Real demand, demanding prices, and growing repayment risk

What “bubble” means here

A bubble means prices or capital commitments sustained by expectations of future economic returns that the available evidence poorly supports, sometimes reinforced by hoped-for resale or continued financing. Rapid growth, a high price or an early loss is not enough to establish one.

An ordinary failed venture, too much productive capacity and a cash-timing crisis are different outcomes, although they can overlap. A tool can work; a customer can benefit; a seller can earn a profit; a new investment can recover its cost; an owner can earn an attractive return on the price paid. Evidence for one step is not proof of the next.

Follow the outside dollar before adding the inside transactions

Households buy subscriptions; enterprises and public buyers pay for useful work; advertisers pay platforms that can use AI to improve their service. Benefits also arise inside companies or through free tools. Not all useful AI therefore appears as a separately labeled AI sale.

Consider the report’s hypothetical example, not a market estimate: an outside customer pays a model service $100; that service pays a cloud company $60; the cloud company buys $40 of equipment attributable to the work. The chain records $200 of sales, but not $200 of independent customer demand. An additional $100 equity investment finances activity; it does not create another $100 of customer revenue.

Linked investment and purchasing can bring forward genuinely useful supply. The risk is correlated dependence: a supplier’s investment, its customer’s purchasing power and the utilization of its new infrastructure may rely on the same eventual outside receipts. The disclosed relationships do not, by themselves, establish fictitious services or fraudulent revenue.

Sources: Amazon June 2026 filing, investment note; Microsoft fiscal 2026 annual filing; NVIDIA July 2026 filing; NVIDIA 17 August financing disclosure.

Genuine demand is stronger evidence than sufficient supplier cash.

Demand layer incorporates the second investigation’s 19 September 2026 retrieval boundary. The observed studies and company periods are earlier.

The evidence for useful applications and paying customers is strengthened by closer examination of the positive coding field study, office deployment and recurring purchases. The question is how those gains become value for buyers and retained cash for suppliers. This is deeper inspection and reconciliation, not a claim that the economy changed in one day.

For general office work, a randomized-access study of 7,137 workers at 66 firms in 2023–24 found about 1.4 fewer weekly hours in Outlook per assigned worker in later months. It did not directly observe employer productivity or profit. Diffuse benefits can justify a moderate license fee without a dramatic payroll reduction—but valuing every saved hour at wages does not establish cash available for an AI supplier.

Sources: Support rollout study; Cui and coauthors’ developer experiments; METR developer experiment; Office randomized-access study. Design, vintage and manuscript-version limits.

A buyer can benefit, keep using AI and pay one provider less

Microsoft’s July call reports more than 30 million paid Microsoft 365 Copilot seats. That is company-disclosed deployment, not a matched renewal cohort or annual revenue at list price. Ramp’s February update reports roughly 4% monthly business-count churn for OpenAI and Anthropic and that 79% of its Anthropic-paying businesses also paid OpenAI. These are panel observations, not national dollar retention. No complete tracing of funding sources or credits was obtained.

Ramp’s September report adds mix-sensitive effective-token-price declines and a fall in its volatile top-spending slice. It does not establish a sector-wide demand collapse or a fixed cloud-contract price cut. More usage is not automatically more cash when price, mix and serving costs change. Payment recurrence, revised observations and their limits.

Sources: Microsoft July call; Ramp February payment update; Ramp September price and spending update.

Census’s November 2025 wording change and firm-versus-worker weighting explain part of the disagreement among adoption headlines. Menlo’s active-adopter estimate is still not an all-business census. The denominator reconciliation partly repairs the baseline’s measurement gap without inventing a current national series or outside-funded dollar total.

Sources: Reuters, 17 August: Anthropic revenue reporting; Financial Times: Anthropic adjusted-profit expectation.

Consumer use, advertising and internal improvements can create value without a separate enterprise AI invoice. They do not establish which share of that value an AI supplier captures. The demand chapter leaves the baseline’s hypothetical margin and asset life as assumptions; it does not re-underwrite private-lab profitability or ownership values.

