The Trillion-Dollar Vintage

Chain of custody. This is the definitive edition of a document with a strict chain of custody: it redesigns Run 2 of 2: Adjudication (10 August 2026), which itself stands on a no-verdict Evidence Pack assembled the same day. Every number, probability, ruling, and conclusion is transcribed unchanged from those two documents. The evidence labels travel with their facts deliberately. Where the original flags a figure as unverified, the flag is kept, the flags are part of the curriculum. Section references of the form §x.y point to the underlying Evidence Pack.

How to read this. Six parts, in order. After each numbered section stands a Mental Model box: three to six sentences built to outlast the details. Threaded through the text are thirteen numbered Patterns, the reasoning moves the analysis runs on, stated explicitly so you can carry them to problems that have nothing to do with GPUs. Dense tables carry a one-line takeaway; Parts III through V open with a bearing line so you always know where you stand. Part VI is the retention layer: twelve claims worth knowing cold in six months, a one-page monitoring card, and the single most important intellectual move in the whole exercise. Read with a pen.

Three rules govern every figure on this page. They are declared once here and never restated, and they are specific to what this analysis is about.

1 · The fork is never closed.

The decisive quantity in the whole exercise, the revenue yield on a vintage dollar, came back from two honest constructions 2.3× apart. Anchor A is drawn in orange, Anchor B in blue, and they are drawn separately everywhere they appear. There is no blended mark anywhere on this page, because no instrument took one.

2 · Solid is a measurement. An open outline is arithmetic.

A filled mark is a quantity some instrument measured or some filer filed. An open outline, drawn at full size and never faded, is a quantity that arithmetic or assumption produced and no instrument observed. It is not a lesser version of the number; it is a different kind of number. That single rule catches the +1.27% expected return, the 44.9% arithmetic ceiling on circularity, the 55.1% midpoint the source itself labels an assumption with no evidentiary basis, and the mid-anchor the analysis refuses to interpolate. Where a figure needs to show a disclosed range rather than a manufactured point, it uses a shaded extension bounded by hairlines. That is a third thing, and it is used only for ranges the source publishes.

3 · Evidence tags travel with the mark.

The source labels every number by how it was obtained, and keeps the label attached even when the number is embarrassing. The labels do work later: when two anchors disagree, which one is built from F and which from S becomes an argument with consequences. So the tags travel here too, in the apparatus rail beside each drawing.

Retrieved fact

Dated and sourced. Tiered from T1, audited SEC filings, down to T7, industry and sell-side data.

Estimate

Assumptions stated.

Inference

Derived from other quantities in the base.

Speculative

Rumoured. Supply-chain chatter, unconfirmed channel checks. A fact that has forgotten how it was obtained is a rumour with tenure.

Not in evidence

The number exists in the discourse and in no primary source. It is refused, and the date its refusal expires is named.

One further note on colour, so nothing is ambiguous. Orange means Anchor A and blue means Anchor B; nothing else on the page is either. Where a graduated blue ramp appears, light to dark, it encodes an ordered quantity named in that figure's own legend: economic life in Figure 1, unlevered hurdle in Figures 2 through 4 and 10 through 15. A ramp never encodes identity, and a ramp never shares a figure with the two anchors.

Every categorical pair and every ordinal ramp used here was checked for colour-vision separation and surface contrast in both light and dark before a mark was drawn. Every figure ships its numbers as a table, so nothing depends on seeing a colour, and the charts are static SVG, so the page reads with scripting switched off.

Between the first weeks of 2023 and the middle of 2026, five technology companies (Microsoft, Alphabet, Amazon, Meta, Oracle) together with a second tier of specialist builders, committed roughly a trillion US dollars of actual cash to the infrastructure of artificial intelligence. The adjudicated figure for this cohort of capital, designated V, is $1,082 billion as originally constructed and $990 billion after one allocation assumption is corrected (low–high bands: $931–1,232B and $822–1,159B respectively). None of this is announced spending, planned spending, or the press-release spending of the week. It is capital already incurred: chips bought, concrete poured, leases signed.

Two questions hang over that money, and nearly everyone who writes about it treats them as one.

The first: will the technology matter? The second: will the capital earn its keep? They sound like a single question wearing two grammars. Historically, they are not even reliably correlated. Britain laid its railways in the mania of the 1840s; investors surrendered roughly a third of their capital, and the network kept growing for some seventy years after the panic that funded it had been forgotten. America electrified; the holding-company pyramids that financed the current collapsed. The internet received its fibre from companies that did not survive the gift. In each case the technology triumphed while much of the capital perished, and the settlement was written in the same clause every time: users retained the network. Users retained the grid. Users retained the fibre. The value of a transformative technology and the return on the capital that builds it are separate variables, and the historical base rate is that the second resolves against whoever owned the fastest-depreciating assets at the highest cost of capital.

The document you are about to study attempts to answer the second question about AI from the least comfortable position available: inside the cycle, in real time, before the outcomes exist, using only what companies actually disclose. The evidence is incomplete by construction. No company reports how much of its capital spending is AI. No company reports how hard its fleet actually works. The most-cited revenue figure of the most-discussed private lab cannot be verified from any primary source. The temptation of that position is to caulk every gap with confidence. The discipline of this analysis is that it declines, every time, and writes “not in evidence” instead.

Its verdict, stated here so you can watch it being earned rather than asserted: H2, the technology is real and the capital returns are inadequate, with the destruction tail concentrated in one owner class and one funding structure rather than smeared across the vintage. The probability-weighted return on a vintage dollar is +1.27% unlevered, set against required returns of 8.13–16.97% depending on who owns the asset. About 76% of the vintage fails its hurdle; only 48% destroys equity outright. The modal outcome is neither boom nor bust but something history photographs poorly: a trillion dollars of real, functioning infrastructure earning less than its cost of capital, a slow, orderly transfer of value from the owners of capital to the users of intelligence.

Why this deserves several hours of your attention: the situation recurs. Every generation gets one or two capital cycles of this magnitude, and the people who read them correctly do so with a specific, learnable toolkit, how to define what you are measuring before you measure it, how to weigh two instruments that disagree by 2.3×, how to reason about evidence you can only partly observe, how to separate what a business is worth from what its price assumes. This document is that toolkit, demonstrated on live ammunition.

