Gravitywell Research · What-If
Scenario Dossier · GWW-2026-001

What if the AI bubble bursts?

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.

P ≈ 20–30%over 24m · vs base rate ~10%
~$710bn
2026E big-four capex, AI-led — the bet
~35%
AI-linked share of S&P 500 market cap
~95%
Enterprise GenAI pilots with no P&L return
−40/−65%
AI silicon & power de-rate if activated (complex ~−30/−40%)
Gravitywell Research Global Technology & Capital Markets Desk · Jul 2026
Medium conviction · Cautious on the tail
Conditional scenario · not a forecast
Gravitywell Research · GWW-2026-001Contents

Scenario DossierContents & conventions

PartSectionPage
The Shock · the activation condition03
Executive scenario summary04
Scenario Quick Read05
I · The Setup§01 Base rate & the reference class · §02 Analogues · §03 Why now · §04 Where the froth sits06–09
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 bid10–15
III · The Impact Map§11 By asset · §12 By sector · §13 By geography & stakeholder16–18
IV · The Dynamics§14 Reflexivity · §15 The branches · §16 The policy-reaction function19–21
V · The Playbook§17 Signposts · §18 Positioning & the convex trade22–23
Pre-mortem · Probability & assumptions · Bottom line24–26
Methodology · Sources · Glossary & Exhibits · Disclosures27–30
About Gravitywell Research

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.

Conventions & data vintage

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.

How to read this dossier

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.

Gravitywell Research02 / 30
Gravitywell Research · GWW-2026-001The Shock

The ShockThe activation condition

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

Activation condition — what must be true

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.

~$710bn
2026E big-four capex (AI-led) — the line that must freeze1
~35%
AI-linked share of the S&P 500 — the concentration5
~95%
GenAI pilots with no P&L return — the ROI gap3
~1/decade
base rate of a ≥40% named-complex burst14
Probability (magnitude ≠ likelihood)
20–30% / 24m · conf. low–med
Base rate ~10%/24m, adjusted up for live preconditions. Movers: capex guidance + AI ROI · rates/liquidity · an infra credit event · the circular financing unwinding.14
Gravitywell Research03 / 30
Gravitywell Research · GWW-2026-001Executive scenario summary

Executive summaryThe scenario in one page

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.

FIG E.1 · The two axes — never collapse themmagnitude × probability
PROBABILITY → MAGNITUDE → Mild 35%−40/−50% Moderate 45%−50/−65% Severe 20%−65/−80% Where the marketprices a disorderly burst: near zero → the hedge is cheap
Bubble magnitudes are the frothy periphery (silicon/power/neocloud) peak-to-trough, conditional on activation; the cap-weighted complex is shallower (the cash-rich core cushions it — see §15). Weights sum to 100% within the scenario ○. Bubble area ≈ conditional weight, not a probability distribution over the level.14
What the market already prices

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.

The read — what it misses

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 one trade

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.

Gravitywell Research04 / 30
Gravitywell Research · GWW-2026-001Scenario Quick Read

Quick readThe cascade in sixty seconds

FIG Q.1 · Guidance cut → de-rate → credit event, in four beatsday 1 → month 6
Day 1capex guide cut Week 1–4silicon + power re-rate Month 2–4FORK: does the credit stack hold? Month 6+air-pocket or bust
The fork is the financing plumbing, not the index level: if internal cash flow funds the buildout, the de-rate stays an equity event ◐.6
Top hedges — cheap now
Silicon put spreads
Merchant-silicon + AI-power · vol suppressed by the dip bid
Infra-credit protection
Neocloud / GPU-lease / DC-REIT spread hedges
Own the non-AI market
Equal-weight + value · what the burst leaves standing
By-stakeholder posture · 30 seconds
StakeholderDo now (cheap insurance)If it activates
AllocatorMeasure 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 / riskStress-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.
PolicymakerMap 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.
In one line

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.

Gravitywell Research05 / 30
Part I

The Setup

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.