The unresolved conjunction: a verified customer outcome, a recurring net payment, complete provider cash and enough duration for investment recovery. Evidence for one link is not proof of every other link. Read the connected demand chapter and buyer/recovery calculations →

Large operating businesses meet an accelerating investment bill

Microsoft, Alphabet, Amazon and Meta already have large cash-generating businesses. The following comparison matches their reporting to the same dates. It measures the funding pressure on the whole parents, not AI-only investment or AI profit.

Scroll sideways to see every column on a narrow screen.

Scroll sideways to see every column on a narrow screen.

Sources: Microsoft fiscal 2026 results; Microsoft December half-year cash statement; Alphabet June 2026 filing; Alphabet 2025 results; Amazon second-quarter 2026 results; Meta second-quarter 2026 results; Meta 2025 results.

The last column is operating cash flow minus gross cash purchases of property and equipment. It is a limited subtraction, not company-defined free cash flow or cash freely distributable to owners. It has not deducted finance-lease principal, acquisitions, equity investments, debt repayments, dividends or the economic effect of dilution. Operating-lease payments already affect operating cash flow; noncash lease additions do not appear in the cash-purchase column.

Amazon’s trailing-year definition adds $4.021bn of property-sale proceeds and incentives to the −$11.625bn gross-purchase residual, giving its reported free cash flow of −$7.604bn. That difference is definitional, not contradictory. A positive residual does not establish new-project payback, and a negative one does not establish insolvency.

How the dates match: Microsoft supplies its June fiscal year and Amazon a trailing-year measure. Alphabet and Meta use calendar 2025 plus first-half 2026 minus first-half 2025. Microsoft’s January–June cash flows are its fiscal year minus July–December 2025, not its fiscal first half. The downloadable inputs and code reproduce these bridges.

Greater investment intensity leaves less operating cash for other uses without additional funding. It does not imply the parents are about to default: reserves, borrowing, equity, leases and future receipts matter. Actual financing is nevertheless part of the explanation, not an afterthought.

Profits exist in important parts of the chain

Scroll sideways to see every column on a narrow screen.

Sources: NVIDIA second-quarter fiscal 2027 results; Amazon second-quarter 2026 results; Alphabet June 2026 filing; Alphabet July earnings call; CoreWeave second-quarter 2026 results; Oracle quarter ended August 2026 results.

Operating income is preferable to spectacular net-income headlines when investment revaluations intervene. Amazon and Alphabet reported substantial investment gains. A valuable ownership mark is not a customer receipt available to pay an electricity bill without realization or financing.

Accounting labels do not change the physical bill

Microsoft’s change concerns estimated lives of buildings and data centers, from 15 to 25 years, effective in fiscal 2027, and prospective lease classification. It does not mean accelerators have a 25-year economic life. Management said the underlying physical investment expectation was unchanged. Alphabet’s $195–205bn calendar-2026 capex outlook, Microsoft’s approximately $175bn and Meta’s $130–145bn including lease principal use differing definitions; they are not added into a precise forecast.

Microsoft also disclosed $329.1bn of leases not yet commenced at June. That is a future commitment, not $329.1bn already borrowed or paid. Asset recognition, a lease liability, a future promise and a cash payment can describe successive stages of the same project.

Sources: Microsoft July earnings call; Alphabet July earnings call; Meta second-quarter 2026 results; Microsoft fiscal 2026 annual filing.

Who owes what—and who bears a disappointment?

Sources: Meta June 2026 filing, debt note; Amazon June 2026 filing, investment note.

Supplier funding is not unlimited or interchangeable with purchases. Anthropic announced more than $100bn of AWS purchases over ten years in April. Amazon’s separate Anthropic facility had an initial $20bn ceiling; $5bn of Series H investment reduced the remaining ceiling to $15bn. Delivery milestones and other conditions govern availability. The remaining maximum is neither immediately available cash nor a verified drawing. Its expiry after a specified liquidity event is not an established loan repayment date. Complete private security, coupon, maturity and other terms were unavailable.

