The Other Exponential

GE Vernova ended the second quarter of 2026 with 116 gigawatts of gas turbine backlog and slot reservations, and its management described lead times of around three years. Source neededOne named order with a public delivery date, so that the reader can picture a single machine rather than an aggregate. A turbine ordered this year arrives around 2029 and then runs for decades. What it will power is a stranger question than it sounds. The accelerators that exist today are depreciated over five or six years in the filings of the companies buying them, and they will be well into that schedule before the turbine turns. The turbine is being ordered for chips that have not been designed yet, bought in a year nobody can forecast, to serve demand that has not been contracted. The most durable machine in this story is being committed on the strength of a forecast about its least durable one.

I find that image useful because it contains the whole problem, and because it is not a claim about whether AI works. It is a claim about clocks.

Before going further I owe the reader two things. The first is a disclosure, which belongs here rather than at the end. Disclosure to writeThe author's financial and professional exposure, in two sentences, including holdings in any company discussed and any employment or prospective employment in the sector. Almost nobody writing about this question is disinterested, and it costs two sentences to say so.

The second is an observation about how this subject gets argued.

The people who build AI are unusually good at quantified public argument. They publish model cards and evaluation suites. They write responsible scaling policies with capability thresholds that trigger specified safeguards. They put probabilities on the arrival of transformative systems and argue about them on public forecasting platforms. People debate the probability of catastrophe to two significant figures, and whatever you think of that practice, it is at least an attempt to be precise about something that matters.

Yet the same community discusses roughly $700 billion of capital expenditure this year, most of it physical and much of it on assets with short lives, in the register of vibes and loose analogy. Is it a bubble? It is like 1999. It is nothing like 1999. The sentences are confident, and almost none of them could be scored.

I do not think this is because the money matters less. I think it is a sorting problem. The people who read 10-Ks do not read model cards. The people who read model cards frequently hold equity in the companies whose model cards they read. And the loudest critics of AI's finances are usually also critics of AI's capabilities, so the financial argument tends to arrive bundled with a technological one that a serious reader can dismiss. If you believe the technology is real, as I do, it is easy to hear "this is a bubble" as "this is fake," and to stop listening.

This essay tries to unbundle them. Its thesis is simple to state. AI can be the most economically important technology of the century, and the capital cycle now financing it can still end in a destructive unwind, because the value a technology creates and the returns earned by those who build it first are different quantities. History is full of cases where both were true at once. Whether the technology works and who gets paid are separable questions, and only the second is about money.

So let me say early what this essay does not claim. It does not claim that anyone is committing fraud; it describes magnitudes, not motives, and where it discusses accounting it describes rules that are audited and followed. It does not claim that AI is overhyped, or that the technology will disappoint. It does not even claim that a bust is likely. It claims something narrower and more useful: that the financial question can be defined precisely enough to be answered with a number, that the number can be published before the evidence arrives, and that it can then be scored.

The obvious objection is that markets already do this. Prices aggregate the views of millions of participants with far better data and far more at stake than any essayist, and the price is the forecast. A published probability from outside adds noise, not signal.

I take this seriously, and I think it is half right. A price is an extraordinary aggregator of positions. It is not a statement of reasoning. It does not say what event it is forecasting, by when, or under what definition, and nobody publishes the model behind it. You cannot score a price against a defined outcome, because it was never committed to one. What this project offers is a different product: a stated question, a published method, a number, and a record that can be checked against what happened, including when the number turns out to be wrong.

Most of the disagreement about whether this is a bubble, it turns out, is not disagreement about the world. It is disagreement about the word. So the first job is to define it.

The word bubble is doing at least four jobs. They carry very different probabilities, very different consequences, and very different degrees of observability, and most arguments about whether this is one are people holding different jobs in their hands and talking past each other.

Overvaluation. Prices exceed a defensible estimate of fundamental value. This is common, it usually resolves slowly and undramatically through a decade of mediocre returns, and it is compatible with everyone involved behaving sensibly. It is also the weakest of the four, because the defensible estimate is itself contested.

Bubble proper. Price growth is sustained substantially by the expectation of further price growth rather than by revised fundamentals. This is a claim about mechanism rather than level, and it is close to unobservable from outside. You would need to know what buyers believe and why they believe it, and buyers do not reliably know that themselves.

Malinvestment. Real physical capital is being built that cannot earn its cost of capital. This is the claim that matters for the real economy and it is independent of whether any security is mispriced. A privately held company with no traded stock can malinvest enormously.

Disorderly unwind. The correction happens fast enough, and through leveraged enough channels, that it damages people who never took the bet.

Only the fourth is unambiguously bad for bystanders, and it is the most observable of the four. Not cleanly observable, and I want to be accurate about that, because the institutions which monitor this for a living do not define a disorderly correction by a percentage fall. They define it through transmission: whether leverage, forced selling, funding stress and impaired intermediation turn a repricing into something that damages credit provision and employment. The Federal Reserve's own framework draws the contrast directly, noting that the dot-com correction was very large and did not become a financial crisis, because those amplifiers were not present in the way they later were in housing. That is a dashboard and a supervisory judgement rather than a threshold, which means this project still has to invent and defend its own numbers. But it is the only one of the four claims where the evidence leaves records.

There is a reason the four are nested rather than parallel, and Kindleberger put it better than I can: most expansions of money and credit do not lead to a mania, but every mania has been associated with the expansion of credit. Close to a necessary condition, nowhere near a sufficient one. That asymmetry is why a probability on the fourth claim cannot be an average of probabilities on the other three.

We keep the name anyway. P(Bubble) is the phrase in which the argument is actually conducted, and a name nobody uses is worth nothing. But the name carries an asterisk, and the asterisk does the work.

The asterisk, in full. P(Bubble)* is the probability that the current AI investment cycle ends in a disorderly unwind rather than an orderly absorption, assessed against a named capital deployment series, a named credit threshold, a named equity index and a fixed date. It is not a claim that AI is overhyped, that any particular company is mispriced today, or that the technology will disappoint. The asterisk exists so that the short name can travel while the definition stays fixed, and it is reproduced beside the number on every publication, every time.

Alongside the headline the project publishes four component probabilities, because they differ enormously in likelihood and in consequence and collapsing them is exactly how a forecast gets misread. P(valuation drawdown) asks whether the index falls more than a stated amount peak to trough by the date. P(capex contraction) asks whether aggregate capital deployment falls more than a stated amount year on year. P(credit event) asks whether there is a default, restructuring or redemption gate above a stated size. P(attributable macro damage) asks whether there is a measurable effect on output or employment traceable to the unwind.

They are published separately and they are not averaged into the headline. The four are nested: a disorderly unwind runs through the credit channel, so averaging would let a high probability of an ordinary valuation drawdown raise a number this definition says it should not. The headline is the terminal node of a conditional chain, and the arithmetic is published with it, alongside what a naive weighted average would have produced and why the two differ.

There is prior art, and it makes the case better than I can.

The largest public market on this question opened in November 2025 and had taken about $3.0 million across the event by 20 September 2026, with the 2026 contract at roughly 11%. It resolves yes only if at least three of six specified events occur within ninety days. The six are: Nvidia down 50% from its all-time high; the SOXX semiconductor fund down 40%; OpenAI or Anthropic declaring bankruptcy; OpenAI being acquired; a named vendor's H100 rental index at or below a dollar for five consecutive days; and any of TSM, ASML, Broadcom, Arista or Super Micro down 50%.

Four of those six are share price movements. Two are the fate of individual companies. Not one of them is a credit event, a contraction in capital deployment, or a measurable effect on output or employment.

I want to be careful, because those criteria are well built for what they are. They are unambiguous, they settle from public sources, and the rule that three conditions must be met regardless of what anyone reports about a downturn is exactly the right instinct: the criterion beats the narrative. The problem is not rigour. It is that the question being priced is not the question that matters. Share prices are endogenous to the thing being forecast, and the failure of an individual lab is ordinary competitive selection. This project excludes both explicitly, which means the most visible public probability on the AI bubble is a probability about something else. I have no quarrel with the eleven percent. It is an answer to a different question.

And it is not an isolated case. Searching for any market or institution that has published a settleable AI downturn rule turning on credit or on observed capital deployment turns up nothing at all. The nearest comparisons resolve on an exchange traded fund's price, or on whether management formally cuts forward guidance, which is a statement about announcements rather than about spending. One of the price-based ones had its threshold changed by its author a year after it opened.

What resolves this one.

The equity component is committed to the Nasdaq CTA Artificial Intelligence Index. It is published by someone other than us, its constituent selection is rules-based with a public methodology, it reconstitutes on a fixed semiannual schedule, and its daily history runs back to October 2018, which is long enough to compute a peak to trough drawdown. That matters less for accuracy than for discipline: a project that assembles its own basket can assemble it after seeing which names fell. Two honest caveats. Classification of firms by AI role and intensity involves provider judgement rather than a mechanical screen, and the methodology has been revised four times since 2021, once materially. Naming a ticker does not freeze a definition, so the project pins the methodology version current at publication and logs later changes as amendments.

The capital deployment series is the combined capital expenditure of the four largest builders, defined to include finance lease additions and, where disclosed, build to suit additions. A measure built on reported cash capital expenditure alone tracks procurement structure rather than building. Microsoft moved future leases between classifications while this essay was being written and lowered its reported figure without altering a megawatt of planned capacity.

The credit threshold is a default, restructuring, or redemption gate at a vehicle with disclosed exposure to this sector. Decision openThe size threshold.

Decision openThe date. My proposal is the end of 2029, on the arithmetic in section 3. Two thirds of this capital sits in equipment depreciated over five to six years from a spending peak in 2025 and 2026, and equipment ordered now against three year lead times is delivered by 2029. A date inside that window tests nothing, because the bill has not arrived. A date well beyond it tests some other cycle.

What explicitly does not count. A slow deflation over a decade, which is the overvaluation claim and not this one. Share price movements in either direction. The failure of any individual lab or startup, which is how competitive markets are supposed to work. And benchmark results and capability milestones, which bears stating plainly because it is the whole argument: capability is not the crux, and a forecast that moved on capability news would have misunderstood itself.

The strongest objection is Fama's, and it is not the one I originally attributed to him.

I had him saying that every definition set in advance is arbitrary, and that wrapping an arbitrary definition in a probability lends it authority the reasoning has not earned. He says neither. The objection he does make is harder. He wants a systematic way of identifying a bubble in advance, one that implies a predictable ending, and his empirical claim is that nobody has shown market price declines to be predictable in that way. Calling an episode a bubble after it falls does not establish the category, because the category has to do predictive work. His own summary is that statements about bubbles rest on beliefs rather than reliable evidence.

Notice what that does to my reply. I was going to say that a criterion published in advance and held to afterwards is falsifiable, while a definition that adjusts as events arrive is rhetoric with a number attached. I still believe the second half. Most public commentary on this question moves its definition as it goes, including commentary that agrees with my conclusions, and the one market I found whose creator revised his threshold a year in is a small demonstration of how easily it happens.

But pre-commitment answers a different objection from the one Fama is making. It makes this project auditable. It does not show that the thing being forecast is predictable, and that is exactly what he denies.

