AI Economics Reality Check

Assumptions

Time horizon Analytical window, not a date forecast

Monetization test

Read this window together with capex trend and physical capacity pressure.

Capex: amplified Capacity: material Revenue: dominant Payback: emerging

Capex trend Physical capacity pressure Adoption depth Level 2 Isolated task use Model moat Level 3 Narrowing lead Implementation friction Level 3 Material drag Supply surplus pressure Level 3 Crowded model market Consumer demand drag Level 2 Wage pressure

Labor yield Mixed reallocation Complementary High substitution

Article cases

Start from a coherent case from the article, then tune the controls. The cases are tests of assumptions, not predictions.

Assumption guide

Control levels

Adoption depth: 1: experiments. 2: individual task assistance. 3: team pilots. 4: workflow dependency. 5: structural transformation.
Model moat: 1: near parity. 2: weak lead. 3: differentiated. 4: durable lead. 5: strong moat.
Implementation friction: 1: low review burden. 2: manageable. 3: material drag. 4: heavy drag. 5: blocking constraint.
Supply surplus pressure: 1: scarce frontier supply. 2: competitive market. 3: crowded market. 4: surplus. 5: commodity glut.
Consumer demand drag: 1: broad income support. 2: wage pressure. 3: visible demand drag. 4: demand contraction. 5: severe income shock.

Evidence map

Capex trend: Goldman Sachs and Sequoia-style infrastructure analyses, SEC companyfacts for hyperscaler capex, revenue, operating cash flow, and capex-to-cash-flow pressure.
Physical capacity pressure: IEA, NERC, Gartner coverage, and data-center power-system research. This is a constraint signal, not a dispatch, grid, or battery model.
Adoption depth: Stanford AI Index, S&P 500 adoption studies, workplace adoption papers, enterprise surveys, and evidence distinguishing general usage from workflow dependency.
Model moat: Official API prices, open-weight performance comparisons, inference-pricing research, and benchmark convergence evidence. Frontier reasoning may still earn premiums.
Implementation friction: Productivity experiments, consulting field studies, verification burden, governance needs, and domain-specific failure risk.
Supply surplus pressure: Provider competition, open-weight alternatives, falling token prices, and the ability of AI-assisted coding to increase model and application supply.
Consumer demand drag: Labor-income research, marginal propensity to consume, automation substitution evidence, and the macro risk that firm-level labor savings reduce aggregate purchasing power.
Labor yield: Labor demand reorganization studies, AI exposure research, hiring changes, apprenticeship pipeline risk, and observed displacement claims versus macro employment evidence.

Time horizon

2026 capacity squeeze: Tests whether announced capex can become usable compute. Physical capacity pressure matters most here: power, chips, cooling, construction, and grid connection decide whether spending can actually arrive on schedule.
2027 monetization squeeze: Tests whether enterprise revenue, workflow adoption, and productivity are catching up with the capex run-rate. Physical capacity still matters if delays keep assets idle or push useful capacity into later periods.
2028 payback gate: Tests whether capex already committed can earn durable cash flow before depreciation, refinancing, and replacement cycles bite. Physical constraints matter if they strand assets or shorten the useful payback window.

Why these assumptions matter

Scoring method: Scores use a weighted multi-criteria scenario model. Weights are judgment-based and source-informed, not econometrically estimated.
Bubble pressure: Rises when capex, valuation stories, and physical buildout race ahead of observed revenue, productivity, and adoption depth.
Psychological premium: Represents the risk that investors price strategic fear, FOMO, and winner-take-most expectations before cash flows are visible.
Domain knowledge: AI can accelerate work, but weak domain review can turn speed into false confidence, especially in ESG, risk, science, data, and architecture-heavy tasks.
Post-normal uncertainty: Following Funtowicz and Ravetz, the tool treats this as a problem where facts are uncertain, stakes are high, values can be contested, and decisions may still be urgent.

Other assumptions

Capex trend: Whether infrastructure investment keeps accelerating, flattens, or is cut back before enterprise revenue catches up.
Physical capacity pressure: Whether global power, data-center, chip, cooling, and grid capacity can keep up with the AI buildout. This is a simple pressure indicator, not a full power-system model.
Labor yield: Mixed reallocation: some tasks and junior roles are compressed while new review, governance, and domain roles grow. Complementary: AI raises output while staffing and wages remain broadly supported. High substitution: firms cut headcount or junior roles faster than replacement demand appears.

How to read this

This is a pressure test, not a point forecast. Set the assumptions you want to test, then read the active scenario as an answer to one question: what has to be true for the AI buildout to pay back?

Scores are directional: compare scenarios, not decimal precision.
Time windows change weights: capacity in 2026, monetization in 2027, payback in 2028.
Evidence is mixed: adoption, revenue, labor, and power constraints are moving faster than clean data.

Scoring method

Weighted multi-criteria scenario model. Weights are judgment-based and source-informed, not econometrically estimated.

The model normalizes each assumption to a 0-100 indicator, combines indicators into three outcome scores (revenue payback, vendor margin pressure, macro risk) using explicit weights, routes the result into a scenario class, and separately exposes evidence confidence, stakes, and urgency for each driver. Weights follow the evidence in the source map below rather than an econometric fit, and the horizon selector re-weights capex, capacity, monetization, and payback rather than changing which sources are used. Full formulas and calibration notes are kept in the project's internal model specification.

