The AI Bubble: Financial Reality vs Speculative Fiction |

01 · Executive summary

Is AI a bubble? The question, asked as a binary, produces only noise, because it welds together four separable claims: that the technology is unreal (false on the deployment record), that revenues are illusory (false; they are the fastest-compounding in enterprise history, and still far below capex), that equity multiples match 2000 (false on the arithmetic; the flagship trades near 30x trailing earnings against Cisco's roughly 130 to 200x forward), and that the episode will end without a repricing (unsupported by any completed historical parallel). Audited separately, the claims resolve into a precise diagnosis.

Our conclusion: the AI build-out of 2023 to 2026 is a genuine infrastructure mania in the classical sense, real technology, real revenues, and capital formation running far ahead of both, with the fragility concentrated not where the public debate looks (equity multiples) but where the 1840s and 1920s would tell you to look: in the credit stack, the circular vendor-financing web, the off-balance-sheet structures, and the dependence of measured US growth on a single investment theme. The equity market of 2026 is defensible; the financing architecture underneath it increasingly resembles late-cycle. Probability-weighted (Paper 4), we assign roughly 75% to a major repricing before 2031, and at the same time roughly 95% to AI operations compounding through 2040 regardless. Both numbers, held together, are the honest answer to the bubble question. Postulated

02 · The verified record

ObservedEvery figure in this section carries its as-of date and source; discrepancies between sources are noted rather than smoothed.

The capital side

Growth rate: hyperscaler capex has compounded at roughly 72% annually since mid-2023 and has quadrupled since GPT-4's release (Epoch AI, from SEC filings). Beyond the listed majors sits the announced tier: the Stargate programme, unveiled in January 2025 at up to 500 billion dollars over four years and expanded in September 2025 to roughly 7 gigawatts of planned capacity and more than 400 billion dollars across named US sites. Projections at the system level: Morgan Stanley estimates roughly 2.9 trillion dollars of global data-centre capex through 2028 with a 1.5 trillion dollar financing gap beyond hyperscaler cash flows; McKinsey's base scenario reaches 6.7 trillion dollars by 2030. All of these are modelled or announced, not spent, and are cited here as the scale of intention, not of certainty. The July 2026 earnings round supplied the first clear demand-discrimination signal inside the intention: one major's shares fell roughly 15% on a capex-heavy report while peers that tied identical spending to cloud-revenue acceleration were rewarded, which is the market beginning to price capex conditionally rather than reflexively.

The revenue side

OpenAI booked 13.07 billion dollars of 2025 revenue against a 20.9 billion dollar operating loss, per leaked audited financials reported by the FT (June 2026), with an annualised run-rate near 20 billion stated by its chief executive in November 2025 and roughly 25 billion reported by early 2026. Anthropic's announced run-rate climbed from 9 billion (December 2025) to a reported 47 billion dollars (May 2026), figures that are company-announced month-times-twelve annualisations, not audited revenue, and we label them accordingly. Set against this: OpenAI's disclosed infrastructure commitments total roughly 1.4 trillion dollars across vendors, confirmed by its chief executive in November 2025. The structural arithmetic has two canonical statements: Sequoia's 2024 estimate that the build-out then implied roughly 600 billion dollars a year of required AI revenue, and Bain's September 2025 estimate that 2030 compute demand requires roughly 2 trillion dollars a year, leaving a projected 800 billion dollar annual shortfall even after crediting AI-driven savings. Revenue is real, growing at historic speed, and still an order of magnitude short of the capital being committed against it.

The concentration side

AI-linked companies reached roughly 45% of S&P 500 weight on Goldman Sachs's April 2026 measure, and "around half," up from about a quarter in 2022, on the Bank of England's July 2026 measure. Morgan Stanley's strategist attributed roughly 75% of index gains, 80% of earnings growth and 90% of capex growth since November 2022 to AI-linked names (October 2025). The IMF observed valuations "heading to levels seen during the dot-com boom" (October 2025); the Bank of England's excess-CAPE measure made the same comparison. The multiples themselves, however, are the bulls' best exhibit: Nvidia, the market's flagship at roughly 5 trillion dollars, traded near 30.7x trailing earnings as of 3 August 2026, because earnings grew into the valuation; Cisco at its March 2000 peak traded near 130 to 200x forward estimates, a different metric and a different order of magnitude.