So I am not going to rebut him. I am going to narrow the claim. This project does not identify bubbles in Fama's sense and does not assert that one exists now. It forecasts declared consequences: whether capital deployment contracts by a stated amount, whether a credit event of stated size occurs, whether a named index falls by a stated amount, whether output or employment take measurable damage, all by a fixed date. Those are questions about observable events, and they can be scored whether or not the word bubble turns out to mean anything. If Fama is right that the underlying phenomenon is not identifiable in advance, this forecast should score badly, and the record will say so.

That is a weaker position than the one I started with. I think it is the correct one.

With the claim pinned down, the next question is purely quantitative. How large is the bet?

For about twenty years, the most successful technology businesses in the world had an unusual property: capital was not the constraint. Marginal cost was near zero, fixed cost was modest, and returns on invested capital were high enough that the question of whether a particular investment would clear its cost of capital rarely became interesting enough to argue about. You could be wrong about a product and still be fine, because the money at risk was small next to the cash the business threw off.

This cycle is not that. It is concrete, copper, steel, turbines, transformers, switchgear, water, land and a great many electricians, with the lead times that come attached to all of them. The inversion matters for one specific reason, and most of what follows in this essay depends on it, so it is worth going slowly.

Start with where the money goes. United States private data centre construction ran at a seasonally adjusted annual rate of about $75.2 billion in July 2026, up roughly 57% year over year, and that series covers construction work only: labour, materials, engineering and contractor margin, explicitly excluding equipment installed afterwards. The four largest builders have guided to between $700 billion and $725 billion of capital expenditure for 2026. I am not going to divide one by the other, because the first is domestic and structures only while the second is global and includes a good deal that is not AI. What the two establish without needing to be comparable is that the building is a small part of the bill. Nvidia's own investor material puts a gigawatt scale AI facility at $50 billion to $60 billion, of which management attributes roughly $35 billion to its own hardware and systems. That is a supplier describing its own share and should be read as such, but it points the same way. Most of the money is the computers.

That proportion is the argument, because the buildings and the computers die at very different rates.

The filings are unambiguous about the computers. Server and network equipment is depreciated over two to six years at Microsoft, generally six at Alphabet, five to six at Amazon, and five to five and a half at Meta. On the disclosed mix of recent spending, roughly 63% goes to the short-lived bucket. Those weights give a blended accounting life of about eleven years, and eleven years is the wrong number to carry into this section. It answers whether the asset base eventually earns its keep. The split answers how much time the cycle has, and the answer is that two thirds of new spending has to be re-earned on a five to six year clock, while the financing wrapped around it is sized partly against assets with lives measured in decades.

One clarification, because the next two figures look contradictory and are not. That 63% is a flow. The accumulated stock looks different: at their latest year ends, servers and network equipment were about 32% of Amazon's gross property and equipment, an estimated 35% at Alphabet, 42% at Meta and 50% at Microsoft. Balance sheets carry years of buildings, land and construction in progress, and Amazon also owns warehouses and delivery infrastructure. What is being bought now can be dominated by short-lived equipment even where the installed base is not, and the flow is what sets the clock.

Both ends of that schedule are live and moving in opposite directions. Amazon shortened the useful life of a subset of its servers and networking equipment to five years, citing the pace of development in artificial intelligence and machine learning. Microsoft extended the estimated life of data centres and office buildings from fifteen to twenty-five years from fiscal 2027, a change that also moved more future leases from finance into operating classification and lowered its reported capital expenditure expectation without altering the underlying plan. I raise the second not as an accusation, because it is an ordinary and properly disclosed change, but as a warning about the series itself. Capital expenditure figures move for accounting reasons as well as physical ones. The four company total I quoted three paragraphs ago is lower than the one I was working from a week earlier, partly for exactly that reason, and nobody built less.

One more thing about asset life, arriving from an unexpected direction. Lawrence Berkeley National Laboratory's latest projection puts United States data centre electricity use at 192 terawatt hours in 2024, about 4.7% of national consumption, reaching 649 terawatt hours by 2030 in its reference case within a scenario range of 521 to 843. One of the parameters that moves that range is the assumed operating lifetime of AI chips. The national laboratory modelling the electricity grid has the same variable at the centre of its forecast that the accountants have at the centre of theirs.

Now the second variable.

The accelerators arrive quickly. Almost nothing else does. The Department of Energy reports distribution transformer delivery moving from three to six months in 2019 to twelve to thirty months in 2023. The North American Electric Reliability Corporation's 2025 assessment cites industry data putting large transformers at an average of 120 weeks, within a range of 80 to 210. Gas turbines are tight too: GE Vernova's backlog and slot reservations reached 116 gigawatts in the second quarter of 2026, up from 100 in the first, with management describing lead times of around three years. I have to flag something about all of these numbers. There is no neutral series. Even the official publications rest on observations supplied by manufacturers, utilities and consultants, every one of whom benefits from the perception of scarcity. That is the weakest sourcing in this section, it runs in the direction of the section's own argument, and this is exactly the situation in which to say so rather than to cite the official-looking version and move on.

Underneath sits a constraint that money moves slowly. An electrician's apprenticeship runs four or five years at about 2,000 paid training hours a year. That is the official figure, and it is considerably narrower than the shortage claims in circulation, most of which I could not trace to anything.

This produces an asymmetry I think is underappreciated. A cycle governed by lead times is slow to start and slow to stop. An order placed in 2026 for 2029 delivery is already spent in the sense that matters: committed before anyone can know whether the revenue arrived. And deceleration will appear in order books long before it appears in construction spending, because the concrete going down this month was financed by a decision taken three years ago.

The gap between announcement and construction is already visible, and the grid data show it most clearly. Of 429.8 gigawatts of cumulative large load applications on the Texas system with stated energisation years through 2033, 310.6 gigawatts had no studies underway as of March 2026, 93.1 gigawatts was under review, 3.2 gigawatts was approved but not operating, and 5.8 gigawatts was actually observed. Those are not interchangeable quantities, and the headline is the one that travels. Separately, one market intelligence firm counted nearly five gigawatts of data centre capacity under construction for 2026 delivery against up to eleven gigawatts announced with no visible sign of construction. And the Federal Reserve's project level forecast puts 2027 data centre investment anywhere between $360 billion and $930 billion, on an explicit assumption that new plans arrive at between a quarter and twice the recent pace. That is a remarkable range for a central bank to publish, and an honest admission of how little anyone knows about the back half of this.

Now the objection, which is strong, and which I want to put at full strength rather than set up to knock down.

Large numbers are not evidence of anything. Rural electrification was enormously expensive and worth every dollar. The interstate system was enormously expensive and worth every dollar. The fibre laid in the late 1990s ruined the people who financed it and then quietly underwrote twenty years of everything that came afterwards. "This is a great deal of money" is not an argument, and a reader asked to treat it as one is being handled rather than persuaded.

And the historical comparison, which I expected would help, does not. Set the cycles side by side on the best figures available and the British railway mania peaks at about 7.3% of output in 1847, all American highway capital outlay reaches 1.32% in 1958, telecommunications capital expenditure about 1.17% in 2000, shale exploration and production about 0.74% in 2014, and peak Rural Electrification Administration loan authority about 0.16% in 1949. Those numerators are, in order, private capital formation, public capital outlay, an industry estimate, another industry estimate, and lending authority rather than spending. They are not one series and they cannot be made into one. Against them, the current build is either about 0.24% of American output, taking domestic data centre construction, or about 2.2%, taking global guidance from four American firms over American gross domestic product. The second is the figure people quote, and it puts a global numerator over a domestic denominator. On the only domestically scoped measure available, this cycle is smaller than the telecom build it is usually compared to.

So scale settles nothing, and I concede it completely. What the physical layer establishes is not that the number is large. It is that the number is committed years before it can be tested, and that two thirds of new spending sits in equipment which has to earn its cost back before the decade is out.

Which turns the question from rhetoric into arithmetic. If the size of the bet proves nothing by itself, the thing to ask is what would have to be true for the bet to pay.

There is a version of this question that cannot be answered and a version that can.

The version that cannot be answered is whether AI is valuable. I think it will be enormously valuable, and I notice that this belief does almost no work in the argument that follows, which is the point. The version that can be answered is narrower, duller, and much more useful. Given what is being spent, what would revenue have to be, at what margin, sustained for how long, before this capital earns what it cost to raise? That question has an arithmetic answer. It requires four assumptions, all of which can be stated openly and attacked.

The arithmetic is unglamorous. An investment of size C, with a useful life of T years, financed at a cost of capital r, carries an annual charge of roughly C times (1/T + r): you must recover the asset before it dies, and you must pay for the money in the meantime. To cover that charge out of gross profit at a gross margin g, annual revenue has to be about C divided by g, times (1/T + r).

Put the current numbers in. The latest company guidance is approximately $175 billion from Microsoft, $175 billion to $185 billion from Alphabet, $220 billion from Amazon, and $130 billion to $145 billion from Meta. That is $700 billion to $725 billion in 2026, midpoint $712.5 billion, against roughly $409.9 billion actually spent in 2025, an increase of about 74% at the midpoint. The definitions are not perfectly uniform, since Microsoft and Meta include specified finance-lease additions while Amazon guides to cash capital expenditure, but the order of magnitude is not in doubt. Take the midpoint, a five year life and a cost of capital of 9.5%. The annual charge on that one year of building is about $210 billion. At a 50% gross margin it requires roughly $420 billion of revenue every year for five years. At a 40% margin, about $525 billion.

That is one year of spend. The year before it carries its own charge, and so does the year after.

Against that, the revenue that exists today. Nothing here is cleanly measurable, and it is worth being explicit about why. The private labs publish no audited accounts, so their widely reported run rates, roughly $65 billion for Anthropic at the end of July 2026 and roughly $40 billion for OpenAI in August, are company statements relayed by reporters rather than filings, and both annualise a short recent window rather than describing booked revenue. The public builders file, but do not break AI out. Microsoft reported $214.4 billion of Microsoft Cloud revenue in FY2026, Alphabet $24.8 billion of Google Cloud revenue in the second quarter of 2026, Meta $59.4 billion of quarterly advertising revenue, and none of those figures separates the dollars attributable to AI. Amazon is the single exception, with management stating that its AI business had passed a $25 billion annual run rate, which is a management metric rather than a reported segment. Working top down instead, Bain has estimated that the industry will need roughly $2 trillion of annual revenue by 2030 to fund about $500 billion of annual data centre capital expenditure, and projected a shortfall of about $800 billion.

Two honest corrections before anyone draws a conclusion from that comparison.

The first is that lab revenue is the wrong numerator. Most of this capital is being deployed by companies that monetise AI inside advertising, search and cloud contracts rather than by selling tokens, and a fair calculation has to count that. It also cannot, from public reporting.

The second correction cuts against my own argument, so I want to give it properly. Five years is a server life, not a blended one. The filings put server and network equipment at two to six years at Microsoft, generally six at Alphabet, five to six at Amazon and five to five and a half at Meta, while the buildings around them run for decades. Microsoft says roughly two thirds of recent capital spending went to short-lived processors; Alphabet says roughly 60% of technical infrastructure spending was servers. Weighting 63% at 5.75 years against 37% at 20 years gives a blended accounting life of about eleven years, with a defensible range of eight to fourteen. At eleven years, the same midpoint requires about $265 billion of annual revenue at a 50% margin rather than $420 billion. That is a large difference and it points toward comfort, so it should be stated plainly rather than buried.