Reading
Pages 06 — 09
Key figure
~35% · index in AI
Base rate
~10% / 24m
What you'll take away
01
The outside view
A named tech or capex complex bursts ≥40% roughly once a decade; even inside a bubble, a disorderly deflation is a minority path, not the default.
02
What rhymes
2000 telecom-fiber, not pets.com: real infrastructure, circular vendor financing, and ROI timing that wipes the financiers before demand catches up.
03
Where the froth is
The mega-cap core is funded from cash flow; the froth is in the merchant-silicon multiple, the neocloud credit and the AI-power names.
Part I · The SetupGWW-2026-001

§ 01 · Base rateThe reference class

In one line · a named tech complex bursts ≥40% about once a decade; conditioned on already being in a bubble, a disorderly 24-month deflation is roughly a 1-in-4 event — not remote, not a coin flip.

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.

FIG 1.1 · The reference class — peak-to-trough drawdowns of named bubbles%, peak → trough
Nasdaq 2000–02Telecom/fiber 2000–02Japan TOPIX 1989–92Chinese internet 2021–22SPAC / ARKK 2021–22Crypto 2021–22Biotech (XBI) 2015–16 −78% −95% −63% −70% −75% −65% −40% 0−50%−100%
Peak-to-trough index drawdowns; telecom/fiber = the equipment + carrier complex ●.11 Roughly one ≥40% named-complex burst per decade in US markets, counting distinct episodes (2000, 2015–16, 2021–22), not each index.14
FIG 1.2 · From reference class to a probability bandP(activation) / 24m
market ≈ 0 base ~10% P ≈ 20–30% / 24m the scenario band · faint = inside-view uplift on the base rate 0%50%100%
Base rate ~10%/24m from the reference-class frequency ◐; the band adds an inside-view uplift for the live preconditions (§03) ○. The band draws over [20%,30%] only — not from zero.14
The read

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.

Gravitywell Research06 / 30
Part I · The SetupGWW-2026-001

§ 02 · AnaloguesWhat rhymes, what's different

In one line · the buildout rhymes with 2000 telecom-fiber, not pets.com — real infrastructure, circular vendor financing, and ROI timing that wipes the financiers before demand arrives.

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

EpisodeWhat rhymes with AI nowWhat's different this time
Telecom / fiber 2000–02Real 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 / 1873A 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 2000Narrative-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 2021–22Circular 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.
FIG 2.1 · The vendor-financing loop, then and nowwho funds the buyer
2000 — equipment vendors funded the carriers that bought their gear Lucent / Nortel CLEC carriers Dark fiberdemand no-show vendor loanscapex 2026 — the chip vendor invests in the buyer that buys its chips Nvidia OpenAI / neoclouds GPU orders investsbuys chips
The structure rhymes: the seller of the picks funds the miners ◐. In 2000, Lucent committed ~$6–8bn and Nortel ~$3.1bn of vendor financing to their carrier customers; ~$2tn of telecom market value and ~$1tn of debt were destroyed in the unwind.11 The 2026 web is detailed in §08.7
The read

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.

Gravitywell Research07 / 30
Part I · The SetupGWW-2026-001

§ 03 · Why nowThe live preconditions

In one line · three preconditions are live at once (thin ROI evidence, record concentration, circular financing), which is what lifts the inside view above the base rate.

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

~95%
enterprise GenAI pilots with no measurable P&L return3
~35%
AI-linked share of S&P 500 market cap5
~$1.4tn
cumulative AI compute commitments announced6
<10%
of that backed by contracted end-demand today ○
FIG 3.1 · The revenue gap — 2026E big-four capex vs AI revenue booked$bn / yr
2026E capex, big-four (AI-led)Industry AI revenue booked, run-rate ~$710bn ~$90bn ○ 0$375bn$750bn the gap the return must close
Big-four 2026E total capex ~$710bn, overwhelmingly AI/datacentre-driven (~$550–650bn AI/DC-specific) ●;1 industry AI revenue booked is a fraction of it ○. Sequoia's David Cahn put the annual shortfall near ~$600bn in 2024 (his "$600bn question"); Bain estimates the industry must find ~$2tn of new annual AI revenue by 2030 and would still fall ~$800bn short.7
Why the gap can persist

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.

