This is a real bubble, built on capex, concentration and circular financing, but funded at its core from real cash flow. That split is the whole story. The equity de-rate is the leg the market can already price, while the second-order credit event and the third-order growth shock are the legs it cannot. Low-ish odds, high magnitude, and convex.
| Part | Section | Page |
|---|---|---|
| — | The Shock · the activation condition | 03 |
| — | Executive scenario summary | 04 |
| — | Scenario Quick Read | 05 |
| I · The Setup | §01 Base rate & the reference class · §02 Analogues · §03 Why now · §04 Where the froth sits | 06–09 |
| II · The Transmission | §05 The chain · §06 The equity de-rate · §07 The credit stack · §08 The circular web · §09 Capex as GDP · §10 Concentration & the passive bid | 10–15 |
| III · The Impact Map | §11 By asset · §12 By sector · §13 By geography & stakeholder | 16–18 |
| IV · The Dynamics | §14 Reflexivity · §15 The branches · §16 The policy-reaction function | 19–21 |
| V · The Playbook | §17 Signposts · §18 Positioning & the convex trade | 22–23 |
| — | Pre-mortem · Probability & assumptions · Bottom line | 24–26 |
| — | Methodology · Sources · Glossary & Exhibits · Disclosures | 27–30 |
Gravitywell Research publishes independent research for professional investors across public and private markets, both institutions and sophisticated individuals, and for policymakers, not regulated ratings or investment advice. The What-If series studies defined shocks conditionally: we state the trigger, weigh it against base rates, trace the mechanism, and end in positioning. We source to primary filings and official statistics, disclose coverage and confidence, and mark every figure to its date.
Currency: USD throughout. Markets as of Jul 2026; company figures to the latest reported quarter.1 "The AI-cap complex" = the AI-capitalisation-weighted equity basket (merchant silicon, hyperscalers, AI-power, memory, networking, neoclouds). Estimates are ranges, not points. Confidence tiers: ● filed/official · ◐ modeled/derived · ○ Gravitywell estimate. Magnitude and probability are separate axes throughout.
Every effect in this dossier comes with a mechanism, a dated and tiered magnitude, an order in the chain (first, second or third), and a source. As you read, keep two questions apart: how big the burst would be, and how likely it is to happen. They are separate, so a large drawdown is never evidence that the event is probable. And this is a conditional study of a defined trigger, a careful look at what would follow if the burst came, rather than a prediction that it will, or any suggestion that the AI build lacks a future.
Almost everyone now agrees that an AI bubble exists, yet almost no one has priced the burst, and that gap is what this dossier is about. Calling something a bubble is a claim about valuation, but a burst is a claim about a mechanism and a date, and only the second is something an investor can actually position around. So instead of treating the shock as a mood, we define it as an observable regime that the market either enters or does not.12
The froth (the AI-cap complex's leadership layer: Nvidia-led merchant silicon, AI-power, memory and the neoclouds) falls ≥40% from its peak and holds ≥40% below peak for 60+ trading sessions; aggregate top-tier AI/datacentre capex is cut ≥25% YoY (the hyperscalers plus Oracle, xAI and the neoclouds pull planned AI spend — a buildout freeze, not a pause); and a marquee credit event lands in the AI-infrastructure financing stack — a neocloud or GPU-lease default, a GPU-backed-ABS mark-down, or a failed marquee-lab raise.
We use three tests rather than one because any single test would be too weak on its own. A 40% fall by itself is just a correction that the dip-buyers reverse, a capex freeze by itself is a growth scare, and one default by itself is idiosyncratic. It is only when all three line up, with the equity signal, the real-economy signal and the financing signal pointing the same way, that the scenario becomes the self-reinforcing regime we are studying. And if that regime cannot be confirmed from capex guidance, Nvidia's data-centre revenue, the equity tape and AI-infra credit spreads, then it falls outside our scope.
The odds of this burst are low-to-moderate, the magnitude if it comes is large, and the two facts are independent of each other. We put the chance of activation at 20–30% over 24 months. If it does activate, the moderate branch takes the frothy leaders down −50 to −65%, which works out to about −30 to −40% across the wider cap-weighted complex once the cash-rich core is folded in, and it pulls the S&P 500 down −20 to −26%.14 What makes the scenario worth acting on is the asymmetry between those two facts. Because the market prices a disorderly burst as though it were near-impossible, insurance on the credit-and-macro leg is cheap relative to what it would pay if the burst arrived.