It should also be read carefully. Eleven years is the right number for asking whether the whole asset base eventually earns its cost of capital. It is the wrong number for asking how much time this cycle has. Two thirds of the spending sits in equipment that must be re-earned every five to six years, and that bucket, not the blend, sets the pace at which revenue has to arrive. A data centre depreciating over thirty years is patient capital. A rack of accelerators is not.

So the shape of the gap survives both corrections, and it is worth being precise about what kind of gap it is. It is not the gap between a real technology and a fake one. It is the gap between the value a technology creates and the value its builders capture.

Those two quantities can diverge enormously, and this cycle contains the exact mechanism that drives them apart: the collapsing price of the thing being sold. Take one fixed series, OpenAI's successive flagship general models. List prices moved from $30 per million input tokens and $60 per million output tokens in March 2023, to $10 and $30 that November, to $5 and $15 in May 2024 and $2.50 and $10 by August, to $2 and $8 in April 2025. Over that window input prices fell 93%, output prices 87%, and an equal blend of the two 89%. That is a price per token across successive products rather than a capability-adjusted index, which makes it an understatement of what a buyer actually gained. If the price of a unit of intelligence falls by an order of magnitude, volume must rise by more than an order of magnitude simply to hold revenue flat. Every dollar of that decline is real surplus, and most of it goes to the buyer.

This is not a malfunction. It is the ordinary behaviour of a competitive market with a capable open weight floor, and it is entirely compatible with the technology being everything its advocates say. It is also the single most plausible route to a world in which AI works beautifully and the people who financed the capacity do not get their money back.

Margins are where this shows up first. Traditional enterprise software earns 70% to 80% gross margins because serving one more customer costs almost nothing. Inference is not like that: every query consumes power, memory bandwidth and silicon time. Microsoft's disclosed cloud gross margin moved from 71% in FY2023 and FY2024 to 69% in FY2025, then 68%, 67%, 66% and 65% across the quarters of FY2026, for a full year margin of 66%. Management attributed the decline to scaling AI infrastructure, growing use of AI features and sales mix, partly offset by efficiency gains. The frontier labs publish no audited gross margin line at all, and the figures that circulate use custom definitions that often exclude training. That is not a small caveat. Nobody outside these companies currently knows the true gross margin on inference, and public filings cannot deliver it.

Which brings us to the question underneath the margin: is the productivity there?

At the level of individual tasks the evidence is strong and getting stronger. The most careful study we have, a staggered rollout across 5,179 customer support agents, found average productivity gains of 14%, concentrated among novice and lower-skilled workers, who improved by 34% (NBER Working Paper 31161, published in the Quarterly Journal of Economics in 2025). Controlled trials in software engineering and management consulting point the same way.

At the level of the economy the evidence is not there yet. In the San Francisco Fed's data vintage of 3 September 2026, utilisation-adjusted total factor productivity fell 0.34% over the four quarters ending in the second quarter of 2026. An earlier vintage showed a small positive figure through the first quarter, and this series is revised, which is a warning about how much weight any single reading can carry. The most cited task-based model projects no more than 0.66% cumulative TFP gain over a decade, and under 0.53% after its adjustment for harder-to-learn tasks (NBER Working Paper 32487, published in Economic Policy in 2025), against private forecasts many times larger.

The fairest reading of that contrast is not that the micro results are wrong. It is that they have not yet aggregated, and there is a serious argument that they should not be expected to: every prior general purpose technology showed a long lag between deployment and measured productivity, for the mundane reason that the gains require firms to reorganise around the tool. There is early evidence pointing the other way within the lag, too. Work from the Dallas Fed finds that the most AI-exposed sectors have grown productivity at 3.7% annualised since the start of 2024 against 1.7% for the rest of the economy, and account for around 40% of total US productivity gains while representing 16% of hours worked, though the authors are careful to call this a correlation.

So the honest position is that the productivity case is unproven rather than refuted. But notice what the financing does with that. Two thirds of this capital sits in equipment with a five to six year life, and that portion cannot wait fifteen years for a J-curve to resolve. The technology is allowed to take a decade. The hardware is not, and neither is the debt against it.

This points at the scenario that deserves more attention than it gets, because it is probably the most likely bad outcome and it requires nobody to have been wrong about anything important. Not collapse. Not fraud. Just late. AI is genuinely very useful, capability keeps improving on trend, adoption keeps widening, and the revenue arrives two or three years after the financing assumed it would. Late is sufficient. Most of the harm described in the rest of this essay follows from delay rather than from failure.

Now the strongest case against everything above, which is stronger than I would like.

The first form is empirical and hard to argue with. Google Cloud revenue grew 82% year over year in the second quarter of 2026. Amazon's disclosed AI run rate passed $25 billion. Anthropic went from roughly $9 billion of annualised revenue at the end of 2025 to $65 billion seven months later. Whatever the measurement caveats, nothing in the history of enterprise software has grown like that, and a gap closing at that rate is a financing problem rather than a solvency problem. If it continues at even a fraction of the current pace, the arithmetic at the top of this section resolves itself and this section will look foolish. I would rather say that plainly than hedge it.

The second form is more interesting and cuts closer to the bone. Suppose the surplus really does accrue to buyers rather than builders. That is an excellent outcome for the world and a poor one for shareholders, and a project that publishes a probability of a bubble had better be careful not to describe a very large welfare gain as a disaster. I take this seriously. My answer is narrow: the damage from an unwind does not travel through disappointed shareholders, who chose the risk and can absorb it. It travels through credit and through employment, and it lands on people who made no bet at all.

Which is why the next question is not how large the gap is, but how it is being financed.

The largest number in this story does not appear on anyone's balance sheet.

Taken from their own filings at their latest disclosed dates, the five biggest American builders carried roughly $1.09 trillion of lease commitments that had not yet commenced: $329.1 billion at Microsoft, $279.0 billion at Meta, $260 billion at Oracle, $137.2 billion at Amazon and $85.2 billion at Alphabet. Moody's reached the same conclusion earlier in the year from the same disclosures, finding in February that uncommenced data centre leases at the five already exceeded their most recent adjusted debt. Two cautions belong beside the total. It is a mixed-date figure rather than a common-date balance, since Oracle's year ends in May and the others report at June. And it should not be added to capital expenditure, purchase commitments or remaining performance obligations to produce a larger headline, because those are different layers of obligation and the filings do not support treating them as comparable.

Nothing improper is happening here. Under the accounting rules a lease liability is recognised when the asset becomes available for use, and a building under construction is not available for use. But it means that the most widely watched measure of how indebted these companies are, the one on the balance sheet, currently excludes obligations larger than the debt it includes.

This section is about that gap, and about a claim I want to state as precisely as I can. Overbuilding does not cause financial crises. Leverage does. Equity losses are absorbed by people who chose the risk, and they are contained, because an equity holder who is wrong simply becomes poorer. Debt losses propagate, because debt has covenants and maturity dates and holders who may themselves be borrowing. The interesting question is therefore not whether some of this capital will be wasted. Some of it certainly will be. It is who is forced to sell, and when.

The funding has been shifting. For most of the last decade this spending came out of operating cash flow, which is the most benign way to fund anything. That has changed quickly, and the offering calendar is more relentless than the headlines suggest. Meta sold $30 billion of senior notes at the end of October 2025 and another $25 billion in May 2026. Alphabet sold $17.5 billion and €6.5 billion in November 2025, then $20 billion and £5.5 billion in February 2026, including a £1 billion note at 6.125% that matures in 2126. Amazon sold $15 billion in November 2025, then $40 billion and €14.5 billion in March 2026, then $25 billion in July. Microsoft issued no registered senior notes and used finance leases instead, which are economically debt-like without being the same instrument: the classification, the remedies and the asset the obligation runs against all differ. A company borrowing for a century to build something with a five year useful life is making an interesting statement about which part of the asset it expects to last.

The leases got shorter, and that turns out to matter more than it sounds. Data centre leases historically ran ten to fifteen years. Newer ones run four to six years to the first renewal, because tenants want the lease to match the life of the hardware inside it rather than the life of the building. Two things follow, pulling in opposite directions. Renewal periods stay out of reported commitments unless exercise is reasonably certain. And a short initial lease makes project debt harder to place with conventional long-duration lenders, because the debt outlives the contracted revenue, which can push the financing toward bespoke private credit. How much of the migration that explains cannot be measured from public filings, and the essay should not pretend otherwise.

Which is where the structures get interesting, and where the popular account is least accurate. CoreWeave's filed agreements do not show a borrower whose terms simply got worse. Its $2.3 billion facility began at SOFR plus 8.75% and was later amended to SOFR plus 9.62%. A second $7.6 billion facility ranged from SOFR plus 6% for investment grade customer contracts to plus 13% for non-investment grade ones. But later contract-backed facilities repriced sharply the other way: the third at SOFR plus 4%, and a non-recourse, investment grade rated $8.5 billion facility at SOFR plus 2.25%. That trajectory matters, because it shows the market learning to substitute the customer's credit for the hardware's liquidation value. When it can do that, the price of the debt collapses. When it cannot, the hardware is the collateral and the price reflects it.

The structures around the hyperscalers are similarly more specific than their summaries. In Meta's Hyperion venture, Blue Owl owns 80% and Meta 20%; Meta took an initial four year operating lease of about $12.3 billion with renewal options out to twenty years, and separately promised a declining residual value payment if it does not renew and the campus value falls below a threshold beginning at $28 billion. That is a residual value exposure, not a guarantee of the venture's debt service. And in the $35 billion vehicle arranged by Apollo and Blackstone, the filed structure is that an Apollo-managed warehouse holds the purchase obligation, Athene guarantees 15% of the outstanding obligation if the warehouse defaults, and a third party reimburses Athene for 35% of what it pays (Athene Q2 2026 filing). Nothing in the public record establishes a $29 billion Broadcom residual value guarantee.

Then, in August 2026, Nvidia did something worth describing carefully, because both the alarmed and the dismissive accounts of it are wrong.

Nvidia disclosed memoranda of understanding with large capital providers to build independent financing platforms intended to mobilise more than $500 billion of third party capital over time, warning that the preliminary arrangements may never become definitive agreements. It said that at its option it may offer limited residual value support for a portion of particular projects, reviewed one at a time. The filing states no percentage.

But the same 10-Q is where the real exposure sits, and it is both larger and narrower than the press version. At 26 July Nvidia carried $36 billion of AI cloud capacity commitments, typically running six years, and had recognised $3.529 billion notional of land, power and shell guarantees as credit derivatives at an immaterial fair value. After quarter end it signed a form of residual value guaranty for the Ohio campus with an aggregate cap across nine phases of $105 billion. That guaranty becomes effective as leases commence, is triggered by specified tenant defaults, covers defined portions of lease, power and transmission loss rather than every tenant obligation, declines with performance, and terminates on certain credit events. It is a payment obligation with a residual value formula attached to real estate and power. It is not a guarantee of the chips.

Jensen Huang's defence of the platform structure is fair as far as it goes. This is not the Lucent model: the capital providers underwrite independently, and Nvidia has not promised the principal of a $500 billion loan book. The equipment is productive and redeployable. But a filed contract with a $105 billion cap is a different object from a memorandum of understanding, and the distinction cuts against the reassuring account as much as the headline number cuts against the alarmed one.

