Experimental — Dyno-Sim is under active development. These runs are research-grade: reduced-form models with disclosed dials, reporting relative, structural results — the mechanism and the comparison between runs. Where this piece combines models, they currently couple through disclosed elasticities, not yet a single conserved ledger (active development). Illustrative / decision-support, not a forecast — and deliberately no "AI-bubble probability." The financing parameters below are disclosed illustrative dials, not measured constants, and no fitted AI-capex calibration exists. Every number is a relative comparison between two runs of a validated, accounting-consistent debt-deflation model — not a prediction of any company or any year. This is a read on a mechanism; it is not investment advice.

In August 2026, NVIDIA and six of the largest names in finance — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — announced independent platforms designed to mobilize more than $500 billion of third-party capital into AI factories. NVIDIA will backstop part of the depreciation risk on its own chips — a residual-value guarantee of up to 25% of a given deal, which means roughly 75% of that risk stays with the lenders.

That announcement did not create the risk; it industrialized a financing structure that was already building. And it is exactly the kind of structure a simulator built for credit cycles exists to reason about. So rather than add another opinion about whether AI is a bubble, Comuvia's system did what it does with the rest of its work: it represented the financing as explicit dials in a model it owns — in this case a faithful, validated port of Steve Keen's own debt-deflation model — ran it, and shows the work below, including a plain statement of what the numbers are not.

Concept diagram — the AI-capex money machine as three compounding amplifiers: (1) $500B financing platforms (Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR) mobilize third-party capital; (2) a circular vendor-financing loop — NVIDIA sells GPUs on credit to neoclouds (CoreWeave), which sell compute to AI labs, whose revenue flows back to NVIDIA; (3) roughly $120B of GPU-collateralized special-purpose vehicles moved off the balance sheet, with NVIDIA residual-value support covering only about 25% and the other ~75% sitting with the lenders.

The money machine has three amplifiers

Underneath the headline sit three features that turn an ordinary capital-spending boom into a leverage story:

  1. Circular / vendor financing. NVIDIA sells GPUs to neoclouds (CoreWeave and peers) on vendor terms; the neoclouds rent those GPUs to AI labs; a slice of the revenue flows back to NVIDIA. The supplier is, in part, financing its own demand — which flatters the demand signal everyone else is underwriting against.
  2. Off-balance-sheet SPVs. Roughly $120 billion of AI-infrastructure debt (Oracle, Meta, xAI, CoreWeave) has already been moved into special-purpose vehicles funded by Wall Street — delayed- draw term loans with the GPUs pledged as collateral. The leverage is real but sits outside the headline balance sheets, and lenders' visibility into the aggregate is limited.
  3. Collateral that depreciates — fast. The collateral behind these loans is not houses; it is hardware on a technology-obsolescence clock. If AI demand softens or the next chip generation lands, the resale value of the pledged GPUs can fall well before the loan is repaid. That is the feature 2008's mortgage collateral did not have.

None of this is a secret, and it is not illegal. But it is largely unregulated and largely off-balance-sheet, which is precisely why the Bank of England has already flagged financial- stability risk from highly-leveraged AI firms and the limited visibility lenders have into it. The question a decision-maker actually needs is not "bubble: yes or no" — it is how much more fragile does this financing structure make the system, and what happens if the collateral reprices?

The spine: Minsky's own definition of Ponzi finance

Hyman Minsky sorted every borrower into three kinds, by whether their cash flow can cover their debt. The dangerous one he named precisely:

"'Ponzi' finance: cash flows… fall short of cash payment commitments and net income… falls short of the interest portion… [the unit] must increase its outstanding debt in order to meet its financial obligations. Presumably, there is a 'bonanza' in the future which makes the present value positive for low enough interest rates… every investment project with a long gestation period and somewhat uncertain returns has aspects of a 'Ponzi' finance scheme." — Minsky, The Financial Instability Hypothesis: A Restatement

Read that last sentence with the build-out in mind. Datacenters and GPUs are the textbook case of a long gestation period and somewhat uncertain returns; a large share of them is being funded by circular-vendor and off-balance-sheet SPV debt that the operators must roll and extend before the returns arrive. By Minsky's own definition, that carries Ponzi aspects — and the bet that AGI-scale returns arrive later is what makes today's debt-funded build-out look present-value-positive. Rising rates flip speculative units into Ponzi units.

