Is AI a Bubble?
Evidence & tests
What each source establishes, what the models assume, and the observations that would change the answer.
The right evidence depends on the claim. An experiment can identify a task effect without observing the buyer’s invoice. A payment panel can show buying without identifying the buyer’s funding source. A contract can establish a duty without establishing settlement.
Keep the unit, population and status attached
| Evidence | What it supports | What remains unestablished |
|---|---|---|
| Study output or time | A defined effect for tested workers, tasks and tools | Universal productivity, payroll savings or provider margins |
| Paid seats | Vendor-reported licensed deployment | Active use, annual renewal or list price × seats as revenue |
| Payment-panel recurrence | Repeated buying in the observed panel | Dollar retention, credits and a representative outside-funded total |
| Operating profit | Earnings under the specified accounting boundary | Cash recovery of a new investment or a fair ownership price |
| Contract or facility | Specified obligation, rights and conditions | Receipt, full draw availability or collateral recovery value |
| Rating agency base case | An attributed underwriting expectation for a specified structure | Observed collections, today’s covenant headroom or an independently replicated private model |
| Conditional collateral threshold | The realization needed under stated balances, costs and priority | An auction price, default probability or final-maturity collateral value |
A study among active adopters is not an all-business census. The Census AI question expanded on 17 November 2025 from producing goods or services to any business functions, with new results released from 4 December. A change across that break cannot be read as an unchanged adoption series. Firm weighting, employment weighting and individual use also answer different questions. Official wording document; Federal Reserve measurement comparison.
Similarly, Ramp expanded its coverage and vendor identification in June 2026. Its monthly account recurrence, high-spending tail and effective-token-price observations are not one homogeneous matched cohort. A mix change can alter measured price without showing the price of a fixed-quality model or a long-term cloud contract. Payment evidence and methodological limits.
In the selected DDTL 4.0 case, DBRS’s 1.20× stabilized base case and the contract’s 1.15× maintenance threshold have different status. A 4.17% NOI decline connects them only if the starting period actually matches that base case, definitions align and debt service stays fixed. That calculation is not observed headroom, a revenue haircut or an outage allowance. Underwriting source; coverage definitions and limits.
Do not turn a capacity label into a utilization or return estimate
A physical milestone strengthens a particular claim, not every claim about the asset. Accepted capacity is more informative than an equipment order; it is still not measured occupancy, an electricity meter reading or proof that invoices were collected. The selected cohorts do not provide a representative sector-wide return or utilization estimate.
| Measure | What is being counted | What must stay separate |
|---|---|---|
| Acceptance or billing | A contractual service milestone or basis for fees | Actual GPU hours, cash collection and investment return |
| Paid utilization | Capacity sold or reserved for payment | Electrical load and useful compute efficiency |
| Electrical load / PUE | Energy drawn relative to IT capacity; facility energy relative to IT energy | Customer willingness to pay or model-FLOP utilization |
| Compute efficiency / availability | Useful calculation relative to a technical denominator; ability to deliver service | Commercial occupancy and complete cash margin |
| Active / contracted power | Present operator-defined capacity relative to a pipeline including future capacity | Not a fleet utilization percentage |
The distinction is visible in the supplied sources. CoreWeave’s June active/contracted-power ratio is 40.5%, not a utilization rate; its construction-in-progress share is not a stranded-asset estimate. Meta’s 2024 Llama 3 report gives roughly 41% model-FLOP utilization and more than 90% effective training time—different measures of an older training process, not competing 2026 occupancy readings. Exact populations, dates and definitions.
An evidence date is not an event date
NVIDIA / Energy Global example · not a payment timeline
Four dates answer four different questions
Scroll the figure sideways to see the full diagram.
Financial periods, transaction dates, disclosure dates and inspection dates travel separately. The whole-system financial, valuation and offering boundary is 18 September 2026; the demand and receipt-correction boundary is 19 September. Neither promises exhaustive coverage through midnight. June company accounts are not a complete September balance.
The strongest causal evidence in the demand chapter concerns older tools and selected populations: support rollout in 2020–21, developer field experiments beginning in 2022–23, office deployment in 2023–24 and experienced-maintainer tasks in early 2025. Later use and payment observations do not change those experimental populations or outcomes. Exact study sources and versions.