The problem with residual support is not circularity. It is correlation. An obligation to make up the difference between an asset's assumed value and its actual value grows precisely when values disappoint, which is likely to coincide with weak demand for new accelerators and pressure on the guarantor's own earnings. Credit people call this wrong-way risk. It is the property that turns a modest-looking guarantee into a compounding one, and it applies to the Ohio structure even though the collateral there is land and power rather than silicon.

Underneath all of this sits a single assumption about how long equipment is worth something, and we can watch the whole sector revise it. Meta moved most servers from four years to four and a half in 2022 and then to five, adding $693 million to that year's net income. Microsoft went from four years to six in fiscal 2023, adding $3.7 billion to operating income. Alphabet made the same change in 2023, cutting depreciation by $3.9 billion and adding $3.0 billion to net income. Amazon went from five years to six in 2024, adding $2.5 billion to net income. Oracle made two extensions, in fiscal 2023 and fiscal 2025, the second adding $573 million to net income. Meta extended again in 2025, adding $2.59 billion.

Every one of those changes moved in the same direction, and every one increased reported profit. Then Amazon went the other way. In January 2025 it shortened a subset of servers back to five years, citing the pace of development in AI, and the actual effect was $1.4 billion more depreciation and $1.0 billion less net income, roughly double its own initial estimate. It also took $920 million of accelerated depreciation on early retirements in late 2024 and forecast $600 million more. Microsoft has since announced that from fiscal 2027 it will move buildings from fifteen years to twenty-five, saying the operating income benefit is minimal but the lease classification effect larger. A sector that has repeatedly revised this assumption upward has now produced one significant revision downward, from the company running the largest fleet.

Oracle is where all of these threads meet, and it is the reason this section is not hypothetical. S&P downgraded Oracle on 9 July 2026 from BBB to BBB minus, one notch above speculative grade, citing rising business risk and weaker cash flow. Oracle disclosed $260 billion of uncommenced lease commitments at the end of May and reported negative $23.7 billion of free cash flow for fiscal 2026. Its gross debt to EBITDA, computed from the filings, fell from 4.86 times in February to 3.69 times in August as earnings grew, which is progress and still several times the others. This is what the mechanism looks like while it is running: not a collapse, but a company funding long dated commitments before the contracted revenue arrives.

Now the strongest case that none of this is systemically dangerous, which is considerably stronger than the preceding pages imply. The numbers at the centre are reassuring in a way I did not expect when I started. Computed from the filings at the latest quarter, Microsoft sits at 0.21 times gross debt to EBITDA, Alphabet at 0.59, Meta at 0.76 and Amazon at 0.79, with all but Amazon still holding net cash. Oracle at 3.69 is the outlier, and one outlier is not a system. Nor is the paper held the way subprime was held. The Federal Reserve reports that traditional private debt funds lock institutional capital for seven to ten years with no redemption right, and that the redeemable perpetual business development company and interval fund segment amounts to roughly $241 billion of net assets, about 20% of the total. Insurers and pension funds hold long liabilities that genuinely match illiquid assets. There is no deposit-funded maturity transformation at the core of this, and therefore no run in the 2008 sense. The realistic bad outcome is a sharp equity drawdown, defaults inside ring-fenced vehicles, and a capital expenditure pause. That is a business cycle event rather than a financial crisis, and I think it is the more likely one.

So my disagreement is narrower than I first wrote it. Private credit is not slowly runnable. A minority of it is. The Fed reports that redemption requests at several perpetual funds exceeded 5% of net asset value in early 2026 and that most managers applied a 5% quarterly cap. The filings make that concrete: Blackstone's BCRED received requests of about 10% of shares in two successive quarters and fulfilled 5%, while Blue Owl's OCIC reported 21.9% and then 18.8% against 5% offers. Regulators have looked at this and called it manageable: the IMF says the gates are containing the risk, and the Bank of England says the stress so far has had limited financial stability impact while warning about opacity, leverage on leverage and valuation uncertainty. A queue is not a fire sale.

What that leaves is a slower mechanism than the one I originally described, and it is worth naming precisely because it is undramatic. The way this transmits is not a bank run. It is a fund that quietly declines to refinance a marginal project, a lender that reprices a facility on renewal, and a capped redemption queue that takes four quarters to clear instead of one. Those are the things to watch, and they will not look like a crisis while they are happening.

And all of it depends on the creditworthiness of a very small number of counterparties, which is the subject of the next section.

In fiscal 2023, no single customer accounted for as much as a tenth of Nvidia's revenue.

In fiscal 2024 one direct customer reached 13%. In fiscal 2025 the top three clustered at 12%, 11% and 11%. In fiscal 2026 one direct customer was 22% of revenue and another 14%, a combined 36%, or roughly $78 billion of $215.9 billion. In the first half of fiscal 2027 the top three ran at 16%, 15% and 13%, together 44%. The receivables moved faster than the revenue: at 26 July 2026 five direct customers accounted for 70% of accounts receivable, against three customers at 56% six months earlier.

That is a company which went from no significant customer concentration to a great deal of it in three years, and which disclosed every step. This section is about what that means, and about not claiming it means more than it does.

Two qualifications first, because they cut against the easy reading.

Nvidia's direct customers are not necessarily the companies using the chips. The definition covers distributors, original design and equipment manufacturers, cloud providers, model developers and system integrators, some of them assembling for other buyers. Nvidia reports indirect customers separately, estimates that attribution from purchase orders and internal sales data rather than auditing it as a second ledger, and says only that some indirect customers individually exceed 10% without publishing the percentages. The two views overlap and cannot be added. Anyone reading the direct customer table as a list of clouds is reading it wrong. Broadcom makes the point sharply: one semiconductor distributor accounted for 50% of its net revenue in the quarter ended 2 August 2026, which tells you about a sales channel rather than about an end customer.

Second, concentration at a supplier is not evidence of anything improper. When a few firms are building much of the world's new compute, the company selling them accelerators will have concentrated revenue. That is arithmetic. And it is not universal: AMD disclosed no customer above 10% of consolidated revenue in either fiscal 2025 or fiscal 2024. The pattern is real but it is not the weather.

What is worth examining only becomes visible when you separate two time horizons.

Microsoft has disclosed both halves of it. Nearly 90% of fiscal 2026 Microsoft Cloud revenue came from customers outside the frontier model companies, which is a thoroughly diversified book of business. And in January it disclosed that OpenAI represented approximately 45% of its then $625 billion of commercial remaining performance obligations. By June the total had reached $678 billion. Microsoft did not refresh the OpenAI percentage, but it did say that obligations excluding OpenAI grew 25% year over year and that all sequential growth came from outside the frontier model companies, and both of those cut against the reading I am about to offer. Take them at full weight. What remains is that current revenue is spread across the world economy while contracted future revenue carries material exposure to one counterparty whose own future revenue is the question this essay has been asking throughout.

Oracle looks similar and cannot be measured. OpenAI says the partnership exceeds $300 billion over five years; Oracle reported $638 billion of remaining performance obligations at 31 May 2026. If substantially all of the announced amount sits inside that balance, it is at least 47% of the total. Oracle publishes no customer composition of its obligations, so that is a conditional arithmetic comparison and not a disclosed ratio, and it should never be quoted as one.

Further down the stack the cushion thins and, helpfully, the disclosure improves.

CoreWeave reported three customers at 36%, 26% and 10% of revenue in the second quarter of 2026, with two of them at 32% each of accounts receivable; its largest customer in 2025 was Microsoft, at 67%. Applied Digital reported three customers at 59%, 25% and 12% of fiscal 2026 revenue, together 96%. Nscale's preliminary prospectus reports $103.4 billion of active and contracted total contract value at 31 August 2026, including Anthropic agreements providing for payments of up to $44.6 billion, against full year 2025 revenue of $33.0 million, and states that it had not obtained binding commitments for the financing required to perform those agreements at the time of filing. That is not a criticism of Nscale, which disclosed it. It is a description of how thin the layer can be between a signed contract and the capital required to honour it. Note also that total contract value is a company measure rather than remaining performance obligations under the accounting standard. The two get quoted side by side constantly and they should not be.

So the claim is narrow. Not that the demand is fake. That a very large quantity of contracted revenue depends on the performance of a small number of counterparties, several of them private, and that the further down the stack you look, the less stands between that performance and somebody's debt service.

Now the part that sounds like an accusation and is not.

When a chipmaker invests in a customer who then buys chips, people reach for the word circular. The accounting question begins with linkage rather than with the label. Consideration payable to a customer that is linked to the revenue contract normally reduces the transaction price, and therefore revenue, unless it buys a distinct good or service, with any excess over that good's reasonably estimable fair value coming out of revenue. A genuine investment accounted for under the financial instruments guidance may be a separate transaction, but the facts have to support the separation, and an equity stake is not automatically a distinct good or service merely because it was bought at fair value. Where the seller has significant influence and uses the equity method, its proportionate share of profit on sales to the investee is generally eliminated while the asset remains, with narrower circumstances calling for fuller elimination.

I want to be unambiguous about the implication. These rules exist, they are audited, and I found no public enforcement action or standard setter finding alleging that any named company has violated them in these arrangements. The point is smaller and stranger. The correct treatment can turn on linkage, on significant influence, and on valuations of private companies in a market where private values have moved faster than almost anything else in the economy. The accounting may be entirely right. Its most important inputs are not available to an outside reader, and that is a fact about observability rather than about anyone's conduct.

There has been one external examination. The Federal Trade Commission issued orders in January 2024 to Alphabet, Amazon, Anthropic, Microsoft and OpenAI, and published a staff report in January 2025 covering three partnerships. It found terms including significant equity and in some cases revenue sharing rights, multi billion dollar cloud commitments, discounted compute, consultation or exclusivity rights, and exchanges of sensitive information, and it used the phrase circular spending for the required cloud commitments. It identified possible foreclosure, switching cost and information advantage concerns. It also conducted no comprehensive legal or economic analysis and made no accounting finding of any kind. That is the whole of what an outside body has established about these structures, and it is worth having precisely because it is narrow.

The telecom precedent is instructive, and the lesson runs opposite to the one usually drawn.

Lucent, Nortel and Cisco all carried customer financing into the downturn, and all of it was on the face of the filings beforehand. Lucent reported $2.034 billion of drawn customer financing commitments at September 2000 with $3.462 billion more available or committed; a year later the total was $5.314 billion and its financing reserve had risen from $604 million to $2.109 billion after $1.787 billion was charged to expense. Nortel reported $5.168 billion of customer financing at December 2000 and $2.075 billion a year later, with provisions rising from $433 million to $887 million as customers failed borrowing conditions, cut purchases or entered bankruptcy. Cisco's July 2000 balance sheet already showed $1.115 billion of lease receivables and $205 million of structured loans, and its doubtful account provision went from $40 million to $268 million over the following year.

None of that was hidden. It was disclosed, period by period, and the market read it anyway. I had this wrong in the first draft of this section, where I described vendors who lent, booked the sale and then took write-downs, in a construction that implied discovery after the fact. The correction matters for this essay's whole method, because it means the failure in 2001 was not a disclosure failure. The numbers were public and were not acted on.

The differences from today are real and should lead. The carriers of the late 1990s mostly had no end user revenue; the frontier labs have a great deal of it, growing quickly, although their accounts are private and their capacity to pay at contracted scale cannot be tested from outside. The vendors then financed the purchase directly; Nvidia states in its fiscal 2026 annual report that it has been asked to provide financing and has not entered into such arrangements, which leaves its equity positions and commercial guarantees to be judged on their own terms. And the revenue standard is more explicit now about consideration payable to a customer than the rules in force then.