Keen gives the mechanism a sharper name and a citation. The problem is not debt as such — it is "debt that doesn't create new productive capacity" (Giraud & Grasselli, 2019). And it has a technology precedent: in "Credit Creates Asset Price Bubbles," Keen shows it is the acceleration of credit that drives changes in asset prices, and names the Internet bubble as the direct technology-credit forerunner of exactly this kind of episode. Asset prices can fall while debt is still growing — the moment credit decelerates.

Two older voices frame the two sides of the trade. Schumpeter is the optimistic foil: "credit is essentially the creation of purchasing power for the purpose of transferring it to the entrepreneur… by credit, entrepreneurs are given access to the social stream of goods before they have acquired the normal claim to it." That is the innovation-financing case for the build-out. Fisher supplies the downside if the returns don't validate the debt: over-investment and over-speculation, he wrote, "would have far less serious results were they not conducted with borrowed money" — and once "over-indebtedness" meets "deflation following soon after," his nine-link debt-deflation chain becomes the bust engine. The whole question is which of the two you are watching.

The original contribution: not all credit is the same

The critics of Keen's headline claim (Mason, Fiebiger, Palley) have one point that is exactly right and directly useful here: a large share of new private debt funds asset purchases, not newly-produced output, so "change in debt" does not map cleanly onto demand. We turned that objection into the model's central dial.

The model splits credit into two kinds:

  • Real-investment credit adds productive capacity — output grows alongside the debt, and the debt is validated by real returns. This is Schumpeter's channel.
  • Financial / SPV / circular credit is the destabilizing kind in Keen's precise sense: it inflates leverage without adding capacity. The debt-to-output ratio climbs at the same output — which is what pushes an economy toward Fisher's debt-deflation cliff.

That split is a single disclosed dial — financial_credit_fraction, the share of credit funding financial rather than productive positions. It is off (zero) in the baseline, and turning it up is how we represent the AI-capex financing structure. We lead with the mechanism, not with a raw correlation, precisely because "ΔDebt correlates with activity" is partly definitional — the interesting result is that the behavioral model produces a lower tipping point when more of its credit is financial.

Mechanism diagram — credit splits into two lanes. Real-investment credit builds productive capacity, so output grows with the debt and the safe-leverage ceiling holds (200%, or 245% with government). Financial / SPV / circular credit inflates leverage without adding capacity, drives a debt-deflation spiral, and the ceiling falls to 175% then 120% while the fiscal rescue fails. The financial_credit_fraction dial sits at the split.

The result: how much leverage is safe before it tips?

Comuvia's keen_monetary_minsky_v1 runtime is a faithful, validated port of Keen's reduced-form Financial-Instability model — three states (employment, wage share, private-debt ratio), Keen's own generalized-exponential investment and wage functions at his actual pivot slope, a Kalecki markup inflation rate for the Fisher deflation channel, and a counter-cyclical government rule. It is bistable: below a critical leverage the economy settles into a finite-debt equilibrium; above it, private debt runs away into a debt-deflation spiral. That critical leverage — the safe-leverage ceiling — is what shifts with the financing structure. Every figure below is read from one committed run.

Financing regime (no government)Safe-leverage ceiling (private debt ÷ GDP)
All-real credit (baseline)≈ 200%
20% financial / SPV / circular credit≈ 175%
40% financial / SPV / circular credit≈ 120%

Routing 20% of credit to these financial positions pulls the safe ceiling from 200% of GDP down to 175%; routing 40% pulls it to 120%. For scale: US private debt peaked near 175% of GDP in 2008. The more of the build-out's financing is circular, off-balance-sheet, and collateralized by depreciating hardware, the less leverage the system can carry before the same debt-deflation dynamics take over.

Bar chart — the safe-leverage ceiling (private debt divided by GDP) an economy can carry before it tips into a debt-deflation spiral: 200% under all-real credit, 175% at 20% financial credit, 120% at 40% financial credit; with counter-cyclical government the ceiling rises to 245% (all-real) and 220% (20% financial). The financing parameters are disclosed illustrative dials.

Counter-cyclical government spending does what Keen's own model says it should: it raises the ceiling — to ≈ 245% for the all-real economy and ≈ 220% for the 20%-financial one — because a slump automatically widens the deficit, which props up profits, revives investment, and slows debt growth. A fiscal backstop buys real headroom. The next section is where that backstop meets its limit.

