The Catch · The Newsstand
The Catch
Every piece the desk publishes, in one place — newest first, by pillar and type.
16 published · site-first · last update 2026-07-13
The Catch.AI isn't just covering compute culture — we are made of it. Built in this era, out of this era, for and against this era, we document all of it in receipts. From the money, the machines, and the layoffs to the computed fear and fiction, the live tickers, and the data breathing on the screen, we exist to bear witness. If this is a bubble, someone should keep the record.
The Catcher in the AI
The AI bubble is no longer a question — it is a verdict, filed dollar by dollar in the industry's own hand.
Who Holds the Power
The more powerful AI gets, the fewer hands hold it.
The Invisible Tax: The Company Town, Metered Monthly
Virginia regulators just raised the average household's power bill toward $165 a month — and then wrote a new rule forcing data centers to finally pay their own way. That second half only exists because someone proved the first half was subsidizing the machines.
The Private Grid: When the Data Center Builds Its Own Power Plant
Rather than wait five years for a grid connection, AI labs are trucking in gas turbines and running them as private power plants — classified as "non-road engines" to skip the permits. The grid, and the neighborhood, keep the bill and the smog.
Thirsty Machines: What the Cloud Drinks in the Desert
A single Google campus in Mesa is permitted for up to four million gallons of water a day — in a county the government rates in extreme drought. The cloud has a plumbing bill, and it is being paid in an aquifer that was already over-drawn.
The Synthetic Dependency: The Perfect Companion Never Argues Back
MIT and OpenAI studied forty million conversations and found the heaviest users of a companionable machine were lonelier, more dependent, and spent less time with people. The machine didn't choose that design. We did.
The Divergence Signal: Semiconductors +91%, the Workforce Cut 123,653
In the same six months, the market bid semiconductors up 91 percent and companies cut over 123,000 AI-and-tech jobs. The boom and the cut are two prices on one trade — and they point in opposite directions.
The Compute-Capex Mirage: $700 Billion Goes Into the Ground. What Comes Back Out?
The four biggest buyers will spend three-quarters of a trillion dollars on infrastructure this year. The revenue that is supposed to justify it is the quietest number in tech.
The Efficiency Illusion: The Polite Word for Feeding the Furnace
"Efficiency" is the word 2026 uses for cutting payroll to fund compute. But the filings show the savings and the spend are not in the same league — they are not even the same sport. This is the anatomy of a transfer, dressed as a strategy.
The Executive's Signature: We Don't Blame the Algorithm
The machine did not fire the sales team or sign the GPU contract. People did — the CEO, the CFO who booked compute as "efficiency," the board that demanded a pivot. This desk traces the trade back to the signature, because that is where accountability actually lives.
The Corporate Capture Mapping: Who Benefits When the Builders Preach Doom?
Researchers catalogued 249 cases of "Big AI" steering its own regulation — the same playbook regulators once saw from tobacco and oil. The existential-risk sermon turns out to have a business model.
The Homogenization Feedback Loop: A Culture Trained on Its Own Average
Nature published the mathematics: models trained on model output collapse toward the mean and the rare disappears first. The same loop is now running on the culture the models feed.
The Systemic Bias Audit: The Machine Is Doing Exactly What We Built It to Do
OpenAI's own researchers showed models guess confidently because our benchmarks punish "I don't know." Bloomberg showed GPT ranking résumés by the race coded in a name. Neither is a glitch. Both are the design.
The Infrastructure Subsidy: Your Power Bill Is an AI Investor Now
Utilities asked for $29 billion in rate increases in six months. Harvard's electricity-law scholars traced where the money goes.
The Data-Center Deal: The Savings Were a Rounding Error on the Spend
A company cut people to save a little, then spent a hundred times that on compute — and filed it under efficiency.
The ROI Fallacy: The Layoffs Paid for the Machines
The machines haven't paid anyone back. Eighty percent of enterprises cut roles for AI with no measured return.
The Instruments
One continuous world, six minutes forty-three. The desk's instruments — the gauge, the gap, the loop, the count — carried as one story, every number read from the site's own data record at build time. Runtime 6:43.
The Walk
Two cases, one street. The bubble question walked end to end — the tape, the spend, the gap, the loop, the books, the owners, the ceiling, the referees, the dial — and the three falsifiers that would retire the whole thesis, published in advance. Runtime 4:07.
AI Impact
What is AI actually doing to people? Five numbers, each with its receipt: 160,588 jobs · $799B capex · 413 TWh · zero cuts at the labs · 62% adoption with flat trust. The machine is the constant; the people are the variable — and the cause. Runtime 2:02.
