If AI Fails or Succeeds: What Changes for Finance

01 · Executive summary

The question "what if AI fails?" is badly posed until failure is defined. The technology cannot now un-happen: measured deployment in fraud detection, document automation and research is real and cost-justified at current capability (Paper 1). What can fail is the investment thesis, the proposition that roughly 2.9 to 6.7 trillion dollars of projected data-centre capital (Morgan Stanley through 2028; McKinsey through 2030, modelled) will earn its cost of capital through transformative productivity gains. Success and failure are therefore economic branches, not technological ones.

Our conclusions run against both optimistic and pessimistic consensus. If AI succeeds economically, finance faces margin compression, labour restructuring and a machine-native demand layer for money, with the productivity dividend accruing mostly to users and consumers rather than to the AI sector's investors; success for the economy and success for today's capital structure are different claims. If AI fails economically, finance absorbs an equity repricing whose stress-test magnitude the Bank of England has already modelled at 45%, plus a credit event concentrated in the newly built private-credit and off-balance-sheet financing web, yet the deployed operational layer of AI survives, exactly as the fibre of 2000 survived the telecoms crash and became the substrate of the internet economy. The deepest finding is the asymmetry of memory: in both branches, the financial system of 2040 runs on AI operations; what differs is who paid for the transition and which decade's investors were destroyed doing so. Postulated

02 · Defining the branches

PostulatedWe define the branches by measurable end-states, so that the definition itself is falsifiable rather than rhetorical.

Definitions: EWC. Anchors: Bain Global Technology Report, September 2025 (modelled); Acemoglu, NBER working paper 32487, 2024 (modelled); Bank of England Financial Stability Report, July 2026 (observed).

The scale at stake is now macroeconomic, not sectoral. Observed The four largest hyperscalers spent roughly 410 billion dollars of capex in 2025, with plans clustering near 725 billion for 2026 (FT compilation, April 2026; definitions vary and the range runs 690 to 770 billion). Investment in information-processing equipment and software accounted for 92% of US GDP growth in the first half of 2025, on Jason Furman's decomposition of official data, with the economy excluding those categories growing at 0.1% annualised. The BIS put combined hyperscaler AI capex above 1 trillion dollars for 2025 and 2026 together and reached for the canal mania, the railway mania and the dot-com era as its comparators (June 2026). Whatever branch obtains, the macro exposure is already booked.

03 · The success branch

What success looks like by 2040

PostulatedSuccess means the Goldman-class estimates (roughly 7% added to global GDP over a decade, 1.5 percentage points of annual US productivity growth, modelled 2023) prove closer to reality than the Acemoglu range. Traced through the financial system, four consequences dominate.

• The cost of intermediation collapses, and with it a business model. McKinsey's modelled 200 to 340 billion dollars of annual value in banking arrives mostly as competition, not margin: AI-era cost curves get competed into pricing, as electronic trading's did (Paper 2). Mid-sized institutions without proprietary data or distribution lose the cross-subsidies that funded their overheads. Expect consolidation of the banking long tail on a scale comparable to the post-1990s US branch consolidation.
• Finance employment restructures from below. The earliest measured labour effect is already visible: a roughly 13% relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations since late 2022 (Stanford Digital Economy Lab, August 2025, observed; causality debated). Success scales that pattern through the middle office. The institutional knowledge pipeline, in which juniors learn by doing the work AI now does, becomes a strategic problem for every firm and a quiet systemic one for the industry.
• A machine-native monetary layer emerges. If agentic commerce compounds from its current beachheads (live but constrained as of mid-2026, Paper 1), software agents become high-frequency users of money, favouring programmable, instantly settling instruments: stablecoins, tokenised deposits, and whatever the BIS's unified-ledger programme produces. Machine demand is indifferent to branding and loyal only to settlement quality, which advantages whoever offers the best legal-finality-per-millisecond, a competition states will not concede to private issuers without a fight. Success accelerates every CBDC and tokenisation timeline in the EWC 2040 atlas.
• The dividend lands with users, not necessarily with today's shareholders. Token-level inference prices for constant capability have fallen at rates between 9x and 900x per year depending on task (Epoch AI, observed). Success built on collapsing unit prices is success in which the consumer surplus is enormous and the producer surplus is contested; the railway precedent, transformative for the economy and ruinous for most railway securities, remains the base case for the split. Positioning inside the supply chain matters more than the aggregate outcome.

