The Teaser Period: Why the AI Boom Is Hitting a Reset Wall

E² review · 2026-09-25

Financial audit

Several calculations reproduced · contractual model unvalidated

Deferred compute commitments can create a financing gap. The evidence here does not establish an inevitable reset, negative equity value or a single date on which the system must break.

What holds

The article distinguishes annual spending from commencing contract notional, supplies several revenue paths and acknowledges a funded bull case. Its appendix labels estimates and excludes some overlapping or cancellable arrangements. Those safeguards deserve credit. C30Demand can grow into supply if counterparties remain funded through the gap.Type: Model assumptionSupport: Not applicableReview: Source checked · C31The appendix separates headline value from spend and excludes several overlapping or cancellable commitments.Type: Reported factSupport: SupportedReview: Source checked

Annual OpenAI compute model

$750B across 2026–30; the appendix totals $749B. C10The appendix totals $1,170.4B of OpenAI headline value and $749.0B of modeled 2026–30 spend.Type: Calculated outputSupport: Not applicableReview: Arithmetic reproduced · C13OpenAI’s modeled annual compute costs total $750B across 2026–30.Type: Calculated outputSupport: Not applicableReview: Arithmetic reproduced

Commencing notional

$352B + $360B = $712B across 2027–28. These are not annual invoices. C14$712B of two-lab contract notional commences in 2027–28.Type: Calculated outputSupport: Not applicableReview: Arithmetic reproduced

Required growth

217.4% annually to reach the modeled 2027 cost. C15Revenue must compound at 217% annually from 2025 to equal the modeled 2027 compute bill.Type: Calculated outputSupport: Not applicableReview: Arithmetic reproduced

Funding gap

The stated $375B does not reconcile with the retrieved chart; the exact series is needed. C17The base case leaves $375B of uncovered compute cost in 2026–30.Type: Calculated outputSupport: Not applicableReview: Unresolved conflict

Runnable audit · Results · Transcribed inputs · Source record

Contract register

These two columns are transcribed from the author’s appendix. “Headline” mixes announced amounts, reports and capacity-priced estimates. “Spend” allocates them to the author’s five-year model. Parent rows are summed once; campus allocations are not added again.

OpenAI

Anthropic

Combined: $1,634.4B headline value and $1,078.7B modeled spend. Matching these sums verifies addition, not contract enforceability.

OpenAI–Broadcom announcement supports 10 GW and a rollout through 2029. AMD–OpenAI announcement filed with the SEC supports a 6 GW program. Neither announcement establishes the appendix’s respective $350B and $90B price estimates. The appendix identifies those estimates; the main text’s “every dollar” signed language does not preserve that distinction. C11The $1.2T OpenAI total consists entirely of signed dollar commitments.Type: InferenceSupport: ConjecturalReview: Contradicted as labeled · C12The Broadcom collaboration covers 10 GW with deployment targeted from 2026 through 2029.Type: Reported factSupport: SupportedReview: Source checked

The full-system $2.4T chart requires a separate entity ledger and intercompany eliminations. The two-lab table cannot reproduce it. C07$470B of reported debt plus $1.66T of commitments gives $2.1T of system obligations.Type: Calculated comparisonSupport: Not applicableReview: Not reproduced · C22The consolidated system has $2.4T of notional commencing across 2026–29.Type: Calculated outputSupport: Not applicableReview: Not reproduced

Funding gap

The base-case line in PDF page 17 is not numerically labeled. We use wide reading bounds instead of inventing exact points. The chart carries a draft/redline note, so changed figures may explain the conflict with the prose.

G = \sum_{t=2026}^{2030}(C_t - R_t)

Using cost minus revenue gives $210–280B. Counting only years with a shortfall gives $213–280B. Neither range reaches $375B. These bounds assess compatibility with the displayed chart; they do not estimate the company’s actual financing needs. C17The base case leaves $375B of uncovered compute cost in 2026–30.Type: Calculated outputSupport: Not applicableReview: Unresolved conflict

Compute less revenue also omits opening cash, prepayments, non-compute costs, capital spending and financing. A full funding model must place each on the same timeline without counting the same payment twice. The reported $400–500B all-in funding need and the $450–500B present value of commitments are different quantities. The latter lacks its schedule and discount rate here. C18$665B of commitments has a present value of $450–500B.Type: Calculated outputSupport: Not applicableReview: Not reproduced · C20Bridgewater’s analysis shows OpenAI exhausting its latest fundraise by early 2027.Type: Reported measurementSupport: ReportedReview: Not independently verified