The first-order re-rating is already largely in the price. Merchant silicon and the AI-power names fall, their multiples compress, and the momentum trade unwinds. All of this is visible on any screen, so owning it gives an investor no real edge.
What the screen cannot show is the second order. A GPU and neocloud credit stack, much of it vendor-financed and wired together by circular Nvidia, OpenAI and Oracle commitments, can turn a demand air-pocket into an outright credit event two or three quarters later. There is a third order on top of that, because AI capex has been carrying an outsized share of both US growth and index earnings, so a freeze reaches well beyond the stock market.6
The trade this points to is to own convexity on that second-order leg rather than to trade the fear. Long-dated put spreads on the merchant-silicon and AI-power names, together with protection on the AI-infra credit stack, are cheap for now because the buy-the-dip bid keeps realised volatility low. You can fund them by staying invested in the non-AI market that the burst would leave standing. At roughly a one-in-four probability, that insurance looks underpriced against a payoff that could reach −50% or more.
| Stakeholder | Do now (cheap insurance) | If it activates |
|---|---|---|
| Allocator | Measure the hidden AI beta in passive holdings; buy the suppressed-vol put spread; rotate to equal-weight. | Add gold + long-vol; short merchant silicon and AI-power; underweight DC-exposed credit. |
| Corporate / risk | Stress-test AI-linked revenue and any GPU-lease or PPA commitments; term-out funding while spreads are tight. | Pull discretionary AI capex; renegotiate take-or-pay; preserve liquidity. |
| Policymaker | Map the DC-credit and private-credit exposure; watch the growth contribution from one sector. | Sequence the response: liquidity to funding markets before a broad rate cut; ring-fence contagion. |
A −40%-plus de-rate of the frothy AI leadership is roughly a one-in-four event over two years, even though the market treats it as near-impossible. The part that is really mispriced sits underneath the fall, in the credit-and-growth leg that almost no one is watching.
Before the mechanism, the conditions. How often a named tech complex actually bursts, what this buildout rhymes with, why the pressure is live now, and the question that decides who gets hurt: whether the froth sits at the cash-rich core or the debt-financed edge.
Strip the AI story and ask only the outside-view question: how often has a named, concentrated tech or capex complex de-rated ≥40% and stayed there? The answer disciplines both sides — it is far from the bulls' "this time is permanent" and equally far from the bears' "it must crash."14 A large drawdown is a fact about the reference class before it is a fact about AI.
The outside view makes a hard burst a minority outcome even from inside a bubble, and the error is symmetric: the bull who prices it at zero and the bear who prices it at certainty are both ignoring the reference class. The number to hold is a band around one-in-four over two years — high enough to insure, low enough that a crash call is unjustified.
If the reference class sets the odds, the closest analogue sets the shape of the burst. And the best analogue here is not the dot-com froth of profitless websites. It is the telecom and fiber capex boom that ran right alongside it, where the infrastructure was real, useful and eventually used, and yet the capital that financed it ahead of demand was destroyed anyway.11
| Episode | What rhymes with AI now | What's different this time |
|---|---|---|
| Telecom / fiber 2000–02 | Real infrastructure over-built ahead of demand; vendor financing (Lucent, Nortel funded the carriers that bought their gear); ~85–95% of laid fiber sat "dark" for years. | Hyperscalers earn real cash flow today; compute demand is growing now, not just promised. The overbuild question is timing, not existence. |
| Railway mania 1840s / 1873 | A genuinely transformative network built faster than the revenue to fill it; the rails outlived the companies that laid them. | Depreciation is far faster: a GPU's useful life is measured in years, not the decades of steel rail — the write-down clock runs quicker. |
| Dot-com equities 2000 | Narrative-driven multiples, retail participation, concentration in a handful of names carrying the index. | The AI leaders are the most profitable companies on earth, not cash-burning start-ups. The froth is in the periphery, not the core. |
| Crypto / FTX 2021–22 | Circular financing — assets pledged as collateral for the entities that created them; a reflexive web where one default propagates. | The collateral here (GPUs, data centres) has real utility and a resale market, however soft in a glut. |
| Japan 1989 | "It can only go up" concentration; a market whose weight became a self-justifying flow. | Earnings are real and global, not a domestic land-and-equity cross-holding loop. |
Real but overbuilt infrastructure destroys the capital of whoever financed it ahead of demand, even when the technology goes on to win outright. The fiber laid in the boom was eventually lit and now carries the internet, but the carriers and their financiers were wiped out a decade before that payoff arrived. The risk for AI is much the same. It is not that the technology fails, but that the demand curve shows up two years after the debt has come due.