“Is AI a bubble?” is not a question. It is a mood with a question mark. You cannot lose an argument about it, which is exactly what makes it worthless: a claim no evidence can wound is a claim no evidence can support.

The analysis begins by dismantling that question and replacing it with decidable ones. Three hypotheses stand trial:

• H1, a productive general-purpose-technology buildout: the capital earns adequate returns.
• H2, the technology is real, but the capital returns are poor.
• H3, a reflexive, circularly-financed cycle ending in material capital destruction.

Notice what the triad accomplishes. It is exhaustive along the dimension that matters, what happens to the money, and it grants the middle hypothesis a name. Most public debate admits only H1 and H3, boom or fraud. The middle world, in which everything works and the capital still loses, is the one history keeps choosing, and it is invisible to anyone whose question admits only two answers.

The hypotheses are then demoted to lenses, and three quantified questions become the actual deliverables:

• Q1. What share of Vintage V, the 2023–2026 incurred capital, fails to earn its owner's unlevered cost of capital? Stated as a probability distribution: P(failure share exceeds 30%), P(exceeds 50%), P(exceeds 70%).
• Q2. How much of the ecosystem's measured revenue is endogenous, funded by other participants of the same ecosystem rather than by outside customers, and, separately, how exposed is future revenue to financed counterparties?
• Q3. Four dated tail probabilities: a capability plateau by a defined date under a defined proxy; external end-user revenue below a defined dollar bar in 2030; demand shortfall conditional on no plateau; fleet validation conditional on a plateau.

The architecture matters as much as the questions. The work was split into two runs executed as deliberately separate exercises. Run 1 assembled an Evidence Pack under a strict no-verdict rule, no probabilities, no allocation views, no synthesis toward a conclusion; classification and retrieval only, every figure labelled by how it was obtained. Run 2, the document taught here, was then confined to that Pack as its exclusive factual base. Anything Run 2 needed that the Pack lacked had to be retrieved and logged in a Gap Log with source and access date, or declared not in evidence. Five retrievals were logged; one of them was retrieved, judged confounded, and explicitly rejected rather than used.

Why bother with the ceremony? Because conclusions curate their own evidence. An analyst who already knows where the argument lands will, without ever noticing, collect the facts that land it there. Freezing the factual base before the verdict phase opens is the analytical counterpart of a clinical trial's pre-registration: it converts trust me into audit me. The same logic produces the supersession check (did anything material publish between the evidence date and the verdict date? here the window was zero days, and it was checked rather than assumed) and the disagreement rule: if the verdict phase finds a defect in the evidence phase, it may neither silently repair it nor silently inherit it. It must flag the defect, quantify the effect, and report both figures.

Before a single computation, the analysis establishes something most research never possesses: a chain of custody for facts.

Every number wears its provenance. F means retrieved fact, dated and sourced. E means estimate, assumptions stated. I means inference. S means speculative or rumoured, supply-chain chatter, unconfirmed channel checks. Sources are tiered from T1 (audited SEC filings) down to T7 (industry and sell-side data). This is not decoration; the labels do work later. When two anchors disagree, which one is built from F and which from S becomes an argument with consequences. A fact that has forgotten how it was obtained is a rumour with tenure.

The demand firewall. The most consequential classification rule in the exercise: money is sorted by origin, and two categories, internal corporate cash redeployed (e1) and capital raised in the markets (e2), are real external money that is never counted as demand. A hyperscaler spending $100 billion on GPUs is not evidence that end users want $100 billion of AI; it is a costly forecast issued by roughly five correlated decision-makers holding superior private information and a plausible bias toward overbuilding. Their capex is admitted as biased evidence about demand, never as demand itself. Once seen, the rule's violation is everywhere: most bullish AI commentary consists, at bottom, of citing purchases by five buyers as proof of demand from everyone else.

The materiality bar for “circularity.” The bear's cherished exhibit is a map of interlocking deals: chipmaker invests in lab, lab commits to cloud, cloud buys chips. The analysis declines to treat the map as the finding. A structure counts as circular only if it can be shown to inflate measured demand, revenue, capex, or asset values relative to what external cash alone would support. Apply that bar honestly and most of the celebrated edges fail it. Oracle–OpenAI fails: a large contract alone is not circularity, and no Oracle equity in OpenAI was retrieved. Meta's $27 billion Blue Owl joint venture fails as circularity, the capital is genuinely external, though it passes as off-balance-sheet leverage, a different problem filed under a different name. Broadcom–OpenAI fails: nothing retrieved. What survives is smaller and sharper, above all a single edge: NVIDIA is contractually obligated to purchase CoreWeave's residual unsold capacity through 2032 ($6.3 billion) and that customer contract is itself pledged as collateral for CoreWeave's borrowing. A vendor manufacturing demand for its customer's output, with the manufactured demand securing the customer's debt. That is what a genuine circular edge looks like, and its rarity in the evidence is itself a finding.

The disagreement rule, demonstrated. Run 2 found a real defect in the Pack. The Pack concedes that Amazon's AI share of capex can only be bounded at 40–70%, weaker than the 70–90% band assigned to the other four firms, and then applies a uniform 80% to all five when constructing V. Amazon is the largest single spender in the vintage. The correction: V's base falls from $1,082 billion to $990 billion (−8.5%). Then the pedagogically important part: the effect on the verdict was computed rather than presumed, and it is nearly nil, the failure share moves by less than 2 percentage points and the headline probabilities do not move at all, because Amazon's capital sits in the same owner class either way. Both figures travel through the document side by side. Flag, quantify, report both, never silently fix, never silently propagate.

“Not in evidence,” even when everyone knows the number. Anthropic's ubiquitous figures (a $47 billion revenue run-rate, a $65 billion Series H at a $965 billion valuation) appear throughout the press and in no primary source anywhere. The company's own newsroom carries neither. What it does carry, dated 1 June 2026, is the announcement of a confidentially submitted draft S-1, a filing that discloses no figures and is not public. The analysis therefore rules the run-rate unverifiable, flags the contamination it causes (external end-user revenue at the cloud layer is partly anchored to it), and notes the dated path to resolution: confidential S-1s become public shortly before a roadshow. To refuse a number every journalist repeats, while specifying exactly when the refusal will expire, that is evidentiary discipline under social pressure.

