AI and Finance: The Most Probable Scenarios to 2040

01 · Method: priors before opinions

EWC scenario probabilities are built in three steps, in the order stated, to keep judgment honest.

• The prior comes from the reference class. The relevant class is major infrastructure investment manias attached to real technologies: British canals in the 1830s, railways in the 1840s, electrification and utilities in the 1920s, telecoms and the internet in the 1990s. The BIS itself invoked precisely this class in June 2026. In every completed case, the technology was real, the investment overshot, a repricing followed, and the infrastructure was absorbed at written-down cost by a subsequent, larger economy. In no completed case did the technology simply vanish; in no completed case did the first wave of investors, in aggregate, earn their cost of capital. That is a strong prior: repricing is the historical norm, extinction and unbroken triumph are both exceptions.
• Evidence updates the prior. Against the reference class, the current episode has documented differences in both directions. Supporting stability: the largest AI firm trades near 30x trailing earnings against Cisco's roughly 130 to 200x forward multiple in 2000, and most capex is funded from operating cash flow of firms with fortress balance sheets. Supporting fragility: revenue remains far below capex (Bain's estimated 800 billion dollar annual shortfall, modelled), leverage has migrated into private credit and off-balance-sheet structures (1.65 trillion dollars of commitments on the Nikkei tally), AI-linked names are roughly half of the S&P 500 (Bank of England, July 2026), and AI-linked investment carried 92% of measured H1 2025 US GDP growth. Net effect: the equity-multiple channel is less stretched than 2000; the credit and concentration channels are more stretched.
• Watchpoints keep the forecast alive. Section 5 pre-registers the observable events that will move these probabilities, with resolution criteria, so revisions are mechanical rather than rhetorical.

02 · Market psychology and sentiment regime

EWC classifies market sentiment into seven regimes (despondency, skepticism, hope, optimism, belief, euphoria, complacency), and the classification is justified with reference to specific indicators, not impressions. As of 3 August 2026, we classify the AI-and-finance complex as belief, with euphoric pockets and complacent credit. The evidence for each clause:

• Belief: the trend is treated as obvious and allocators chase it. AI-linked names are roughly half of S&P 500 capitalisation (Bank of England, July 2026), passive mandates transmit the concentration automatically, and capex guidance increases were rewarded through most of 2025 to 2026 wherever demand evidence accompanied them. Observed
• Euphoric pockets: valuation untethered from current cash flows in specific segments rather than the index core: private AI-lab marks (an 830 billion dollar valuation against 13 billion of booked revenue at the leader), vendor-financed neocloud structures, and warrant-linked deals priced for flawless deployment schedules. Observed
• Complacent credit: record AI-linked issuance absorbed smoothly (225 billion dollars in H1 2026) even as the same market priced distress in the leveraged tier and Moody's warned on the majors' credit quality; long-dated debt continues to fund short-lifecycle assets at spreads that do not obviously price the depreciation dispute. Observed

Two features distinguish this regime from the 1999 template and matter for the probabilities below. First, the warnings are coming from inside: the sector's own leaders have called it a bubble on the record (the leading lab's chief executive in August 2025; the Amazon founder's "industrial bubble" formulation in October 2025), which was not a feature of 1999 and suggests narrative saturation rather than narrative innocence. Second, official-sector warnings (IMF, Bank of England, the Federal Reserve chair's valuation remark) have been issued and absorbed without deleveraging, which historically indicates a market that has decided the authorities will not act on their own warnings. Behavioural diagnosis: anchoring on the post-2022 trend, herding through index structure, and base-rate neglect on infrastructure-mania outcomes. The regime classification feeds directly into the probability weights: belief regimes can run for quarters or years, but they do not resolve into despondency without passing through the repricing this forecast centres. Postulated

03 · The four scenarios

PostulatedProbabilities as of 3 August 2026. Basis: reference-class prior, updated on the evidence recorded in Papers 1 to 3.

Base case, 50%: absorption with a violent repricing

The technology keeps working and keeps spreading; the capital structure does not survive intact. At least one drawdown exceeding 30% in the AI-linked equity cohort occurs before 2031, accompanied by visible distress in the leveraged tier of the build-out (neoclouds, GPU-collateralised credit, selected SPVs) but no systemic banking crisis, because the core of the financing sits on cash-rich balance sheets and the banking system's direct exposure remains bounded. Post-repricing, surviving infrastructure is absorbed at lower cost, adoption continues on the operational economics documented in Paper 1, and measured productivity gains emerge in the 2030s in the band between the Acemoglu floor and the Goldman ceiling, closer to the middle: roughly 0.5 to 1.0 percentage points of annual productivity contribution at maturity.