Why it might not

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.

The read

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.

Gravitywell Research08 / 30
Part I · The SetupGWW-2026-001

§ 04 · ValuationFroth at the core, or the edge?

In one line · on cash flow the mega-cap core is stretched but not 2000-crazy; the froth is in merchant-silicon multiples, neocloud credit and the AI-power names — so a burst re-rates the periphery hardest.

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

FIG 4.1 · The AI-cap complex by layer — multiple vs cash-flow supportfroth rises outward
HyperscalersMerchant AI siliconMemory / networkingAI-power utilitiesNeoclouds / GPU-leasePrivate AI labs cash-flow funded cyclical, concentrated order-book exposed re-rated on AI demand debt-financed no earnings floor ~25–35× fwd earnings, but funded by ~$300bn+ operating cash flow ~30–50× on peak-cycle earnings; revenue concentrated in a few buyers HBM / switching priced on the buildout continuing 2–4× re-rating on AI-power demand; long-dated PPAs levered to GPUs whose resale value falls in a glut marked on the last primary round, not a market
Froth heat: teal = cash-flow supported · amber = cyclical/exposed · crimson = no earnings floor ◐. Multiples indicative, latest reported basis.1
FIG 4.2 · How the buildout is funded — the fault lineshare of AI capex
Aggregate AI/datacentre capex, by funding source ○ internal cash flow (hyperscaler core) bond issuance debt / equity, periphery ~55–65%~15–20%~15–25%
The externally-funded ~third is the fault line: it cannot wait for ROI and carries the credit risk mapped in §07–08 ○.
The read

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.

Gravitywell Research09 / 30
Part II

The Transmission

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.

Reading
Pages 10 — 15
1st order
the equity de-rate
2nd / 3rd
credit · growth
What you'll take away
01
Three channels, three speeds
Equity multiples reprice in hours; the credit stack in quarters; the growth hit over a year — and only the first is priced.
02
The credit stack & the circular web
A >$20bn neocloud debt load and a >$800bn interlocked financing loop turn a demand pause into a credit event.
03
Capex is now GDP
Information-processing investment was ~92% of H1-2025 US growth, so a freeze is a macro shock, not a sector one.
Part II · The TransmissionGWW-2026-001

§ 05 · The chainShock to credit event, by order

In one line · the market prices the first-order equity de-rate in hours; the mispriced legs are the credit event a few quarters later and the growth shock underneath it.

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

FIG 5.1 · The transmission — capex freeze to credit event1st → 3rd order
SHOCK1ST ORDER · hours2ND ORDER · quarters3RD ORDER · a year Capex guide cutROI disappoints Silicon + powerre-rate −40 to −60% Concentrationpassive bid reverses Credit eventneocloud / GPU-lease default Growthshock, earnings freezefunding jamcapex = GDP
Gravitywell Research scenario model ○; magnitudes branch-conditional (see §15). Arrows are transmission channels, not certainties.9
What is already priced — own the unpriced column
NodeMechanismPriced?
Equity de-rateMultiple compression as the growth narrative breaks.half-priced
Concentration unwindCap-weighted funds sell the complex proportionally.half-priced
Infra credit eventGPU-lease / neocloud default; collateral value falls in the glut.unpriced
Circular unwindOne broken link re-rates the whole interlocked loop.unpriced
Growth & earningsCapex was carrying US growth and index EPS; a freeze cuts both.unpriced
The read

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.

Gravitywell Research10 / 30
Part II · The TransmissionGWW-2026-001

§ 06 · First orderThe equity de-rate

In one line · the de-rate hits merchant silicon and AI-power hardest because their forward revenue is the most concentrated in the hyperscaler order book — and one stock, Nvidia, is ~8% of the index.

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

FIG 6.1 · The most concentrated index in fifty years% of S&P 500 market cap
Magnificent-7, 2017Magnificent-7, 2026Nvidia alone, 2026 ~12% ~35% ~8% 020%40%
Mag-7 ~34–35% of S&P 500 cap, roughly tripled from ~12% in 2017; Nvidia alone ~8% — highest single-stock weight since 1981 ●.5
Why the periphery falls further than the core

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.