Microsoft’s fiscal-2026 filing recognizes $24.1bn revenue from OpenAI, including revenue sharing, and $6bn receivables at June. Revenue is not same-period cash collected. A supplier’s genuine commercial relationship with an investee is not independent evidence of final-customer funding.

Sources: Anthropic 20 April Amazon compute commitment; Amazon June 2026 filing, investment note; Microsoft fiscal 2026 annual filing.

CoreWeave: the June schedule is a snapshot, not a September balance

CoreWeave generated $3.663bn operating cash in January–June and paid $14.117bn for property, equipment and capitalized software: a $10.454bn difference before other funding uses. Its June principal schedule—not debt carrying value—was:

Including actual principal cash payments of $5.219bn gives a −$15.673bn first-half residual after PPE. The detailed bridge adds back expensed cash interest before subtracting total interest and removes capitalized interest from the remaining PPE deduction, so it is counted once. Principal payments can include refinancing; this is a historical financing requirement, not a recurring default gap. Consolidated cash reconciliation and advance-payment limits.

$10.597bn, or 29.8%, was scheduled through the end of 2027. Old report 04 used a March snapshot: $11.718bn through 2027 out of $25.149bn principal, or 46.6%. The lower June share is not a like-for-like reduction in risk: the remaining calendar window is shorter, balances and financing changed, and the denominator is larger.

Later transactions prevent calling June a complete cutoff balance. August events include a $2.6bn DDTL 5.5 facility with $1.2bn drawn, plus a $1.2bn revolver draw. The 17 September $3bn convertible proposal progressed to $3.7bn priced on 18 September, with settlement expected 22 September, subject to conditions. At the financing investigation’s 20 September cutoff, proceeds remain prospective. A $0.5bn purchaser option is not assumed exercised. Up-to-35-million-share distribution authority is not proof of sales. Pricing update.

Sources: CoreWeave June 2026 filing; CoreWeave 17 September financing proposal.

One executed facility, DDTL 4.0, shows why the borrower and customer both matter. CoreWeave Compute Acquisition Co. VIII, LLC reported $2.837bn principal at 30 June, against an $8.5bn ceiling. Lenders have specified borrower assets and contractual cash-flow rights. DBRS’s 3 April analysis identifies Meta’s take-or-pay contract and expects full amortization from contracted operating cash, without refinancing or re-leasing. That is underwriting, not this investigation’s observation of a current cash certificate. Original rating rationale.

The defined 1.15× maintenance test starts after the applicable commitment-termination rule and transition periods. It excludes initial special amortization and final-maturity principal. Passing it alone does not establish full repayment; the exclusion does not mean a material balloon is expected, because full amortization is the issuer/DBRS base case. Actual installments, eligible collections, costs and reserve support are still missing. The parent guaranty covers defined conduct-related obligations, not every ordinary project shortfall; other CoreWeave facilities differ.

Sources: CoreWeave DDTL 4.0 credit agreement; CoreWeave DDTL 4.0 parent guaranty; CoreWeave June 2026 filing.

NVIDIA: current profit can coexist with longer-dated contingent exposure

NVIDIA disclosed a $105bn main ceiling for phase-dependent residual-value support associated with the initial 4.25 GW at PORTS-Pike and OpenAI occupancy. Lease commencements are expected from 2028; the additional capacity option is separate. The ceiling is not cash already paid, expected loss, today’s activated exposure or an unconditional guarantee of every missed invoice. August disclosure.

A supplier can profit on current equipment sales while taking on future customer-performance risk. Support can help a useful project become financeable; an indemnity is less protective when the tenant is unable to pay. That is a possible loss mechanism, not a prediction of a guarantee call.

The filed lease and residual-value guaranty forms add delivery and ready-for-service gates, contractual rent relief, reserves and alternative remedies. Assumption or re-letting may preserve value; a specified deferral can leave carrying costs and interim funding needs. The main loss formula and phase values are redacted. Reimbursement from the same distressed tenant is not equivalent to collected cash. Lease form; guaranty form; loss-transfer analysis.