An analogue only rhymes if its preconditions are present now, and here three of them are, all at once. Enterprise returns on AI spending are still hard to find, the index has never been more concentrated in a single theme, and the financing has turned circular, so the same dollars now show up as revenue in more than one place.3
There is a genuine land-grab logic to it. If compute is the scarce input to the next platform, then over-building now can be entirely rational for a hyperscaler that funds it from cash flow and can afford to be early.
But that logic only holds for players funding the build internally. The neocloud and marquee-lab layer funds it with debt and equity raised against demand that has not arrived, and it cannot wait two years for the returns.
No single precondition is a trigger on its own. Together they leave the buildout with no margin of safety if enterprise returns disappoint, because concentration has removed the diversification, circular financing has removed independent price discovery, and the revenue gap has removed the fundamental floor. The system is primed rather than lit, which is the condition in which a routine guidance cut can cascade.
With the preconditions primed, the next question is the one that decides who actually gets hurt: is this a bubble at the cash-rich core, or at the debt-financed edge? That distinction runs through everything downstream. The hyperscalers trade richly but earn enormous cash flow, so the real froth, the part with no earnings floor beneath it, sits one layer further out.1
So the burst is first a re-rating of the externally-funded periphery, and not obviously a solvency event at the internally-funded core, and that single structural fact routes the rest of the report. The credit risk sits at the debt-financed edge, which is the second order, while the growth risk sits in the core's own capex line, which is the third, and the equity de-rate runs through both. The setup is loaded, and because the trigger is a capex guide, the capex channel is where the cascade begins.
How a capex air-pocket becomes a market event — channel by channel, first order to third. The equity de-rate is fast and visible and the tape prices it in hours. The credit event a few quarters later and the growth shock underneath it are the legs the market cannot see until they move.
The transmission is a mechanism you can trace explicitly, channel by channel. A capex air-pocket runs down three of them at three very different speeds. Equity multiples reprice in hours, the AI-infrastructure credit stack reprices over quarters, and the hit to growth and earnings plays out across a year or more. Of the three, the market prices only the first with any real conviction.1
| Node | Mechanism | Priced? |
|---|---|---|
| Equity de-rate | Multiple compression as the growth narrative breaks. | half-priced |
| Concentration unwind | Cap-weighted funds sell the complex proportionally. | half-priced |
| Infra credit event | GPU-lease / neocloud default; collateral value falls in the glut. | unpriced |
| Circular unwind | One broken link re-rates the whole interlocked loop. | unpriced |
| Growth & earnings | Capex was carrying US growth and index EPS; a freeze cuts both. | unpriced |
The two rows that matter are the unpriced ones. Everyone can see the multiple compress, but almost no one is positioned for the credit event and the growth shock that follow it, because both lag the equity move by a quarter or more and neither shows up on an equity screen. So the pages ahead walk the chain in order: the priced equity leg first, to size what the tape already holds, and then the two unpriced rows where the real mispricing sits.
Start with the fast, visible leg. The equity de-rate is largely arithmetic. Multiples that were set on the buildout continuing simply revert once it pauses, and the damage to any given name is proportional to how much of its forward revenue rests on a handful of buyers.5 The most concentrated exposure is also the most crowded, because a single chipmaker now carries around 8% of the S&P 500, the heaviest one-stock weight since records began in 1981.2
The hyperscalers earn the cash flow, so their de-rate is really the loss of a growth premium, which is painful but bounded. Merchant silicon, AI-power and memory have no such floor beneath them. Their multiples price the slope of the buildout, so a freeze compresses the multiple and the forward estimate at the same time.
This leg is already on every screen and in every risk model. It is a re-rating rather than a solvency event, and a diversified index holder can ride it out. The trouble is that owning or hedging only the equity de-rate leaves the two legs that actually surprise, credit and growth, completely unhedged.
The equity de-rate is the part the market has already rehearsed, which is why owning it gives no advantage. Price it as arithmetic, roughly −40 to −65% at the frothy periphery and −15 to −45% at the cash-rich core in the deepest branch, and then move on to the legs the screen cannot show. The question worth asking is not how far the silicon falls, but what its fall does to the debt secured against it.