Vintage V is defined before it is measured, because the definition is the largest single lever on every downstream number. In: accelerators, networking, servers, data-centre shells, land and grid interconnection held by market participants, and merchant power dedicated to AI load, capital whose owner bears residual and obsolescence risk. Out: regulated utility capex (ratepayers bear that risk; it is reported separately, roughly $1.3 trillion forecast for 2026–2030); announced-but-uncontracted commitments (not incurred); and guided future spending (V is actuals only).

Now open a vintage dollar, because its composition is its fate.

Fifty-seven cents of every dollar sits in an asset that evaporates in five or six years and carries high technology risk. Nineteen cents endures for decades. Five cents may appreciate. When you hear “data centre,” you picture the building; the building is a fifth of the money. Which is why the asset-life question will prove to be one of the two hinges on which the entire verdict turns.

One further structural fact before the machinery starts: 28–40% of V: $303–433 billion, is construction in progress. Capital spent, resting in assets not yet in service, whose return clock has not started. Only two of the five firms disclose the figure at all (Alphabet: $122.8 billion; Oracle: $40.0 billion, up from $16.5 billion in a single year); the band for the rest is modelled, and no firm discloses what event triggers “placed in service.” A third of the vintage consists of planes in the air that have not landed, and, as the adjudication will show, this fact cuts against both the bull and the bear in ways neither side welcomes.

Ownership, finally, because returns will be judged owner by owner.

Here is the move that separates this analysis from nearly everything written about AI economics: the same physical asset is required to earn different returns depending on who owns it, not as a modelling preference, but as an observed market fact.

Each owner class receives an unlevered cost of capital built from its own market beta and balance sheet (10-year Treasury 4.69%, equity risk premium 4.28%, both dated and sourced).

Read the extremes. A REIT and a neocloud can own physically similar capacity, and the market charges one 4.56% for senior debt and the other 9.625%, a spread of 506 basis points on the same class of collateral. The equity market performs the equivalent operation through beta. Nothing about the silicon differs; everything about the claim structure around the silicon does, leverage, customer concentration, contract duration, corporate history.

Now re-lever, and meet the number that will organize the entire verdict: the unlevered return at which each owner's equity return crosses zero.

A hyperscaler, nearly unlevered, destroys equity only if the vintage earns less than 0.37%: almost any positive outcome preserves principal. A neocloud, levered at 41.5% debt costing 9.625%, destroys equity below 4.00%, more than ten times the threshold, on the identical asset. Leverage does not change what the asset earns; it changes who survives what the asset earns.

Everything to this point is scaffolding. Here is the load-bearing problem.

To model the return on a vintage dollar you need one number above all others: the gross revenue yield, annual revenue per dollar of installed capital. The Evidence Pack constructed that number two independent ways, and the two constructions refuse to agree.

Anchor A, realized, from a filer. CoreWeave, the one company that discloses both revenue and active capacity, generated an annualized $8.99 million per active megawatt (the average of $8.73–9.25M across three quarters: F, from filings). Divide by an all-in build cost of $50M/MW (S, no filer discloses it) and the yield is 18.0% per year.

Anchor B, constructed, from market prices. A GB200 NVL72 rack at $3.0 million (S, supply-chain rumour) drawing 132 kW implies $22.7M of accelerator per MW: $39.9M/MW all-in at the 57% component weight; rent the GPUs at $3.50/GPU-hour around the clock and revenue reaches $16.7M/MW/year. Yield: 41.9% per year.

The ratio between the anchors is 2.3×. That is not a rounding disagreement. It is two instruments reporting different worlds. Feed each anchor through the same cash-flow model, component weights from §3, a 60% cash margin (E, unsourced; no filer discloses AI-fleet operating margin either), residual values at the accelerator's end of life, three accelerator lives, three utilization/revenue scenarios, and you obtain the pair of grids on which the entire verdict turns. Failure is judged against each owner's hurdle from §4; a cell's failure share is the slice of V whose owner-class hurdle exceeds that cell's IRR.

Stare at this table until it becomes uncomfortable. Under Anchor A, 100% of the vintage fails in every single cell, even at five-year lives and high utilization; even after applying the most favourable published sensitivity (+6.5 points), A's best cell reaches only +6.1%, still beneath the lowest hurdle in the set (8.13%). Under Anchor B at the five-year life, the base case earns +13.9%, clearing the hyperscaler hurdle by more than 300 basis points. The bucket assignment for the entire trillion dollars flips on which instrument you believe.

Which should you believe? Now the evidence labels earn their keep. Anchor A's revenue is F, a real company's filed numbers. Anchor B is S on both sides: rumoured rack price, marketplace rental rate. Strict source discipline says A. And yet the Pack designates B as the base case, on three dated, independent observations that are very hard to reconcile with a world in which A's level is correct:

1. Credit committed real money against these exact assets at investment grade. On 31 March 2026, Blackstone-anchored insurance capital lent $8.5 billion secured by GPUs and customer contracts at SOFR+225bp with a Moody's A3 rating, roughly 750bp tighter than the same borrower's first facility. Lenders underwriting a first-lien claim do not price uniformly negative unlevered returns at investment grade.
2. Six-year-old hardware refuses to die. A100 accelerators (2020 silicon) are “completely sold out,” never retired at AWS, still carrying over $20 billion per quarter of NVIDIA's data-centre revenue in prior-generation product, with rental rates rising through 2026. An asset class that failed to earn its cost of capital in its first vintage does not reprice upward in year six.
3. The one realized per-unit revenue series held flat through a 70% capacity expansion, precisely where oversupply economics predicted compression.

Meanwhile, A's likely defect is identifiable but not curable from public data: its “active power” denominator is probably facility power rather than fully-deployed IT power (correcting at a 70% IT share closes roughly half the gap), and its contracted take-or-pay pricing sits below marketplace spot. So the better-sourced anchor is probably mismeasured; the worse-sourced anchor is probably nearer the economic truth; and neither is resolvable. The tension is stated openly rather than dissolved.

And now the move this entire study edition exists to teach. The reflexively “sophisticated” response is to interpolate, split the difference, blend a ~30% yield, build one tidy grid. The analysis refuses: “A defect that is half-corrected argues for a mid-anchor, and no mid-anchor cell exists in the Pack. I do not interpolate one.” A blended anchor would be a measurement no instrument took, data invented by arithmetic, wearing moderation as a costume. Instead the fork is carried into the verdict intact: 60% weight on B, 40% on A, both grids alive, the downside branch treated as live rather than discarded.