Sources: NVIDIA 17 August financing disclosure; NVIDIA July 2026 filing.

A costed capacity case makes the recovery boundary more concrete

Physical delivery evidence reinforces commercial substance but sharpens the distinction between first equipment payback and full investment recovery. IREN’s Microsoft deployment at Childress provides a separately costed case: $5.8bn of equipment and ancillaries plus $2.8–3.2bn of facilities, planned for 200 MW IT. Horizon 1 was reported accepted in August; the other phases remained delivery targets. Those milestones do not reveal actual paid GPU hours or collected cash. Cohorts and measurement boundaries.

The original approximately 85% project EBITDA estimate includes direct site costs but does not recover the building or represent complete owner cash. In a synchronized five-year service-equivalent model, the $8.8bn combined midpoint requires $2.81bn of net end value at a 10% hurdle even when that project measure is treated as cash; an assumed five-point extra cost allowance raises the requirement to $3.42bn. These are conditional continuation-value thresholds, not appraisals, measured losses or actual company returns. Full-cost calculation and sensitivity figure.

The customer contract improves the case in a concrete way: service dates extend to preserve aggregate fees after a delayed start, and the advance improves collection timing without creating additional revenue. But capital may be committed earlier and the financing has its own payment calendar. A protected fee total does not make time or short-term funding free. Contractual clocks.

The retained site can supply a second life. One finite replacement scenario clears the whole investment hurdle with high paid utilization and sufficient future prices; a weaker case generates positive operating cash without recovering the successive investments at the assumed rate. The fixed cohort also imposes a limit on the claim that volume can always offset cheaper prices. Neither current scarcity nor loan repayment settles that future operating proposition. Replacement costs, utilization thresholds and alternatives.

This narrows rather than reverses the whole answer: real demand and accepted capacity coexist with uncertainty about complete cost recovery. Contract-backed DDTL 4.0 remains a serious favorable repayment case. Its maintenance, spares, service availability and power protection matter precisely because physical assets must keep supporting the promised cash; projected coverage is not an observed owner return. No newly chosen residual value is substituted for actual collateral evidence.

What must an investment earn over its useful life?

Revenue, profit, repayment and economic recovery answer different questions. EBITDA means earnings before interest, taxes, depreciation and amortization; “adjusted” EBITDA also makes company-defined adjustments. It excludes important costs, including asset consumption and financing, and is not cash left after asset purchases. Even positive operating income does not establish that the amount and timing of cash compensate for the capital employed and its risk.

Sources: CoreWeave June 2026 filing.

The sensitivity is the economic argument. A one-year delay, retaining four operating years, raises required annual revenue from $8.907bn to $9.798bn before any construction overrun. At the same life and margin, a 10% annual decline in realized revenue requires $10.232bn in year one; a 20% decline requires $11.760bn. These are revenue paths, potentially combining price and utilization—not token-price forecasts.

With constant margin, revenue 20% below the flat base creates a $2.823bn net-present-value shortfall. Fixed costs can make the actual margin deteriorate as utilization falls, worsening the result. Longer useful life or stronger cash capture lowers the hurdle; rapid replacement and delayed service raise it.

This model does not compare a hypothetical new cohort with twice the entire company’s first-half revenue and call the difference a funding gap. Zero residual value can be stringent for buildings and more plausible for displaced hardware; the actual asset mix matters. Maintenance is already in the assumed margin, replacement beyond the modeled life requires new capital, and interest is not deducted again when the discount rate represents all capital.

Cheaper compute can help customers and hurt older assets at once

Lower costs and better capabilities can make new tasks worth purchasing. Higher volume can outweigh lower unit prices. But substitution can also reduce what an earlier asset earns while its financing bill remains fixed. Open-weight alternatives such as Mistral 3 provide another supply option; they do not make production hosting or engineering free.