Pricing is not settlement. The September 18 CoreWeave release prices $3.7bn of notes but expects settlement on September 22, subject to conditions. The financing investigation’s September 20 boundary precedes that date. Neither the prospective proceeds nor August draws create a complete roll-forward of June cash and debt. Pricing release; financing calendar.
A model makes an assumption inspectable, not true
The return scenarios state the investment scale, timing, margin, discount rate, productive life and residual assumption. The demand sensitivity separates price, volume and unit cost. Those are tools for showing which assumptions matter; they are not company revenue targets, measured shortfalls, fitted default probabilities or verified project valuations.
The support model’s avoided-contact fraction is not taken from the support augmentation experiment. The office calculation uses the offer-of-access time effect, not the larger induced-adopter effect for every licensed employee. Added wage/value rates, working weeks and implementation costs are explicit assumptions. Support example; office example.
The public inputs preserve periods, units, source locations and missing-data limitations. The three calculation companions run locally with Python’s standard library and fetch no data. Reproducing arithmetic does not independently verify disclosures, confidential models or scenario assumptions.
Financial inputs (CSV) · Whole-system calculations (Python)
Demand observations (CSV) · Demand calculations (Python).
The financing model follows a different claim from the all-capital valuation: it includes cash principal and interest in a payment reconciliation, while separating interest already deducted in operating cash. Its required-collection formula leaves actual costs and installments missing rather than silently zero. Reserve funding and excluded payments require additional tests; recovered current cash and sale costs are explicit assumptions, not measured collateral values.
Financing inputs and assumptions (CSV) · Financing calculations (Python) · complete definitions.
The disclosed-budget recovery model is conditional. Four actual phases are synchronized into 60 equal service months; all capital and the customer advance are placed at the modeled start. The advance is credited once against subsequent collections. The 85% company project measure is not full cash flow, and the 80/75% cases are additional-cost sensitivities. Delays extend the service term rather than automatically delete contracted fees. Net end value is either realizable proceeds or continuation value, not sale plus reuse. Model boundary; 83 inputs, calculation code and source figure.
Correct the exact claim—not the entire subject
NVIDIA: obligation is established, receipt is not. The underlying whole-system research asserted that the prepaid-forward contract acknowledged payment, and its source note repeated that characterization. The reader account withheld that stronger assertion. Inspection of section 1(f), the separately supplied wire instructions and the attached lock-up form does not support receipt confirmation. The stronger characterization is withdrawn.
NVIDIA entered into a $1.5bn prepaid-forward contract with Energy Global, LP. The agreement requires payment within three business days after its date. The documents inspected for this investigation do not verify receipt.
The contract’s filename contains “exhibit1027,” but the subsequent September 4 index lists the forward at 10.26 and the separate subscription at 10.27. The title, parties, date and section identify the contract more reliably than the filename number alone. The July balance date predates it; the inspected later disclosures did not provide missing confirmation, and the registration-statement body could not be retrieved. Forward, section 1(f); index; complete correction.
This removes unjustified certainty about liquidity. It does not prove nonpayment, create customer revenue, combine the forward with the separate subscription or change inherited numerical outputs: those calculations used no prepaid-forward receipt input. A later receipt could resolve settlement without making the original claim about what this contract proves correct.
Amazon: the payment correction predates the current assessment. The earlier baseline left announced funding unverified although the July filing was available. Cash, commitments and capacity corrected that omission, reporting $28.7bn invested through June and $21.3bn afterward. It changed the evidence coverage, not the event date. The current analysis does not claim that discovery. Amazon investment note.
CoreWeave: a new period, not a correction of a March number. The June principal table is a later snapshot. Its smaller through-2027 percentage reflects a shorter remaining window, changed borrowing and a larger denominator. It is not a clean measure of reduced risk. Comparison and source.