To aggregate the grids into an answer, weights must be assigned, and here the analysis imposes a constitutional rule on itself: no implicit weights anywhere. Every weight is declared before use, tied to evidence, and stress-tested under alternatives.

The declared weights. Anchor: A 0.40 / B 0.60: B is the designated base case, held near the middle because A is better sourced and B's designation rests on triangulation rather than measurement. Accelerator life: 5y 0.55 / 3y 0.33 / 2y 0.12, disclosed depreciation lives cluster at 5–6 years across all five reporters, and the two-year story is contradicted by both the disclosure record and six-year-old hardware still clearing $5,400 on the executed secondary market. Scenario: base 0.50 / high 0.28 / low 0.22: “sold out” statements from four operators argue against the low case, but commercial-commitment language is not a utilization measurement, so the up-case is capped for resting on the weakest species of evidence in the base.

Your eye should snag at once: the probability is identical at all three thresholds. That is not indolence. Return to the grid: its cells produce failure shares of 0%, 8.9%, 97.8%, or 100%, nothing between 8.9% and 97.8%. The distribution is bimodal. Either nearly all of the vintage fails (the A branch, plus B's weaker cells) or nearly none of it does (B at the five-year life). Any threshold drawn between 8.9% and 97.8% slices the same two clumps of probability. The honest answer to “P(more than 30%)? more than 50%? more than 70%?” is one number, three times, and the analysis says so plainly: manufacturing three different probabilities here would be false precision.

The expected unlevered IRR under these weights is +1.27%. Notice what species of number that is. The nearest cells in the entire grid are +0.9 and +1.5: the mean of a fork lands in the valley between its modes, a forecast of a world almost no scenario produces. The figure is reported, it powers the equity arithmetic, but the verdict never pretends the vintage will earn 1.27%. It will earn roughly −13% or roughly +14% at the branch base cases; the mean is merely the fork's centre of mass.

The two levers that matter are the anchor and the asset life, both supply-side questions about what the asset is worth and how long it endures. The utilization-and-demand scenario, the subject of the entire public debate, moves the verdict by ±2 points. This inversion, the answer lives where the discourse isn't, is one of the analysis's quietest and most valuable results.

The stress test. Under the required alternative weighting, evidence-tier-first, which mechanically privileges the better-sourced anchor (A to 0.55), the failure share rises to 84.0% and the probability to 0.831. Under a bull-tilted base-case-dominant weighting (A to 0.25, the five-year life to 0.65): 62.1% and 0.600. The full envelope across all three defensible weightings: failure share 62–84%, probability 0.60–0.83. The headline moves; its sign does not. Every defensible weighting concludes that a substantial majority of the vintage fails its hurdle, with probability comfortably above a coin flip.

And now the second number, which matters more than the first. Failing a hurdle is not destroying capital. Re-run the same machinery against §4's equity-destruction thresholds instead of the hurdles.

• Share of V failing its hurdle: 75.6%.
• Share of V destroying equity outright: 48.0%, and conditional on Anchor B alone, just 13.3%; conditional on A, 100%.
• The band between the two tests: 27.6% of V, roughly $299 billion (original basis; $273 billion adjusted), earns a positive return below its cost of capital.

That band is the modal world. Not a crash: functioning data centres, paying customers, positive cash flows, and returns that quietly fail to compensate the capital that built them. This is H2's signature, and it is why “bubble: yes or no?” misses the likeliest outcome entirely. Purgatory does not photograph well, so nobody predicts it.

The bear's favourite word is “circular”: the money travels in a ring and the revenue is an apparition. The analysis's contribution is to dissolve the word into two different quantities that the discourse chronically conflates, and then to bound each one honestly.

Why “lower bound”? Because disclosed financing structures are observable and undisclosed ones are not, the measurement error is one-sided. Every future revelation moves the number up, never down. The bear brief leaned its full weight on exactly this, and the adjudicator's ruling is a sentence worth memorizing: the bear is entitled to the direction and nothing else. One-sided error establishes the sign of a revision, never its size. The honest posture is an interval (at least 31.8%, at most 44.9% at the cloud layer) with no point estimate placed inside it, because no evidence locates one.

Three constraints then cage the bear's expansion of these numbers. Even at the ceiling, 55.1% of cloud-layer revenue is non-endogenous. The unclassifiable bucket (d) is no reservoir of hidden circularity, its stated contents (Google Cloud AI, Azure's non-OpenAI AI) are, on any reading, mostly enterprise cash. And the supply-chain reconciliation points the wrong way for the bear: measured chip-vendor revenue is smaller than customers' silicon-bound capex, the opposite of what channel-stuffing produces.

Then arrives the most instructive empirical moment in the document. The Pack's top-ranked gap was Broadcom's AI revenue, an entire missing vendor. Run 2 retrieved it: $10.8 billion in the quarter ended 3 May 2026, +143% year-over-year (with guidance to $16.0 billion, over +200%, the following quarter). Adding it enlarged the identified revenue base from $327.7 to $370.9 billion, collapsed the reconciliation residual from 26–56% to 11.4–38.0%, and moved the semiconductor layer's endogenous share down, from 14.0% to 12.4%. Sit with that. Closing the scariest data gap made the scary number smaller. The Pack had predicted precisely this (it diagnosed the residual as missing suppliers, not hidden inflation) and the prediction paid. When a framework tells you in advance which way an unknown will break, and it breaks that way, the framework is earning its keep.

Forward dependence, the separate quantity. Of the cloud layer's identified $78.3 billion revenue base, how much comes from counterparties that are themselves financed by the ecosystem? The floor, cleanly identified: 30.8%: Microsoft's $24.1 billion of FY26 revenue from OpenAI, set against $11.6 billion of cumulative funding into OpenAI, with roughly $6 billion of it standing as an uncollected receivable. The ceiling, if every undisclosed counterparty share breaks maximally bad: 85.8%. The midpoint the original displays (55.1%) is explicitly labelled an assumption with no evidentiary basis, italicized in the source as decoration, not data. The honest statement: the width of that range is a disclosure fact, not an analytical one.