Why it is only half the story

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 read

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.

Gravitywell Research11 / 30
Part II · The TransmissionGWW-2026-001

§ 07 · Second orderThe neocloud & GPU-lease credit stack

In one line · the unpriced leg is a debt stack secured against GPUs whose resale value falls in the very glut that triggers the default — collateral that is worth least exactly when it must be sold.

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

FIG 7.1 · The AI-infra credit stack — what funds the buildout's edge$bn, 2025–26
Hyperscaler AI bonds, 2025Meta "Hyperion" SPV (off-b/s)GPU / DC securitisation, 2025Neocloud sector, GPU-backedCoreWeave alone, all debt ~$121bn ~$27bn ~$27bn >$20bn >$21bn
Hyperscaler AI-related bond issuance ~$121bn in 2025, >4× the 2020–24 average ●;10 GPU/DC securitisation ~$27bn (JPMorgan sees $30–40bn/yr in 2026–27); CoreWeave's all-source debt >$21bn (from <$8bn in 2024) ●.9 Meta's ~$27bn data-centre JV was flagged a "critical audit matter."10 Total debt exceeds the sector's GPU-backed-only estimate because it includes corporate facilities — the two bars are different measures.
FIG 7.2 · The collateral doom-loop — why the default is reflexiveprocyclical LTV
Demand pause GPU resale ↓ LTV breach →forced sale More supply → ↓ the loop feeds itself — collateral is worth least exactly when it must be sold
GPU collateral is procyclical: a glut cuts resale value, which breaches loan-to-value, which forces sales into the glut ◐.9
The read

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.

Gravitywell Research12 / 30
Part II · The TransmissionGWW-2026-001

§ 08 · Second orderThe circular financing web

In one line · the same dollars appear as revenue in several places at once — a >$800bn interlocked loop where the chip vendor funds the buyer that buys its chips, so one broken link re-rates the whole ring.

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

FIG 8.1 · The circular web — who funds whomannounced commitments, $bn
Nvidia~$5tn OpenAI~$25bn rev CoreWeave OracleMSFT · AMD $6.3bn backstop + ~11% invests $100bn LOI→$30bn buys GPUs $300bn (Oracle) · $250bn (MSFT) ~$22bn compute clouds run on Nvidia GPUs
Amber = vendor funds customer; grey = customer buys/commits. Nvidia's marquee ~$100bn OpenAI commitment was a letter of intent, never signed, and was restructured to a ~$30bn equity stake by early 2026.8 OpenAI: ~$25bn revenue run-rate against ~$1.4tn of compute commitments.6
Why "circular" is the risk, not a slur

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

The read

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.

Gravitywell Research13 / 30
Part II · The TransmissionGWW-2026-001

§ 09 · Third orderWhen capex is the economy

In one line · information-processing investment was 4% of GDP but ~92% of US growth in H1 2025 — so a capex freeze is a macro shock and an earnings shock, not a one-sector drawdown.

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

FIG 9.1 · US GDP growth in H1 2025, with and without the buildoutannualised real growth
GDP growth as reportedEx information-processing capex ~92% of it was this ~0.1% annualised 0
Furman (Harvard), on BEA data: information-processing equipment & software ≈ 4% of GDP but ~92% of H1-2025 GDP growth; ex those categories, ~0.1% annualised ◐.4 An accounting attribution, not a counterfactual — the author notes cheaper rates and power absent the boom might recover about half.
~46%
of the S&P 500's 2025 total return from the Mag-75
~2.8pp
of the S&P's ~17.9% 2025 gain from Nvidia alone5
~100bp
est. add to US real GDP growth from DC capex4
~$710bn
2026E big-four capex — the line that would freeze1
The read

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.

Gravitywell Research14 / 30
Part II · The TransmissionGWW-2026-001

§ 10 · Third orderConcentration & the passive bid

In one line · record index concentration turns the passive bid that inflated the complex into a passive offer — the flows that never chose the AI trade become its forced sellers.

