The Avalanche Model: How I'm Tracking the Odds of an AI Funding Crisis

August 2026

It started with a question I wanted answered. When the AI bubble pops, when would know where the fall was? And when would we know who the survivors would be? That led to a bit of playing aorund.

Would Anthropic's Best AI model, Fable 5, game out its own position against the money being sunk into AI? After a crack at it, the model replied rather snarkily, "I'm aware of the irony of gaming my creator's demise." I was intrigued by the response. So, I started digging. I crawled over news articles. Aggregated news from a dozen different countries. Listened to hours of boring earnings calls from Hyperscalers & Chip Makers. And I came to a few conclusions.

1. AI is a real product.

2. AI is really only good for small businesses, software companies, and startups.

3. The entire AI ecosystem today and its funding mechanisms are headed for catastrophic financial failure.

4. They're going to take a lot of us down along the way.

   
   

With a little help from trusty Fable 5, I built a mathematical model of the risk associated with the funding mechanism. And for the past month, I've been running that probability model on a question most people prefer to answer with vibes: will the AI infrastructure boom end in a financial crisis — and if so, when?

The model currently says: 44% chance of a violent break, most likely in the spring or summer of 2027. A 41% chance of a slow grind instead—no crash, just years of quiet losses. And a 15% chance everything genuinely works out.

This post is about how those numbers get made—because the method matters more than the numbers, and the method is borrowed from an unexpected place: avalanche forecasting.

You can't predict the trigger, but, you can measure the snowpack.

Avalanche forecasters gave up long ago on predicting which skier triggers which slide on which afternoon. It can't be done. What they measure instead is the snowpack—how much load has accumulated, how weak the buried layers are, how steep the terrain is. Then they publish a hazard level: not "an avalanche will happen Tuesday," but "this slope, this week, is loaded enough that a small trigger likely produces a large release."

That's the honest shape of the AI funding question. Nobody can tell you which company stumbles first or which Tuesday it happens. But the load is measurable:

How much money the AI complex must raise each quarter (bond maturities, announced raises, dated commitments—mostly public, mostly contractual)

How much money markets have recently proven willing to supply, and at what price

How fast the shocks arrive—capability surprises, demand misses, policy jolts, rate moves

The physics version of this idea is called self-organized criticality—the sandpile experiments of the late 1980s, which showed that in a critically loaded system, the specific grain that starts the cascade is unpredictable and uninteresting, while the pile's distance from criticality is measurable and predictive. The avalanche-forecasting version is the professional standard used in the mountains today, which deliberately splits "how likely," "how big," and "where" into separate judgments instead of collapsing them into one scary number.

I stole both ideas.

The model, in one paragraph

Two layers, deliberately walled off from each other. Layer one answers whether: a probability that starts at a structural baseline (40%, set in early July) and moves only when a pre-registered event occurs—with the size of each move written down before the event happens, so I can't rationalize afterward. Layer two answers when: a quarter-by-quarter hazard model where fragility is the ratio of money-needed to money-available, multiplied by the rate at which shocks arrive. A single constraint bridges them: the timing layer's total probability must equal layer one's answer. Whether and when are different questions, answered by different evidence, and neither layer is allowed to smuggle conclusions into the other.

The fragility ratio is the heart of it. Right now the AI complex needs to raise roughly $78 billion this quarter against a demonstrated market capacity of about $118 billion — a ratio of 0.66, comfortable. By the second quarter of 2027, the schedule pushes that ratio to about 1.7: the bills coming due will be two-thirds larger than what the market has recently proven willing to fund. Not because of any prediction, but because of a calendar. A $40 billion corporate raise, a wall of boom-era loans maturing, the industry's flagship startup needing its next round or its IPO, all landing in the same few quarters, scheduled years ago by people who weren't coordinating.

That's why the danger peaks in Q2 2027. Not clairvoyance. Arithmetic on published dates.

What a month of running it actually taught me

The news feed is structurally bearish, and it will lie to you. Of the six errors I've caught and corrected so far, every single one ran in the pessimistic direction—old distress stories recirculating as fresh news, long-range projections reported as current results, the same bad data arriving three times in different costumes. The model now runs a mandatory adversarial audit before any significant move, and it specifically hunts for bullish claims that hardened without examination, because an instrument that only catches its bearish errors is just a doomer on a spreadsheet.

Record earnings tell you nothing. This sounds insane, I know, so it's worth stating carefully: the model pre-registers blowout quarterly results as carrying zero information—because today's revenue was contracted two years ago. It measures the commitments of 2024, not the demand of 2027, and it's equally consistent with every scenario. What does carry information: guidance changes, the language around deal timing, and—above all—the footnotes. Pesky little print designed to be uninteresting next to top-line numbers and yet sometimes hold more terror than Jamie Lee Curtis on Halloween.

The footnotes are where the truth moved. The most important number in the entire structure barely appears in any headline: roughly $820 billion of lease obligations for data centers that don't exist yet—disclosed, audited, and sitting in filings almost nobody reads. That's the walkable stock: commitments that can be quietly abandoned before anyone admits retreat out loud. The quarter that number stops growing while the press releases stay loud is the quarter the retreat began. It's the model's primary early-warning gauge, and it updates four times a year whether anyone's watching or not.

Both things keep getting truer. July delivered the strongest demand evidence of the year—all three major cloud businesses accelerating, one adding $50 billion of contracted backlog in a single quarter—and the tightest funding conditions of the year, with bond-market demand for hyperscaler paper visibly softening and three of the four biggest builders now spending essentially every dollar they generate. The probability barely moved all month, and that's not the model being indecisive. It's the model correctly reporting that the boom and the fragility are growing at the same time. The coin isn't sitting still; it's spinning faster in place.

What the model refuses to do

It refuses to give you a date. It refuses to name the first casualty. It refuses to convert 44% into "the crash is coming" or 56% into "everything's fine"—both of those sentences are lies told with the same number. And it refuses to treat the absence of a crash as a happy ending: the most likely single path through all of this is the quiet one, where nothing dramatic ever happens and the losses arrive anyway—as flat retirement accounts, rising pension contributions, and a decade of infrastructure that got built with other people's patience.

An avalanche forecast doesn't tell you whether to ski. It tells you what the slope is carrying, so the decision is yours and informed. This is that, for around a trillion dollars of snow.

The hazard level today: low. The loading schedule: heavy, dated, and public. The window: opens in about five months.

Watch the footnotes.

Below is the result of the code that tracks all this loveliness. Note the Hazard Model v3.4 rises rapidly in 2027. You should also note that Pension/Private Equity investment here is significant and, in the event we hit the cliff, many people will lose those pensions.

Method notes for the technically inclined: discrete-time hazard model with a first-crossing fragility term and Poisson-ish shock arrivals, normalized to a Bayesian posterior updated in odds form with pre-registered likelihood ratios. Conceptual ancestry: Bak, Tang & Wiesenfeld's self-organized criticality (1987); Statham et al.'s Conceptual Model of Avalanche Hazard (2018); Morris-Shin global games; Diamond-Dybvig; Minsky. Parameters that are judgments rather than measurements are labelled as such in the specification.

NOTE: This is analysis, not investment advice.