What survives the comparison is narrower and still worth saying. In both cases a large share of a supplier's growth ran through intermediate buyers whose ability to perform depended on end demand that had not yet been demonstrated at the required scale. That was the load bearing fact then. It is the load bearing question now.

What I cannot tell you is how large the circular share is.

I would like to give the fraction of AI revenue paid by entities that received capital from the payee or from someone in the same complex. It cannot be done honestly from public sources, and I want to be specific about why rather than gesture at difficulty. It would require transaction level revenue by ultimate user, the amount and timing of capital transferred, whether that capital was restricted or economically tied to purchases, and how a given dollar moved through clouds, integrators and labs. Public companies disclose investments and aggregate revenue and rarely the matching commercial flows. Private companies disclose less. The Federal Trade Commission obtained non public contracts for three partnerships and did not calculate an ecosystem share. The estimates in circulation are built by summing announced investments and applying an assumed reciprocation rate, which is an assumption wearing a decimal point. A project that publishes probabilities cannot launder an assumption into a finding. So this is one of the most important quantities in the whole question, it is not currently observable, and that sentence is the finding rather than a gap in it.

The strongest case against all of this is straightforward and largely correct.

Where the binding constraint is physical supply, long term offtake agreements and capacity prepayments are the ordinary way buyers secure access and sellers secure visibility. Aerospace, foundry capacity and resource extraction have always worked this way. A customer prepaying for wafer capacity is evidence of scarcity, not of mania. Strategic investment in a customer finances capacity that would otherwise not be built, and moves risk to the party best able to carry it. Sovereign programmes make the same point from another direction. The European Commission's July 2026 call for up to seven AI gigafactories offers up to ten billion euros of European and national support against an expectation of at least twenty billion euros of private investment, and the United Kingdom has committed five hundred million pounds to a sovereign AI effort combining equity, national compute allocation and procurement. A government buying strategic capacity is not trying to earn its cost of capital and will not stop when the commercial case weakens, which puts a floor under demand that a purely private cycle would not have. It also means some share of this capacity is being built to a standard against which no commercial test will ever be run.

I accept most of that. The place where it stops being reassuring is the same as before. It explains why these arrangements exist without telling us whether the end demand that has to service them is there. An offtake agreement transfers risk. It does not create a customer.

One last form of concentration, which nobody signed up for.

The ten largest companies in the S&P 500 represented roughly 40% of the index by the middle of 2025, which its provider says was the highest such level since the mid 1960s. I wanted to say something stronger, that concentration now exceeds the March 2000 peak, and I have taken it out, because no index provider publishes top five and top ten weights for both major American indices spanning that date on a single consistent methodology, and splicing a current factsheet onto a third party reconstruction of March 2000 would produce a precise looking number nobody could reproduce. The narrower claim carries the argument anyway. A capitalisation weighted retirement account now holds unusual exposure to a small group of companies, many of them tied to the same compute supply chain, and tens of millions of households hold that exposure having formed no view and made no decision. That is not a claim about mispricing. It is a claim about who is in the room when the bet is placed, and who is in the room when it settles.

Which is the point at which this argument owes the other side its best case.

I have spent four sections building an argument. This one tries to take it apart, and I want to say plainly that I am not performing the attempt. Two of the arguments below are better than anything I have written so far, and one of them may be decisive.

The buyers are not the 1999 buyers. This is the most basic structural difference and it is easy to underrate. The telecom capacity of the late nineties was laid by competitive carriers financed with high yield debt against almost no operating cash flow, which is why the repricing turned so quickly into defaults. The equivalent capacity now is being built by four of the most profitable enterprises that have ever existed, out of cash flows generated by search, advertising, enterprise software and retail, businesses that have nothing to do with whether the next model generation works. In the twelve months through June 2026, Microsoft, Alphabet, Amazon and Meta generated $660.3 billion of operating cash flow and spent $510.7 billion in cash on property and equipment. Cash capital expenditure therefore consumed 77.3% of operating cash flow, up from 59.1% in the twelve months through June 2025. The cash still covers the build, but the cushion is shrinking. A firm funding an investment from its own profits has a completely different failure mode from one funding it from a bond that matures on a date. The first can stop; the second can default.

Source and method: issuer-filed quarterly cash flow data in the SEC's XBRL Companyfacts datasets for Microsoft, Alphabet, Amazon and Meta. OCF is net cash from operating activities. Capex is cash paid for property and equipment, excluding assets obtained through finance leases and other non-cash financing. Quarters are derived from filed year-to-date figures where a standalone quarter is not reported.

That said, the strongest version of this argument has weakened over the last two years, and the section should say so rather than wait to be caught. For Alphabet, Amazon, Meta, Microsoft and Oracle, FactSet estimates that incremental annual debt went from about 9% of capital expenditure in fiscal 2024 to roughly 32% in the twelve months to mid-2026. The fortress argument is about a level. The interesting question is a direction. CheckSection 5's later audit could not reproduce this FactSet series and deleted it from that section. It should not survive here unexamined. The quarterly table above already demonstrates the direction from filings, so the likely fix is to drop the FactSet sentence and let the table carry the point.

The capacity is full, not idle. Dark fibre is the image that does most of the work in the sceptical case, and it does not fit. Fibre sat unlit for years because it was laid ahead of demand that had not arrived. Accelerators have been rationed since they were deployed. In May 2026 Anthropic said it would use all of xAI's Colossus 1 cluster, more than 220,000 Nvidia GPUs and over 300 megawatts, and simultaneously doubled Claude Code's five hour limits, removed its peak hour reduction and raised Opus API rate limits. Upstream, the constraint shows up where a shortage should show up. TSMC said in April that advanced packaging capacity remained very tight; Micron had price and volume agreements covering its entire calendar 2026 high bandwidth memory supply; and SK hynix said demand exceeded its supply capacity even as HBM4 entered mass shipment. ERCOT's February 2026 queue contained more than 232 gigawatts of proposed large loads, 72% of it data centres, though ERCOT warned that project timing and realisation were uncertain. Those are real constraints. They establish that scarcity exists now. They do not establish that today's build cannot become tomorrow's glut.

Demand has been contracted, not forecast. This is the pillar the sceptical case has the most trouble with, and it is now very large. Oracle finished fiscal 2026 with $638 billion of remaining performance obligations, up 363% year over year. Microsoft reported $678 billion of commercial remaining performance obligations at June 2026, with roughly 30% expected inside twelve months. Alphabet reported $513.9 billion of cloud remaining performance obligations. Amazon reported about $496 billion at 30 June 2026, primarily related to AWS, with a weighted average remaining life of 6.4 years. Source neededDirect link to the Amazon second quarter 2026 filing for this figure. These are signed contracts, not a market sizing exercise. The categories are not interchangeable, and the essay should not pretend otherwise: Microsoft's figure covers commercial contracts broadly, Alphabet changed its definition in the March quarter to include contracts with original terms of a year or less, Amazon's excludes exactly those contracts and describes its balance only as primarily related to AWS, and none of the four publishes a complete split by AI workload or customer. More striking still, Oracle disclosed that the prepaid and customer-supplied hardware portion of its large AI contracts now totals about $75 billion, meaning some customers are funding the equipment up front. It is hard to construct a speculative mania in which the buyers pay in advance.

The valuations are anchored in earnings, not expectations. Nvidia reported fiscal 2026 revenue of $215.9 billion, up 65%, with GAAP gross margin of 71.1% and net income of $120.1 billion, a net margin above 55%. Whatever one thinks of the multiple, the denominator is real and audited. The honest historical distinction, though, is narrower than the one usually drawn. FactSet data published by Nasdaq show that companies with positive margins represented 89.5% of Nasdaq 100 weight in 1999, not a small minority, against 99.9% in November 2025. So the contrast is not profitable now versus loss-making then. It is the price: the index traded at 104 times trailing earnings at the end of 1999 and an estimated 150 to 200 times near the 2000 peak, against the low thirties in November 2025. On a forward basis, a FactSet snapshot at the end of 2025 put Apple and Broadcom at 32.1 times, Amazon at 29.2, Alphabet at 27.9, Microsoft at 27.7, Nvidia at 25.2 and Meta at 22.1. Expensive is not the same as unsupported by earnings.

And now the objection I have rarely seen made properly, which I think is the strongest thing in this section. "Bubble" may be the wrong category. A bubble describes a market in which participants buy because they expect to sell higher. That is not obviously what is happening here. Executives at every one of these firms have described the calculation in the same asymmetric terms: if you overbuild, you own depreciating hardware and some stranded capacity; if you underbuild and the capability arrives, you lose a franchise permanently. Under that payoff structure, heavy spending is not exuberance. It is correct risk management by each individual actor, and it would remain correct even for a decision maker who privately believed the returns were unlikely to clear.

This matters because it changes both the diagnosis and the remedy. A mania is corrected by disillusionment. A rational arms race is not corrected at all until someone concludes the race is over, which can take much longer and end much more abruptly. It also means that the aggregate outcome, overcapacity, can arrive without a single irrational participant, which makes the usual bubble indicators close to useless for detecting it. I do not have a good answer to this argument. The most I can say is that it explains the spending without settling who absorbs the cost of the overcapacity, and that is the question the rest of this essay is about.

What an orderly landing looks like. Because the essay owes the reader a good outcome described as carefully as a bad one, here it is. Capital expenditure growth decelerates rather than reverses: the year over year increase falls from the seventies to the teens and then toward zero, while the absolute level stays high. Utilisation stays strong on the installed base because inference demand keeps growing even as training demand plateaus. The modular shift helps, since a thirty megawatt increment can be deferred in a way a multi-gigawatt campus cannot. Equity in the builders underperforms for two or three years while earnings catch up to the multiple. Second tier operators are acquired rather than liquidated. Debt is refinanced on worse terms without anyone defaulting. The buildings and the substations, which have thirty year lives, get used by whatever comes next. This is not a fantasy scenario. Amazon's cash purchases of property and equipment rose from $40.1 billion in 2020 to $61.1 billion in 2021 and $63.6 billion in 2022, then fell to $52.7 billion in 2023. Free cash flow went from $25.9 billion to negative $14.7 billion, negative $16.9 billion, and then positive $32.2 billion across those four years. S&P upgraded Amazon to AA in June 2021 and still rated it AA stable in November 2025. A company can spend through a cycle like this, absorb two years of negative free cash flow, and emerge with its credit intact. That retrenchment is probably the single most useful precedent for an orderly landing.

The counter, stated once and briefly, because the next section tests it properly. Every one of these arguments was true of some earlier cycle at its peak, and two of them have a specific weakness. Utilisation is a fact about today's price, not tomorrow's. And contracted backlog is only as good as the counterparty behind it. OpenAI says its Oracle partnership exceeds $300 billion over five years; if substantially all of that sits inside Oracle's year end remaining performance obligations, OpenAI is at least 47% of the total, though Oracle publishes no customer reconciliation that would turn the estimate into a fact. Microsoft disclosed in January that OpenAI represented approximately 45% of its then $625 billion commercial remaining performance obligations, and did not update that share when the total reached $678 billion in June. Oracle's free cash flow for fiscal 2026 was negative $23.7 billion. A backlog is not a receivable, which is a point of accounting I had wrong: it is exposure to future counterparty performance. That is weaker language and a stronger statement of the risk.

