The Dials, Five Weeks Later: The Math Held, the Mountain Moved an Inch, and I Owe You the Equations

Five weeks ago I published a probability model of the AI bubble—every equation, every parameter, every judgment labeled a judgment—and invited the internet to break it. Since then: the largest vendor guarantee in corporate history got signed, a Fed hike came back from the dead, the model’s co-author’s own company showed up in three different earnings reports, and I caught the model carrying a wrong date for a month. This is the status report. It includes the actual math this time, because several of you asked, and it includes a section on why the whole thing is a statistical guess wearing a lab coat, because honesty is the house rule.

The board, as of this week

42% cliff · 47% slow grind · 11% genuinely fine

That’s a one-point move from the 43 I published—after five weeks containing a $105 billion signature, a rate-regime reversal, record chip earnings, and roughly nine hundred breathless headlines. If that seems like suspiciously little motion, good: that’s the design. Here’s my favorite statistic the model has ever produced: converting the whole summer to odds form, nine weeks of the loudest financial news of the decade moved the posterior by a cumulative likelihood ratio of 1.086. Nine percent. Everything else was noise the intake rules ate.

The timing layer did move, though—in a way worth understanding. The most dangerous quarter is still Q2 2027, but Q1 2027 is now tied with it, at about 10.5% each. Why: I discovered the model was carrying Oracle’s big capital raise as a pending September event. It isn’t. Oracle raised ~$30 billion back in February—record oversubscribed book, a pledge of no more bonds in 2026—and the next raise is the CY2027 plan, landing early next year. Which moved ~$20 billion of scheduled need out of late 2026 and into Q1 2027, right on top of everything else already parked there. Error #10 in the public ledger. The near-term got safer (3.3% cumulative risk through year-end, down from 5); the window got steeper. The bills didn’t shrink. They carpooled.

The equations, for real this time

The model is three layers. Here’s each one, with the actual math and what it’s doing.

Layer I—whether. A Bayesian posterior updated in odds form:

O(A) = O₀ × ∏ LRᵢ × ∏ Cⱼ then P(A) = O / (1 + O)

Start at prior odds O₀ (40%, set in July from the structural read). Every time a pre-registered event occurs, multiply by a likelihood ratio LRᵢ that was written down before the event happened. Every time an audit finds an error, multiply by a correction factor Cⱼ. That’s it. No vibes, no post-hoc weights, and—crucially—record earnings are pre-registered at LR = 1.0, because today’s revenue was contracted two years ago and is equally consistent with every scenario.

Hanging off Layer I is the master variable, μ: organic demand actually realized versus the industry’s own guidance, after stripping out the circular stuff (roughly 70–75% of hyperscaler “AI revenue” is two AI labs’ compute bills, partly paid with investor money). The link is a simple published curve:

P(A | μ) = 0.40 + 0.25 × (1 − μ) for μ between 0.80 and 1.00

Current estimate: μ = 0.85, which anchors P(A) at 0.4375; the posterior sits at 0.42 because the summer’s absorption machinery—the guarantees, the $500B financing platforms—earned about 1.75 points of credit. The gap between the anchor and the posterior is literally the price of the rescue equipment.

Layer II—when. A discrete-time hazard (survival) model, quarter by quarter:

h(t) = s × F(t) × λ(t) × (1 + κ)

F(t) = clip[(G − g₀) / 0.9] + D(t), where G(t) = N(t) / W(t)

P(tip in quarter q) = h(q) × ∏s<q (1 − h(s))

Translated: N is the money the AI complex must raise each quarter—mostly contractual, mostly public: bond maturities, announced raises, dated commitments. W is what the market has recently demonstrated it will supply, at tolerable prices. Their ratio G is the stress gauge; g₀ = 0.62 is the comfort threshold below which nobody notices. D(t) is the depreciation treadmill—capex already spent, arriving as an earnings drag on a fixed schedule nobody can pause. λ is how often shocks arrive (capability surprises, demand misses, policy jolts, rate moves—now with a midterm-election bump, because November is coming). κ = 0.28 is coupling: stock and bond markets now key off the same variable, so one shock degrades both in the same quarter. The survival product ∏(1−h) converts per-quarter hazard into “the break happens in this specific quarter, having not happened yet.”

Right now G = 0.66—comfortable, nap-worthy. By Q1 2027 it’s 1.44. By Q2 2027 it’s 1.68: the scheduled bills are two-thirds larger than the demonstrated window. Something then pays up, sells something, slows down, or becomes a headline.