Against the exposure stands a defended central recapture rate of 15–25%, central 20%, the fraction of exposed revenue protected by third-party credit support, anchored to observable contract features. The explicitly supported exposures (Google's $1.4B + $1.8B lease backstops, NVIDIA's $6.3B residual obligation and $3.5B in guarantees, Meta's residual-value guarantee) are small beside the unsupported ones (Microsoft–OpenAI; Oracle's undisclosed OpenAI share of a $638 billion backlog; CoreWeave's $99.4 billion backlog). Net: 24.6%–68.6% of the identified base is exposed and unsupported, roughly four-fifths of whatever is exposed carries no backstop at all.

A measurement footnote that generalizes far beyond AI: a run-rate is one month multiplied by twelve; recognized revenue is what actually happened; and the two are routinely a large multiple apart. The only hyperscaler that disclosed an isolated AI run-rate at all was AWS (>$25 billion). Whenever anyone quotes an AI revenue figure at you, your first question is now reflexive: run-rate or recognized, and says who?

Vague fears cannot be adjudicated. The analysis compels its four tail risks into dated, proxied, internally consistent probabilities, and shows every derivation.

Tail 1: P(frontier capability plateau by 31 Dec 2028) = 0.15 (band 0.08–0.25). Since no capability benchmark exists in the evidence base, the proxy is defined before the probability is assigned, and it is an economics-of-delivery proxy, flagged openly as the thinnest of the four. A plateau requires both legs: frontier API pricing failing to halve (to $1.25 per million input tokens from $2.50), and per-GPU inference throughput failing to double (to ~11,700 tokens/s/GPU from a measured 5,842), by the date. Either leg would demand a sharp deceleration from trend: pricing fell 92% across the prior three years ($30 → $2.50); measured throughput rose 4.66× across roughly two hardware generations; and ASML is adding 30% EUV capacity in each of 2027 and 2028, the supply chain is being tooled for continuation.

Tail 2: P(external AI end-user revenue below $475 billion in calendar 2030) = 0.70 (band 0.55–0.85). The bar is derived, not asserted, follow the annuity logic, because this is the transferable skill. For V alone ($1,082B) to return its cost of capital at the hyperscaler hurdle (9.86%) over a five-year life, it must yield $1,082B × 0.2629 (the annuity factor) = $284 billion of cash margin per year, which at the 60% cash margin requires $474 billion of revenue per year, rounded, $475B. Current external end-user revenue, the only category the firewall admits, is $26.3 billion annualized. Reaching the bar by end-2030 demands a 93.0% compound growth rate sustained for 4.4 years with zero deceleration; grant a generous $50 billion true base for undisclosed lines and the requirement is still 66.8%. Current growth genuinely runs in that neighborhood (AWS's AI line at triple digits, paid enterprise adoption up from 35% to 50.4% in a year) but against it stands the best-sourced demand evidence in the base: Census data (~20,000 firms per wave) showing 17–20% of firms using AI at all, 57% of users running it in three or fewer business functions, and a median paid spend of $11.38 per employee per month against a top-1% figure of $7,449, a 654× depth gap. Breadth without depth, racing a 93% required compound rate.

Tail 3: P(demand shortfall | no plateau) = 0.66 (band 0.50–0.80). Derived for coherence rather than assembled from mood: with P(plateau) = 0.15 and P(shortfall | plateau) declared at 0.90, arithmetic forces (0.70 − 0.15 × 0.90) / 0.85 = 0.665. That it sits a mere 3.5 points below the unconditional 0.70 encodes a thesis: the binding constraint is diffusion and monetization, not capability. Models improving does not compress an enterprise integration cycle.

Tail 4: P(installed fleet validated | plateau) = 0.18 (band 0.10–0.30), higher than the naive 0.10, for a quantified reason: a plateau is two-sided for installed capital. No next generation means no obsolescence; the fleet's life extends, and at an eight-year life the revenue bar drops 29%, from $474B to $336 billion (annuity factor 0.1865), trimming the required growth rate to 78.5%. A plateau caps the demand path while protecting residuals and utilization, the same mechanism as the A100s sold out in year six. The two effects substantially offset. Most commentary treats “plateau” as pure downside for AI capital; the arithmetic says it is a trade.

Finally the tails are audited against Q1 from an independent direction. The grid route implies P(vintage validated) = 1 − 0.743 = 0.257; the revenue route implies 1 − 0.70 = 0.300. Two paths, built from different sections of the evidence and never tuned to each other, land within 4.3 points. When independent derivations of the same quantity agree, the number begins to deserve your trust.

An adjudication is worth only as much as the strongest version of the case it rules against. Both briefs are argued in good faith, from the evidence base, with every claim leading on a figure. The bull's seven exhibits:

B1: Old chips refuse to die. A100 silicon is six years old, “completely sold out,” never retired at AWS, still generating over $20 billion per quarter of NVIDIA data-centre revenue in prior-generation product, with rentals rising in 2026 and executed secondary sales at $5,400. If year-six hardware earns like this, the five-year-life column is the correct one, and Anchor B's base cell (+13.9%, above hurdle by 300 basis points) is the world we are in.

B2: The one measured yield series did not compress. CoreWeave's revenue per active MW held at $9.25 → $8.73 → $8.99 million across three quarters while active power grew ~70%; backlog per contracted MW rose 48%. The oversupply thesis issued exactly one testable prediction in the whole evidence base, per-unit yield compression, and in the one place it could be tested, it failed.

B3: Contracts with cash attached. IREN receives ~45% customer prepayment of GPU capex on ~85% of a >$4B target. Applied Digital holds $36.2 billion of fifteen-year non-cancellable take-or-pay against 175 MW live. Microsoft's commercial backlog stands at $678 billion. Buyers do not prepay for capacity they doubt.

B4: The physical constraint is the tell. GE Vernova holds 116 GW of contracted gas capacity, deposits paid, against roughly 20 GW per year of production; grid equipment books at 1.57× billings; Meta signed twenty-year nuclear agreements with prepayment. You cannot simultaneously have a supply throttle this tight and an oversupply.

B5: The same firms ran this play before, benignly. Cloud capex 2010–2020: the same companies, the same asset class, life-extensions then too, and the returns came. Alphabet's capex even fell two years running (2019–20) with no destruction event.

