AI looks more like a real technological revolution than a pure speculative mania, but parts of the market are fragile: solid revenue and adoption trends underpin a meaningful slice of AI valuations even as concentrated private bets, fast-rising capex and exuberant narratives produce localized bubble dynamics.

Main Finding

AI in May 2026 is best characterized as a real, economy‑relevant technological revolution that exhibits localized bubble dynamics rather than as a single, homogeneous speculative mania or a bubble‑free productivity miracle. Some layers (large public AI‑infrastructure firms, leading semiconductor suppliers) show substantial fundamental support; other layers (private foundation‑model valuations, thin‑profit application startups, speculative data‑center projects) display bubble‑like fragilities. The authors therefore recommend a segmented, multi‑method diagnostic approach rather than a binary “bubble/no‑bubble” verdict.

Key Points

• Conceptual framing
  • Bubbles must be evaluated against an asset‑pricing benchmark (stochastic discount factor / state‑price framework). Price = fundamental value + bubble component; the bubble component must survive after modeling fundamentals.
  • The diagnosis is a joint‑test problem: expensive prices can reflect omitted growth options, mis‑specified risk premia, or genuine exuberance.
• Segmentation of the AI ecosystem
  • The paper treats AI as a multi‑layer “stack”: semiconductors, cloud platforms/hyperscalers, data centers & power, foundation models, applications & agentic AI, and venture‑backed startups. Bubble risk can be layer‑specific.
• Multi‑method diagnostic framework
  • No single test suffices. The authors propose integrating (i) fundamental valuation and residual‑exuberance measures, (ii) econometric explosive‑root tests (SADF/GSADF), (iii) LPPL / HLPPL pattern diagnostics, (iv) sentiment / narrative / issuance metrics (including ML‑based text measures), and (v) capex‑payback & sustainability analysis.
• Evidence mosaic (as of May 2026)
  • For fundamentals: realized revenue growth, enterprise adoption, and some productivity signals support nontrivial valuation in major public infrastructure firms (example: NVIDIA’s record data‑center revenue).
  • For fragility: rapid capex growth outpacing monetization in some layers, concentrated private valuations (few firms), investor narratives capitalizing expected future productivity before cash flows appear, and energy/capacity bottlenecks that raise uncertainty.
• Practical diagnostic construct
  • Introduces a residual‑exuberance metric REi,t = log(Pi,t) − log(bFi,t) (price vs. modeled fundamental), combined with price‑dynamics and sentiment indicators to flag segments for closer scrutiny.

Data & Methods

• Theoretical foundations
  • Asset‑pricing framework centered on stochastic discount factors, martingale valuation, and learned pricing kernels. Emphasizes state‑contingent payoff valuation and the Hansen‑Jagannathan bounds.
• Fundamental valuation approaches
  • Discounted cash‑flow (DCF) / residual income / SDF‑based valuation tailored to AI specifics (capex, depreciation, energy/inference costs, customer concentration, margins, terminal values and real option value).
  • Residual‑exuberance measure to quantify deviations: REi,t = log Pi,t − log bFi,t.
• Econometric / time‑series diagnostics
  • Recursive right‑tailed unit‑root tests (SADF and GSADF) for explosive behavior (Phillips et al. literature).
  • LPPL (log‑periodic power law) and HLPPL (hype‑augmented LPPL) pattern recognition to detect faster‑than‑exponential growth with oscillations.
• Text, sentiment, and ML tools
  • Machine‑learning classifiers and text‑based narrative measures to quantify hype, attention, and issuance momentum (drawing on recent HLPPL and ML in finance work).
• Capex‑payback & physical constraints
  • Capex sustainability checks: investment growth vs. monetization, useful‑life and utilization assumptions, energy and data‑center bottlenecks (IEA projections of data‑center electricity use, Goldman Sachs capex estimates).
• Empirical design (illustrative)
  • AI exposure mapping (assign firms to stack layers), compile public equity prices, private‑market valuations, capex & energy data, revenue/cash‑flow series, issuance & sentiment measures.
  • Tests: cross‑sectional residuals, time‑series explosive tests, LPPL/HLPPL fits, sentiment correlation, capex vs. payback and utilization analysis, plus robustness checks (alternative discount rates, specification bounds, sample splits).
• Key data examples cited
  • Stanford AI Index private AI investment ($285.9B in 2025), IEA data‑center electricity projections, Goldman Sachs supply‑side capex forecasts (~$765B in 2026 baseline), NVIDIA fiscal revenues (Q1 FY2027 figures).

Implications for AI Economics

• For valuation and research
  • Necessity of segment‑level valuation: models must incorporate layer‑specific capex intensity, useful life, energy costs, customer adoption lags, and competition/commoditization dynamics.
  • Improve SDF/pricing‑kernel estimation using information from narratives, attention, and machine‑readable signals (learned pricing kernels) to reduce model misspecification.
  • Use multi‑method diagnostics (residuals + price‑dynamics + sentiment + capex sustainability) to reduce false positives/negatives in bubble identification.
• For investors
  • Tailor risk assessment by layer: large incumbents with realized cash flows and bottleneck rents differ materially from private foundation‑model firms or speculative data‑center plays.
  • Monitor capex payback horizons, customer monetization, private‑market funding structure, and concentration risk (few winners carrying high private valuations).
• For policymakers and financial stability
  • Focus oversight on systemic amplification channels (debt financing of capex, circular private valuation rounds, collateral/linkages to broader credit markets), energy and grid constraints, and infrastructure bottlenecks that could transmit distress.
  • A segmented monitoring framework is preferable to blanket policy responses; some parts of the AI ecosystem may require targeted interventions (e.g., power grid planning, data‑center permitting, disclosure standards for private valuations).
• Open methodological challenges
  • The joint‑test problem remains: distinguishing mispricing from omitted legitimate growth/options and correctly specifying risk premia is hard.
  • Measuring and valuing intangible option values (network effects, complementary adoption) and integrating ML‑driven narratives into formal SDF models are active research frontiers.

Overall takeaway: treat “AI bubble” as a nuanced, layer‑dependent empirical question. Use an asset‑pricing grounded, multi‑method toolkit, map exposure across the AI stack, and combine fundamentals, price dynamics, sentiment, and capex sustainability before drawing conclusions about overcapitalization or systemic risk.