That question, what the silicon's fall does to the debt secured against it, is the second order, and it is a mechanism the equity screen simply cannot show. A fast-growing AI-infrastructure credit market has been built on the assumption that the buildout, and the value of the GPUs behind it, keep rising.9 It runs through neocloud borrowings, GPU-collateralised loans and off-balance-sheet data-centre vehicles. One marquee neocloud already carries more than $20bn of debt at a blended cost near 11%, secured against chips whose resale value is estimated at only about 45% of new by their third year.9
The collateral is procyclical, and that is really the whole mechanism. A GPU is worth most when demand is strong and you would never dream of selling it, and worth least in the glut when a covenant forces you to. Investment-grade wrappers and off-balance-sheet vehicles have made this stack look safer than the cash flows underneath it, which is the same maturity-and-mark mismatch that official-sector watchers have started to call "shadow borrowing".10 That mismatch is what turns an equity air-pocket into a credit event.
The credit stack is dangerous because it does not stand on its own. It is wired into a circular financing web that removes the market's ability to price any single link independently. The chip vendor invests in the model lab, the lab commits hundreds of billions to the clouds, the clouds buy the vendor's chips, and the vendor then backstops the cloud's unsold capacity.8 Estimates put the interlocked commitments above $800bn.8
When a vendor funds its own customer, the customer's purchase is partly the vendor's capital coming back as revenue. That flatters the demand signal on the way up and reverses it on the way down, because if the funding stops, the revenue and the collateral fail together, having been the same dollars all along. OpenAI shows how fragile this can be. A revenue run-rate of around $25bn stands against roughly $1.4tn of compute commitments and a cash burn that leaked investor documents put in the hundreds of billions through 2030.6
Circularity removes independent price discovery, which is the one thing a stressed market needs most. On the way up it manufactures the very demand that justifies the next round of spending. On the way down a single broken link, a stalled raise or one lab that misses a payment, re-rates every commitment in the ring at once, because each was underwritten against the others. So the web is the amplifier that turns the local credit event of §07 into a systemic one.
Even if the financing holds, the third order is unavoidable, because the buildout has become load-bearing for the economy itself. Investment in information-processing equipment and software was only about 4% of US GDP in the first half of 2025, yet it accounted for roughly 92% of the growth, and strip it out and the economy grew at just a 0.1% annual rate.4 That same spending is also the market's earnings engine.5
The same capex line is at once the market's earnings and the economy's growth, and that collapses the usual firewall between a stock correction and a recession. A freeze cuts index earnings and GDP at the same moment, and the two then feed on each other, because weaker growth compresses the multiple further, that tightens financial conditions, and tighter conditions slow growth again. This is why the moderate branch carries a mild recession rather than just a bear market. The third order routes the equity shock straight into the real economy.
The third order closes a loop back into the first, because concentration makes the de-rate self-reinforcing through the mechanics of the index itself. A market-cap-weighted fund holds the AI-cap complex in proportion to its weight, so the same passive flows that bid it up on the way in are forced to sell it down on the way out. And at around 35% of the index, there is no non-AI position of comparable size to take the other side of that trade.5
The loop is not unbounded, though. The hyperscaler core keeps generating cash and can keep buying back its own stock, a deep buy-the-dip reflex has been trained into the market by fifteen years of V-shaped recoveries, and price-insensitive sovereign buyers of compute can support the private layer. Which of these forces wins, the amplifier or the limiter, is the subject of Part IV, and it is what separates a −40% air-pocket from a −70% bust.
Concentration is the amplifier that ties the three orders into a single system. The equity de-rate cuts the index weight, forced passive selling deepens the de-rate, and the growth and credit legs remove the fundamental buyers who would normally step in. Because the same flows are both the accelerant and the largest holders, this de-rate could easily overshoot the fundamentals, which is exactly why the next Part turns to the question of which force wins.
Where the burst actually lands — across assets, sectors, geographies and the stakeholders who hold the risk. The damage is not where the index weight sits: the mispriced nodes are the false safe-havens dressed as defensives and the passive holders who are long the theme without knowing it.
Map the credit-and-macro leg onto asset classes and the mispriced nodes come into view. The obvious casualties, merchant silicon and the hyperscalers, are on every screen already. The ones that are not so obvious are the AI-power names that re-rated as defensives and the investment-grade credit that quietly funded the whole buildout.14
Two nodes are mispriced. The AI-power utilities were bid up as a way to own AI demand inside a "defensive" bond-proxy wrapper, but their re-rating was really AI beta rather than utility beta, so they fall with the theme instead of cushioning it. Investment-grade credit, meanwhile, looks safe until the market notices how much of the recent issuance actually funded the buildout. That credit hit lags the equity hit by a quarter or two, which is precisely why it is still cheap to hedge today.
If the asset map shows who falls, the sector ledger shows why. A sector's exposure is simply the share of its forward revenue or valuation that rests on the buildout continuing, so the deepest cuts land where a single customer set, the hyperscaler order book, carries the top line.1 The entries worth reading first are the non-obvious ones: the surprise casualty and the surprise shelter.