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

FIG 10.1 · The passive reflexivity loopwhy concentration amplifies
AI-cap falls index weight ↓cap-weighted funds sell no non-AI bidat comparable size falls more the flows that never chose the trade become its forced sellers
Cap-weighted passive flows are procyclical by construction; concentration removes the diversifying bid ◐.5
The limiter on the other side

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.

The read

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.

Gravitywell Research15 / 30
Part III

The Impact Map

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.

Reading
Pages 16 — 18
False haven
AI-power utilities
Hidden long
every passive S&P holder
What you'll take away
01
By asset
Merchant silicon and AI-power fall hardest; cash, gold, long-vol and the non-AI market are the shelter.
02
By sector
The utilities bought as bond-proxies are the surprise casualty; the beneficiaries are whatever did not re-rate on AI.
03
Who holds it
Concentration means the risk sits in passive index funds and pensions that never chose the AI trade.
Part III · The Impact MapGWW-2026-001

§ 11 · By assetWho gets hit, who gets paid

In one line · the false safe-haven is the AI-power utility bought as a bond-proxy; the shelter is cash, gold, long-vol and the equal-weight market the burst leaves standing.

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

FIG 11.1 · Impact heatmap — asset × horizon (moderate branch)% move, scenario-conditional
0–1m1–6m6m+ Merchant AI siliconAI-power utilitiesMemory / networkingHyperscalersAI-infra creditGold / long-volCash / T-bills · non-AI eq −35% −55% −45% −30% −55% −40% −30% −45% −35% −15% −25% −18% −5% −25% −30% +8% +15% +6% +2% flat/+ +10%
−40% or worse−20 to −40%−5 to −20%0 to +10%+10% or better
AI-cap peak-to-trough by asset, moderate branch, conditional on activation ○. AI-infra credit lags equities — the second-order leg (§07) prices last, not first.14
The read

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.

Gravitywell Research16 / 30
Part III · The Impact MapGWW-2026-001

§ 12 · By sectorThe winners & losers ledger

In one line · the casualties cluster where forward revenue is concentrated in the buildout; the beneficiaries are whatever the AI trade left cheap — value, non-US, and the balance sheet itself.

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.

Casualties — ranked by buildout dependence
−50/−65%AI-power & electricalfalse safe-haven: utilities, gas-turbine and grid names re-rated 2–4× on data-centre demand and long-dated PPAs that a freeze strands
−45/−60%Merchant AI siliconforward order book concentrated in a handful of hyperscaler buyers
−40/−55%Memory (HBM) & networkingpriced on the buildout continuing at the current slope
−30/−50%Neoclouds / DC-REITslevered to GPU resale value and take-or-pay leases
−15/−25%Hyperscalerscash flows cushion; the capex write-down and lost growth premium do not
Beneficiaries — what did not re-rate on AI
+ / flatValue & equal-weight equitythe ~two-thirds of the market that carried none of the AI premium
+10/+20%Gold & long-volthe flight-to-safety and the de-concentration trade
+Cash & short Treasuriesthe option value of dry powder when everything correlated falls
rel. +Defensives ex-AIstaples, healthcare, real (non-DC) assets — genuine, not AI-proxy, defensives
The trap in the ledger

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.

The read

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.

Gravitywell Research17 / 30
Part III · The Impact MapGWW-2026-001

§ 13 · StakeholdersBy geography & who holds the risk

In one line · the exposure is most dangerous where it is least chosen and least liquid: in passive index funds that never decided to own AI, and in the private venture, credit and LP books that can only mark the burst down a quarter or more after the public tape has moved.

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

By geography
GeographyExposureRelative hit
United StatesEpicentre: hyperscalers, merchant silicon, AI-power, the index concentration and the growth contribution.High
East AsiaTaiwan's advanced-node silicon and Korea's HBM memory, plus Japan's equipment and its leveraged AI-investment sponsors.High
Gulf sovereignsPrice-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.)
Who holds it — public vs private
HolderExposure & how it transmitsSpeed
Passive / index fundsAbout 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 fundsAI 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.14Lagged 1–3 qtrs
Private credit fundsThe data-centre and GPU lending of §07, a market Morgan Stanley sizes near $800bn.9Mark-based
Sovereign & crossoverGulf 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 read

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.