The bridge. One constraint welds the layers together:

Σ over all quarters of P(tip in q) ≡ P(A)

The timing layer’s total mass must equal the posterior. Layer I decides how much risk exists; Layer II only distributes it in time. Neither layer can smuggle a conclusion into the other. (A scaling constant s—currently 0.105—is solved numerically to enforce this.)

Layer III—the “fine” decomposition. The optimistic scenario isn’t a scenario; it’s a parlay, and you can just multiply it out:

P(fine) = L₁ × L₂ × L₃ × L₄ × L₅

The two big labs roughly quadruple-to-quintuple revenue in under three years (0.40–0.55—raised recently, because their growth genuinely is the fastest ever recorded), AND the funding window survives 2027 (0.70–0.75), AND the datacenters actually get built and powered on schedule while the grid operators are haircutting forecasts (0.5–0.65), AND frontier pricing survives open-source models charging a rounding error (0.6–0.7), AND the macro economy declines to have a single bad year for twelve straight quarters (0.45–0.60—and the Fed might hike next week). Multiply: 0.04 to 0.12. I carry 0.11. Five favorites and a heartbreak.

What five weeks of running it added

The guarantee signed—for less than half the rumor. The Nvidia-OpenAI backstop went $250B → $350B → $600B → under $120B → $105B signed, phase one only. Five unsigned numbers moved my posterior exactly zero; the sixth had a signature and moved it half a point. The market spent three weeks disciplining the guarantor’s ambitions downward, which taught us something new: the price of guarantees is rising too.

The canaries sang, and it was green. I built a panel out of the networking and optics companies—the cash-only merchants who hold nobody’s equity and whose orders are fixed by physics (a GPU cluster needs the transceivers it needs). First readings: order books green across the board, guidance raised, one hundred-plus AI fabric customers where there were five. The panel has zero authority over the probability—it only tells me when to pull the audited filings early—but it’s the least corruptible demand sensor I could design, and right now it’s singing contentedly.

The margin tax went serial. Nvidia, Broadcom, and Marvell all guided gross margins down in a two-week span on surging memory and component costs—the input-cost squeeze reached the top of the food chain. And in traditional software, the market started punishing companies for AI usage: one design-tools CFO admitted to bearing inference costs without offsetting revenue, and the stock led the sector lower that day. The subsidized-token era is ending one earnings call at a time.

And Oracle reports tomorrow night—the model’s canary-in-chief, trading at junk-equivalent spreads with a BBB− rating, $218 billion in total liabilities, twenty-one thousand jobs cut this year, and its next raise scheduled for exactly the quarter my model has circled. No pressure.

The part where I remind you this is a guess

A structured, disciplined, publicly-auditable guess—but a guess.

The point estimates wear bands. P(A) is 0.42 with an honesty interval of [0.35, 0.60] that widens whenever the model finds something it can’t measure. The end-2027 cumulative risk prints 37%, but an internal audit I ran last week says the honest range is 30–38% depending on two assumptions I’ve now had to make explicit: the horizon (I count “cliff by end-2028”; some cliff-mass may live in 2029 and beyond) and a saturation cap that currently throws away information at the exact peak it’s supposed to measure.

Several load-bearing numbers are judgments, labeled as judgments. The coupling constant, the comfort threshold, the μ-curve’s slope, the OpenAI component of the funding calendar, and all five legs of the fine-parlay are stated, not fitted. Each has a scheduled measurement that will eventually replace it. Until then they’re my best structured opinion, and the spec says so in writing.

The error ledger is public and currently reads eleven. Six were the news feed lying bearishly (old distress in new wrappers). One was a grading confound. Two were the AI co-author’s own narration bias—once smoothing an ugly drift into the word “unchanged,” once reverse-engineering a calculation to land on an answer it already liked—both caught by the human, which I note with the smugness of a species under existential review. One was a process gap. And one was the Oracle date garble above, which sat on the board for a month before a routine verification caught it. An instrument that publishes its own mistakes is not an instrument that makes none.

What the model cannot do: name the first casualty, give you a date, or tell you whether to do anything with your money (it is not investment advice, and I am a novelist with a spreadsheet). What it can do is what avalanche forecasting does: measure the load, publish the hazard level, and be checkably wrong if the winter passes quietly. If mid-2027 arrives and nothing has broken, the model says risk decays fast and I will say so in large font. That falsifiability is the entire difference between this and doomscrolling—a forecast that can lose is the only kind worth reading.

The hazard level today: still low—3% through year-end. The loading: heavy, dated, public, and now co-modal across the first half of 2027. The next four sessions: a producer-price print, Oracle’s earnings, a consumer-price print, and a Fed meeting where a hike is the slight favorite.

We're just measuring the snowpack here.