Every "AI beneficiary" that was sold as a diversified way to own the theme is in fact a leveraged bet on the slope of the buildout rather than its level. That includes the picks-and-shovels utility, the cooling supplier, the copper miner and the "AI real estate" REIT. In a burst, none of them diversifies the AI position; each one concentrates it. The only genuine diversifier is something whose cash flows never touched the theme in the first place.
Sort every holding by a single question: what share of its forward value assumes the buildout keeps running at the current pace? That number, and not the sector label, is the real loss estimate. A "utility" at 90% is a merchant-silicon proxy, while a "tech" name at 10% is a shelter. The ledger is a re-underwriting exercise, and most books have not yet done it.
Concentration decides who holds the AI theme, and the ownership splits in two. Public holders carry it in liquid form and can sell in a day, while private holders carry it in venture, growth and credit funds that can only reprice it slowly, through down-rounds and write-downs. The geography is US-centred, with an East-Asian supply chain and a Gulf financing sleeve, but the ownership is what decides who actually bears the loss.5
| Geography | Exposure | Relative hit |
|---|---|---|
| United States | Epicentre: hyperscalers, merchant silicon, AI-power, the index concentration and the growth contribution. | High |
| East Asia | Taiwan's advanced-node silicon and Korea's HBM memory, plus Japan's equipment and its leveraged AI-investment sponsors. | High |
| Gulf sovereigns | Price-insensitive financiers of both the compute and the labs, holding large primary stakes marked to the last round. | Med |
| Relative shelters: China is walled off from the Western capex loop; India and Europe are exposed to AI demand more than to AI capex. | Low (rel.) | |
| Holder | Exposure & how it transmits | Speed |
|---|---|---|
| Passive / index funds | About a third of the index is the AI-cap complex, so a cap-weighted tracker is a concentrated AI bet held by savers who never chose it. | Same-day |
| VC & growth funds | AI took more than half of US venture dollars in 2025, and the labs raised huge equity rounds on top: OpenAI alone took in ~$40bn, led by SoftBank and MGX.6 Those stakes sit at last-round marks that only reprice in a down-round.14 | Lagged 1–3 qtrs |
| Private credit funds | The data-centre and GPU lending of §07, a market Morgan Stanley sizes near $800bn.9 | Mark-based |
| Sovereign & crossover | Gulf sovereign and crossover funds holding primary stakes marked to the last round. | Last-round |
| LPs (pensions, endowments) | Public index beta plus private AI and VC marks, hit twice as the denominator effect then forces rebalancing. | Slowest |
The dangerous exposure is the one nobody chose, and the slowest to escape. If the thing that cracks is a marquee lab, through a failed raise or a sharp down-round, the loss lands first on the venture, crossover and sovereign funds that own its equity, and then on their LPs, concentrated in whichever fund vintages leaned hardest into AI. Those private marks reprice a quarter or more after the public tape, and the denominator effect then drags the whole fund into selling good assets to rebalance. There is no policy put in private markets and no same-day exit, so the only defence is to run private marks against public comps now and slow commitment pacing.
What sets the magnitude. A burst spans a distribution of outcomes, and where it lands is a tug-of-war between the forces that amplify it (passive concentration, circular financing, procyclical collateral) and the forces that limit it: core cash flow, the buy-the-dip reflex, and a policy reaction that bends the path but lags it.
Which force wins is what sets the magnitude, so the question left open in Part II, amplifier or limiter, deserves an explicit ledger. Each amplifier turns a fall into a bigger fall, and each limiter puts a floor under it. The severity of the burst comes down to which side pulls harder once the shock lands.14
Passive selling and the circular unwind are fast and mechanical, and they do not wait for a committee to decide anything. In the first weeks the amplifiers have the field to themselves, which is why the initial leg can overshoot before any limiter has a chance to engage.
Cash flow and the dip reflex are real and large. Given enough time, the hyperscaler core can keep spending and buying back stock, and price-insensitive buyers can set a floor under compute. So the limiters decide where the fall stops, even if they do little to slow how fast it gets there.
Timing is really the whole game. The amplifiers are fast and the limiters are slow, so the path is an overshoot that gets retraced later, rather than a smooth glide down to fair value. That shape is what makes the convex trade work, because the market travels further in the first three months than the eventual resting point would justify, and the hedge is paid on the overshoot rather than the destination.