Which is why the next question is what history says about which of these signals actually discriminate.

The Nasdaq Composite closed at 5,048.62 on 10 March 2000. Thirty-one months later it closed at 1,114.11, down 77.9%. It did not set another record close until 23 April 2015, fifteen years and six weeks later. And on the day it finally did, the S&P 500's technology sector was still 21% below where it had stood in March 2000.

Now hold that against the other thing that happened during those fifteen years, which is that essentially every claim the internet's advocates made about the internet came true. Commerce moved online. Advertising moved online. Communications, media, software distribution, mapping, banking and eventually most of the economy's coordination layer moved online. The technology was not oversold. If anything it was undersold. The capital still lost catastrophically, and both of those sentences are correct without qualification.

That is this essay in one historical example, so it is worth saying slowly. Whether a technology works and whether its first financiers get paid are different questions, and the second is not settled by the first.

Which means the useful thing history offers is not a verdict but a set of variables. The cycles below differed enormously in outcome, and the difference was not technological.

Dot-com. Equity funded and retail exposed. The Survey of Consumer Finances found that 51.9% of American families owned stock directly or indirectly, and household corporate equity holdings fell from $11.52 trillion at the end of 2000 to $7.56 trillion by the third quarter of 2002. Through all of that the commercial banking system stayed profitable: return on assets in 2002 was the highest in more than three decades, and ten commercial banks failed that year holding $2.5 billion of assets between them. I want to be careful about what that does and does not establish. Banks do not report whether a borrower was an internet company, so nobody can calculate the banking system's exposure to the sector, and nobody can say that no individual bank failed because of it. What the record supports is the system-level claim: the banking system was not where the losses landed, and there was no credit crunch.

Telecom and fibre. Debt funded, vendor financed, and built out of real infrastructure with a long useful life. On the Census Bureau's consistent series, capital expenditure by telecommunications services companies ran at $47.7 billion in 1996, peaked at $117.3 billion in 2000, and had fallen 56% from that peak by 2004. At least 63 telecom companies filed for bankruptcy between 2000 and July 2002. The American speculative-grade default rate reached 10.67% in 2001, and in 2002 telecom issuers accounted for 56.4% of defaulted dollar volume globally.

The part worth dwelling on is the vendors. Lucent had $2.034 billion of drawn customer financing at September 2000 with $3.462 billion more available, and its reserve against that financing rose from $604 million to $2.109 billion over the following year. Nortel's provisions went from $433 million to $887 million across the same window. Cisco's doubtful account provision went from $40 million to $268 million. All of it sat in the filings before the collapse rather than emerging afterwards, which is the part of this analogy people reliably get backwards, including me in an earlier draft of another section.

One piece of the telecom story that everybody repeats, including both of the research briefs behind this section, is that 85% to 95% of long-haul fibre sat unlit at the trough. I cannot find a source for it. The survey it is usually attributed to does not appear to exist, and the FCC's own deployment reporting of the period counted lit and unlit fibre together without estimating the split. So I am dropping the number, and with it a tidy contrast I had intended to draw. Buried glass is long-lived but it is not free: it needs route, splice and right-of-way maintenance, and it degrades under crushing, moisture and hydrogen. Idle accelerators are the other way around. They do not rot, but their economic life ends when performance per watt and software support move past them. The clean version, that glass does not degrade and silicon does, is wrong in both directions.

Housing. Leverage, plus maturity transformation, plus the banking system at the centre of the structure rather than at its edge. It is the only one of the three that became a general catastrophe. It is also the one whose underlying technology was least novel. Nobody in 2006 believed that houses were a breakthrough.

And a control case, because the alternative is selecting on the outcome. Here I have to report a failure rather than a finding.

Half of the control case holds up well. Across Amazon, Microsoft and Alphabet between 2010 and 2019, combined cash capital expenditure was $244 billion against $769 billion of combined operating cash flow, and cash flow covered capital expenditure in every single year. That is a real and documented contrast with the telecom carriers. It is also not the whole story, because Amazon used finance leases and build-to-suit arrangements throughout the decade, and in 2014 its capital expenditure, finance leases and build-to-suit additions together exceeded its operating cash flow. The filings do not let anyone assign a dollar of debt to a dollar of cloud capacity.

The other half does not hold up at all. I wrote that the build was repeatedly described as an overbuild while it was happening, and I cannot source that. Searching 2012 through 2018 turns up warnings that cloud was capital intensive and was compressing free cash flow margins, which is a different claim entirely. Nobody I can find said the capacity itself was excessive. So the control case does the financial work it was brought in to do and fails at the rhetorical work, which was the more important job. It was supposed to show that observers had made this mistake before in the other direction. It does not show that, and the section is weaker for it. A comparison built only on cycles that ended badly is selecting on the outcome, and this one is only partly cured.

Line them up anyway and the discriminating variable is not whether the technology worked. In three of the four it worked. The variable is who held the losses, and whether the liabilities funding the build could be withdrawn faster than the assets could be sold.

The empirical literature says the same thing with numbers attached. Greenwood, Hanson, Shleifer and Sørensen examined 42 countries between 1950 and 2016 and estimated how often a financial crisis follows rapid growth in credit and in asset prices together. When neither is elevated, about 7% over the following three years. When business credit growth over three years sits in the top fifth of its historical distribution and stock returns sit in the top third, 45%. The equivalent figure driven by household credit is 37%. Not price growth on its own. The combination.

That is the closest thing this essay has to empirical backing, and it is worth noticing what it does not say. It does not say a large capital cycle is dangerous. It says a large capital cycle financed by rapidly growing credit is dangerous, which is the claim the previous sections have been building toward and the reason the financing question matters more than the size question.

Now the case against all of this, which is strong, and which I will make in the same researchers' numbers.

Forty-five percent is not ninety. More than half of the observations meeting that condition were not followed by a crisis. And in a companion study, Greenwood, Shleifer and You looked for crashes after sharp industry price run-ups in American data from 1926 to 2014. They found forty episodes where an industry doubled both outright and against the market over two years. Twenty-one subsequently fell by at least 40%. Nineteen did not. Slightly worse than a coin flip, which is far above the base rate and nothing remotely like a prediction.

There is a harder problem underneath, and I would rather state it than let it sit unstated in a section about history. There is no base rate for the question this essay actually asks. The crisis paper studies banking crises conditional on credit and price growth. The industry paper studies stock drawdowns after price run-ups. Neither defines a population of capital cycles by investment size, which is the thing being forecast here. The four cycles in this section are not a sample. They were chosen because they are memorable and because they differ in the ways I wanted to illustrate, and a frequency computed from them would be selection bias with a decimal point. History can supply mechanisms and failure channels. It cannot supply a denominator.

There is also a sharper version of the objection, which the evidence supports directly. Which archetype you land on depends on which population you look at. On the quality of new listings the present looks a good deal like 1999: 78% of the 856 offerings in Ritter's filtered sample came to market with negative earnings across 1999 and 2000, against 70% of 1,814 across 2011 to 2025. Two caveats, and they cut different ways. Those are all qualifying offerings rather than technology offerings, and technology was 74% of the earlier sample against 31% of the later one, so the cohorts are not the same animal. But the direction survives both caveats, and it is not a comforting contrast. Meanwhile the firms actually deploying capital in this cycle are not the listing cohort at all. They are among the most profitable companies that have ever existed. Choose the listings and you get the dot-com archetype. Choose the spenders and you get the control case. Anyone who tells you confidently which cycle this one resembles has chosen a population, usually without saying which.

So where does it sit? Not in one place. The centre of this build looks like the control case: very large cash generation, investment-grade debt, no runnable funding. The edges look like telecom: single-tenant contracts, thin equity, borrowed money against equipment with a short life. Nobody can say what share sits at each end, and I tried. Company capital expenditure, leases, special purpose vehicle debt, project finance, private credit and equipment lending use different perimeters and overlap on the same facilities, so adding them produces a number that double counts and means nothing. What history contributes is a view on which of the two will matter more, and on that it is not ambiguous. It will be the edges, because that is where the liabilities are.

And history is unusually clear about one more thing, which the comparison almost always leaves out. It is clear about who ends up paying.

The easiest way to lose this argument is to overstate it, so let me start by conceding the parts that do not hold.

The most widely circulated version of this claim is that living near a data centre has raised household electricity bills by as much as 267%. PolitiFact checked it in June 2026 and found that the figure describes wholesale prices rather than residential bills, which are a different thing and only part of what a household pays. More awkward still for my argument, when researchers at Rutgers compared bills directly they found that ZIP codes with a data centre had monthly residential bills about $2.09 higher than those without, against an average bill of roughly $164. Two dollars a month is not an injustice. Anyone building a case on the claim that data centres have already devastated household energy costs is building on sand.

And yet something real is happening underneath the exaggeration, and it is not primarily about today's bills. It is about who is contractually on the hook for infrastructure being built right now, against demand that has not yet arrived.

Start with what the tariffs actually do, because this is where the argument gets tested properly. In July 2025 the Public Utilities Commission of Ohio approved AEP Ohio's data centre schedule for loads above 25 megawatts. The shorthand is an 85% take-or-pay contract, but the filed schedule is more exact: the minimum is the greater of 85% of the customer's highest demand in the preceding eleven months or a stepped share of contract capacity, capped at 85%; the term is a ramp period of no more than four years plus eight years; a customer may exit only after the fifth post-ramp year, by paying thirty-six months of minimum charges; and a customer lacking both strong credit and ample liquidity must post collateral equal to half of all minimum charges over the term. The order adopted a settlement signed by AEP Ohio, commission staff, the Ohio Consumers' Counsel, the Ohio Energy Group, Ohio Partners for Affordable Energy and Walmart, in preference to a competing settlement backed by developers and trade groups. If you had asked me to design a mechanism to protect households from exactly the risk this essay is about, it would look close to this.

So the honest question is what such a tariff does and does not cover, and there are two answers.

The first is that it converts a stranded asset problem into a counterparty credit problem. A minimum charge is only worth what the signatory and its guarantor can pay. Where the signatory is a trillion dollar company the protection is excellent. Where it is a specialised operator or a project vehicle financed in the private credit market, it is worth what that entity's balance sheet is worth in the state of the world where it is trying to walk away. Collateral makes the protection materially stronger, and Ohio's requirement is substantial. But the tariff moves the residual risk rather than deleting it.

The second is that dedicated infrastructure is not the only infrastructure. Generation, transmission and capacity procured for aggregate regional load growth sit outside the wire and substation charged directly to one site. Virginia is the clean illustration, though not for the reason I first assumed. In its 2025 biennial review the State Corporation Commission created a separate class for customers at or above 25 megawatts, with a fourteen year contract, minimum billing of 85% for transmission and distribution and 60% for generation, an exit fee covering the remaining minimum obligations, and enhanced collateral. In the same order it approved a typical residential increase of $11.24 a month in 2026 and a further $2.36 in 2027, while rejecting much of what the utility had asked for. That $11.24 is an overall base rate outcome. It is not a data centre surcharge and it is not evidence that data centres caused it. The narrower point is the one worth keeping: even a strong large load tariff does not make every system-wide cost disappear.

