On Tuesday a Columbia finance professor's paper, published through Brookings, put a date on something the industry had been avoiding: in 2026, for the first time, the five largest AI builders — Microsoft, Amazon, Alphabet, Meta, Oracle — will spend more on data centres than their combined operations generate in cash. Internal money no longer covers the build. From here it runs on borrowed money or someone else's balance sheet, and the paper's point is that this is the mechanical setup for a cut: the moment the industry turns cash-flow negative is the moment capex starts getting trimmed.

Wells Fargo said the same thing in trader's language: investors will start treating 2027 as the peak spending year, probably right after third-quarter earnings and the midterms, and "most investors now agree" the spending slows meaningfully in 2028. Pantheon's forecasts for 2027 capex growth are a fraction of 2026's. The SpaceX chip financing — forty billion dollars from Apollo to buy Nvidia hardware — was the week's illustration of what "someone else's balance sheet" looks like.

Three weeks ago the industry's own leaders said slow down. This week Wall Street said you will whether you like it or not. That is two different crowds agreeing from opposite directions, and if your thesis has a Shift II — AI compressing the value of a salary — it is the week your thesis is supposed to be able to say what would make it wrong.

Mine couldn't. I admitted that last Sunday. Tonight is the repair.

A kill fact written before the event is a promise. A kill fact written after the event is, until proven otherwise, an excuse — because the easiest thing in the world is to write a kill fact that the event you just survived doesn't trigger. "AI capex slows" was a candidate for Shift II's kill fact a month ago; this week, with the slowdown arriving, it is suddenly very tempting to write "AI capex falls two years in a row" instead, and call that honest. It isn't. It's the thesis buying itself two more years with a pen.

So the problem isn't finding a fact. It's finding one that passes a test the author can't game. That's the tool.

A kill fact added after the event is admissible only if it clears all four.

Gate one — it would have been reasonable to write it before. Ask: if I had sat down in March with no knowledge of this week, would this fact have been on the shortlist? If the honest answer is "no, I'd never have thought of it," then the news is writing your thesis, and that's the failure the kill list exists to prevent.

Gate two — it is about the mechanism, not the mood. Every shift in this thesis is a mechanism: money debasing, work decoupling from output, rules restructuring. The kill fact has to be a failure of the mechanism, not a change in how people feel about it. Capex is a mood — it's how much the builders are willing to bet. Deployment is the mechanism — whether the machines are actually replacing hours of work. Shift II is killed by the second, not the first. A world where the build stops but the models already shipped keep eating knowledge work is a world where Shift II is still true. A world where spending continues and nobody uses the output is a world where it's false.

Gate three — it is observable, public, and dated. Not "AI disappoints." A number, from a source you don't control, with a frequency. If you can't say what you'd look at and when, you have an opinion, not a fact.

Gate four — it goes in the ledger with the date and the reason. "Added October 11, 2026, after the capex-peak week. Previous list had no Shift II fact." The entry is the confession. It's permanent, it's public, and it's the thing that lets a reader in 2028 check whether you were fixing a gap or fitting the news.

Then one rule above the gates, and it's the whole point: you get one amendment a quarter. If the thesis needs patching more often than that, it isn't incomplete. It's wrong, and the kill list has already fired — you just haven't let it.

Here is the actual work, with the rejects shown, because the rejects are where the honesty lives.

Candidate A: "Hyperscaler capex falls year-over-year." Gate one: yes, obvious in March. Gate two: fails. Capex is the mood. Spending can fall because the build is finished, or because financing got expensive, with deployment untouched. Rejected — and this is exactly the one the week was trying to get me to write.

Candidate B: "AI companies stop releasing better models." Gate one: plausible. Gate two: fails the other way — Shift II never required the next model, only the last one. Gate three: "better" isn't observable. Rejected.

Candidate C: "The share of US businesses using AI in production stops growing." The Census Bureau's Business Trends and Outlook Survey asks this every two weeks, publicly, of hundreds of thousands of firms; it was around 4 percent of businesses in late 2023 and has roughly doubled since. Gate one: yes — I'd cited this series in Letter #007, so it was already on the table. Gate two: yes, it is deployment, the mechanism itself. Gate three: public, biweekly, from a source I don't control. Gate four: entered tonight. Passes.

Candidate D: "The productivity–pay gap closes." The gap Letter #007 is built on — output rising, real pay flat — measured by the BLS. Gate one: yes. Gate two: yes, arguably the purest. Gate three: yes, but quarterly and noisy, and it can close for reasons that have nothing to do with AI — a tight labour market, a policy change. Passes, but weaker than C on its own. It goes in as the second condition.

So the kill fact for Shift II, as written tonight:

Both halves, together. Adoption stalls and the gap closes. If adoption stalls alone, the mechanism paused. If the gap closes alone, something else happened. If both, Shift II is dead and I say so.

Ledger entry: Added 2026-10-11, after Wells Fargo/Brookings capex-peak week. Previous kill list had no Shift II fact. Amendment 1 of the quarter. Next permitted amendment: January.

The kill fact is new; the three questions aren't.

The capex peak itself. Kill list, now that there is one: no — adoption is still rising, the gap is still open. Your three numbers: if your Upside layer holds AI infrastructure directly, yes; inside a total-market index fund, barely. Ninety days: yes. One or two yeses. Sunday item. The new kill fact is precisely what tells you this headline is about the mood and not the mechanism.

