2030, while realised AI end-user revenue sits near $12B. That is an observation about how the case is constructed: the gap between the two lines is what has to be financed.
Signal 5 — The Recycling Ratio
The desk's signature measurement, and the one to check hardest. Committed compute across the largest deals totals $539.5B. The denominator is deliberate: not that committed figure, not announcements, but supplier equity funded in cash from the same filings — the real equity actually funded into the sellers' counterparties, about $34.8B. The ratio is 15.5×; the 1.0× line is what genuine end-customer demand would read.
The funded-versus-committed distinction is itself the audit. This was first published at 26×; a line-item re-check of the filings on 2026-07-02 raised the real funded equity and pulled the ratio to 15.5× — lower, and a tighter loop. The desk did not quietly restate it. The full move, 26× → 15.5×, is dated on the receipts ledger. A number you can audit is the point.
Signal 6 — The Loops
“Recycling” names specific, filed transactions in which a dollar leaves a chipmaker, reaches a lab, and returns as demand for that chipmaker's own compute. Three, each cited to its SEC filing.
Drawn together, the same structure repeats: equity out, compute back as demand. Of the 33 financing edges in the ledger, 22 are SEC-primary-filed and 11 are reported; where a figure is reported rather than filed, the exhibit says so.
Signal 7 — The Depreciation
A quieter lever, running through reported earnings. The useful life assumed for AI hardware has been extended — from three or four years toward five or six. Lengthening the assumed life lowers each quarter's depreciation expense, which raises reported profit with no additional revenue. The forensic names who extended and by how much, and sets booked life against economic life. It is an observation about accounting, not intent; if a generation strands early, the extension will read differently in hindsight.
Signal 8 — The Debt
Where the spend meets the cash to pay for it. Internal cash flow stopped covering capex; the gap between the two is the point at which external financing enters. Quarter over quarter the funded capex line pulls above operating cash flow, and bond issuance steps in behind it — the capex-vs-cash-flow gap as the debt-demand driver.
Signal 9 — The Memory
The build-out has precedent, and price cycles leave shapes. Overlaid on its own history, the AI-era index is laid against the dot-com (2000) and the 2008 tapes from each cycle's start. The overlay is a fact about shape, aligned by month — not a claim that history repeats. The reader can see where the current line sits against the two it is drawn beside.
Signal 10 — The Narrow Point
Where the weight concentrates. The compute-and-infrastructure layer — the chips, the servers, the power gear — is the point the whole stack balances on. The dossier scores each name in that layer across the same indicators, so the concentration is visible rather than asserted.
Signal 11 — The Corroboration
The desk publishes the signals that support its own composite, labelled as corroborating rather than independent: the share of capex not covered by cash flow (Microsoft and Amazon above 65%), discretionary insider selling, the AI share of announced layoffs, and the earnings added by lengthening asset lives. Naming them as corroboration, not proof, is part of the discipline.
The Composite — Six Ways It Can Crack
The eleven observations reduce to six independent stresses, each scored 0–100 from the filings: capex vs demand, organic demand, circular financing, insider selling, energy, depreciation. No single bar is the argument; the argument is convergence — how many are elevated at once. The composite reads 49/100.
The Measurement
Here is the only conclusion the desk draws, and it is a number, not a verdict. Set what the market has priced against what the filings and the arithmetic support, quarter by quarter, and the distance between the two lines is the Divergence. It is negative when price lags the fundamentals and positive when it leads them; today it reads near its widest on record. That gap is the whole instrument's output — everything above is how it is computed.
Drive It Yourself
None of this asks for trust; it asks to be checked. The same six filing-sourced indicators behind every exhibit are wired into a screener over the scored universe — filter and sort by convergence, depreciation, capex, insider selling, financing, energy or demand, and find the names that read as fragile to you.
It is an instrument, not a document, and it is meant to move: as new filings land — the next material read is the late-July earnings window — the scores update and the lines redraw. A living count, not a snapshot. The whole structure, every domain radiating from the center with the best and worst cases drawn as its two forked tails, is on one map.
Every figure above traces to a filing or a named source; every estimate is labelled as one; the best and worst cases are drawn side by side on purpose. The desk holds no position in anything named here and sells nothing. Find an error and it goes on the receipts ledger, credited to you — that is where the 26× → 15.5× correction lives, and where the next one will. Read the full proof, every map interactive, in The Catcher in the AI. This piece is sealed into today's Bitcoin-anchored manifest — verify the date yourself.
05 · The Variable
The machine is the constant. The people are the variable.
Follow the loop all the way around and it closes on a person — the one at the start, cut to free the payroll that bought the compute that filled the data center that strained the grid.
The whole system is built to treat that person as the adjustable number. The thing you trim when the model needs more.
This desk was built to put the number back where it belongs. On the record, name by name, defended. Not because it moves the market. Because somebody should keep the count.
Somebody should keep the count. So we do.
Author’s note
Built by a human and a machine — the way the whole argument says it should go. The human did the thinking; the machine did the lifting. Every figure is dated, carries its evidence class, and is left checkable.
We publish the floor. Argue with the ceiling.
THE CATCH.AI · figures as of July 13, 2026