Custom assumptions

Active scenario

Mixed transition state

2027 monetization squeeze: recurring revenue has to begin catching up with infrastructure commitments.

Uncertain

Revenue payback 47 Higher is stronger

Likelihood that enterprise revenue and productivity can justify the capex burden.

Vendor margin pressure 61 Higher is riskier

Pressure on model providers from falling API prices, open-weight parity, and high inference costs.

Macro risk 55 Higher is riskier

Risk from capex overhang, weak adoption, labor-income drag, or financing fragility.

Main sensitivity drivers

Why this result moved

The current result is mainly driven by capex burden, adoption depth, and physical capacity.

Capex burden High

Aggressive infrastructure spending raises the burden of proof.

Adoption depth Medium

Workflow integration is not yet deep enough to prove durable revenue.

Physical capacity Elevated

Power, chips, cooling, and construction can delay usable compute.

Claim compatibility

No major tension

No major claim tensions flagged.

High enterprise monetization

Low enterprise monetization

Low capex

High capex

Q1 Capex buildoutHigh value, high capital intensity

Q2 Sustainable GPTHigh value, lower capex burden

Q3 Commoditized utilityUseful, low vendor pricing power

Q4 Infrastructure bubbleHigh capex, weak monetization

Buildout validity

Commodity pressure

Wage-demand drag

Supply surplus

System stress loop

Pricing power erosion Low

Physical scaling strain Elevated

Consumption drag Material

Break-even viability Stretched

Valuation pressure Elevated

This state reflects moderate adoption with material implementation drag, crowded model supply, and wage pressure.

Uncertainty and stakes

Post-normal zone

Facts are uncertain, stakes are high, and decisions are time-sensitive. Treat the score as a structured judgment, not a settled measurement.

Capex burden

Evidence: High
Stakes: High
Urgency: Immediate

Physical capacity

Evidence: Medium-high
Stakes: High
Urgency: Immediate

Adoption depth

Evidence: Medium
Stakes: Medium
Urgency: Near-term

Model moat

Evidence: Medium
Stakes: Medium
Urgency: Near-term

Labor and demand

Evidence: Low-medium
Stakes: Very high
Urgency: Medium-term

Source-data calibration

Not a forecast

Capex pressure High confidence

SEC filings and credit research support the direction: capex is large enough to raise the payback burden. AI-only returns remain undisclosed.

Adoption depth Medium confidence

Census and Fed survey data support uneven adoption. Usage is visible, but workflow dependency and measured returns are still harder to observe.

Productivity and labour Medium-low

NBER, BIS and ILO evidence supports gains with friction, complementarity and reallocation. It does not support simple replacement-rate scoring.

Physical capacity Medium-high

IEA and EPRI support local power and delivery pressure. This remains a pressure signal, not a dispatch or battery model.

The detailed source-data table sits behind the article in research/ai_economics/source_data_calibration_table_2026-07-17.md. These anchors calibrate direction and confidence, not probabilities.

Practical implication

Start with workflow mapping, controlled pilots, and ROI measurement before large platform commitments.

Source map

CapexSEC companyfacts, S&P credit research, Goldman, Sequoia, hyperscaler filings

Power and capacityIEA, EPRI, NERC/FERC, data-center power research

AdoptionCensus BTOS, Federal Reserve notes, Stanford AI Index, firm adoption papers

Model pricingOfficial API prices, open-weight benchmarks, inference-pricing research

Productivity and frictionNBER, BIS, ILO, task studies, consulting experiments, governance evidence

Labor and demandILO, IMF, Fed, BEA, labor-demand and income-shock literature

Evidence anchors

Full source list — numbered to match the article's footnotes.

Goldman Sachs Top of Mind: capex scale, payback skepticism, and productivity uncertainty.
SEC EDGAR APIs: official XBRL companyfacts for capex, cash flow, revenue and depreciation panels.
S&P Global Ratings: hyperscaler capex as a credit and cash-flow pressure signal.
U.S. Census BTOS and Federal Reserve adoption notes: adoption differs by survey unit, firm size, sector, and function.
NBER executive survey, BIS firm evidence, and ILO empirical review: productivity gains are real but uneven, often delayed, and not simple replacement rates.
Goldman/Morgan Stanley capex-risk coverage: higher 2027-2028 capex expectations and valuation risk.
Sequoia/Cahn: revenue bridge between infrastructure spend and end-user AI revenue.
Stanford AI Index: adoption, cost, benchmark, and investment context.
AI Adoption in S&P 500 Firms: deep enterprise integration remains much narrower than general AI discussion.
IEA Energy and AI, IEA Key Questions, EPRI and NERC: electricity demand and large-load reliability constraints.
Gartner data-center power coverage: AI server power demand and project delays from power and water constraints.
Generative AI and labor-demand reorganization: firms adjust hiring and task design before broad macro job losses are clear.
Income-shock consumption research: wage losses do not disappear from the demand side.

Caveats

Scores are scenario outputs, not forecasts.
Headline adoption is not the same as workflow dependency, EBIT impact, or durable productivity.
Task-level productivity does not automatically imply firm-level or macro productivity.
Labor substitution can improve margins while weakening the consumer market that later has to buy services.
The tool treats investor psychology as a risk channel, not as proof of fraud or inevitable collapse.
Physical constraints matter globally, but detailed dispatchability, batteries, Dunkelflaute/Hitzeflaute, and grid modeling belong in Track 2.