The energy and sovereign layer of success

ObservedSuccess is also a physical event with a financial shadow. The IEA projects data-centre electricity consumption more than doubling from roughly 415 terawatt-hours in 2024 (about 1.5% of global electricity) to roughly 945 terawatt-hours by 2030, more than Japan's total consumption, with AI the principal driver; US data centres alone would account for nearly half of US electricity-demand growth to 2030. In the success branch these curves steepen further, and the financing of generation, grid and siting becomes a structural fixed-income and infrastructure asset class in its own right. Compute and the power behind it become strategic resources in the way refining capacity was for the twentieth century, which pulls AI into the sovereign-risk frameworks EWC applies elsewhere: export controls, siting nationalism, and energy-security coupling between the AI complex and national grids. Finance's exposure to AI, in this branch, increasingly runs through utilities, power purchase agreements and sovereign industrial policy rather than through technology equity alone. Postulated

Success stresses the system too

PostulatedThe success branch is not the safe branch; it is the fast branch. It compresses the labour adjustment into a decade, hands supervisory-perimeter questions (model concentration, correlated behaviour, agentic payments fraud) to regulators at maximum speed, and, if productivity gains lift equilibrium interest rates, repriced duration becomes the sting in the good news. A success that arrives faster than institutions can govern it converges, in stability terms, with failure.

04 · The failure branch

The exposure map, as it stands

ObservedFailure transmits through four documented channels, each of which has grown an order of magnitude since 2024.

• Equity concentration. AI-linked stocks approached 45% of S&P 500 weight on Goldman Sachs's April 2026 count and "around half" on the Bank of England's July 2026 count. Morgan Stanley attributed roughly 75% of index gains since late 2022 to AI-linked names (October 2025). A failure-branch repricing is not a sector event; it is an index event, transmitted instantly to every pension scheme and passive vehicle in the world.
• Credit and the new shadow structure. The five hyperscalers were 3% of outstanding US investment-grade debt at end-2025 but over 15% of year-to-date issuance by May 2026 (BoE). Moody's warned in July 2026 that AI spending now threatens the credit quality of the majors, who are shifting from asset-light to asset-heavy models. Beneath them sits the leveraged tier: roughly 24.9 billion dollars of debt at the largest neocloud, much of it GPU-collateralised private credit, and record-scale SPV structures such as the 27 billion dollar Meta and Blue Owl data-centre joint venture. The BoE flags the circularity explicitly: AI revenue forecasts "may reflect circular financing arrangements." Failure here is a private-credit and insurance-balance-sheet event, in the least transparent corner of the system.
• Macro dependence. With AI-linked investment carrying most of measured US growth (92% of H1 2025 growth on Furman's decomposition), a capex stop mechanically subtracts from GDP while the wealth effect of the equity repricing subtracts from consumption. The BIS states the historical pattern plainly: such investment episodes "ended with an eventual reversal in investment, inducing economy-wide recessions."
• Duration mismatch. Long-dated debt is financing assets whose useful life is contested: the depreciation dispute (Burry's estimate of 176 billion dollars of understated depreciation across the majors for 2026 to 2028, against evidence that older accelerators remain revenue-productive for years) is at bottom an argument about whether the collateral behind the new credit stack holds value for six years or three. The BoE flags the same mismatch in official language.

The stranded-asset question

PostulatedFailure adds a category the 2000 precedent only partially rehearsed: physically massive, rapidly depreciating, power-hungry assets whose salvage value depends on the very demand scenario that failed. Fibre from 2001 stored well; accelerators may not, and the open depreciation dispute (Section 4) is in essence a stranded-asset argument conducted in accounting language. The counter-evidence deserves equal weight: prior-generation accelerators have remained revenue-productive for four to six years, and collapsing inference prices create demand at every price point on the way down. Our judgment: partial stranding concentrated in the leveraged tier, with the hyperscaler-owned core absorbed internally, which is precisely why the ownership structure of the compute, on balance sheet versus SPV versus vendor-financed, matters more than its aggregate quantity. The failure branch's true novelty is that it would be the first infrastructure bust in which the collateral's useful life was contested at the moment of underwriting, in public, by the underwriters' own auditors and critics.

What failure does not do

PostulatedThree things survive the failure branch, and the historical record is the basis for each. Deployed operational AI survives, because at collapsed inference prices the fraud, AML and automation economics of Paper 1 clear their hurdle rates even in a capex winter. The infrastructure survives and cheapens, as dot-com fibre did; distressed compute becomes the feedstock of the next cycle, and the Amazon precedent (down roughly 94% in the crash, then the defining company of the following two decades) is the standing reminder that a bubble can be right about the technology and wrong about the price. And the state's agenda survives: CBDC, tokenisation and payments modernisation programmes are policy-driven, not valuation-driven, and a private-capital winter historically strengthens the relative position of official infrastructure. Failure, in other words, redistributes the transition's cost; it does not cancel the transition.

The two branches, phased on the EWC timeline

PostulatedMapped onto the site-wide phasing convention (2026 to 2030, 2030 to 2035, 2035 to 2040), the branches diverge earliest in credit and latest in operations.