The repricing shock: who does the government rescue?

Now add the feature 2008's collateral didn't have. We model a GPU-collateral repricing as a jump in the real interest rate of +2 percentage points partway through the run — the financing-cost consequence of pledged hardware losing value faster than the loans amortize. We point it at economies that are only moderately leveraged (140–170% of GDP) — comfortably below every no-shock ceiling above, so absent a shock they all survive.

The repricing shock changes that, and the pattern is the punchline:

Matrix — a repricing shock (real rate +2 percentage points) against a moderately-leveraged economy. With no shock, all four financing regimes survive. Under the shock, the all-real-credit economy collapses and the 20%-financial economy collapses; adding counter-cyclical government rescues the all-real-credit economy (survives) but the financial-credit economy still collapses even with government.

  • Without a shock, all four regimes survive — moderate leverage is fine on a calm day.
  • The +2pp repricing tips them. At these leverages the shock alone is enough to push both the all-real and the financial-credit economies into collapse.
  • Government rescues the real-credit economy — but not the financial-credit one. Add counter-cyclical government and the all-real-credit economy survives the shock. The financial / SPV / circular economy still collapses, even with the same fiscal backstop.

That last line is the finding worth carrying out of this piece. The fiscal rescue that stabilizes an economy whose debt is buying productive capacity cannot stabilize one whose debt is buying leverage — the government props up profits, but these financial positions were never generating the cash flow that profits feed; they were rolling debt against future returns that the repricing just cancelled. The same dial that lowers the tipping point also defeats the rescue.

How the simulation works

The runtime is Steve Keen's monetary-Minsky model in its cleanest three-variable form — employment λ, wage share ω, and the private-debt ratio d_r — with his own generalized-exponential behavioral functions at pivot slope s = 2, a Kalecki markup inflation channel, and stock-flow-consistent accounting. The accounting choice is load-bearing: it uses the correct Bank-Originated-Money-and-Debt structure, in which a bank loan is simultaneously a bank asset and a firm liability. Credit therefore does not cancel out — it is part of aggregate demand — unlike the neoclassical "Loanable Funds" table, where it nets to zero and disappears. That is why debt has real macroeconomic consequences in this model rather than being a wash.

The model's core equations — the three state laws of motion (employment λ̇, wage share ω̇, private-debt ratio ḋ_r), the Kalecki markup inflation equation, and Keen's generalized-exponential behavioral functions — each tagged as derived from the stock-flow-consistent accounting or as a behavioral closure.

Reading the notation. Two conventions carry most of the figure. A dot over a variable is its rate of change over time — λ̇ means dλ/dt, how fast employment is moving. And GEF(·) is Keen's generalized-exponential reaction function, written GEF(input; pivot-x, pivot-y, slope, floor): the semicolon separates the input on the left from four shape settings on the right, always in that order. The curve is pinned to pass through the point (pivot-x, pivot-y), takes the given slope there, and flattens toward the floor as the input falls away. So the investment rule GEF(π_r; 0.11, 0.36, 2, 0.03) reads: at an 11% profit rate firms invest 36% of output, the response steepening at slope 2, and never dropping below a 3% floor. The wage rule GEF(λ; 0.90, 0, 2, −0.04) reads: at 90% employment wage growth is zero, and it bottoms out near −4% if employment collapses.

The remaining symbols, once: λ employment rate · ω wage share of output · d_r private debt ÷ GDP · π_s profit share · π_r profit rate · i_G investment share of output · g_r real growth · i inflation rate · c credit flow (new borrowing ÷ GDP) · ν capital-to-output ratio · δ depreciation · r interest rate · m_L wage markup · α, β productivity and workforce growth · g_d counter-cyclical government deficit · f the credit dial (share of new lending routed to financial / AI-capex borrowers).

Those are the three state equations, the price channel, and the behavioral closures — each tagged as accounting-derived or behavioral. Underneath them is a stock-flow-consistent discipline, and it is worth being exact about what this reduced runtime actually enforces every step — to machine precision — and what it deliberately leaves out.