Power and delivery affect the earning window. The IEA’s April 2026 outlook estimates 485 TWh of global data-center electricity use in 2025 and projects 950 TWh in 2030 in its central case. These are all-data-center figures, not U.S. generative AI alone. The projection is not delivered capacity. Grid bottlenecks can support high near-term utilization while delaying receipts and shortening a chip’s competitive earning life.

Global supply and funding also matter: NVIDIA’s disclosures identify dependencies involving Taiwan and South Korea. U.S. corporate spending does not locate every component, customer or ultimate risk in the United States.

Sources: OpenAI 8 September commercial thesis; Mistral 3 release and licensing; IEA April 2026 energy and AI outlook; NVIDIA July 2026 filing.

A successful business can still be an expensive investment

Ownership prices ask how much future cash belongs to the investor relative to the price paid—not merely whether the company has a useful product. The following are reverse-valuation scenarios, not fair-value estimates or trading guidance.

Public-company example: a conditional Microsoft equity value

Use the research’s approximately $3.692tn equity value as a scenario input. Its underlying finance-feed observation was timestamped 17 September 2026 at 19:29:50 UTC, not a closing price. Research had no durable vendor page and did not reconcile a same-instant diluted share count. This page does not present that input as a verified current market quotation.

A simple fiscal-2026 cash proxy is $182.935bn operating cash minus $115.948bn gross cash property purchases and $3.101bn lease principal: $63.886bn. An alternative deducts $12.405bn stock-compensation expense, leaving $51.481bn. That is a rough allowance for dilution, not an exact repurchase bill. Neither proxy is a complete forecast of owner cash.

Sources: Microsoft fiscal 2026 annual filing.

Scroll sideways to see every column on a narrow screen.

At the middle assumptions, the first path reaches roughly $448bn annual cash in year ten; about 69% of present value comes from cash after that year. Profitability alone does not establish such a path. But it is not plainly impossible: margins, revenue growth and lower investment intensity could contribute.

The exercise omits separate credit for net nonoperating assets and Microsoft’s OpenAI holding, and does not precisely model acquisitions, borrowing, dilution, tax changes or investment timing. It values the whole equity claim, not an AI segment. Treating temporary reinvestment as permanent or assuming away required replacement can both mislead.

Private rounds price ownership, not cash already earned

OpenAI announced $122bn committed in March at an $852bn post-money equity value. Anthropic announced a $65bn May round at $965bn post-money, including $15bn of previously committed hyperscaler funding. That included capital must not be added again. These are dated financing-event values, not enterprise values, current public market capitalizations or amounts of customer revenue.

March OpenAI monthly revenue of $2bn annualizes to $24bn; Anthropic’s May run-rate claim was $47bn. The dated equity-value/run-rate ratios are 35.5 and 20.5 times, respectively—not September sales multiples or comparable audited annual ratios. August reporting put OpenAI near $40bn annualized and Anthropic above $65bn at end-July. Later reported discussions of an OpenAI round above $1.2tn were not an established completed transaction.

Sources: OpenAI 31 March financing announcement; Anthropic 28 May Series H announcement; Axios, 17 August: lab revenue run rates; Reuters, 17 August: Anthropic revenue reporting; Wall Street Journal: possible OpenAI financing.

Changing the required return from 8% to 12% shifts the OpenAI revenue hurdle from approximately $368bn to $953bn and Anthropic’s from $417bn to $1,079bn. Interim distributions would lower the endpoint requirement; further cash needs and dilution can raise the burden on an existing holder. The reported Anthropic operating-profit inflection matters because it could bring cash generation nearer—but complete accounts are needed to determine when distributions are feasible.

The bullish case needs durable value capture after inference, continued research, distribution, capital commitments and competition. The bearish case must explain why those cash paths are implausible, not merely point to a high sales ratio. Private preference rights and full cash statements are not available here.

An offering has stages; a filing is not proceeds

These are documentary stages established by the research’s 18 September retrieval boundary, not a live offering calendar. A confidential submission is not a public prospectus, approved deal, fixed launch date or completed cash raise.

Scroll sideways to see every column on a narrow screen.