A single path would be a fantasy; a distribution is the analysis. So we branch the counterfactual into three, defined by how far the amplifiers get before the limiters engage, and weight them from the outside view. Most bubbles deflate somewhere between mild and moderate, and the internally-funded core here makes the systemic tail the minority case.14
| Branch | What has to be true | Froth periphery | Complex, cap-wtd | S&P 500 | Weight |
|---|---|---|---|---|---|
| Mild | Air-pocket: core cash flow holds, no credit event, the dip is bought within two quarters. | −40/−50% | −20/−28% | −12/−18% | 35% |
| Moderate | Capex recession: the freeze ripples to semis, power and memory; a neocloud/GPU-lease credit event; growth stalls. | −50/−65% | −30/−40% | −20/−26% | 45% |
| Severe | 2000 plus a credit leg: the circular web and GPU-collateral stack crack; contagion to private credit and DC-exposed lenders; recession. | −65/−80% | −48/−58% | −30/−38% | 20% |
Froth periphery = merchant silicon, AI-power, memory and neoclouds. The cap-weighted complex is shallower because the cash-rich hyperscaler core (which re-rates only −15/−45% by branch) carries most of the basket's market value ○.
Multiply the two axes together and the number that sizes the hedge appears. A roughly 25% chance of activation, times a roughly 20% severe branch, gives about a 5% unconditional probability over two years of a −48 to −58% de-rate across the cap-weighted complex, or −65 to −80% through the froth. That is a small number, and yet the tape prices it at essentially zero. The moderate branch is the base case within the scenario, but the convex trade is underwritten against the severe tail rather than the modal path.
The branch you land in depends heavily on the policy reaction, and here that reaction lags. A central bank does not ease simply because AI stocks fall. It eases when the fall starts to threaten funding markets or growth, which are the second and third orders, and those arrive a quarter or two after the equity leg.10 That delay is why the mild branch needs the dip-buyers, and not the Fed, to do the early work.
Unlike a liquidity crash, the buildout itself was disinflationary, and the labour market is not the trigger here. The Fed's reaction is keyed instead to funding stress and growth data, both of which lag the equity move, so the reflexive early leg runs without a backstop underneath it.
Once rung 3 is reached, meaning a genuine funding jam in the credit stack, the response can be fast and large, because that is squarely within the mandate. A decisive facility is what caps the severe branch and pulls the path back toward moderate.
The policy put is real but back-loaded, and that is precisely what preserves the convexity. The market has to travel through the unbackstopped early leg before help arrives, so the option pays on the gap between the overshoot and the eventual floor. To time the reaction, follow the credit signals rather than the equity tape, because the ladder moves when spreads move, well before the index does.
From analysis to action. The scenario converts into two things a reader can keep: a dashboard of leading indicators with trigger levels, so it becomes a live monitor rather than a one-off warning; and a convex position — cheap to hold now, large to own if the credit leg cracks.
The dashboard leads with the credit signals, because they are the ones that move the reaction ladder first. Each comes with a trigger level you could confirm from data and a current reading, and these second-order signals, the spreads, the guidance and the funding, all flip ahead of the equity move that everyone is already watching.10
No single row is the trigger on its own. Activation is confirmed only when the three signals align: capex guidance down two quarters, a neocloud or GPU-lease credit event, and the froth ≥40% below peak and holding there. It is the activation condition from page 03, restated as things you can actually watch tick over.
The credit signals lead the equity signal by a quarter or more, which is what makes this a dashboard rather than a rear-view mirror. Neocloud spreads, CDS and capex guidance will move while the index is still near its high, so an allocator who watches the funding tape can buy protection before the volatility that would price it properly, and a risk officer gets a quarter's warning that the equity screen never gives.
If signposts tell you when, positioning tells you what to do. The trade follows directly from the analysis. Because the equity de-rate is already priced, protection is cheapest on the credit-and-growth leg the tape ignores, and it can be funded by staying long the two-thirds of the market that carried none of the AI premium.14
| Stakeholder | Now (cheap insurance) | If it activates |
|---|---|---|
| Allocator | Measure hidden AI beta in passive books; buy 12–18m put spreads on silicon + AI-power while vol is suppressed; rotate toward equal-weight and ex-US; run private AI and VC marks against public comps and slow commitment pacing; hold a gold / long-vol sleeve. | Add to the credit-stack hedge; short AI-power as a false haven; keep dry powder for the retrace. |
| Corporate / risk | Stress-test AI-linked revenue, GPU-lease and PPA commitments; term-out funding while IG spreads are tight; avoid take-or-pay against unproven demand. | Pull discretionary AI capex; renegotiate leases; preserve liquidity ahead of the funding jam. |
| Policymaker | Map DC and private-credit exposure and the growth reliance on one sector; pre-position liquidity tools for the funding channel. | Sequence the response: liquidity to funding markets before a broad rate cut; ring-fence contagion in private credit. |
The idea is to own convexity rather than fear. A put spread beats simply owning the fall because its cost is bounded and known, while its payoff scales into exactly the moderate-to-severe branch that the tape treats as a non-event. At around a 25% activation probability, insurance that costs a small premium and pays a multiple on a −50 to −80% move is underpriced. What we are recommending is that structure and its price, and not a crash call.
