Evidence & tests

Keep the unit, population and status attached

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.

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.

Design vintage, service continuation and current value need different records

The continuation investigation strengthens an actual favorable mechanism while keeping its measurement limits. An older design can win a new contract; a previously used facility can serve another tenant. Neither dates specific machines or establishes the original owner’s return. The Austin calculation is deliberately historical, compiled from amended disclosures inspected later.

The Austin expense calibration already contains rent and noncash compensation. A signed correction is needed to bridge it to cash; setting that correction to zero is not evidence that missing costs are zero. The undated later fee bucket and associated costs remain outside the finite test, not written off. Formula, assumptions and source dates.

The source amendment date matters: a March 2025 model anchor does not mean the later amended information was known then. The exercise is neither a backtest nor a current market valuation. Inspected sources and limits.

An observation, a proxy and a requirement are different evidence

The new valuation results combine measured inputs with analytical choices. Calling them all “current figures” would conceal the important distinction.

IREN’s original-recovery end values are not market appraisals. Microsoft’s prior cash-growth and new revenue-growth rates are different model outputs. The frontier funding call is not the burn forecast minus committed financing. All model and acquisition limits.

Public price, count and statement dates are deliberately not collapsed. The September 24 daily history is frozen; IREN’s shares are August 14, Microsoft’s July 23 and NVIDIA’s rounded count August 21. The older financial statements do not become complete September balance sheets. No like-for-like investment return or market repricing is calculated by comparing different snapshots and models.

Do not promote a cash ceiling into a contract right

Sources, equations and disclosure limits accompany the new chapter. The reported cash accounts are facts at their source dates; the cash-growth, policy, financing-rate and buffer choices are scenario assumptions. A negative ending resource is an incompatible modeled budget—not a default date.

The $50bn buffer is explicitly a research convention. Microsoft’s required overlap/valid adjustment is not an observed cancellable amount. Meta’s favorable 2027 case has little remaining modeled cushion, but unobserved timing and new uses can change it in either direction. Neither an equation nor a larger ledger fills the missing operating-versus-capital allocation. The relevant observable tests

This investigation mainly makes existing statements and terms more informative through matched-period reconstruction and fuller interpretation. It does not turn June into September or claim discovery of Meta’s previously noted zero buybacks. It retains the later CoreWeave closing established by the ownership chapter and the separate NVIDIA unverified-receipt boundary. The account of changes

A positive result needs its boundary as much as a negative one

The September 29 investigation strengthens commercial persistence and possible value capture. These observations cannot be relabeled as an AI-only renewal or cash-margin dataset. Each important result has a positive meaning and a limit, rather than a generic “uncertain” label. Complete source notes.

Buying, benefiting and paying require different records

The buyer chapter combines an executed order, actual service observations, contractual persistence rules and reported expense. The positive inference is purposeful buying and plausible budget substitution. None of those evidence types alone supplies a verified invoice, causal productivity estimate or complete supplier return. Methods and achieved source access.

The NNC notice conflicts on VAT labeling; the signed fee is £48,000 before contractual VAT. The £57,600 spending entry remains a secondary transcription without the original CSV/invoice join. The later £51,000 five-month broader package remains secondary award evidence; changed scope and term prevent a like-for-like price comparison. Missing primary confirmation is not nonpayment, and an automatic-renewal clause is not proof that renewal occurred. The complete status ledger.

The conditional 14% model assumes causality, sustained effect, constant resources and comparable quality/case mix beyond the buyer’s observed statement. It solves two different questions: no higher unit cost while expanding service, and covering costs through adjustable resources at fixed output. Their benefits cannot be added; neither supplies observed payroll savings. The whole package fee, actual utilization, review/implementation and tax must retain their boundaries. Two thresholds with full-cost sensitivities.

Klarna’s 37% of $7m is management attribution, yielding $2.59m for first-quarter 2024 versus first-quarter 2023. Its agency decline and annual estimate are not independent additive savings. The 2024–2025 marketing endpoints yield an $86m increase versus the issuer’s printed $87m; Q2 2026 rounded subtraction yields $110m versus adjusted $111m. Those differences stay visible. Neither the annual decomposition nor the quarterly adjustment series identifies AI cash costs. Attribution and periods; rounding and cost reconciliation.