• 2026 to 2030: capex growth decelerates; the revenue gap forces consolidation among model providers; the repricing event lands in this window on the base path; the EU high-risk AI regime takes effect; first Calibration-style supervisory frameworks for AI concentration appear.
• 2030 to 2035: absorbed-infrastructure economics: cheap inference industrialises the operational layer of finance; agentic payments cross from pilot to material niche; productivity effects become statistically visible.
• 2035 to 2040: AI-saturated operations, human-governed authority (Paper 1's 2040 map); market structure consolidated around machine intermediaries (Paper 2); the episode enters the textbooks beside the railways, as the mania that built the grid.

Bull case, 20%: the compounding decade

Required conditions: AI revenue growth sustains its 2024 to 2026 trajectory (the leading labs' combined run-rates multiplying year over year) long enough to close the Bain gap by roughly 2030; agentic commerce scales to a genuine machine-economy demand layer; productivity gains print at 1.5 percentage points or more annually in official data by the early 2030s. In this branch the capex proves rational ex post, the equity concentration partially justifies itself through delivered earnings, and finance experiences the fast, compressed restructuring described in Paper 3's success branch: consolidation, entry-level labour displacement, machine-native money rails scaling with machine demand. The repricing risk does not vanish; it shrinks to sector rotations inside a rising aggregate. We hold this at 20% rather than higher for one documented reason: every completed reference-class episode included the repricing, and "this time the investment wave clears its cost of capital on schedule" has no historical precedent to anchor on.

• 2026 to 2030: lab revenues sustain triple-digit compounding; agentic payments cross 5% of e-commerce transactions in at least one G7 market; productivity prints begin exceeding trend; capex remains cash-flow-dominant and the issuance share stabilises.
• 2030 to 2035: the Bain-scale revenue requirement is substantially met; AI-infrastructure debt seasons into an investment-grade asset class; financial-sector restructuring runs at maximum speed.
• 2035 to 2040: the build-out is retrospectively judged rational; concentration, labour adjustment and machine-money governance replace valuation as the dominant policy questions.

Bear case, 25%: the long write-down

Required conditions: the revenue gap fails to close and financing conditions turn before the productivity dividend arrives. The sequence runs: capex guidance cuts, then distress in the leveraged tier (the market was already pricing roughly 50% default odds on the largest neocloud's debt in late July 2026, a figure we cite with a verification caution), then write-downs propagating through private credit, insurance balance sheets and the off-balance-sheet web, then the macro round trip as the investment engine that carried most of measured US growth goes into reverse. The equity drawdown in this branch approaches the Bank of England's 45% stress geometry, and the recession the BIS warned of arrives. Finance's AI operations survive on collapsed cost curves, but AI as an investment theme enters a winter of a half-decade or more. We hold this at 25%: the credit and dependence evidence has been strengthening through 2026, but the core funding remains cash-flow-based, which distinguishes the episode from debt-built manias like the railways.

• 2026 to 2030: the sequence runs from guidance cuts through leveraged-tier defaults to the index-level repricing; recessionary impulse follows as the dominant investment engine reverses; policy easing cushions but does not prevent the credit losses.
• 2030 to 2035: write-downs propagate through private credit and insurance; distressed compute is absorbed at deep discounts; AI-as-theme enters an expectations winter while AI-as-operations keeps compounding on collapsed costs.
• 2035 to 2040: the surviving infrastructure underpins a cheaper, more concentrated AI economy; the episode's financial history is written as a credit event that equity commentary spent years mislocating.

Tail case, 5%: regime break

Low probability, high impact, both directions. Negative variants: a correlated-model market event with real-economy transmission before governance matures (Paper 2's adversarial breakdown); a geopolitical compute shock centred on Taiwan repricing the entire hardware complex; an AI-enabled cyber or manipulation event that breaks confidence in market integrity. Positive variant: a capability discontinuity that makes current capex look small, forcing an economic re-planning problem rather than a financial one. We deliberately keep the tail thin: tails are where analysts hide unquantified drama, and the discipline of this series is to keep drama priced.

Probability accounting

PostulatedFor the record and for later grading, the arithmetic behind the headline weights. The reference-class prior, before any 2026 evidence, would assign roughly 65 to 70% to the full mania arc (our base plus bear), 20% to a benign absorption, and the remainder to regime breaks, because that is approximately the distribution of outcomes across the completed episodes. The current evidence moves us off the prior in two directions. The cash-flow funding of the core and the earnings delivery of the leaders shift weight from bear toward base (a repricing that the system absorbs, rather than a systemic unwind). The 2026 credit acceleration shifts weight back toward bear, and is the reason bear holds 25% rather than the high teens. The bull's 20% is above the reference-class base rate of roughly zero for "no repricing at all"; we grant that premium on a revenue compounding speed that has no enterprise-software precedent, and we cap it there because demand growth without pricing power has never yet rescued an investment mania's financiers. Readers who dispute a weight should dispute one of those three moves; that is what stating the accounting is for.