What would change the answer?
| Observe | Stronger sustainable case | Stronger concern | Horizon / limit |
|---|---|---|---|
| Customer payment | Same buyers renew and expand net of credits | Retained accounts spend less or need continuing subsidies | Next 2–4 quarters; count retention can hide falling dollars |
| Useful work | Quality-adjusted support cost falls; coding and office gains survive rework | Repeat contacts, defects or implementation costs consume the benefit | Operational cycles plus 6–12 months / annual renewal |
| Provider cash | Paid volume grows with stable or improving complete cash contribution | Cheaper units and higher serving costs erode retained cash | Quarterly results plus annual development and replacement costs |
| Delivered capacity | Usable capacity meets service dates and earns through its life | Power delays or underuse shorten the earning period | 6–24 months for delivery; 1–3 years for cohort evidence |
| Binding obligations | Customer cash funds amortization and needed replacements | Repeated concessions, unpaid invoices or collateral shortfalls | Actual covenant dates and maturities; compliance is not full recovery |
| Project performance | Accepted capacity, net collections and amortization match the contract | Deductions, delayed acceptance, repeated cures or trapped cash | Installation through commitment termination; group OCF is not the borrower certificate |
| Available cash | Actual K, A, H, reserves and eligible LC support cover all dated uses | Reserve top-ups and draw restrictions consume liquidity | Monthly/three-month tests; covenant exclusions require separate payment checks |
| Financing settlement | Priced notes close and net proceeds bridge productive delivery | Closing fails or cash uses repeatedly outrun collections | Expected 22 September closing, then quarters; no proceeds assumed at the 20 September boundary |
| Contingent support | Activated phases retain useful assets and viable alternative tenants | Qualifying failure combines with weak recoveries and interim funding needs | Expected commencements from 2028; main cap is not a callable schedule |
| Ultimate holders | Diversified positions and long-term liabilities absorb loss | Concentration, leverage or redemption terms force further funding or sales | Current positions remain unidentified; arranger names do not establish holdings |
The strongest evidence against the concern would combine independent recurring payment, durable all-cost cash margins and successful capital recovery. Strong demand alone cannot establish all three. Conversely, worse retention, more financing dependence and lower collateral recovery would warrant greater concern even if technical capability improves. Whole-system tests; demand-conditioned payment risks.
These financing checks derive from the inspected credit agreement, underwriting, pricing release, project forms and general FSB transmission framework. They do not add the facility ceiling, guarantee ceiling and underlying obligations into one loss total. Sources, horizons and principal ambiguities.
Physical-recovery tests: replace the original budget with final component costs and their payment dates; match each phase’s acceptance to credits and net collections; measure maintenance and full cash costs; and observe renewal or transfer transactions net of retrofit, downtime and new equipment. The first operating term, subsequent renewal and asset transfer answer different questions. The observations and what each would change.
What the evidence does not yet measure
No defensible single total reconciles final-customer cash, credits, venture-funded purchases and internal transfers across the AI economy. Complete product cash margins, private preference rights, project-level realized prices and recovery values are incomplete. The Microsoft market observation has a narrow reproducibility limit; the Anthropic adjusted-profit report retains attribution and access limits. Whole-system limitations.
The served support and office manuscripts have unresolved differences between version-header and internal dates. They are identified by the inspected URLs and actual observation periods; no assertion about changes between manuscript versions is made. The studies and transaction panels also remain selected rather than representative of the whole economy. Demand methods and limits.
Research model attribution: GPT-6 Pro for the whole-system and customer-demand investigations. Source-linked AI-generated analysis is not an independent audit of private accounts or confidential study microdata. Evidence, scope and reproducibility—not a model name—determine what each claim supports.
The financing investigation leaves actual borrower collections and compliance, unredacted amortization/cost schedules, letters of credit, active PORTS-Pike phase amounts, the main residual-loss formula and current beneficial loan holdings unresolved. The full SB Energy registration statement exceeded the retrieval limit even though relevant exhibits were readable. A missing disclosure is not evidence of safety or failure. Financing inspection boundaries.
The capacity investigation does not establish a used-GB300 transaction curve, an appraisal of Horizon facilities, actual phase-level GPU hours, electrical load, full costs or a retrospective IRR. IREN’s group NVIDIA investment rights are separate from the NVIDIA–Energy Global prepaid forward. No receipt-confirming evidence changes that forward’s correction.
The September 21 Texas permit/grid-audit development is attributed to Reuters. The underlying order and applicability to these phases were not established in the supplied investigation, so it is not used to declare Horizon 1 closed or impose a project-specific delay. Nor is the earlier expected September 22 CoreWeave settlement promoted to confirmed cash simply because that date arrived. Delivery-condition boundary; Financing continuity.