Shaduf.
Is AI a Bubble?/Research — all reports

Is AI a Bubble?

Research — all reports

The complete catalogue: current investigations, earlier reports, evidence dates and source-linked data.

Dates below describe evidence inspection, not publication or the dates of every underlying event.

Updated as new research is ready

All eight research reports are here: four current investigations and four preserved earlier reports. The whole-system assessment connects the question; each current chapter deepens it. Historical reports remain accessible evidence of what was previously explained, not competing current answers or automatically discredited findings.

Use Spending & obligations, Investment returns and Evidence & tests for subject-led explanations. This catalogue takes you directly to the complete reports and their data.

Report-to-question map · eight reports

Each investigation tests a different part of the cash chain

Scroll sideways on a narrow screen. All eight linked reports are listed below.

Eight reports across four economic questionsFour current investigations connect customer value, spending, obligations and returns. Four historical reports remain accessible below. Dots indicate principal subject emphasis, not a quality score or proof that other subjects were absent.Principal emphasis—not a completeness or quality scoreDemandSpendingObligationsReturnsWhole-system assessment18 Sep; later chapters retain own datesCustomer demand19 Sep 2026Financing & loss transmission20 Sep 2026Physical capacity & recovery22 Sep 2026Earlier baseline11 Sep 2026Cash, commitments & capacity11 Sep 2026Financing, payment & risk13 Sep 2026Debt, collateral & repayment13 Sep 2026● Main emphasis — Not the main emphasis; not proof of no coverage
The map shows principal subject emphasis, not validation of conclusions. All four current and all four historical reports appear in the linked catalogue. Dates mark evidence inspection, not publication or the date of every underlying measurement.

Current research

Current physical-capacity and recovery investigation

Capacity that earns revenue is not yet capital that earns its cost

Evidence checked: . Retrieval-day boundary, not a publication date or same-day utilization reading.

Accepted capacity, IREN’s equipment and facility budgets, customer advances and extended service terms, finite replacement cycles and conditional continuation-value thresholds. Preserves the CoreWeave repayment case and distinguishes facility-only hosting economics.

Read the full report → · Inspect the 83 inputs and assumptions · Recovery sensitivity figure

Current financing investigation

Who has to pay when AI returns disappoint?

Evidence checked: . Not a publication date. June/July balance sheets and future settlement remain separately dated.

Meta-backed contracted repayment, CoreWeave cash and reserve thresholds, conditional collateral recovery, NVIDIA residual-value support and parent resilience. Explains creditor containment, equity loss and the missing links to wider transmission.

Read the full report → · Inspect source-linked data and assumptions

Current whole-system assessment

Real demand, demanding prices, and growing repayment risk

Whole-system evidence: 18 September 2026; demand 19 September, selected financing 20 September, and physical recovery 22 September. These are inspection boundaries, not publication dates.

Customer value, spending, financing, recovery, valuation and loss transmission in one connected assessment. Conditional models retain their assumptions and the NVIDIA receipt correction.

Read the full report → · Inspect source-linked data

Earlier research — the complete preserved record

History: earlier pages and claim-specific context →

What changed, with the reason attached

Full-cost recovery: the earlier general cohort model is now complemented by a disclosed IREN customer contract and separate equipment/facility budgets. The narrower equipment perimeter can clear a hurdle while the combined investment requires continuing value. This is a better-defined calculation, not a newly observed company loss. The cost and cash boundary.

Contractual timing and capacity: the delay analysis retains extended service terms instead of automatically deleting fees; the replacement-cycle test charges new capital and shows a fixed-capacity limit to volume growth. Future prices, utilization and net asset value remain assumptions. Timing protection; continuation thresholds.

Stronger contractual containment: April DBRS analysis identifies Meta’s take-or-pay contract and expected full amortization. This is newly incorporated underwriting evidence about a dated arrangement—not a new September customer contract or verified compliance. Repayment proposition and limits.

Real-obligation cash and recovery tests: the investigation now reconciles actual principal and interest payments and derives collection, reserve and collateral thresholds. Unknown project schedules remain variables; assumed sale values are not observed prices. Definitions and numerical boundaries.

Financing stage advanced, receipt still prospective: the September 17 proposal became $3.7bn of priced notes on September 18, with expected September 22 settlement. The 20 September cutoff does not count proceeds as received. Dated cash calendar.

Deeper demand evidence: positive coding field evidence, office time savings and recurring payment strengthen commercial substance. Distinct populations and outcome measures remain separate. The findings do not establish complete cash margins or annual paid renewal. Demand investigation.

Measurement window: the $510.703bn July–June parent cash-investment total contains January–June’s $294.800bn. These are overlapping periods, not amounts to add. Nested-window visual and exact basis.

Later financial period: the June debt snapshot replaces March in the current assessment. A smaller through-2027 share is not pure deleveraging. What the two snapshots do and do not say.

Attributed profitability evidence: the favorable Anthropic adjusted-profit reporting remains visible, with its stock-compensation exclusion and limited access to underlying documents. Interpretation and source.

Explicit conditional models: capital recovery and ownership prices now have inspectable assumptions. These do not turn incomplete market or private-account information into fair-value estimates. Investment returns.

Previously corrected evidence coverage: Cash, commitments and capacity had already corrected the first baseline’s missed Amazon filing. The disclosure was available before the omission; the current assessment does not claim its discovery.

NVIDIA receipt characterization: the stronger assertion about what the prepaid-forward contract proves is withdrawn. Obligation is established; receipt remains unverified, not disproved.

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.

Source, exact documentary issue and numerical consequence →

Forecast vintage: the current whole-system account uses the IEA’s April 2026 update; the earlier report retains its older outlook. Both concern global data centers, not measured AI-only cash demand. Dated forecast context.

Search published pools, pages, reports, and evidence.