The 48-record public ledger preserves all values, units, dates, evidence statuses and qualifications. Its scenario source links point into the chapter. The public code preserves the calculation and validation logic with reader filenames and presentation adaptations. Arithmetic reproduction is not verification of the outcome, invoice, attribution or unobserved cost allocations. Source notes retain the original retrieval limitations.

A position, a capital buffer and a compliance certificate are different evidence

The new investigation locates an actual holder and tests its funding, rather than inferring current ownership from an arranger’s name. It also retains source gaps instead of turning a reproduced ratio into an official certificate. Methods and achieved access.

The accounting reconciliation matters near the edge. The face-debt/net-assets proxy reproduces GECC’s reported rounded ratio but is not its supplied legal computation. The unadjusted-book comparison is lower because of the face/carrying debt difference. The research does not accuse the issuer of misreporting compliance; it withholds the claim that a full selected write-off certainly fits every test. Inspect the two computations and their limits.

The one-day holder/borrower DDTL V maturity difference remains identified. The later GECCI redemption notice is not silently treated as settlement or as failure. NVIDIA’s January subscription is not charged again to the half-year that begins later. Its interactive cash-flow table’s duplicated purchase label is not used to invent a category allocation. Data-quality and date qualifications.

A matched calculation can combine sources without merging their status

The connected test is explicit about what comes from each document. It retains April’s operating illustration, substitutes May’s executed coupon and maturity, and uses June’s assumed first installment. It is a conditional compilation, not an updated issuer forecast or current compliance certificate. The complete source and payment reconciliation.

The public data keeps April’s 8.25% coupon classified as an issuer illustration; the executed May coupon is 7.75%. This is a source-status distinction, not a change in the numerical input. The original research’s PDF screenshot retrieval failed, so the source note does not claim successful visual inspection of that external PDF. Methods and sources.

A complete inference still needs the actual payer allocation, credit/cost split, eligible resources, distributions and pre-maturity receipts. More decimal places or a reproducible script cannot supply those observations. Nor should a funding mismatch be relabeled a measured loss merely because the cash arithmetic is exact.

A funding condition is not a current account or certificate

The completion study combines a reported construction match, contractual cash permissions and conditional tests. They do not have the same evidential status. The useful result is the condition under which a build can be funded—not a new October balance or an all-covenants certificate. Source dates and boundaries.

The construction reserve is a remaining stock, not an extra expense. Estimated remaining reserve earnings in the formula are not also a current inflow. Modeled room after completion does not grant distribution permission. Other debt permissions exist even though the specified post-Final-Commencement basket cannot be assumed to fund its own milestone. Equations and unresolved reconciliation; the actual debt gate.

Qualifying favorable paths must be preserved as carefully as the limits. The first-tranche advance exception is distinct from the 97% delivery route; both retain other restrictions. The $50m sixty-day spare-covenant cure requires eligible sources and no resulting customer credits; it is not a monthly penalty for all delay. The full chapter keeps those protections separate. Funding and performance conditions.

The next discriminating record is matched remaining bills, effective site rights, reimbursements, draw releases and a dated reserve determination. A certificate missing from public evidence does not mean the certificate does not exist. What remains unmeasured.

An evidence date is not an event date

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 timing result, a quality score and a profit claim need different evidence

The document-conversion study provides a concrete source reconciliation. The ACL paper’s printed $165-per-10,000-page label conflicts with its 2,207-second batch at the historical $2.69/hour assumption. Those inputs yield about $1.65 per 10,000 pages, or $165 per million. The same paper’s Mistral table shows a parallel denominator conflict. The investigation considers a mislabeled denominator likely, but no author clarification was obtained. The printed values remain tagged as conflicts, not silently repaired observations. Full arithmetic and source treatment.

The listed host-deprecation notice is also bounded: its numeric date is ambiguous and the served page is cached; no endpoint was tested. It does not establish a successful shutdown, worldwide OCR demand or company cost recovery. The research inspected the principal ACL cost/runtime tables visually, but its earlier PDF-image access failed and was not silently counted as successful. No experimental replication or new source audit is part of this reader integration. Methods and access limits; specific source notes.

A receipt, a deduction and a complete return need different records

Completed-transaction evidence can establish genuine realization without a private bank audit or a full IRR. The final CoreWeave prospectus and subsequent accounts establish different parts of the IPO, tenders, cost timing and claim conversion. Read the identity, economic date and recipient before interpreting a total. The reconstructed offering and cash statements.