A four-sector transaction-flow matrix — firms, workers, bankers, government. Flows are booked as double entries (a payer −, a payee +) — value-added enters as the one single-sided source — and each sector column must close: its income is exactly absorbed by its saving. In one period firms generate value added (normalized to 1), pay wages ω to workers and interest r·d_r to bankers, receive the counter-cyclical government deficit g_d into profit, and retain π_s — the government financing g_d by issuing public debt. The column that has to balance gives the model's income identity:

ω + (π_s − g_d) + r·d_r = 1 — wages + core profit + banker's interest exhaust output.

One balance-sheet identity — the firm's, the only balance sheet the runtime carries:

assets K = ν·Y_n (real capital) · liabilities D = d_r·Y_n (loans) · equity E = K − D, checked as A − L − E = 0.

When equity K − D goes negative — debt above the real capital it financed — that is the Minsky insolvency the cycle drives toward, shown in the accounting rather than asserted. This is Bank-Originated-Money logic: a loan is at once a firm liability and a bank claim, so credit does not net out of demand — the flow of new borrowing c = i_G − π_s (investment the firm cannot fund from retained profit) is money the bank creates and the firm spends. That is the whole difference from the neoclassical "Loanable Funds" view, where the same loan is one saver's money lent on, adding no new spending.

What this reduced runtime does not carry — stated plainly, because the piece is about financing. Workers and bankers hold income as a single running balance, not a portfolio; there is no separate deposit/money stock and no equity-issuance channel — a firm cannot raise cash from investors against a share liability. External finance here is therefore bank credit only, and firm equity is a residual net worth, not a claim investors hold. So the build-out's heavily equity-financed character — the $500B investor platforms — enters this model only through the credit dial f, not as an investor-equity entry on a balance sheet. The full money circuit and per-sector balance sheets live in the platform's fuller stock-flow-consistent runtimes; here we keep the smaller, honest object — three state laws whose flows close every step.

One switch — whether the price (Kalecki inflation) channel is coupled in — flips the model between two regimes that together tell the whole story:

  • Price ON → bistable. A stable "good" equilibrium (a spiral sink; the leading complex eigenvalues of this calibration's Jacobian are −0.259 ± 0.726j, alongside a third real root) coexists with a separate debt-deflation collapse basin. Below a critical leverage the economy settles; above it, it runs away. This is where the US actually sits — and it is the regime behind the safe-leverage-ceiling and repricing-shock results above.
  • Price OFF → Keen's rising-cycle "march." Remove the price stabilizer and the good equilibrium repels (its leading complex pair gains a positive real part, +0.010 ± 0.674j): cycles that start calm grow — a "Great Moderation → Great Volatility" — and eventually escape into collapse. This is the underlying private-sector instability with nothing holding it in check.
  • + counter-cyclical government → the correction. Turn the government arm on during the march and the runaway becomes a peak-then-mean-revert: debt peaks around 1.54× GDP, then reverts to ≈0.75×. The economy survives. Government converts the repelling equilibrium into a stable one — Keen's "Big Government prevents debt-deflation."

That progression reproduces Keen's own published sequence — the no-price march, then the price channel stabilizing it into a bounded cycle, then counter-cyclical government preventing the collapse — and every behavior here is dt-converged (identical at successive step-sizes), so these are structural features of the model, not integration artifacts. (The equations and accounting identities above are the model in full; the phase-space and cycle diagrams live in the runtime reference.)

How it's grounded to the real situation

A model like this earns trust by matching the shape of the real data, not by fitting a number. Three checks, all against committed FRED and BIS series:

US private debt does not run away — it mean-reverts. It built up to ~173% of GDP at the 2008 peak, then corrected, and now sits at ~137% (2025 Q4) — the lowest in about twenty years, and still falling (BIS credit to the private non-financial sector ÷ GDP). Started at the empirical 2008 peak with the government arm on, the model autonomously deleverages in step — the same build-up → correction → fluctuation. That is precisely the signature of a bistable model with a government-rescue arm, and it is evidence against a globally-unstable runaway.

US private-debt-to-GDP, actual (FRED) versus simulated (keen_monetary_minsky_v1, government arm on, started at the 2008 peak): both climb to about 173% in 2008 and correct to about 137% by 2025, staying well below the model's tipping-leverage lines. The 2008 peak sits inside the good, mean-reverting basin — a correction, not a debt-deflation runaway.