The reference class, quantified

ObservedBecause the prior carries half the weight of this forecast, its inputs belong on the record rather than in a footnote.

Sources: Advisor Perspectives compilation of GDP-intensity estimates, October 2025; DQYDJ drawdown records; Furman/BEA decomposition. The AI row's intensity range reflects two published estimates.

04 · What the weighted forecast implies

PostulatedReading the four scenarios as one distribution produces three probability-weighted statements.

• A major AI-equity repricing before 2031 carries roughly 75% probability (it occurs in the base, the bear and parts of the tail). The productive question is not whether but through which channel: orderly in the base, credit-propagated in the bear.
• The probability that finance in 2040 runs on substantially more AI than today approaches 95% across all scenarios, because the operational layer is branch-invariant (Paper 3). The absorption thesis is the closest thing this series has to a certainty.
• The distribution of investor outcomes is wider than the distribution of technology outcomes. The technology's 2040 range is narrow: somewhere between important infrastructure and transformative infrastructure. The capital range runs from compounding triumph to a 45%-class drawdown with credit losses. Most public argument conflates these two distributions; keeping them separate is this paper's contribution.

05 · Cross-asset implications matrix

PostulatedA scenario forecast that stops at the technology sector is incomplete; the weights above ripple through every major asset class. The matrix states EWC's central qualitative expectation per scenario. It is direction and mechanism, not price targets, and the prescriptive translation of these directions is outside the public layer of EWC research.

Matrix: EWC postulation, direction and mechanism only. The digital-assets row links this series to the EWC Future of Money scenario architecture (EWC-TT-2026-0007).

06 · Steel-manning the opposition

The strongest case against our base case comes from opposite directions, and both deserve their best form.

The bull's best argument: revenue is not failing; it is compounding faster than any enterprise technology in history, with the leading labs' run-rates multiplying within single years, and inference demand growing even as unit prices collapse. Jevons-style demand elasticity could close the Bain gap on schedule. Our response: the argument is strong on demand and silent on margins; collapsing unit prices are documented, pricing power is not, and the reference class turned on exactly that distinction. We weight it into the 20%.

The bear's best argument: the system's stated defence, that capex is cash-flow-funded, is eroding in the data. Debt issuance is accelerating (over 15% of year-to-date US investment-grade issuance by May 2026), off-balance-sheet structures are growing at 8x over four years, and depreciation assumptions are contested. The financing is quietly migrating toward exactly the structure that made historical manias systemic. Our response: correct as a trend, not yet as a level; the hyperscalers were 3% of the outstanding investment-grade stock at end-2025. The trend is why the bear holds 25% and why the credit watchpoints below carry the largest pre-registered probability deltas.

07 · Pre-registered watchpoints

PostulatedEach watchpoint has a resolution criterion, a deadline and a pre-committed direction of probability revision. This is the v2.0 discipline: updates are declared before the evidence arrives, not fitted after it.

Revisions are stated in percentage points, applied at the next quarterly review after resolution, and logged. Where two watchpoints resolve in conflict, the credit watchpoints dominate, per the path-dependency note above.

08 · Conclusions

First, the honest summary of the distribution: the most probable single future (50%) is that the AI build-out follows the full historical arc of its reference class, real technology, capital overshoot, violent repricing, then absorption into a larger economy, with finance experiencing the repricing as an index-level event and the absorption as a permanent operating upgrade.

Second, the asymmetry that should organise attention: a repricing is roughly three times as likely as an unbroken bull path, but survival through the repricing is the historical norm for the technology and the exception for the leveraged capital. Between now and 2031, watching the credit stack tells you more than watching the models.

Third, the commitment: this forecast is stamped, logged and graded in public. When a watchpoint resolves, the probabilities move by the pre-committed amounts, and when the record shows we were wrong, the record will say so in our own pages. That, and not any single number above, is the methodological point of this paper.

EWC Investments Think Tank publishes independent research for educational purposes. EWC does not manage client capital, does not operate an investment fund, and does not provide regulated portfolio management or personalised investment advice. All probabilities in this paper are subjective analytical judgments with their basis stated; they are not market-implied odds and not guarantees. Forward-looking statements are inherently uncertain. Figures carry their as-of dates and may have changed. © EWC Investments, 2026.