Use the reported April option cash amount, not a quietly repaired base-terms product. Preserve the cumulative preferred-dividend cash while leaving its conflicting quarter allocation open. Do not invent a deduction to repair the October 2024 tender row. These limits concern the precision of particular joins, not the existence of the completed IPO. Option and dividend comparisons; the excluded tender row.

A conditional allocation must declare its starting point

The historical company-sale comparison freezes March 31, 2025 capital and assumes an equity budget after all non-equity claims, costs and needed funding. It changes the then-live right while conserving that same budget. September’s actual termination is shown beside it. Below the protected floor the code rejects the calculation; it does not invent an insolvency allocation. The examples are not current liabilities, observed losses or company valuations. Domain and assumptions.

The $1.88845bn net-parent reference uses assumed closing cash fees equal to recorded costs and a closing-reserve proxy equal to June’s later stock. Other uses are excluded at a zero reference, not observed to be zero. Paid bridge interest and fees remain real; actual reserve or fee timing can change the reference in either direction. It is therefore neither an observed payout nor a measured bound. Internal transfers cannot be added again to consolidated cash. The full conditional closing identity.

The missing records that would permit more precise returns are complete seller bases and taxes, fund-investor distributions, current fully diluted claims and the closing-account allocations. More precise arithmetic alone cannot supply those fields. They limit recipient IRRs and current value conclusions, not the documented transaction findings. Methods and remaining evidence limits.

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

26 September: later closing evidence, not an earlier factual error

A later current report confirms $4.2bn of CoreWeave notes issued on 22 September 2026, including the exercised $0.5bn option. About $4.137bn net of purchaser discounts, less $566.2m capped-call cost, leaves $3.5708bn before other expenses. This is completed financing, not operating revenue or a complete September cash balance. The September 20 investigation correctly left the then-expected closing prospective. Closing disclosure.

The former expected-settlement status remains an accurate account of the September 20 inspection. Likewise, the richer IREN subsidiary reconciliation is new inspection of June accounts, not a new September transfer. The Microsoft model change is a fuller economic convention—not a correction of every earlier number or proof the market repriced. Date and claim identification.

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.

4 October: an illustration’s status is corrected—not the earlier debt arithmetic

The completion ledger labels April’s 8.25% assumed coupon issuer_illustration, resolving the earlier joint-cash row’s contractual label. May’s 7.75% is the separate executed input, already used by the earlier debt calendar. Historical files and numerical results remain unchanged. This is a source-status correction, not the new construction-matching or financing-eligibility investigation. The exact correction and its consequence.

What would change the answer?

For these two buyers, seek the next matched order/invoice, actual utilization, equal-quality output and staffing, all incremental review/change costs, and genuinely avoided outside bills. Improved service after full costs strengthens the favorable account; weaker quality or costs that consume the gain weaken it. A larger total marketing line alone is not a verdict. Specific observations and their interpretation.

For older capacity, matched contract terms, actual deployment ages, cash maintenance, refit costs and retained rights would sharpen continuation. A new design launch or a higher portfolio margin would not fill those fields. Specific favorable and adverse tests.

For payer funding, obtain the actual lease-start calendar, commitment overlap, documented reductions and net cash after prepaid service is delivered. This would test whether resource capacity is achievable while preserving service, rather than merely changing the buffer until a scenario turns positive. Funding tests and time boundaries

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.

Price-to-owner-cash tests: reconcile remaining IREN costs and distributions, Microsoft growth and full investment, private funding and rights, and actual offering proceeds. New cash evidence can replace an assumption; another capacity announcement cannot. Observations and interpretation boundaries.

The new holder tests are actual updated positions, cash realizations, borrowing-base and covenant calculations, confirmed note redemptions/refinancing, and matched payment and recovery records. Continued collections can strengthen containment; wider common funding pressure could weaken it. A manager logo, larger ceiling or stock-price fall is not a substitute. Discriminating observations for both cases.

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.

Holder coverage has advanced, not become exhaustive. GECC’s selected June holdings and financing terms are now established in the dated holder investigation. Its actual current borrowing-base and coverage certificates, all ultimate creditors/investors, and NVIDIA’s later original holdings table remain unavailable. No full current exposure network or sector loss probability is inferred. Exact scope and access limits.