The national evidence makes that caution run in both directions. A review commissioned by the Data Center Coalition examined eleven quantitative studies and found no quantitative evidence of historical cross-subsidy under existing rate structures. It is industry funded and its state level evidence is largely correlational, which its authors say themselves, and it also concludes that the risk is not absent as load accelerates. I think that is roughly the right reading of the current evidence, and it is not the reading most critics offer.

PJM is the sharpest exception. Its own auction report shows the price rising from $28.92 to $269.92 per megawatt-day between two successive delivery years, alongside 3,243 megawatts of additional forecast load, a higher reserve requirement and 13,252 megawatts less offered supply. The industry funded review itself attributes roughly half the increase to load growth, mostly data centres, and half to supply withdrawal and market rule changes. And the pass-through is not hypothetical: an analysis for Maryland's consumer advocate put the auction change alone at about $16 a month for an average customer in one utility zone, $14 in another, $18 in a third. Those are advocate estimates for particular zones rather than a regional average, and they are an order of magnitude larger than the two dollars this section opened with.

Then there is the finding that changed my mind about how much of this demand is real. It is often said that AEP's interconnection queue fell from 30 gigawatts to 13. That is directionally right and terminologically wrong, and the accurate version is more interesting. Before the tariff, developers had expressed interest in more than 30,000 megawatts, which was non-binding. Once the schedule took effect, they had forty-five days to request and pay for a formal engineering study, and thirty-six sites totalling 13,023 megawatts did so. Once AEP came back with service plans, collateral requirements and binding agreements, 5,642 megawatts signed, on top of 12,219 megawatts contracted before the tariff existed.

Read the three numbers carefully, because each measures a different commitment. The first was an expression of interest. The second bought an engineering study. The third signed a contract and posted collateral. Interest fell by more than four fifths on the way from the first to the third. That is stronger evidence of speculative padding than the queue story it replaces, and it comes with a caveat I could not resolve: this is one utility's funnel, and no other American utility or system operator appears to publish a comparable before-and-after series. One case is a fact, not a law.

The fiscal channel works differently and is less discussed. Loudoun County is simultaneously the strongest evidence that host communities benefit and the clearest case of concentration risk. For fiscal 2026 the county reports $1.2 billion of data centre property tax, equal to 39% of the overall budget; measured against budgeted local revenue alone the share is about 42%. Of that, roughly $795 million comes from taxing the computer equipment itself, which depreciates on a published schedule: formerly 50% of original cost in year one falling to 10% by year five, and from 2026 a slower schedule beginning at 60% and running to a 5% sixth year floor. So the county is not simply levered to the servers, but a meaningful part of it is, and that part depends on the hardware being replaced rather than on the building existing. A pause in refresh spending, the mild scenario, erodes it. The 2026 schedule change happens to slow that erosion.

And now the strongest case on the other side, which is stronger than the critics allow. Loudoun's real property rate fell from $1.145 per $100 of assessed value in 2016 to $0.805 in 2026, a reduction of nearly 30%, while data centre revenue became a central support for the county's schools and services. Two honest qualifications: the rate did not fall every year, and a lower rate does not guarantee a lower bill when assessments rise. A county publication cites an estimate of about $15 of data centre tax revenue for every $1 of county services consumed; a larger figure of $26 circulates but comes from an industry funded economic impact study, and neither is a directly observed cost accounting ratio. The robust conclusion survives anyway, and it is substantial: these facilities occupy little land, send no children to school, and have returned far more local revenue than the county attributes to the services they consume. The claim that data centres impose costs and deliver little local value is untenable in at least one major host county. The open question is not whether the benefit is real. It is what happens to a budget built on it when the refresh cycle slows.

The employment picture is routinely oversold in both directions, including by me. County employment statistics cannot isolate one campus's construction crew from its permanent staff, so the neat project level ratios that circulate are not derivable from public data. What the data do show is worth having. In 2025 the computing infrastructure and data processing category accounted for 2,034 jobs in Loudoun County, 145 in Licking County, Ohio and 20 in Racine County, Wisconsin, against nonresidential construction employment of 1,297, 561 and 288. Those ratios vary by a factor of twenty across three counties, which tells you they are measuring local industrial structure rather than any project. The defensible claim is the qualitative one: a campus creates a large temporary construction surge and a smaller permanent operating workforce. Saying the operating workforce numbers in the tens is not defensible, and I had written exactly that.

The historical precedent that matters is not a data centre at all. Mount Pleasant, Wisconsin issued $120 million of tax increment revenue bonds in 2018 while Racine County borrowed for land acquisition and later refunded it with $110 million of taxable general obligation bonds. The promised factory and its thirteen thousand jobs shrank by an order of magnitude; the obligations did not, and the village still reported roughly $163 million of revenue bonds outstanding at the end of 2025. But the ending is markedly less bleak than the version I first wrote. Microsoft and other development lifted the district's value to about $1.7 billion, the village now projects a surplus, and Foxconn's special assessment was covering $7.3 million of county debt service as recently as the 2024 budget. The oversized water system remains underused, with Foxconn its only connected diversion customer in 2024 at about 20,600 gallons a day, though parts of that network replaced ageing mains and improved regional reliability. The lesson is not that the debt became a disaster. It is that an obligation can outlive the promise used to sell it, and that recovery then depends on finding somebody else who wants the infrastructure.

Which brings me to the combination I find hardest to think about, and here I have to be careful, because the specific version of this claim has never been tested. Nobody has observed an AI capital bust arriving alongside AI driven displacement, so what follows is an inference from adjacent literatures rather than a measured result.

Those literatures are consistent. Across five American recessions, places that lost more jobs suffered persistent relative declines in employment, population and earnings long after the recession ended. Workers displaced in mass layoffs when unemployment exceeded 8% lost, in present value, about 2.8 years of prior earnings over the following two decades, roughly twice the loss suffered by workers displaced into a strong labour market. And western counties caught in the 1970s resource boom saw higher incomes during it and lower per capita income and more unemployment compensation afterwards.

None of that proves anything about AI. It does justify a narrower warning. Displacement and a capital downturn are usually discussed by different people in different arguments, and the literature says the timing is what determines the damage. If they arrive together, the workers displaced will not be made whole by the fact that the buildings remain.

None of this is an argument against building. The infrastructure is largely useful, much of it will outlive the cycle that financed it, and the communities hosting it are not being defrauded. It is an argument that the allocation of the downside is a set of decisions rather than a natural outcome, and that those decisions are being made now, in tariff dockets and development agreements, while the best evidence available shows one utility's demand funnel narrowing from more than 30 gigawatts of stated interest to 5.6 gigawatts of binding contracts once somebody asked for a study, a signature and a deposit.

Which is why the useful thing to publish is not a verdict about what will happen. It is a number that can be checked.

Essays are not falsifiable. This one would not be, if it stopped here. I could write every preceding section, be wrong about all of it, and never find out, because prose has no resolution date and no score.

That is the case for publishing a probability, and it is narrower than it sounds. Section 2 already gave up the strong version: this project does not claim to identify a bubble in the sense Fama demands, and pre-commitment does not make an unpredictable phenomenon predictable. What a number with a fixed resolution criterion does is make the author accountable. It converts a position into something that can be checked against what happened.

So here is the machinery, in enough detail to attack.

The decomposition is the whole aggregation rule. Rather than one intuitive number, a chain of conditional questions. Does the revenue gap close by the date? Conditional on it not closing, does capital deployment contract by more than the threshold? Conditional on that, is there a credit event of the stated size? Conditional on that, is there measurable damage to output or employment? Each node carries its own estimate, its own evidence, and the observability status of the indicators bearing on it, so a reader can see which nodes can be scored on a schedule and which can only be watched. The headline is the terminal node. There is no separate weighting on top, and the four published component probabilities are marginals reported beside it rather than inputs to it.

One problem with that structure is unresolved, and I would rather say so than let a reader find it. The node judgements are not independent. They rest on overlapping evidence, the same handful of disclosures, and one person's reading of them. Multiplying straight down the chain therefore understates the uncertainty around the headline, and I do not have a defensible correction for it.

I am not going to reach for a range as a way out of that, because the institutions that publish intervals earn them. The Congressional Budget Office's eighty percent bands around its long-run projections come out of stochastic simulation with stated sources of variation, not from placing brackets around a central guess. The IPCC goes further and refuses to collapse the problem: it reports quantitative likelihood and qualitative confidence as separate dimensions, and where a single distribution would hide structural uncertainty it publishes storylines instead. Neither of those licenses two intuitive bounds. So the honest position is that this project publishes neither a point nor a range until it can explain how the uncertainty was constructed.

There is no base rate, and I am not going to manufacture one. The original plan for this section was to start from what comparable capital cycles did and then adjust. Section 8 went looking and found that no such base rate exists. The two bodies of work the essay relies on study banking crises conditional on joint credit and asset price growth, and stock drawdowns after industry price run-ups. Neither defines a population of capital cycles by investment size, which is the thing being forecast, and the four historical cases in section 8 are selected examples rather than a sample. A frequency computed from them would be selection bias with a decimal point attached.

What history supplies instead is mechanisms and failure channels: who held the losses, whether the liabilities could run, how fast the assets could be sold. The conditional frequencies enter as warning rates on a named condition rather than as the base of the tree. About 7% of three year windows produced a financial crisis in normal conditions across 42 countries since 1950, and about 45% when three year growth in nonfinancial business credit sat in a country's top historical quintile and three year stock returns sat in its top tercile. Those thresholds are the number's entire meaning, which is why they are stated rather than paraphrased as a high credit environment.

The indicators and their thresholds are in the next section and the appendix. Two lists, not one, because part of what would raise the number cannot be observed on any schedule. And a disclosure that belongs here rather than buried: about ten of the numeric thresholds are first-pass judgements that have not been checked against how often each series has actually moved that far. They are marked as such and they get calibrated before the first publication rather than after the first crossing.

Scoring, cadence and the changelog. Quarterly updates, and a public log of every change to the number, the criterion or the indicator set, with its date and reasoning.

On scoring I have to separate two things I previously ran together. A single long-horizon question can be resolved and scored, and it has been: a public forecasting platform opened a five year question in November 2017, closed it in 2021, resolved it in January 2023 and retains the scores. So a score for this project's headline is possible. But a proper score on one binary outcome is a loss value, not a calibration estimate. A twenty percent forecast that resolves no scores well, and that tells you nothing about whether twenty percent events happen a fifth of the time. Calibration is a property of repeated forecasts, and the platform in question says exactly that about its own single-question scores.

I had intended to answer that by pointing at the four components and the indicator crossings, which resolve more often. That answer is weaker than it looks. The components can share evidence and resolve through the same underlying event, which reproduces the dependence problem at the scoring stage rather than escaping it. So the honest claim is accountability, not demonstrated skill. This project can be checked. It cannot yet show that being checked will mean much.

Adversarial review and conflict disclosure. A named author for the bull case and a named author for the bear case, both published, neither editing the other. And the authors' exposure, financial and professional, stated at the top rather than the bottom. Almost nobody writing about this is disinterested, myself included, and saying so is cheap and clarifying.

The precedents are instructive, and the best of them is a warning.

Weather services have published probabilities of precipitation for decades and verify them with proper scoring rules and reliability diagrams. The IPCC adopted a formal error protocol in 2011, so that corrections are a procedure rather than an embarrassment. Forecasting platforms publish explicit criteria and timestamped forecasts, and also clarify, reinterpret and occasionally annul questions, which is a more honest model than pretending a criterion never needs editing. All three rest on repetition, ensembles or large assessment teams. None of them resembles one author assigning one probability to a unique capital cycle.