The credit and circularity side

This is where the 2026 record departs most sharply from 2024's. AI-related bond issuance reached 225 billion dollars in the first half of 2026 (S&P Global), with the five hyperscalers at over 15% of year-to-date US investment-grade issuance by May against 3% of the outstanding stock at end-2025 (Bank of England). Nikkei tallied 1.65 trillion dollars of off-balance-sheet GPU and data-centre commitments at five US tech giants, an eightfold rise in four years, exceeding their on-balance-sheet debt; Moody's counted 1.2 trillion of off-balance-sheet liabilities and warned in July 2026 that the spending now threatens the credit quality of the majors. The circular web is documented deal by deal: Nvidia's up-to-100-billion-dollar OpenAI arrangement, later recharacterised by Nvidia's chief executive as "an invitation," "never a commitment," after reports it had stalled; OpenAI's roughly 300 billion dollar Oracle compute contract, which inflated Oracle's disclosed backlog past 455 billion; AMD's warrants granting OpenAI up to roughly 10% of the company at a cent a share, tied to purchase milestones; Nvidia's contracted purchase of 6.3 billion dollars of its own customer's unsold cloud capacity through 2032 (CoreWeave 8-K). The largest neocloud carried roughly 24.9 billion dollars of debt by mid-2026, much of it GPU-collateralised private credit, with late-July secondary reporting, which we could not verify to Tier 1 standard, putting credit-default-swap-implied default odds near 50%. The Bank of England's July 2026 wording covers the whole cluster: AI revenue projections "may reflect circular financing arrangements."

The measurement problem

ObservedBefore any verdict, an honest audit must state how slippery this episode's core numbers are, because the slipperiness is itself evidence about the cycle. Four measurement traps recur. Run-rate revenue, the standard AI-lab disclosure, annualises a single month by twelve; it is not audited revenue, it embeds no churn, and the gap between the leader's 25 billion dollar run-rate and its 13.07 billion of booked 2025 revenue shows the scale of the difference. Backlog figures such as Oracle's 455-billion-dollar-plus remaining performance obligations count contracted future spending by counterparties whose own funding is not yet secured; a backlog is a promise resting on a promise. Capex totals vary by hundreds of billions depending on whether finance leases and which set of firms are included, which is why our 2026 figure is stated as a range with its definitional caveat. And depreciation schedules, the denominator of every profitability claim in the compute economy, are the subject of an open dispute with a nine-figure earnings impact. None of this is fraud; all of it is the characteristic accounting fog of a capital cycle running faster than its measurement conventions, and the same fog was reported, in period, around 1999 telecom swap deals and railway construction accounts. When the numbers themselves become contested territory, that is a cycle signature, not a footnote. Postulated

03 · The fiction file, audited in both directions

ObservedA research house earns its keep by rejecting bad evidence on its own side of an argument. The claims below circulate widely; each fails an audit.

04 · The Kindleberger ledger

The classical anatomy of a mania, displacement, credit expansion, euphoria, distress, revulsion, is a checklist, and checklists exist to be filled in with dated evidence rather than invoked as atmosphere. The ledger as of 3 August 2026:

PostulatedReading the ledger as a position: the episode sits between stages 3 and 4, with the unusual feature that stage 2 is still accelerating while stage 4's first signals are already visible. In the completed historical cases that overlap, credit still expanding as distress signals appear, the expansion phase bought two to four further years in the railways and roughly eighteen months in telecoms. We cite that range as context, not as a timing forecast; Section 5 states why we refuse the timing trade entirely.