Sources: OpenAI 8 June confidential submission announcement; Fortune, 12 September: OpenAI offering timing; Anthropic 1 June confidential submission announcement; Reuters, 4 September: reported Anthropic timetable; SE Global Holdings 4 September S-1/A exhibits; Reuters, 15 September: planned Japanese share tranche; CoreWeave 17 September financing proposal.

A primary issue supplies cash to the business net of costs and dilutes existing ownership. A secondary sale primarily pays an existing holder. Both can feature in a large IPO headline, but only the former directly funds new activity. SB Energy’s NVIDIA subscription makes this distinction contractual: qualifying proceeds exclude existing-holder sales.

An IPO may expand funding access and improve disclosure. It cannot make an uneconomic project economic merely by changing owners. A postponement is not proof of insolvency either: other funding and company-specific constraints matter. Failure to establish public terms is not proof that no confidential process exists.

Sources: NVIDIA offering-linked subscription agreement.

The CoreWeave financing status above uses the selected 20 September inspection. The other offering stages retain the broader 18 September review; this update does not certify a new exhaustive search of all issuers.

Which arguments survive the evidence?

Sources: OpenAI 8 September commercial thesis; NVIDIA second-quarter fiscal 2027 results; CoreWeave second-quarter 2026 results; Sequoia, June 2024: AI’s $600B Question; Ed Zitron, February 2026: NVIDIA financing critique; Amazon June 2026 filing, investment note.

The arguments should not receive equal weight simply because both camps are vocal. “No real value anywhere” is poorly supported; “existing profits validate every commitment” is poorly supported too. The unresolved contest is between cash-generating demand catching up with investment and useful demand that still fails to deliver the returns embedded in commitments and prices. Historical booms illustrate how useful infrastructure and investor losses can coexist, but supply no crash calendar.

An unwind need not look the same for everyone

A cloud operator may pay a chip supplier in full and later impair its investment. A diversified parent may absorb a project loss while a thinly capitalized borrower cannot. New equity can avert a cash crisis but transfer much of the remaining upside away from earlier owners. Collateral value depends on recovery in the actual market, not the original equipment price.

For savers, profitable technology exposure, expensive equity exposure and contractual credit exposure are different risks. Several names in a portfolio can depend on the same customer cash. This report has assessed no reader’s holdings and recommends no trade.

Outside portfolios, construction employment, suppliers, committed utility infrastructure and customer continuity are possible channels. Their aggregate losses are not measured here. A systemic-crisis claim needs holdings, leverage, maturity mismatch, guarantees and forced-selling evidence. The BIS’s January 2026 discussion treated then-observed macro-financial risks as moderate while identifying a shift toward debt; that dated assessment is neither a September all-clear nor an inevitable-crisis forecast.

Sources: BIS, January 2026: financing the AI boom.

Containment is now supported by a specific contractual case

DDTL 4.0 is not best understood as a loan that must sell old GPUs or refinance a large balloon in its base case. Accepted service, Meta payments and scheduled amortization can keep creditors whole even when Meta or the operator’s equity earns less than hoped. The ordinary guaranty’s limits do not erase that repayment proposition. They matter if cash protection fails.

For wider transmission, the missing link is the actual holder. Blackstone Credit & Insurance’s anchoring role and named bank agents identify origination or administration—not the amount retained by a particular fund, insurer or corporate balance sheet. A well-capitalized closed-end holder may absorb a loss; layered leverage or redemption requirements at an exposed holder may transmit it. The FSB’s May framework describes those general channels, not this loan’s holdings. Official private-credit framework; Ultimate-holder evidence and limits.

What would change the answer?

The tests are symmetric. Growing usage alone is not enough to confirm recovery, and refinancing alone is neither a failure nor proof of repayment capacity.

Scroll sideways to see every column on a narrow screen.

The strongest counterevidence to the present concern would be independent paid demand, durable all-cost cash margins and successful amortization and replacement across several investment cohorts. Weak retention, increasing financing dependence and lower collateral recoveries together would justify a stronger unwind assessment—even if technical benchmarks improve.