A scenario with no stated way to be wrong is really just propaganda, so it is worth naming the falsifiers plainly. We do it in two steps. First a pre-mortem that assumes the burst has already happened and explains it in hindsight, which tends to surface the real chain better than forward speculation does. Then the strongest case that it does not happen at all, given real airtime, followed by an honest reconciliation with our headline probability.
It started with a routine line in a Q2 call, when one hyperscaler held its capex flat and admitted that inference returns were lagging the spend. Nvidia then guided its data-centre line below the whisper number, and the AI-power utilities cracked first, because they had the least cash-flow support for their re-rating. Within weeks a neocloud missed a GPU-lease payment and a GPU-backed ABS was marked down, and the market began to re-read the whole circular web, the vendor that had funded the buyer and the lab whose commitments dwarfed its revenue, as a single exposure rather than many separate ones. Passive funds sold the complex into a market with no non-AI bid anywhere near its size, and the growth data soon confirmed what the tape already knew, that the capex which had been carrying the economy was gone.
Unlike 2000, the buildout's core is funded from real cash flow, not fresh issuance: the big four earn hundreds of billions in operating cash flow and can keep spending through a narrative break, so a de-rate need not force a capex freeze. Compute demand is real and growing, not a promise — inference and agentic workloads are scaling now. If enterprise ROI inflects even modestly, the "revenue gap" closes over time exactly as dark fiber was eventually lit, and the multiple compresses without a solvency event anywhere in the chain. In that world this is a drawdown, not a burst.
The steelman is powerful, but it caps the severity rather than the odds of activation. Every disconfirming reason, whether internal funding, real demand or the eventual absorption of the build, limits how far the burst goes and how much credit contagion it carries. None of it stops a de-rate from starting, because the trigger is ROI disappointment and a concentration unwind, and those can fire regardless of who funds the core. So the steelman shifts weight from the severe branch toward the mild and moderate ones, but it does not push the probability of a ≥40% de-rate below our band. Two things would move that number: a clear, sustained inflection in enterprise-AI revenue, or capex that keeps rising with margins intact. Until one of those appears, the band holds.
A scenario is only as strong as its weakest unstated premise, so it is worth stating them and marking which ones are load-bearing. The whole chain rests on a small set of assumptions, and if any of the load-bearing ones breaks, the cascade breaks along with it.14
| Assumption | Load-bearing? | If it breaks |
|---|---|---|
| Enterprise-AI ROI stays thin through the horizon | Yes | A clear inflection removes the trigger — bubble grows into itself. |
| The neocloud / GPU-lease credit stack is large & interconnected enough to transmit | Yes | If contained, no second-order credit leg — a pure equity de-rate. |
| AI capex remains a large share of US growth & index EPS | Yes | Broadening growth severs the third-order macro leg. |
| Passive concentration amplifies via index mechanics | Partly | Active de-risking ahead of the move dampens the overshoot. |
| No decisive policy backstop in the first months | Partly | An early, large facility caps the severe branch. |
| The hyperscaler core keeps funding capex from cash flow | Buffer | If the core also retrenches, severity rises toward the tail. |
One driver dominates all the others, and that is whether enterprise-AI returns inflect. It is both the likeliest falsifier and the widest swing in the outcome, which is why the signpost dashboard leads with ROI evidence and capex guidance, since those are the cheapest early reads on the assumption that carries the whole scenario. Everything else merely adjusts the severity. ROI decides whether there is a scenario at all.
Put the two axes back together one last time, without letting them collapse into one. The magnitude is large and the probability is moderate, and the value sits in the gap between what the analysis implies and what the market actually prices, which is a disorderly burst treated as though it were essentially a zero-probability event.14
A 25% probability of a −30 to −40% cap-weighted de-rate, or more than −50% through the froth, priced by suppressed volatility as though it were near zero, is the very definition of a convex opportunity. The cost of the hedge is small and bounded, while its payoff scales into the moderate-to-severe branch. You do not have to believe the burst is likely in order to want the insurance.