05 · The reflexive loop: psychology as transmission mechanism

Between the two branches sits the mechanism that will choose between them, and it is reflexive in the strict Soros sense: prices are changing the fundamentals they are supposed to reflect. The loop currently runs as follows, each link documented. High AI-linked equity prices lower the cost of capital for the build-out; the build-out is itself the largest component of measured US growth (92% of H1 2025 growth on Furman's decomposition); measured growth validates the earnings outlook that supports the equity prices; and the equity prices collateralise, directly and through sentiment, the expanding credit stack that funds the next round of capex. The loop also runs through compensation, vendor financing and the government's own fiscal projections. Observed

Reflexive loops are self-reinforcing until a link breaks, and self-limiting conditions are already visible at two links. The first is the market's response function to capex itself: in July 2026, one hyperscaler's shares fell roughly 15% in days on a capex-heavy report while two peers rose 8 to 10% after tying identical spending to demand evidence. The market has begun demanding revenue attribution for capital expenditure, which is how every investment mania's financing window starts to narrow. The second is the official-sector narrative: when the ECB lists disappointing AI news as a stability trigger and the BIS reaches for the railway analogy, the reflexive loop acquires supervisory observers whose warnings themselves move the loop. Observed

PostulatedThe branch selection, in other words, will not be made by a laboratory result. It will be made by whether the credit link or the revenue link breaks first: revenue closing the gap sustains the loop into the success branch; financing conditions closing first collapses it into the failure branch. This is why Paper 4's watchpoint hierarchy ranks credit indicators above capability indicators, and why sentiment, usually a decorative chapter in technology forecasting, is load-bearing here.

06 · Pre-mortem: reading 2031 from the failure branch

PostulatedThe EWC discipline for large theses includes a pre-mortem: assume the failure branch obtained, stand in 2031, and write the post-crisis consensus backwards. The exercise is postulation by construction, and its value is that every sentence corresponds to something measurable today.

The 2031 retrospective would likely read: the warnings were public and dated. Issuance data showed AI borrowers taking over 15% of investment-grade supply while holding 3% of the stock; off-balance-sheet commitments had grown eightfold in four years to 1.65 trillion dollars; the largest neocloud's credit was pricing distress a full cycle before the write-downs; the depreciation dispute was an open argument about whether the collateral was worth half its book value; and the sector's own leaders had called it a bubble on the record. The consensus would conclude, as it did after 2001 and 2008, that the information was available and the incentives to act on it were not, because every intermediate actor, vendor-financiers booking sales, private-credit managers deploying committed capital, index investors holding by mandate, was individually rational inside a collectively fragile structure. The pre-mortem's practical yield is a short list of things a research house should therefore track in public, quarterly, with dates: the issuance share, the off-balance-sheet total, the neocloud credit pricing, and the depreciation schedules. Paper 4 formalises exactly that list.

07 · The asymmetry, actor by actor

PostulatedThe two branches, read side by side, produce an actionable asymmetry table. The reader should note which cells are identical across branches; those are the robust positions.

08 · Conclusions

First, "AI fails" and "AI succeeds" are claims about capital, not capability. The measured capability layer already pays for itself and is therefore branch-invariant; the 1-trillion-dollar-a-year investment layer is not, and the gap between those two layers is where the decade's financial history will be written.

Second, the failure branch's transmission map has moved. In 2000 the exposure sat in transparent public equity multiples; in 2026 the equity multiples are more defensible (the market's largest AI name trades near 30x trailing earnings against Cisco's roughly 130 to 200x forward at its peak) while the leverage has migrated into private credit, SPVs and off-balance-sheet commitments that supervisors can only partially see. The next stress event will be discovered, not announced. Postulated

Third, the view from Europe, where EWC writes: the branch asymmetry is sharper here than in the United States, and in an uncomfortable direction. Europe holds a small fraction of the AI capital stack, so the success branch delivers the productivity dividend largely through imported technology while the labour adjustment and the concentration risk arrive undiluted; the failure branch, transmitted through globally correlated indices, private-credit exposure at European insurers, and the trade cycle, arrives with almost no offsetting ownership of the surviving assets. The policy conclusion follows directly: for jurisdictions without hyperscalers, the rational strategy is aggressive adoption of the branch-invariant layer (operational AI, payment rails, verified-data infrastructure) combined with active supervisory attention to imported credit exposure, which is, in compressed form, the agenda the EU's regulatory architecture is groping toward. Postulated

Fourth, the success branch deserves more fear than it receives, and the failure branch less. Success compresses labour and governance adjustment into years; failure mostly repeats a well-rehearsed historical script whose survivors are known in advance. Paper 4 of this series assigns the probabilities across these branches; Paper 5 tests the bubble question directly against the evidence.

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. Conditional scenarios are analytical constructions, not predictions; the Bank of England's 45% scenario is a stress test, not a forecast, and is cited as such. Figures carry their as-of dates and may have changed. © EWC Investments, 2026.