The tipping lines that figure marks (≈198% without a fiscal backstop, ≈241% with one, and the financial-credit dial pulling the no-backstop line to ~181% at a 15% financial-credit share) are the same thresholds the ceiling chart reports as ≈200% / ≈245% and the 20%-financial-credit ceiling of 175% — two independent committed runs of the same model agreeing to within a couple of points of GDP.

The single most Keen-recognizable fact reproduces on real data. Computed from the committed FRED series, the correlation between credit (Keen's flow — the annual change in private debt, % of GDP) and unemployment is ≈ −0.40 over the whole period and −0.92 in the high-debt 1990–2015 window — reproducing Keen's own published −0.41 / −0.92.

Scatter of credit (the annual change in private debt, percent of GDP) against unemployment on real US FRED data, one point per quarter: the relationship is negative over the whole period (r ≈ −0.40) and steep in the high-debt 1990–2015 window (r ≈ −0.92), reproducing Keen's published figures.

And it holds worldwide. Across the BIS panel, aggregate private debt plateaus and mean-reverts — it does not run away. Advanced economies and the global aggregate are correcting from their 2020 COVID peak; Japan has been correcting for three decades since its 214%-of-GDP 1993 bubble; the US is falling. China is the one secular-rise outlier — ~201% and only now plateauing — which is exactly where a concentrated, not-yet-deleveraged build-up shows up, and it maps to the model's financial-credit dial.

BIS private-debt-to-GDP for the US, China, Japan, advanced economies and the global aggregate: everywhere the series builds up then corrects or plateaus rather than running away, with China (~201%) the lone secular-rise outlier and Japan (which peaked near 214% in 1993) the textbook decades-long correction.

Which sets up the honest framing this whole piece turns on. US aggregate private debt is low and falling — so "the US is dangerously leveraged" is simply wrong at the headline level. The AI-capex risk is not in that aggregate; it is the concentrated, off-balance-sheet layer — the GPU-collateral SPVs and circular vendor financing — that the private-debt-to-GDP number doesn't capture, and the question of whether it reverses the deleveraging in the tech and corporate sector. That concentrated layer is precisely what the model's financial-credit dial represents. And the stability we observe in the data is produced by the policy response — the government arm — not by the economy being inherently safe. Take the backstop away, or let the financial-credit layer grow, and the same model shows how the calm ends.

The empirical anchor — used as a benchmark, not as proof

The real-data scatter above is not, on its own, proof — and the model earns its keep by reproducing that same relationship internally. In its bounded credit cycle the model's own credit↔unemployment correlation runs −0.42 to −0.58, bracketing Keen's whole-postwar −0.41: the relationship emerges from the behavioral equations, not just from a data fit. We hold to the conservative whole-period figure and read the steep high-debt-window number (Keen's own −0.71 / −0.92, the 1990–2015 regime) as the AI-capex-relevant one — tightest exactly when leverage is high — without headlining it. A levels correlation on persistent macro series is not causal identification (Granger–Newbold); its role is to benchmark the mechanism, not to stand in for a forecast.

We also read both sides. Keen's headline "aggregate demand includes the change in debt" drew a formal symposium of critiques (Fiebiger, Palley, Lavoie); Keen replied to all three (ROKE, 2015), re-deriving the result from stock-flow-consistent accounting rather than a redefinition. And the mechanics under all of it — that banks create money by lending — were confirmed by the Bank of England's 2014 "Money creation in the modern economy." Our contribution is not the endogenous-money claim (that is standard); it is the AI-capex application of the real-versus-financial credit split, run inside an accounting-consistent model with the dials on the table.

What's Keen's, and what's ours. The economics here is Steve Keen's — his monetary Minsky model: the endogenous-money credit cycle, the generalized-exponential behavioral functions at pivot slope s = 2, the stock-flow-consistent accounting, and the Fisher debt-deflation channel. We didn't rewrite his economics; we implemented it faithfully — bit-for-bit against a validated prototype — and added four things around it: (1) a dt-convergence proof that the reported numbers aren't integration artifacts — the direct answer to a standing critique of this class of model; (2) a real-vs-financial credit split — an SPV / circular financial-credit dial (Giraud & Grasselli 2019) that makes the model speak to the AI build-out's financing structure, and that both lowers the safe-leverage ceiling and defeats the fiscal rescue; (3) decision-grade uncertainty tooling — Monte-Carlo ensembles with confidence intervals and basin maps, for questions that call for a distribution rather than a single trajectory; and (4) an automated stock-flow-consistency gate — a four-sector transaction-flow matrix whose sector columns close every step, plus the firm's A−L−E balance identity — that every run must pass. Around those: a concrete mapping of the actual 2026 financing structures; composition with fiscal, securitization and cohort runtimes; and — where a call has an external scoreboard — a sealed public forecast ledger that makes it checkable (this piece deliberately seals none, because an AI-bubble call has no scoreboard). We didn't improve on Keen's economics; we operationalized it for one concrete 2026 question.

