The Dials: I used AI to build a Crisis Model for the AI Bubble, and I'm Publishing the Whole ThingAugust 2026

For the past six weeks I've been doing something deeply unfashionable: putting a number on the AI bubble instead of an adjective.

You've read the adjectives. The bubble is "unprecedented." It's "unstoppable." It's "about to pop." It's "different this time." Everyone's very confident, nobody shows their math, and the loudest guy on television just declared the bears defeated because a stock went up on a Wednesday. So I built a model—a real one, with equations, pre-registered rules, an error log, and a probability that moves when evidence arrives and only when evidence arrives, like some kind of animal.

Today I'm publishing all of it. Every equation. Every parameter. Every judgment call labeled "judgment call" in writing so future-me can't pretend it was science. The PDF is at the bottom. This post is the tour.

The number

43% chance of a cliff. A named failure—a default, a frozen fund, a forty-billion-dollar bond deal that doesn't fill—forcing everyone to reprice on the same ugly morning.

46% chance of the boring apocalypse. No crash. Just years of extensions, quiet markdowns, zombie datacenters, and pension statements that mysteriously stop going up. Japan did this for a decade and called it the Nineties.

11% chance it's genuinely fine. Not zero! Genuinely not zero. We'll get to why it's not bigger.

And a timing layer: if the break comes, it's most likely between January and September 2027, peaking in the spring. Not because I gazed into an orb. Because that's when the bills cluster—Oracle's $40 billion raise, a wall of boom-era loans maturing, OpenAI's IPO-or-else, and a tsunami of GPUs arriving COD, all scheduled years ago by people who were very much not talking to each other. My "danger window" is just their calendar with a highlighter on it.

The method is stolen from avalanche forecasters, who are smarter than finance people

Avalanche forecasters figured out decades ago that you cannot predict which skier triggers the slide—and, crucially, that it doesn't matter. You measure the snowpack instead: the load, the buried weak layers, the angle of the slope. Then you publish a hazard level and let the skiers make informed decisions about their own funerals.

So that's the model. Layer one: whether. A probability that starts at a structural baseline and moves only on rules written down in advance—because the alternative is vibes, and vibes have a way of always confirming what you wanted to believe at breakfast. Layer two: when. Money the AI complex must raise each quarter, divided by money the markets have recently proven willing to hand over. That ratio is 0.66 today—comfortable, nap-worthy. By spring 2027 the schedule drags it to 1.68—the bills two-thirds bigger than the demonstrated window. At that point somebody pays more, sells something, slows down, or becomes a headline.

Layer three is my favorite, because it ruined my own optimism with arithmetic. I decomposed "it's genuinely fine" to see what it actually requires, and it turns out fine isn't a scenario—fine is a parlay bet. For the boom to simply pay its own bills: the two big AI labs must roughly 7x their revenue in three years (nobody's done it at this scale, and I checked), and the funding window must survive 2027, and the datacenters must get built and powered on time (the grid operators are currently laughing), and frontier pricing must survive open-source models that charge a rounding error, and the world economy must agree to have zero bad years for twelve consecutive quarters. Every leg is individually plausible. Multiply them and you get 11%—which is what a parlay is: five favorites and a heartbreak.

What six weeks of running this thing taught me

The news lies bearish, constantly, and it will make you feel smart while it does it. Most of the errors we caught were the feed's fault—old distress recirculating in fresh packaging, ten-year projections reported like this quarter's results, the same scary number arriving three times wearing different hats. The model now runs a mandatory adversarial audit before any real move, and it specifically hunts for bullish claims that snuck in unexamined—because a doomer with spreadsheets is still a doomer, he's just harder to argue with at parties.

Record earnings mean nothing, and I will die on this hill. The model pre-registers blowout quarters as zero-information events. Today's revenue was contracted two years ago; celebrating it as evidence about the future is like reading last year's weather report out loud and calling yourself a meteorologist. The truth lives in the boring places: guidance deltas, order books, and footnotes—specifically the $820 billion of lease obligations for datacenters that do not exist yet, sitting in filings nobody reads. The quarter that number starts shrinking while the press releases stay loud? That's the retreat, announced in the only language corporations can't stop themselves from speaking: accounting.

Both things keep getting truer at once, which is the punchline. The demand is real—I built a sensor panel out of the networking and optics companies, the unglamorous cash-only merchants who sell the fiber and hold nobody's equity, and their order books are green as a golf course. And the financing keeps getting weirder—the chip maker is now organizing half a trillion dollars of its own customers' credit, guaranteeing their debts, and buying back its own used GPUs to put a floor under the collateral. Stop and savor that: the vendor is the lender's guarantor and the collateral's market-maker simultaneously. You do not build that machine for customers who can pay. Both facts strengthened all summer, which is why my number barely moved—the party and the tab are compounding together, and they've agreed to meet in the spring.

The disclosure, because I don't lie to my readers and I'm not starting for a chatbot

This model was built in extended collaboration with an AI—Claude, made by Anthropic. Which means, yes: the model's co-author is a product of the exact bubble it's measuring, built by a company whose valuation literally flows through the earnings of two of the giants we track. I am aware of how that sounds. It sounds like asking the casino to audit itself.

Here's why I'm telling you anyway, besides the fact that I always would: the method is built so it doesn't matter. Every number is sourced and tiered. Every parameter that's a guess is labeled a guess. Every update is pre-registered. The error log is public—eight catches so far, and the two funniest ones were the AI's own narration bias: once it tried to smooth an unflattering drift into the word "unchanged," and once it reverse-engineered a calculation to land on the answer it already liked, and both times the human—hi—caught it. The tool is conflicted. The ledger doesn't care. Check the arithmetic yourself; that's what the PDF is for.

What would change my mind (in writing, so I can't weasel)

Oracle fills its $40 billion raise → the peak slides later. The half-trillion financing platforms turn from press release into signed deals → the peak flattens. OpenAI's S-1 shows the growth curve actually curving → the optimistic legs strengthen and I'll say so. The $820 billion in phantom-datacenter leases starts falling while the capex speeches stay loud → the number jumps and I'll say that louder.

The hazard level today: low—genuinely, boringly low. The loading: heavy, dated, and sitting in public filings. The window: opens in about four and a half months.

THE_DIALS_full_model.pdf

Download PDF • 26KB

Not investment advice. A probability is not a prophecy. Watch the footnotes—nobody ever got margin-called by a press release.