Gravitywell Research18 / 30
Part IV

The Dynamics

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.

Reading
Pages 19 — 21
Branches
35 / 45 / 20
Severe tail
~5% / 24m
What you'll take away
01
The tug-of-war
Three amplifiers against three limiters; the balance decides whether the de-rate overshoots the fundamentals or is bought.
02
Mild · Moderate · Severe
Weighted 35 / 45 / 20 given activation, anchored to how past bubbles resolved and the cash buffer at the core.
03
The policy put lags
The Fed reacts to the credit-and-growth leg, not the equity leg, so the backstop arrives later than in a pure sell-off.
Part IV · The DynamicsGWW-2026-001

§ 14 · ReflexivityWhat amplifies, what limits

In one line · the magnitude is a tug-of-war — three amplifiers (passive concentration, circular financing, procyclical GPU collateral) against three limiters (core cash flow, the dip reflex, price-insensitive sovereign buyers).

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

FIG 14.1 · The tug-of-war that sets the magnitudeamplifiers vs limiters
the shock ← LIMITS the fall DEEPENS the fall → Passive concentrationCircular financingProcyclical GPU collateral Hyperscaler cash flowBuy-the-dip reflexSovereign compute buyers net pull decides mild vs severe · bar length ≈ relative force ○
Relative force is a Gravitywell judgement, not a measured quantity ○. The core's cash flow is the strongest single limiter; concentration the strongest amplifier.14
Why the amplifiers can win early

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.

Why the limiters win eventually

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.

The read

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.

Gravitywell Research19 / 30
Part IV · The DynamicsGWW-2026-001

§ 15 · The branchesMild · Moderate · Severe

In one line · weighted 35 / 45 / 20 given activation — derived from how bubbles actually resolve, the cash buffer at the core and the starting valuation, not asserted.

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

BranchWhat has to be trueFroth peripheryComplex, cap-wtdS&P 500Weight
MildAir-pocket: core cash flow holds, no credit event, the dip is bought within two quarters.−40/−50%−20/−28%−12/−18%35%
ModerateCapex recession: the freeze ripples to semis, power and memory; a neocloud/GPU-lease credit event; growth stalls.−50/−65%−30/−40%−20/−26%45%
Severe2000 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 ○.

FIG 15.1 · Cascade timeline & the three-way forkday 1 → year 1
Day 1capex guide cut Week 1–4silicon + power re-rate Month 2–4FORK Mild · dip bought Moderate · capex recession Severe · credit contagion the credit stack decides which branch
All three branches terminate on the timeline; the split point is whether the credit stack (§07–08) holds ○.14
The read

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.

Gravitywell Research20 / 30
Part IV · The DynamicsGWW-2026-001

§ 16 · ReactionThe Fed's ladder — and why it lags

In one line · the policy put bends the path but arrives late, because the Fed reacts to the credit-and-growth leg, not the equity leg — a falling stock market alone does not move it.

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.

FIG 16.1 · The reaction ladder — what triggers each rungescalation by stress
1 · Jawbone 2 · Rate cuts 3 · Liquidity facility 4 · Systemic backstop growth-scare talkgrowth & labour softenfunding markets jama systemic lender at risk each rung needs a bigger trigger — and the equity de-rate alone reaches only rung 1
Escalation is triggered by funding and growth stress, not by the equity fall alone ◐. Rungs 3–4 require funding-market or systemic stress, which the mild branch never reaches.10
What makes the put later here

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.

What makes it eventually decisive

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 read

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.

Gravitywell Research21 / 30
Part V

The Playbook

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.