B6: Depreciation is conservative. Four life-extensions since 2021 track hardware that demonstrably still earns at year six; if true economic life is 6–8 years rather than five, the grids understate IRR by harvesting residuals too early.

B7: Credit, the leading indicator, is calm at investment grade. High-yield spreads at 2.71%, investment-grade at 0.78%, both tightening; GPU-and-contract collateral rated A3 at SOFR+225. Markets expecting uniform capital destruction do not price its collateral this way.

R1: The funding mix broke from the benign analogue, measurably. Across 2010–2020 the same firms' capex/OCF peaked at 90.6% and never crossed 100%; funding was internal throughout. Now: Oracle at 174.1%, Amazon at 107.2%, 31% of cumulative big-5 capex issuance-funded, Alphabet 26% → 71% in a single year, Meta 0 → 51%, and $1,090 billion of leases signed but not yet commenced sitting outside every balance sheet, $329.1 billion of it at Microsoft alone. The one quantified, same-firm, same-asset contrast with the benign precedent has crossed its line.

R2: Deflation strands capital without any demand disappointment at all. Next-generation pricing (GB300 at $83–90k per GPU against GB200 at $39–47k, for 1.45× the throughput) implies worsening performance per dollar; ASML adds 30% EUV capacity in 2027 and again in 2028; the memory complex sits at historic peak margins (Micron 84.6% gross, SK hynix 76% operating) that mean-revert. Falling replacement cost lowers the installed fleet's rental value with demand fully intact. Rentals did fall 24–33% through 2025.

R3: The backlog is concentrated exactly where the credit is thinnest. CoreWeave: a single customer (Microsoft) supplied ~67% of revenue; the headline “backlog” is broader than the accounting-standard figure, its increment management-estimated; cancellation terms are undisclosed by all five backlog reporters. Microsoft's $24.1 billion from OpenAI includes roughly $6 billion recognised but uncollected, with no third-party support.

R4: Deployment is running years ahead of diffusion. The probability-sample evidence (Census, ~20,000 firms per wave): 17–20% usage; 57% of users confined to three or fewer functions; AI-driven employment change in 2% of firms; median paid spend $11.38 per employee per month. Against ~$580B+ of guided 2026 capex across four firms. Breadth without depth is the signature of a technology being bought faster than it can be absorbed.

R5: The calm credit headline conceals its own composition. The same physical asset funds at 4.56% (Equinix) and 9.625% (CoreWeave): 506 basis points apart. Data-centre securitisation swelled from $4 billion to $61 billion in six years and prices 35–70bp wide of comparably rated structures. And the Lucent precedent instructs: its vendor-financing exposure ratio deteriorated from 5.0% to 27.2% in twelve months chiefly because the equity collapsed underneath it. Spreads gapped after the equity, not before.

R6: Every measurement error points the same way. The endogenous share is a floor by construction; each undisclosed structure, when revealed, can only raise it. A system that can only surprise in one direction has been read at its most flattering.

Now the part that teaches judgment: every exhibit ruled on individually, against the fixed factual base, with the reasoning in the open.

Three rulings deserve promotion to permanent equipment.

The allocation question is where amateurs and professionals permanently part company. The amateur asks: which layer is strongest? The professional asks: where is the gap between structure and price widest? A structurally impaired asset is a fine investment if its price embeds something worse; a magnificent business is a poor one if its price embeds perfection. Structure and price are independent axes. Alpha lives in their disagreement.

Same asset. Same scenario grid. The neocloud's outcome range is 24 points wider than the hyperscaler's and its expected equity return 5.6 points lower, the whole difference is capital structure. Second: how much of each equity price is the at-risk vintage (AI capex at cost as a share of enterprise value: Alphabet 4.9%, Microsoft 6.5%, Amazon 6.6%, Meta 9.6%, Oracle 14.0%) and CoreWeave ~37.6%. Third: what each price requires. Reverse-DCFs put implied perpetual growth at +282bp over the long bond for Microsoft, +385bp for NVIDIA, +558bp for Arista, while for CoreWeave, Equinix, and Constellation the model has no valid solution at all: at their current returns on capital (2.1%, 6.5%, 8.3%), no constant growth rate justifies the price. The price requires the business to become something else.

The four calls, each argued from both axes:

Best risk-adjusted: hyperscaler equity. Equity impairs only below 0.37% unlevered, the modal purgatory outcome preserves principal. The vintage at cost is 4.9–6.6% of enterprise value, so even a total write-off is a single-digit event. The legacy engine (22.0% ROIC) funds the wager internally: Microsoft issued zero external debt across FY25–26 while annual capex reached $115.9 billion. And the embedded-expectations gap is the narrowest in the set, +282bp. The true risk is not impairment but the pace of spending, which is why the monitoring card watches capex/OCF and uncommenced leases, not the depreciation footnote.

Second: semiconductors, selling the shovel, with one honest caveat. NVIDIA converts the cycle to cash at 76.8% ROIC while reinvesting only 11.8%, bears no depreciation or construction-timing risk on the fleet, and trades at the bottom of its own five-year valuation range (EV/IC 33.9× against a 33.9–86.8× history), the de-rating has already happened. The caveat is structural: 68.6% of semiconductor-layer revenue traces to corporate cash and capital markets rather than to end users. The layer's revenue is the derivative of precisely the funding decision the bear brief attacks, first to move if the funding mix cracks. And within the layer, price discipline still governs: AMD at 82.0× EBITDA sits above its own historical range, the expensive expression of the same facts.

Third: data-centre REITs, the inverse error. The best structure in the set, lowest hurdle, 4.56% debt, 20–40-year assets, impairment only below 0.99%, and a price that already knows it: no valid reverse-DCF solution, and the landlord captures roughly one-fifth the revenue per MW of the GPU owner. Attractive risk, unattractive price. The mirror image of the usual mistake.

Worst: neocloud equity. Every disadvantage compounds. The highest hurdle (16.97%) and the dearest debt (9.625%, 506bp over the REIT) on the fastest-depreciating asset; equity impaired below 4.00% unlevered, more than ten times the hyperscaler's threshold; the vintage at cost near 37.6% of enterprise value; ~67% single-customer concentration where the customer is also a competitor; a headline backlog broader than the accounting standard; no valid constant-growth solution at a 2.1% ROIC. Above all: this is the layer whose equity is most levered to the anchor question, the least-resolved input in the entire analysis. Maximum sensitivity to maximum uncertainty is the definition of a poor risk-adjusted bet, independent of any view on demand.