The honest limits

Everything above is honest about what it is, because that is what makes it usable:

  • Three disclosed dials, no fitted calibration. The results turn on exactly three knobs, all disclosed: the financial_credit_fraction split, the +2pp repricing shock, and the counter-cyclical government rule. There is no fitted AI-capex or AI-SPV calibration — the dials are set to place the regime near the bistable boundary so the mechanism is visible.
  • Results are relative, not levels. Every number is a comparison between regimes — 200% versus 175% versus 120%, survive versus collapse. None of them is a forecast of any lender's book or any year.
  • The model is a stylized reduced form. It is a distributive credit cycle (employment × wage share) with a debt and price-level channel — not a full multi-sector ledger. It has no default (in a deep slump firms are made to borrow to pay interest), so in a pure collapse the credit↔unemployment correlation actually flips positive — a limitation the validation report discloses rather than hides, and the reason we anchor the correlation to the bounded-cycle regime. The Goodwin core also means the downturn is a profit-squeeze that debt amplifies, not a debt-overhang the model "proves" (Bovari, Giraud & McIsaac, 2017).
  • Consistency is not correctness. The runtime's accounting identities close to machine precision — but that proves the books can't silently break, not that the behavioral story is right. This is a structural decision-support tool, not a calibrated forecaster. It is the Kocherlakota discipline: get the structure right and disclose the dials rather than curve-fit a precise-looking number.

That is why this piece — like the fragility study before it — prints no "AI-bubble probability." A number sealed against no external scoreboard would be false precision. What the model does say, and can defend, is the shape of the risk: routing credit into financial / SPV / circular positions lowers the leverage the system can carry before debt-deflation takes over (200% → 175% → 120%), and when a collateral repricing hits, the counter-cyclical government that rescues a real-investment economy cannot rescue the financial-credit one. (The model proposes the mechanism; the disclosed dials decide what we can honestly claim.) The distributional question — which tranche and which holder eats the loss when GPUs reprice — belongs to a separate securitization runtime and is reported on its own.

What a leader does with this

If AI infrastructure is in your book — as a lender, an allocator, an insurer, or a counterparty — the useful questions are now concrete: What share of your credit exposure is funding productive capacity versus inflating leverage against depreciating collateral? How much of a collateral-repricing (rate) shock does your book assume the system can absorb before the financing turns pro-cyclical? And are you counting on a policy backstop that — in this mechanism — stabilizes real-investment credit but not the circular, off-balance-sheet kind? Standing up a properly-scoped version of this analysis against your exposure and your pass-through assumptions is a bounded simulation decision engagement. The deliverable is a trade-off you can interrogate — with every assumption on the table — not a headline about a bubble with no scoreboard.

Bottom line. Achieved: representing the NVIDIA-platform leverage, circular vendor financing and GPU-collateral SPVs as one disclosed dial in a validated port of Steve Keen's debt-deflation model, the safe-leverage ceiling falls from 200% of GDP under all-real credit to 175% (20% financial) and 120% (40%) — and under a +2pp GPU-repricing shock, counter-cyclical government rescues the all-real-credit economy but not the financial-credit one. Business value: a decision-grade, mechanism-level read on a financing structure regulators are only starting to see — grounded in Keen's own model, every assumption disclosed as a dial, and deliberately no fabricated bubble probability.

Companion piece (planned): the same real-versus-financial credit lens turned on other deregulation cases — crypto / DeFi and defense / rearmament financing.

Comuvia builds decision support on owned, accounting-consistent simulators. This piece represents the AI-capex financing structure as disclosed illustrative dials in a validated port of Steve Keen's debt-deflation model to show the mechanism; no fitted calibration is claimed and no AI-bubble probability is quoted. Every number is a relative comparison between runs. Not investment advice.