05 · Sentiment, and the strange maturity of this mania

Under the seven-regime EWC taxonomy, Paper 4 classifies the complex as belief with euphoric pockets and complacent credit, and this paper adds the feature that makes the present episode historically strange: it is a mania conducted with the lights on. The IMF's managing director invoked the dot-com comparison in October 2025. The Bank of England modelled the 45% drawdown in public. The BIS wrote the recession warning into its annual report. The Federal Reserve chair called equity prices fairly highly valued in September 2025 and markets sold off for an afternoon and resumed. The leading AI executive agreed, on the record, that investors are overexcited. In 1999, scepticism was a career risk; in 2026 it is a consensus position that has so far cost its holders performance. Observed

PostulatedTwo readings of that strangeness compete. The comfortable reading: transparency is protective, the excesses are known, monitored and therefore containable. The uncomfortable reading, which we weight more heavily: universal awareness without behavioural change is itself the late-cycle signature, because it means every holder believes they will exit before the others, the precise condition under which exits jam. Narrative saturation also removes the traditional early-warning value of official commentary: when everything has already been said, the eventual trigger will be an event, not a speech. This is why the EWC watchpoint set (Paper 4) tracks balance-sheet quantities rather than rhetoric.

06 · Verdict: what kind of bubble, and where it breaks

PostulatedAssembling the audited record into a judgment, stated with its basis:

• On the technology: no bubble. The deployment economics of Paper 1 clear their hurdles at current capability and collapsing cost. The infrastructure being built will be used, at some price, by someone; this is what distinguishes it from tulips and NFTs (basis: measured adoption data; reference class of used infrastructure post-mania).
• On equity pricing: stretched, concentrated, and not 2000. Roughly half the index in one theme is a historic concentration risk (basis: BoE, Goldman data), but earnings delivery to date separates the leaders from dot-com structure. The equity risk is a violent repricing within a functioning market, on the order of the BoE's stress geometry at the bad end, not a Nasdaq-style 78% wipeout of the core (basis: multiple comparison, cash-flow funding).
• On the financing architecture: the classical mania pattern, forming in real time. Circular vendor financing, collateral whose useful life is contested, off-balance-sheet growth at 8x in four years, private credit and insurance balance sheets absorbing the leveraged tier, and issuance accelerating while the revenue gap stays open. Every completed reference-class episode broke through this layer, not through the famous stocks (basis: 1840s railway credit, 1920s utility holding structures, 2000 to 2002 telecom debt; current data per Section 2). We assign roughly 60% probability, basis: the trend data above, that the eventual stress event is first visible in private credit or a vendor-financed structure rather than in the public equity of the majors.
• On timing: unknowable, and we say so. The 2026 record shows the market already repricing capex announcements negatively in single names within days. Manias can run years past their first warnings; this one's first official warnings (IMF, BoE, Fed chair, and the sector's own leaders calling it a bubble in their own words) are already two years old. We hold no timing view beyond the watchpoint machinery of Paper 4, and we distrust anyone who does.

The historical comparison, quantified where the record permits:

Sources: Advisor Perspectives (GDP intensity comparisons, October 2025); DQYDJ and market records (drawdowns); BIS (analogy framing). The final row's survival column is EWC postulation.

07 · Conclusions

First, the bubble question dissolves under disaggregation, and the residue is precise: real technology, real and inadequate revenue, defensible equity multiples on the leaders, and a financing architecture reproducing the classical late-cycle pattern in the least supervised corner of the system. Anyone offering a one-word answer to "is AI a bubble" is selling narrative, in either direction.

Second, the reality-versus-fiction test this series proposes is transferable: date every number, name every source, refuse decompositions without primary sources, and watch the credit stack rather than the demos. Applied quarterly through our watchpoints, it will either confirm this diagnosis or force us to revise it in public, and the forecast log will show which.

Third, for the 2040 horizon: the most probable reading (Paper 4's 50% base case) is that 2026 sits mid-mania in the oldest pattern in financial capitalism, the one that built the railways, the grid and the internet at ruinous cost to its financiers and enormous benefit to everyone after. The research task between now and the repricing is not to call the top; it is to know, in advance, which claims are load-bearing, which are decorative, and which were fiction all along. That is what this series has tried to put on the record.

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. Probabilities are subjective analytical judgments with stated basis, not guarantees. Run-rate revenue figures are company annualisations, not audited revenue, and are labelled as such. Figures carry their as-of dates and may have changed. © EWC Investments, 2026.