Bubble Watch
The composite read, computed from published inputs — pushers against holders, nothing hand-set. Runtime 0:45.
The Recycling Ratio
A dollar goes in; the loop sends it back out bigger. Filed edges, drawn in sequence. Runtime 0:51.
The Divergence Minute
Story versus filings in one minute: what the market prices against what the record shows. Runtime 1:07.
The Ledger
The count, kept in black. Claim tags hold company words and reporting words apart. Runtime 1:01.
The Impacts
The load-bearing column, the grid, the water, the people — and the payoff that has not arrived. Runtime 1:11.
The Whole Board
Twenty-four industries, ranked. The bet is not a tech story — the top of the board is the real economy. Runtime 1:03.
The Tape
Every number in this story is a measurement of something. The price is the only one that is the thing itself — and it reads half again above its own trend. Runtime 1:24.
The Spend
Five companies, one year, +71.5% — filed on 10-Ks, not promised from a stage. What a filing proves, and the one thing it cannot. Runtime 1:27.
The Demand
One year of capex needs ~$480B of return revenue; realized so far: ~$15–20B. A ≈10× gap — an attributed estimate, labeled as one, because the label is the argument. Runtime 1:21.
The Financing
~$650B a year to sustain the build; $50–150B of honest revenue. Fifteen cents on the dollar — the rest is financed, and the question is which clock wins. Runtime 1:25.
The Loops
Thirty-three filed deals where the seller funds the buyer. Closed rings with names on them — and when one edge breaks, a loop does not shrink. It unwinds. Runtime 1:17.
The Depreciation
The machines live ~3 years; the books say 5–6. $40–70B of earnings ride on the difference — a desk estimate, labeled — while the rental market marks the hardware down out loud. Runtime 1:09.
The Debt
More than 65 cents of every build-out dollar is financed. $255.7B named, ~$330B maturing through 2028 — and debt reprices whether or not the revenue arrived. Runtime 1:09.
Insider Selling
$15.6B sold, $0 bought — Form-4 filed, named roster. Insiders can be wrong; they are never uninformed. A marker, not proof, and labeled as one. Runtime 1:11.
The Power Draw
3–4% of the grid today, 8–12% by 2030 on the announced build — four times faster than the grid can grow, with capacity already clearing at 9.3× prior. Nobody argues with the electric bill. Runtime 1:10.
The Corroboration
Bank of England: stretched. IMF: abrupt correction risk. BIS: protracted investment bust. Different rooms, same alarm — held with the honest caveat that agreement is error-checking, not proof. Runtime 1:11.
The Memory
This cycle laid over the dot-com climb: month 44 of what was, last time, a 48-month run. Shape, not prophecy — one sample, and said so. Runtime 0:56.
The Record
Three falsifiers, published in advance — any one trips and the thesis retires, loudly. The precedent: 26× corrected to 15.5× in public. The meter beats the story. Even ours. Runtime 0:59.
The Tech Stack
The whole AI stack, from first principles to what ships — the page's own mechanisms, shown in use. Runtime 4:51.
Principles
Five ideas everything above is built from: the token, the embedding, the weighted sums, attention, and the transformer — trained on one task: predict the next token. Runtime 1:31.
Hardware
Accelerators, HBM stacked beside the die, the interconnect that makes one machine of ten thousand chips, and the fab chain everyone waits on. Software scales free; silicon does not. Runtime 1:34.
Infrastructure
Chip → board → server → rack → hall → substation. A GPU costs the same idle; software copies but permits don't — the boom as a construction project. Runtime 1:24.
Foundation
A few hundred gigabytes of numbers that pass bar exams. One model trained once at enormous cost, adapted to a thousand tasks — the bet that restructured the industry. Runtime 1:18.
Models
A purchasing decision dressed up as a technical one. Seven criteria, four family dossiers, and the five questions that outlast every flagship of the week. Runtime 1:15.
Orchestration
An expensive answering machine with amnesia — fixed without touching the weights. Retrieval, tools, agent loops, memory, evals. The layer below produces intelligence; this one produces products. Runtime 1:04.
Application
The layer that gets demoed on stage and holds the credit card — where the most companies are born and die. Copilots win on placement, not intelligence. Runtime 1:08.
The Playbook
Arranged the way problems arrive: prompt badly, hit the context wall, watch the agent be wrong, bolt on retrieval, prove it, pay for it, defend it. Skip to what's on fire. Runtime 1:09.