The one trade is to own convexity on the second-order leg, through long-dated protection on silicon, power and the AI-infra credit stack, funded by staying long the non-AI market. Activation will show first in the credit tape, a step ahead of the equity screen. This is not a crash call. It is a cheap, dated and priced hedge against the tail the tape ignores.
This is a conditional scenario analysis rather than a forecast. The method follows the Gravitywell rational-scenario framework, which means we define the shock as an activation condition, anchor a base rate, keep magnitude and probability separate, trace the transmission order by order, map the impact, branch the counterfactual, and end in positioning.
Built to the desk's institutional bar: base-rate anchoring against a stated reference class, uncertainty shipped as bands, coverage and confidence disclosed, magnitude and probability held on separate axes, and every effect carrying a mechanism and a source. Confidence is set by coverage depth and source tier, not optimism. House view · Medium conviction · Cautious on the tail.
Every in-text n resolves here. Primary sources (filings, official statistics, central-bank and academic studies) weighted over secondary; private-company internals flagged. Figures marked to the dates shown; tiers: ● filed · ◐ modeled · ○ estimate.
| FIG E.1 | The two axes | 04 |
| FIG Q.1 | Cascade in four beats | 05 |
| FIG 1.1 | Reference-class drawdowns | 06 |
| FIG 1.2 | Probability band | 06 |
| FIG 2.1 | Vendor financing, then/now | 07 |
| FIG 3.1 | The revenue gap | 08 |
| FIG 4.1 | Complex by layer | 09 |
| FIG 4.2 | How the buildout is funded | 09 |
| FIG 5.1 | Transmission chain | 10 |
| FIG 6.1 | Index concentration | 11 |
| FIG 7.1 | The AI-infra credit stack | 12 |
| FIG 7.2 | Collateral doom-loop | 12 |
| FIG 8.1 | The circular web | 13 |
| FIG 9.1 | Capex as GDP growth | 14 |
| FIG 10.1 | Passive reflexivity loop | 15 |
| FIG 11.1 | Impact heatmap | 16 |
| FIG 14.1 | Amplify vs limit | 19 |
| FIG 15.1 | Cascade timeline & fork | 20 |
| FIG 16.1 | The reaction ladder | 21 |
| FIG 18.1 | Convex payoff | 23 |
| FIG 25.1 | Tornado / sensitivity | 25 |
| FIG 26.1 | Weighted expected de-rate | 26 |
The exhibit series behind these figures ships as ai-bubble-bust-data.csv — one row per datapoint, with tier and source ids, so a reader can rebuild the scenario inputs.
This is a conditional scenario analysis of a defined trigger, not a forecast, prediction, or investment advice. Magnitude and probability are stated separately throughout. A high magnitude is not evidence of high likelihood, and nothing here is a view that the AI build lacks long-run value.
The Gravitywell Research desk responsible for this report certifies that the views expressed accurately reflect its independent judgement about the subjects discussed, and that no part of its compensation was, is, or will be directly or indirectly tied to the specific views expressed herein.
As of the publication date, Gravitywell Research and its analysts do not hold positions in the securities or assets discussed. Gravitywell Research has no advisory, banking, or commercial relationship with the entities named. This report was not commissioned or reviewed by any issuer named.
For professional investors across public and private markets, both institutional and sophisticated individual, and for policymakers. Not for general retail distribution, nor for readers who lack the expertise to assess the assumptions. Intended recipients may not redistribute without attribution. Availability in some jurisdictions may be restricted; recipients are responsible for their local rules.
House view · Medium conviction · Cautious on the tail. Conviction reflects coverage depth and source tier; stance reflects the direction of the house view on the opportunity — here, the asymmetry of the hedge, not a crash call. Prior calls: first edition — no prior Gravitywell call on this topic to score. Errata: material errors are corrected in a dated erratum and noted in the next edition; the permalink serves the current version. Permalink: gravitywell.research/whatif/ai-bubble-bust. Version: GWW-2026-001 · v1.0 · as of 21 Jul 2026.
Gravitywell Research Global Technology & Capital Markets Desk. Independent research for professional investors and policymakers; not regulated ratings or investment advice. © 2026 Gravitywell Research.
What if the AI bubble bursts? A ~25%/24-month chance of a −40%-plus de-rate the tape prices near zero — real on capex, concentration and circular financing, but funded at its core from cash flow, so the mispriced legs are the credit event and the growth shock, not the fall. A convex hedge, not a call.