Reading
Pages 22 — 23
Watch
the credit tape, not the index
Trade
convexity, not fear
What you'll take away
01
The dashboard
Capex guidance, Nvidia DC growth, neocloud spreads and the ABS market — each with a checkable trigger and a current reading.
02
The convex trade
Long-dated protection on silicon, power and infra credit, funded by owning the non-AI market the burst leaves standing.
03
By stakeholder
A now / if-active split for the allocator, the risk officer and the policymaker.
Part V · The PlaybookGWW-2026-001

§ 17 · SignpostsThe dashboard to keep

In one line · the scenario is monitorable — six leading indicators, each with a checkable trigger and a current reading, so this is a live monitor rather than a one-off warning.

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

Hyperscaler capex guidancethe buildout's own signal · now +50–60% YoY for 2026
2 down-quarters
Nvidia data-centre revenue growthnow ~+90% YoY (~$75bn/qtr)
< +25% YoY
Neocloud debt spreadmarquee 9.25% notes traded to ~90.6¢
< 85¢ / +200bp
Hyperscaler / Oracle CDSOracle already BBB−; 5yr CDS ~139bp
> 200bp
Enterprise-AI ROI evidence~95% of pilots show no P&L return today
no inflection in 2 qtrs
PPA / lease cancellationswatch for stranded power take-or-pay
> 2GW pulled
Private-AI down-roundsthe private leg reprices here first, lagging the tape
a marquee lab down-round
What confirms activation

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 read

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.

Gravitywell Research22 / 30
Part V · The PlaybookGWW-2026-001

§ 18 · PositioningThe hedges & the convex trade

In one line · the trade is convexity on the second-order leg — long-dated protection on silicon, power and infra credit, funded by owning the non-AI market — cheap while the dip bid suppresses volatility.

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

StakeholderNow (cheap insurance)If it activates
AllocatorMeasure 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 / riskStress-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.
PolicymakerMap 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.
FIG 18.1 · The convex payoff — put spread vs the underlyingpayoff at expiry
AI-cap complex level at expiry → P/L → owning the fall (linear) put spread — capped payoff small premium at risk −40% −65%
Illustrative: a 12–18m put spread struck ~−40% / −65% risks a small premium and pays a multiple if the moderate-to-severe branch lands ○. Vol is suppressed by the dip bid, so the premium is cheap relative to the payoff.
The read

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.

Gravitywell Research23 / 30
Pre-mortem & Red-teamGWW-2026-001

FalsificationWhy it happens — and why it might not

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.

Pre-mortem · assume it happened

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.

The disconfirming view — the strongest case it does NOT happen

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.

Why we still weight it ~20–30%

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.

Gravitywell Research24 / 30
Probability & assumptionsGWW-2026-001

ProbabilityThe assumption ledger & sensitivity

In one line · the scenario rests on a handful of load-bearing assumptions, and one, whether enterprise-AI ROI inflects, moves the outcome far more than the rest.

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

AssumptionLoad-bearing?If it breaks
Enterprise-AI ROI stays thin through the horizonYesA clear inflection removes the trigger — bubble grows into itself.
The neocloud / GPU-lease credit stack is large & interconnected enough to transmitYesIf contained, no second-order credit leg — a pure equity de-rate.
AI capex remains a large share of US growth & index EPSYesBroadening growth severs the third-order macro leg.
Passive concentration amplifies via index mechanicsPartlyActive de-risking ahead of the move dampens the overshoot.
No decisive policy backstop in the first monthsPartlyAn early, large facility caps the severe branch.
The hyperscaler core keeps funding capex from cash flowBufferIf the core also retrenches, severity rises toward the tail.
FIG 25.1 · Tornado — what moves the outcome mostswing in froth peak-trough, pp
base ~−58% (froth) Enterprise-AI ROICredit-stack sizePolicy-put timingConcentration / passiveRates & liquidity −25pp← contains · base · worsens →+30pp
Swing in the AI-cap peak-to-trough from moving each driver across its plausible range, others held at base ○. ROI evidence swings the outcome roughly ~2× the rates driver.14
The read

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.