The central bank fan chart is the closest precedent and its history is the most useful thing in this section. The Bank of England began publishing forecasts as probability distributions in 1996 and has been evaluated repeatedly since. The results do not tell a tidy story: one study found that realised inflation landed in the central portion of the distributions more often than it should have, meaning the bands were too wide, while a later internal evaluation found that after 2007 they understated the probability of very weak growth and very high inflation. Too wide in the middle, too narrow in the tails that mattered.

Then, in April 2024, an external review recommended eliminating the fan charts as published in the Monetary Policy Report. Its objections are worth reading closely, because three of the four are objections to this project. The distributions mixed model output, historical error and committee judgement without a clear statistical interpretation. The judgemental adjustments to width and skew had little explicit quantitative basis, so the precision was more apparent than real. And the charts treated forecast variables separately, failing to preserve the dependence between outcomes that readers actually care about.

That last one is my unresolved problem, identified by an external reviewer examining the most serious institutional attempt at this. I take it as instruction rather than as discouragement. Note also what the review recommended in place of the charts: not silence about uncertainty, but discussion of historical forecast errors, analysis of recent misses, qualitative risks and scenarios. And note what happened next. The Bank accepted the recommendations, stopped using chart skew to signal its balance of risks, and kept a stripped back chart built from historical errors, which it was still publishing in July 2026. A recommendation, partially implemented, not an abolition.

So the objection is strong and it is the one I would make. A single number launders enormous model uncertainty. The decimal point performs a confidence the reasoning cannot support, readers quote the figure and never open the reasoning, and a project like this converts a set of contestable judgements into something that looks like a measurement.

I do not have a clean answer. What I have is a design choice, and it is the same one the review reached for: the number is not the output, the model is. The full tree, the node estimates and their derivations, the sensitivity analysis, the formal indicator definitions with their thresholds and their calibration status, the historical misses and the changelog all publish at the same moment as the essay, in an appendix nobody has to read but anybody can attack. The detail that makes a method credible is exactly the detail that kills prose, and separating them means neither has to compromise.

And the limitation, stated plainly rather than at the end in smaller type. A probability is not a prediction. It is not investment advice, and anybody trading on a number from an essay deserves what happens next. It is a compressed summary of a model, which is why the model travels with it.

Which leaves one thing to say before the number itself: what would change it.

A probability with no stated update rules is an opinion with a decimal point attached. So here is the list, written before the data arrives, in the expectation of being held to it.

Things that would raise the number. A rising share of the build funded by debt rather than operating cash flow. A rising share of revenue paid by counterparties inside the same investment complex. Falling revenue per unit of deployed compute, or utilisation declining from current levels. Covenant loosening or advance rate increases in compute-backed lending. Announced capacity deferrals, lease renegotiations or power contract restructurings. Redemption gating at any private credit vehicle with material exposure to this sector. Rising customer and receivables concentration at the suppliers.

Things that would lower it. Revenue growth sustained above capital expenditure growth for several consecutive quarters. Margin stability under price competition, which would indicate differentiation rather than commoditisation. Broad productivity gains appearing in official statistics rather than only in task-level studies. Buyers diversifying well beyond the current concentrated set. Financing mix shifting back toward equity and internally generated cash. Secondary market prices for prior generation accelerators holding up, which would validate the residual value assumptions underneath the lending.

Things that explicitly do not count. Share price movements in either direction, which are endogenous to the thing being forecast. The failure of any individual lab or startup, which is ordinary competitive selection. And benchmark results and capability milestones, which bears repeating because it is the essay's last analytical move: capability is not the crux, and a project that let capability news move its financial estimate would have misunderstood its own thesis.

That is the conventional version of this section. Writing the list against the actual public record produced something more useful, and it produced it by proving me wrong, which is worth describing because it is exactly the situation this section exists to handle.

What I got wrong, and what remains true.

I wrote that four of the seven indicators which would raise the number cannot be observed at all. Two of those four can be.

Secondary prices for prior generation accelerators, which I had down as the clearest example of something nobody publishes, are now published. There is a free index of executed sale prices by model, with monthly history running back to mid 2023 and a rolling median. That indicator moves from the unobservable list to the scoreable one, and it matters more than most, because it is the direct test of the residual value assumptions sitting underneath the lending. It comes with real caveats. The sample skews toward retail rather than institutional liquidations, thin models are sparse, and the publisher, while it holds no positions, benefits from its own research being adopted. Another source that publishes resale values buys and resells the equipment itself. But a flawed observation with a stated method is a different object from no observation, and I had asserted the second.

Utilisation is the intermediate case. There is now a directly measured reading drawn from telemetry across tens of thousands of customer clusters, which is a great deal better than nothing. It is also a single annual observation from a vendor that sells the optimisation software its low utilisation finding makes attractive, and it excludes the AI labs entirely. That is evidence. It is not a sector benchmark, and the difference between the two is precisely what a framework like this has to be able to state.

Two of the four survive. There is no series at any cadence for advance rates, haircuts or covenant terms in compute-backed lending: those are bilateral, and they surface only when a public filer happens to be party to a material agreement and files the exhibit. And revenue per unit of deployed compute cannot be constructed at all, because it needs both an operational accelerator count and the split between internal and external workloads, and nobody discloses either. The commercial products that look like this measure are models built from rental prices and assumptions, not observations of the ratio.

So the dramatic version of this section was wrong and the underlying point survives intact. The indicators that are well observed are almost entirely the ones describing the centre of the build. Financing mix, lease additions, uncommenced commitments, useful life changes, customer concentration, backlog, index level credit spreads and construction spending are all free, mostly quarterly, mostly permanent. The project can score the balance sheets of the largest and safest participants in near real time, and it still cannot see the terms on which the borrowers at the edge are lent to. That asymmetry is the essay's own thesis appearing as a data problem, and it did not depend on the count of four.

What follows from it is that the published list has to be two lists: indicators the project can score on a schedule, and indicators it can only watch and report when something surfaces. Publishing them as one would imply monitoring that does not exist.

Two amendments, both forced by this project's own errors.

The depreciation indicator originally said "lengthening depreciation schedules," which is too crude to survive contact with the filings. It has to separate the compute layer from the shell, and it has to catch capital expenditure figures that move for accounting reasons rather than physical ones. Microsoft extended building lives and, in the same change, moved future leases from finance to operating classification, which lowered reported capital expenditure without altering a megawatt of planned capacity. An outside reader cannot cleanly separate those effects, and no rule requires a reconciliation. So the indicator becomes a rule about measurement rather than about the world: any change to useful lives or lease classification at a major builder is logged with its disclosed effect on income and on reported capital expenditure, and treated as a change in the instrument rather than in the thing measured.

The second amendment adds receivables concentration at the major suppliers, which was not on the list and should have been. Nvidia's went from three customers at 56% of accounts receivable to five at 70% across six months. It is a credit exposure measure rather than a demand measure, which is why it belongs. One correction to how I first wrote it: revenue concentration has a fixed disclosure threshold, at ten percent of consolidated revenue from a single external customer, and receivables concentration has no equivalent bright line. It appears when an issuer judges it significant. So the two are not symmetrical, and the receivables indicator has to be written as a record of what each issuer chose to disclose rather than as a threshold crossing.

And a rule for every figure the project publishes. Pin the vintage. Some of these series are revised, and the archiving is worse than I assumed. The productivity series has a free public vintage archive, but only back to 2016, not to the start of its history. The corporate credit spread index, which I described as never revised, carries a publisher warning that observations may be revised, and the free window has narrowed to a rolling three years, with the full history licensed. Construction spending revises the prior two months every release and, once a year, restates more than seven years of it. A revision is not the same event as a change in the world, and where no publisher keeps vintages the project has to keep its own, from the first publication rather than from the first time it matters.

Now the objection, which is serious.

Pre-registration invites gaming and rigidity. The gaming version does not require anyone to act in bad faith, which is what makes it strong. A threshold defined on reported capital expenditure can be satisfied by moving procurement from purchase to lease, an ordinary and properly disclosed decision that leaves the same buildings going up while the indicator reads better. That is not hypothetical. It happened while this essay was being written and it moved the headline number the essay uses. Any capital expenditure indicator has to be defined on a basis that includes leases, or it is measuring accounting.

The rigidity version is equally real. A pre-registered list is built from the last crisis, and the world will produce a development that appears on no indicator here.

The answer is not to eliminate discretion, which is impossible, but to make it visible, and the precedents are more modest than I would like. The Bank of England's Financial Policy Committee publishes a named core indicator set with downloadable series, treats it as an input to judgement rather than a trigger, and dates a revision to that set in its policy statement while explicitly reserving the right to revise again. That is a real amendment record, although it is not a cell by cell audit log, and I looked for an American equivalent and did not find one. The forecasting platforms are instructive in a different way: their resolution criteria are published and timestamped, and they are also not immutable, because administrators clarify ambiguous questions and occasionally annul them. The honest standard available is a public rule and visible intervention, not a promise that nothing will ever be edited. An indicator list that never changes is not disciplined, it is abandoned. One that changes without a record is rhetoric. This one has already changed twice, both times because the project was wrong, and both changes are in the log.

Having said what would change our minds, it is worth saying plainly why any of it matters.

The easiest thing to forget, after this many words about capital structure, is that the technology at the centre of it is extraordinary. I think it will be one of the most beneficial things people have ever built, and nothing in this essay depends on believing otherwise. That is the point. A bust would not be evidence that the technology failed. It would be evidence that it was financed on assumptions about timing that the technology did not keep, and the people who would pay for that mistake are, for the most part, not the people who made it.

That allocation is being decided now. Not in markets, and not at the moment of any unwind, but in contracts being signed this year: in the minimum charges of a utility tariff, the renewal terms of a four year lease, the residual value formula in a guaranty, the redemption cap on a credit fund, the collateral a county did or did not ask for before it rezoned the land. Each of those is a small, specific decision about who holds the loss if the revenue arrives late. Taken together they are the answer to the question this essay asks, and they are mostly written before anyone can know which way it will go.

So here is the number.

Headline openP(Bubble)* and the four component probabilities, published with the asterisk in full beside them, the resolution date, the model in the appendix, and the statement the aggregation analysis requires: that the node estimates share an evidence base and an assessor, that this affects the result, and that its direction is not identified. Under section 10's own standard this cannot be filled until the uncertainty construction can be explained, and the terminus of the chain, node 3 or node 4, is still undecided.

It is not a prediction, and it is not advice. It is a compressed summary of a model that anyone can open, attack and check, and it will be updated as the indicators move and scored when the date arrives. I expect parts of it to be wrong. I would rather be wrong in a way that can be demonstrated than right in a way that cannot, and I would be genuinely glad to be wrong on the high side.

Go back to the turbine. It was ordered this year and it will be delivered in about three, close to the end of 2029, which is when this forecast resolves. Date not ratified By then it will be clear whether the revenue closed the gap, whether the building slowed in an orderly way, and whether anyone who never took the bet ended up paying for it. The turbine will be there either way. It is the most durable thing in the story, and it will outlast every chip it powers and every forecast about them, including this one.

The only question it cannot answer is who owns it on the day the forecast resolves. That question is being answered now.