Two demonstrations of the two-axis discipline, because they generalize:

• Oracle: impaired, but priced. The worst funding profile of any large owner, capex/OCF 174.1%, net debt 4.53× EBITDA, 99% of capex externally funded, AI capex 14.0% of EV, and a valuation near the bottom of its own historical range. Structure and price have partially met. Not a recommendation (expected equity −0.66%, downside −28.4%, and the counterparty concentration behind a $638B backlog is unobservable), but the cleanest live proof that “structurally impaired” and “avoid” are different sentences.
• Arista and AMD: sound, but pricing perfection. Fine businesses at +558bp implied perpetual growth and valuations above their own historical ranges. Nothing wrong with the structure; everything demanding about the price.

And the two places likeliest to be paying peak-cycle prices for value that ends up with users: first, the neocloud layer, owner of the melting asset at the highest capital cost, with no pricing power over its own output (rental rates are set by a market it does not control) and a customer who is also its competitor; second, the memory-and-foundry complex: Micron at 84.6% gross margin and SK hynix at 76% operating margin sit at the top of a cyclical margin range, TSMC trades at the top of its own valuation history, and the equipment complex is adding 30% capacity a year into it. Buying cyclicals at peak margin, into a disclosed capacity wave, is the oldest mistake with the best data against it.

Why does the value end up with users at all? The one welfare measurement in evidence (flagged not in evidence, unverified on re-check, keep the flag): US consumer surplus from AI of $172.3 billion against a revenue comparator of roughly $14.2 billion, an order of magnitude of value created and not captured. That ratio is the mechanism behind the historical refrain. The technology can be everything its believers claim while its capital earns purgatory returns, because the surplus escapes to the demand side, which cannot service debt or depreciate GPUs.

The final discipline is the one that separates a view from a mood: every verdict ships with the observations that would kill it, dated, quantified, source-named. The original's standard is explicit: “if demand weakens” is unacceptable. A falsifier you cannot schedule is not a falsifier.

The full twelve-indicator card lives in §16, with the two reversal gates beside it. Its philosophy first, because the philosophy is the transferable part. Each indicator carries five properties: a current value (or an honest “not publicly observable”), a numeric trigger in both directions, a frequency, a direction-and-size of verdict update, and a named source. Two examples of the craft:

• The single highest-value indicator prints one day after the analysis is dated. CoreWeave's Q2 2026 report (11 August 2026) carries revenue, active megawatts, and backlog in one filing, the pivot variable of the anchor fork. Below $7.50M per active MW for two consecutive quarters, the fork swings toward Anchor A; above $10.50M, toward B. One number, printed quarterly, arbitrates the trillion-dollar question. (This edition preserves the analysis as of its date; the reader checks the print against the card.)
• The earliest-warning indicator is one almost nobody watches. Microsoft's $329.1 billion of leases signed but not yet commenced. Demand deterioration surfaces first as quiet non-commencement, invisible in every income statement, legible only as that disclosure balance shrinking. A sequential decline of more than 10% (~$33 billion) without matching commencements is the tell. The best indicators live where deterioration arrives before the financial statements.

Equally important is the anti-dashboard, the ledger of what cannot be monitored because it is not publicly observable: fleet utilization (the actual measurement, billed hours over available hours); loan covenant packages; GPU appraisal and advance rates beyond one issuer's own deck; cancellation and termination provisions for all five backlog reporters; churn and net revenue retention for any AI product; realized revenue per token for any provider. Writing this list is not a confession of defeat. It is the map of where the ambushes will come from.

Then the reversal, the analysis pre-committing, in public, to what would change its mind. Figure 17 draws both gates.

Flip to H1 (returns adequate) requires all four, by their dates: CoreWeave ≥ $11.00M/MW for two consecutive quarters by 31 March 2027 (raising Anchor A's own yield toward ~22% and narrowing the fork to ~1.9×); executed secondary H100 prices ≥ $18,000 on 25+ observed sales by 30 June 2027; two of the five majors back below 60% capex/OCF by their Q4 2027 filings; and Census depth, the share of users in three or fewer functions, below 45% in the 2027 supplement. Under all four, the failure share drops below 40% and the probability below 0.45. Flip to H3 (material destruction) requires all four: Microsoft's uncommenced leases below $260 billion by 30 June 2027 without commencement; a ≥$5 billion AI impairment or a second useful-life shortening by end-2027; high-yield spreads above 4.50% for 20 consecutive trading days; H100 rentals below $1.50 for two months by mid-2027. Under all four, the failure share exceeds 92% and the equity-destroying share 80%.

Refusing to average the fork.

The situation: two honest constructions of the decisive quantity, the revenue yield on a vintage dollar, came back 2.3× apart. One (18.0%) rests on filed revenue and a rumoured cost; the other (41.9%) on rumours at both ends that happen to triangulate with credit markets, hardware longevity, and the one realized yield series. Under the first, 100% of a trillion dollars fails in every scenario. Under the second, the base case clears its hurdle by 300 basis points. There is no third instrument.

Nearly every professional in that position produces a blend, call it a ~30% yield, and erects one smooth model upon it. The blend feels rigorous. It is the opposite: a measurement no instrument took, a world with no evidence behind it, manufactured because single numbers are socially comfortable. And it silently deletes the most decision-relevant fact available, that reality is currently one of two very different places, and we do not yet know which.

The analysis instead carries both grids to the end, weights them explicitly (0.60/0.40, the weights themselves argued from evidence and stress-tested to 0.55/0.45 the other way), and lets the fork's shape flow undiluted into the final answer, which is why P(failure > 30%), P(> 50%), and P(> 70%) are all the same number, 0.743. The distribution is bimodal because the evidence is bimodal. The expected IRR of +1.27% is reported and then explicitly disbelieved as a forecast: the vintage will not earn 1.27%; it will earn roughly −13% or roughly +14%, and the mean is merely the fork's centre of mass, resting in a valley almost no scenario produces.