Gravitywell Research25 / 30
Bottom lineGWW-2026-001

Bottom lineThe probability-weighted read

In one line · a ~25%/24m chance the froth falls ≥40% and holds, dragging the cap-weighted complex −30 to −40% and the S&P −20 to −26%, with a ~5% severe tail the tape prices near zero. A convex hedge, not a call.

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

FIG 26.1 · Probability-weighted expected de-rateAI-cap complex, conditional on activation
Mild · 35%Moderate · 45%Severe · 20% −24% −35% −53% 0−45%−90% weighted expected cap-weighted de-rate ≈ −35% · bar opacity ≈ branch weight
Cap-weighted AI-cap complex, conditional on activation: weighted expected de-rate ~−35% ○; the frothy periphery runs deeper (−40 to −80%). Unconditional severe tail (P × severe) ≈ 5% / 24m.14
The asymmetry

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

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.

The AI build may well be the most important capital cycle of the decade. That is entirely consistent with a −50% detour along the way, because the railways and the fiber were real too, and both ruined the people who financed them ahead of demand.Gravitywell Research · Scenario Desk
Gravitywell Research26 / 30
MethodologyGWW-2026-001

MethodologyHow this was built

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.

Scope & coverage
How the key numbers were derived
Standards

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.

Gravitywell Research27 / 30
Source RegisterGWW-2026-001

SourcesRegister

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.

Gravitywell Research28 / 30
Glossary & ExhibitsGWW-2026-001

ReferenceGlossary & list of exhibits

Glossary
Activation condition
The precise, observable trigger that defines the scenario; everything downstream is "given this."
AI-cap complex
The AI-capitalisation-weighted equity basket: merchant silicon, hyperscalers, AI-power, memory, networking and the neocloud/AI-software beneficiaries.
Hyperscaler
A large cloud operator funding the buildout from its own cash flow (Microsoft, Alphabet, Amazon, Meta).
Neocloud
A specialist GPU-cloud provider (e.g. CoreWeave) that rents compute, typically debt-financed against its chips.
Circular financing
When a vendor funds the customer that buys its product, so revenue is partly the vendor's own capital returning.
GPU-lease / GPU-backed debt
Borrowing secured on the GPUs themselves — collateral whose resale value falls in a glut.
Reflexivity
A feedback loop where price moves change fundamentals that then drive further price moves (here, passive flows and collateral values).
Convexity
A payoff that gains far more than it loses — a small, bounded cost for a large, scaling gain if the tail lands.
List of exhibits
FIG E.1The two axes04
FIG Q.1Cascade in four beats05
FIG 1.1Reference-class drawdowns06
FIG 1.2Probability band06
FIG 2.1Vendor financing, then/now07
FIG 3.1The revenue gap08
FIG 4.1Complex by layer09
FIG 4.2How the buildout is funded09
FIG 5.1Transmission chain10
FIG 6.1Index concentration11
FIG 7.1The AI-infra credit stack12
FIG 7.2Collateral doom-loop12
FIG 8.1The circular web13
FIG 9.1Capex as GDP growth14
FIG 10.1Passive reflexivity loop15
FIG 11.1Impact heatmap16
FIG 14.1Amplify vs limit19
FIG 15.1Cascade timeline & fork20
FIG 16.1The reaction ladder21
FIG 18.1Convex payoff23
FIG 25.1Tornado / sensitivity25
FIG 26.1Weighted expected de-rate26
Data pack

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.

Gravitywell Research29 / 30
Disclosures & GovernanceGWW-2026-001

GovernanceDisclosures

Scenario, not forecast

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.

Certification (Reg AC style)

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.

Positioning & conflicts

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.

Distribution

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 rating & governance

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.

Desk

Gravitywell Research Global Technology & Capital Markets Desk. Independent research for professional investors and policymakers; not regulated ratings or investment advice. © 2026 Gravitywell Research.

Gravitywell Research30 / 30
Gravitywell Research · What-If

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.

GWW-2026-001 · Jul 2026 · Gravitywell Research Global Technology & Capital Markets Desk
Conditional scenario analysis. Not a forecast, prediction, or investment advice. Figures marked to the dates shown and may be revised.
© 2026 Gravitywell Research.