Bitcoin's week. Down 2.8 percent, a Thursday low near $80,300, about $730 million out of the ETFs in two days, a billion dollars of leveraged positions liquidated in a day, the largest since August. Kill list: no. Numbers: no — the slice is sized for 80 percent, this was three. Ninety days: no. Zero. The arithmetic from Letter #021 says this is the kind of week you watch. Watch it.

The Fed minutes. Another hike "likely appropriate by year end"; the ten-year at 5.23 percent; Brent over $100. Kill list: no — one more hike isn't three years of real yields above two. Numbers: the floor earns a little more. Ninety days: yes. One yes. Ledger line, nothing else.

Four headlines. One thesis amendment, done properly. Zero trades.

It is not a licence to add kill facts whenever the news gets uncomfortable. The one-a-quarter rule is the fence. A thesis with a growing list of ever-more-specific conditions for its own death is a thesis that has already died and is being kept on a ventilator.

It is not a way of dismissing the capex story. The build slowing is real, it is probably coming in 2027, and if you own the builders directly you should be reading it as a number and not a headline. The tool tells you it isn't a thesis event. It doesn't tell you it isn't a portfolio event.

And it is not a substitute for having written the fact in the first place. This letter exists because of a mistake. The four gates make the repair honest; they don't make it free.

First — gate one is unverifiable, and I'm the one grading it. "Would I have written this in March?" is a question about a counterfactual version of me, and that version is very cooperative. The best I can do is point to Letter #007, where the Census series already appears, and let you decide whether that's evidence or a convenient coincidence. It's some of both.

Second — the two-part kill fact may be too hard to trigger. Requiring adoption to stall and the pay gap to close is two conditions, and two conditions are always less likely than one. A sceptic could say I've built a fact that will never fire. The defence is that each half alone has a non-AI explanation and the thesis shouldn't die of something unrelated. The honest counter is that this is exactly what a thesis-protecting author would say.

Third — "adoption" measured by survey is a weak proxy for the mechanism. A firm can say it uses AI in production and mean a chatbot on the help page. The Census series counts self-reports, not hours of knowledge work displaced. There may be a better series — the BLS occupational data, task-level studies — but they're slower and I'd be choosing among them after the fact, which is gate one again.

Fourth — the one-a-quarter rule is arbitrary, and the first quarter is already spent. I picked a quarter because it's long enough to stop news-fitting and short enough to be realistic. But it means that if something genuinely new appears in November, the rule says wait until January. A rule that stops you from learning is a rule that will eventually cost you. I'm keeping it, and I'm telling you that it will, one day, be wrong.

The Blind have no kill list, so there is no gap to notice. Every headline is either confirmation or noise, decided after the fact by how the price moved.

The Scared have a kill list and rewrite it monthly. Each rewrite feels like rigour. Over a year the list has grown from three facts to eleven, every one of them tuned to not-quite-fire on the last scare, and the thesis has become unfalsifiable without anyone deciding it should be.

The Prepared found the hole on a Sunday, said so in print, waited a week, wrote four gates, ran four candidates through them, threw out the one the news wanted, kept the one the mechanism demanded, and logged the date. One amendment. Next one permitted in January. Then they closed the ledger and went back to work — and the price that week was neither here nor there.

Four storms now: hands, mind, nerve, honesty. The page has a kill list with four facts on it instead of three, and a rule for how it's allowed to grow.

If this one was useful, three things:

One — audit your own list against the four gates. For each fact: would you have written it before? Is it the mechanism or the mood? Is it observable and dated? When did it go in, and why? Any fact that fails a gate gets rewritten or removed — this week, while you're calm.

Two — start the amendment ledger. One line per change to the thesis: date, what, why. If it has more than one line a quarter, that's your answer, and it isn't the one you wanted.

Three — reply with the shift your own kill list was weakest on: money, work, or rules. One word. I read every one, publish none, and the distribution tells me which mechanism this readership has the least honest grip on — mine, it turns out, was work.

Next Sunday, the fifth storm: The Euphoria — the week it rips, the bands feel stupid, the group chat is all-in again, and the hardest sell of the whole arc: trimming into green. A drawdown breaks your hands; the noise breaks your mind; the crowd breaks your nerve; the gap breaks your honesty; the euphoria breaks your discipline. We'll build the tool before you need it, which is the only time it can be built.

— Bill2Billion

P.S. — Not financial advice, and not a market call in either direction. The one-to-ten-percent slice is the canon; no leverage, no timing, no all-in, ever — a week with a billion dollars of leveraged positions liquidated is the standing reminder of why. Live numbers, re-verified this weekend: bitcoin near $82,100 after a −2.8 percent week and a Thursday low of about $80,300, which is roughly 1,218 satoshis to the dollar; spot ETF outflows about $730 million across Wednesday and Thursday; three-month bills at 4.10 percent against core PCE of 3.0 percent for August, so the silent tax stays at +1.1 percent; the federal funds range at 3.75–4.00, with September's minutes saying another hike is "likely appropriate by year end"; the ten-year at 5.23 percent, Brent above $100. The Census adoption series and the BLS productivity–pay gap are now the two halves of Shift II's kill fact, entered tonight, dated, with the reason. The amendment ledger has one line on it, which is the only endorsement this letter will ever make.