The demand investigation sharpens these tests to net same-customer spending after credits, quality-adjusted buyer outcomes, supplier unit costs and cash available on contractual dates. Demand-conditioned loss pathways and observation horizons.

For the selected project, inspect acceptance, net deposited collections, actual amortization, reserve/LC support and cure use. For the priced notes, inspect settlement and use of net proceeds. For PORTS-Pike, inspect activated phases and remedies; for wider contagion, identify retained positions and holder funding before assigning exposures. Financing-specific observables and horizons.

The new physical comparison makes the tests more specific: actual Horizon component costs and capital dates; net collections after phase acceptance and credits; full cash service costs; major renewal spending; and realizable, transferable facility or cluster value. The earlier broad parent-company ledger is not refreshed by those selected cohort disclosures. Physical-recovery evidence to obtain.

Methods, boundaries and reproducible data

This foundation combines the whole-system investigation’s 18 September 2026 retrieval boundary with the demand layer and exact NVIDIA correction from the second investigation’s 19 September boundary. It does not refresh every company, valuation or offering. Filings and executed agreements establish reported figures and terms, not an entire economic thesis. Studies retain their populations, designs and tool vintages; vendor claims, attributed reporting and calculations remain distinct.

Publication dates, transaction dates, financial periods and research dates are distinct. An unestablished payment, field or balance is unknown—not zero, false or unpaid. Multiple statements from counterparties to one deal may share an underlying source rather than provide independent corroboration.

Research model: GPT-6 Pro for the whole-system and customer-demand investigations. Source-linked findings and conditional calculations are not independent audits of private accounts or confidential microdata; the scope and source limitations determine what each claim supports.

What remains unresolved

The largest gap is independently financed final demand that persists at complete margins sufficient for the committed capacity. Full frontier-lab cash-flow statements and private preference and facility terms were unavailable. The Anthropic profitability account was inspected through accessible reporting or preview material, not its underlying investor documents or the full paywalled article. Oracle receives lighter, rounded treatment.

There is no exact June-to-cutoff CoreWeave debt roll-forward, full project-level recovery assessment, current covenant-certificate test or independently measured collateral valuation. The prepaid-forward exhibit does not establish receipt. The public-market input lacks a durable vendor page and share-count reconciliation and is used only as a scenario, not a current quotation.

Customer studies remain selected, not a current economy-wide productivity synthesis. The demand investigation incorporates the positive developer field study and documents the Census question change and weighting distinctions. A complete current Census series, matched annual paid-renewal cohorts and full product cash margins remain unavailable. The support and office texts also have unresolved header/internal-date discrepancies. Demand-specific methods and limits.

Inspect the arithmetic

Selected source-linked financial inputs (CSV) · Reproduce the comparisons and scenarios (Python).

The CSV is a reader subset of the financial inputs, with original units, periods, statuses and source URLs. It is not an AI-sector total or a duplicate narrative report. The script uses Python’s standard library, needs no network access and checks the matched windows, Amazon’s definition bridge, principal schedule, recovery sensitivities and reverse-valuation identities. Save both files together and run python whole-system-calculations.py. These checks reproduce arithmetic, not the validity of every source or model assumption.

The earlier reports remain accessible as dated context. All eight investigations and historical context preserve the record, including the Amazon funding correction already made in Cash, commitments and capacity.

Demand data and calculation companion reproduce the second investigation’s buyer hurdles and extended recovery sensitivities separately; the original financial companion is unchanged.

The financing investigation’s source record uses the 20 September retrieval-day boundary and separates original contracts, partly populated forms, a solicited rating expectation and prospective funding. It adds no observed borrower coverage, empirical resale price or complete retained-holder schedule. Financing data and calculation code reproduce the new tests separately; earlier model inputs remain unchanged.

The capacity and recovery investigation has a 22 September 2026 retrieval-day boundary and supplies its own costed case, 83-record ledger, finite-cohort calculation and sensitivity figure. It is not a new nine-company balance-sheet comparison or a market-price appraisal. Read the substantial capacity chapter; data, code and source boundaries.