And name it once, so you see it everywhere: this is Pattern №5 (carry irreconcilable instruments), Pattern №6 (report the distribution you actually have), and Pattern №7 (one-sided observability yields bounds, not points) revealed as a single epistemic character exercised at three scales, the same refusal to purchase comfort with invented precision, applied to a fork, to a distribution, and to a floor.

Why the move is rare: every institutional incentive opposes it. Clients pay for points, not intervals; committees average dissent into consensus by reflex; a bimodal answer sounds like hedging to anyone who has not understood it; and “the answer is one of two numbers, and here is what will tell us which” demands more nerve than “the answer is 30%.”

Why it is valuable, three compounding reasons. First, it preserves decision structure: how you size, hedge, and stage commitments under a fork differs fundamentally from how you do so under a bell curve, and blending destroys the information those decisions require. Second, it converts debate into monitoring: a preserved fork has an observable pivot (here, one company's revenue per megawatt, printed quarterly, with tripwires at $7.50M and $10.50M) whereas a blend leaves you litigating a phantom mean forever. Third, it is the discipline that runs the whole document at every scale: bounds instead of points for circularity, “not in evidence” instead of the number everyone repeats, direction without magnitude for one-sided errors. One move, applied consistently, amounts to an epistemic character: when the evidence is two-valued, the honest answer is two-valued, and the productive question becomes which, not how much.

What you should still be able to state, cold, in six months.

1. The verdict is H2: real technology, inadequate returns, purgatory, not collapse. 75.6% of the vintage fails its hurdle (P = 0.743), but only 48.0% destroys equity; the modal ~$299B slice earns positive-but-sub-hurdle returns.
2. The evidence is bimodal: no scenario yields a failure share between 8.9% and 97.8%. The real question is which branch, not how much, which is why all three thresholds return the identical 0.743.
3. The verdict hinges on two supply-side unknowns, the revenue anchor and the asset life, not on demand. Shifting 0.20 of anchor weight moves the probability by ~0.086; the demand scenario moves it by ±0.02.
4. Hurdles attach to owners, not assets: the same collateral funds at 4.56% and 9.625%: 506bp apart. The neocloud destroys equity below a 4.00% unlevered return; the hyperscaler below 0.37%, a >10× survival gap on identical hardware.
5. Circularity is real, bounded, and smaller than the discourse: at least 31.8% but at most 44.9% of cloud-layer revenue is endogenous, and one-sided observability licenses the direction of future revisions, never their size.
6. Closing the biggest data gap made the scary number smaller: retrieving Broadcom's $10.8B/quarter AI revenue cut the semiconductor layer's endogenous share from 14.0% to 12.4% and collapsed the reconciliation residual to 11.4–38.0%, the direction channel-stuffing cannot produce.
7. The funding mix is the cleanest measured break from the benign 2010–2020 precedent: capex/OCF peaked at 90.6% then and never crossed 100%; it is 174.1% (Oracle) and 107.2% (Amazon) now, with 31% of cumulative capex issuance-funded.
8. $1,090 billion of signed-but-uncommenced leases is where demand deterioration will surface first, outside every balance sheet, visible only as a shrinking disclosure.
9. Validating the vintage requires external end-user revenue of $475B/yr by 2030, a 93.0% CAGR from the identified $26.3B base (66.8% from a generous $50B). P(shortfall) = 0.70, resting on measured depth: 57% of business users run AI in ≤3 functions at a median $11.38/employee/month.
10. A capability plateau is two-sided for installed capital: it caps demand but kills obsolescence, extending life and cutting the revenue bar 29% (to $336B at an 8-year life), hence P(validated | plateau) = 0.18, not 0.10.
11. Structure and price are independent axes: Oracle is impaired-but-de-rated; Arista (+558bp implied growth) and AMD (82.0× EBITDA, above its own range) are sound-but-demanding; the ruinous quadrant is the fastest-melting asset at the highest funding cost with the vintage at ~37.6% of enterprise value, the neocloud corner.
12. The technology working and the capital earning are independent questions, and history settles them separately: users retained the network, the grid, and the fibre, and the measured surplus-to-revenue gap ($172.3B vs ~$14.2B, flagged unverified) is the mechanism by which value escapes to users again.

Return here every quarter. Base verdict: P(failure share > 30%) = 0.743, at 10 August 2026. Each row shows what moves it, and in which direction. Update pairs map to the thresholds in order.

Not monitorable, because it is not publicly observable, and therefore where the ambushes come from: actual fleet utilization; covenant packages; GPU appraisal and advance rates beyond one issuer's deck; cancellation terms on all five backlogs; churn and retention for any AI product; realized revenue per token.

The thirteen reasoning moves, one line each, the part of this document that transfers to everything. Each links back to where it was earned.

1. Convert the vibe into a random variable, quantity, threshold, date, before arguing.
2. Trace every dollar to its origin, revenue is demand only if the payer stands outside the system.
3. The unit of account decides the answer, define the cohort before measuring its returns.
4. Hurdles attach to owners, not assets, decompose every cycle by who owns what, at what survival threshold.
5. Carry the fork, never average irreconcilable instruments into a world nobody measured.
6. Report the distribution you actually have, if the evidence is bimodal, so is the honest answer.
7. One-sided observability yields bounds, not points, the direction of revision, never its size.
8. Date the tail, define the proxy, derive the bar, and force your probabilities to cohere.
9. Never spend evidence twice, credit each datum on exactly one axis.
10. Commitments are not measurements: “sold out” is language; utilization is a ratio nobody publishes.
11. Rule on the best version; name the defeater, disagreement is earned datum by datum.
12. Score structure and price separately; trade the gap, the off-diagonal quadrants are where the money is.
13. Write the reversal before you need it, dated, quantified tripwires in both directions, drafted while calm.

Plus the meta-rule that governed the entire exercise: flag, quantify, carry both, on finding a defect upstream, never silently fix it and never silently inherit it.

Not predictions, reading habits, each anchored to something this analysis measured.

Beneath all of it lies the frame that outlasts the cycle: whether the technology matters and whether the capital earns are different questions, usually with different answers, settled at different times, the railway network peaked seventy years after the mania that funded it, and someone owned every year of that gap. And when the settlement of this cycle is finally written, expect the oldest clause in the record to appear once more:

Users retained the network. Users retained the grid. Users retained the fibre.