Growth and timing

g = \left(\frac{132}{13.1}\right)^{1/2}-1 \approx 2.174 = 217.4\%

The 217% claim and the approximately 203% compute-to-management-revenue ratio reproduce on the printed inputs. Both compare modeled future values. C15Revenue must compound at 217% annually from 2025 to equal the modeled 2027 compute bill.Type: Calculated outputSupport: Not applicableReview: Arithmetic reproduced · C16Modeled 2027 compute costs exceed 200% of management-plan revenue.Type: Calculated outputSupport: Not applicableReview: Arithmetic reproduced

Adding 24–36 months to signings across 2025–26 permits 2027–29 commencements. A peak in 2027–28 needs weights and dates. The original AWS agreement gave immediate access and targeted full deployment before end-2026. The appendix also marks some agreements already in delivery. C05A 2025–26 signing boom plus 24–36 months mathematically guarantees a 2027–28 commencement boom.Type: Calculated outputSupport: Not applicableReview: Contradicted as labeled

Accounting and liquidity

RPO measures revenue still to be recognized. It includes some previously invoiced amounts; it does not identify all unpaid AI obligations. Microsoft’s definition, CoreWeave’s prepayments and Oracle’s prepaid-hardware disclosure make that distinction material. C02The $2.3T cloud backlog represents compute contracts that are not yet billing.Type: Reported factSupport: ConjecturalReview: Contradicted as labeled

A fixed service commitment can be risky without being a lease. FASB’s lease criteria additionally require control of an identified asset. Credit adjustments, accounting recognition and legal priority need separate analyses. C06Every take-or-pay compute contract is economically a lease and should be treated as senior debt.Type: InferenceSupport: ConjecturalReview: Unresolved interpretation

Depreciation reduces accounting profit. The original asset expenditure and debt payments belong in cash flow; depreciation is not an additional invoice. A low utilization rate can still destroy profit: in an explicit illustration, revenue 120 minus fixed cost 80 minus variable cost 20 gives profit 20. Reducing revenue and variable cost by 30% gives a loss of 10. The article’s possibility claim holds; its 80% fixed-cost share remains an assumption. C23Depreciation behaves like take-or-pay once a data center enters service.Type: InferenceSupport: ConjecturalReview: Unresolved interpretation · C24About 80% of a representative leveraged cluster’s monthly costs are fixed.Type: Model assumptionSupport: Not applicableReview: Not independently verified · C25A 30% utilization decline can eliminate a leveraged facility’s operating profit.Type: Calculated outputSupport: Not applicableReview: Arithmetic reproduced

Similarly, compute consuming about 60% of revenue does not establish whole-firm solvency. Nor do gross purchase commitments alone establish that equity is underwater. Both require the missing cash flows and balance-sheet terms. C19OpenAI’s equity is effectively underwater because of its compute commitments.Type: InferenceSupport: ConjecturalReview: Not established · C21Anthropic is structurally solvent because modeled compute costs peak near 60% of revenue.Type: InferenceSupport: ConjecturalReview: Not established

Historical test

The mortgage analogy has a useful core: obligations can outpace income. Its stronger historical premise fails a contemporaneous check. In March 2007, Federal Reserve testimony already described early payment defaults and lender failures. Payment stress did not have to await a reset. Current performance can be incomplete evidence without being information-free. C01Subprime defaults were the scheduled consequence of rate resets.Type: InferenceSupport: ConjecturalReview: Contradicted as labeled · C29A teaser-period economy cannot produce bad credit data before the reset.Type: InferenceSupport: ConjecturalReview: Contradicted as labeled

Revision tests

1. Release a dated contract register with disclosed dollars, price estimates, minimum purchases, prepayments, acceptance milestones and termination rights in separate fields.
2. Publish annual cash schedules and revenue series. Reconcile the $375B gap, the private-statement present value and the consolidated system total.
3. Test a staggered-buildout case, slower demand, lower prices, delivery delays and a funding interruption. Report minimum cash balances and the timing of financing needs.
4. Set observable thresholds: accepted capacity, cash compute payments, annual revenue, unrestricted cash, committed funding and renegotiated minimums. Record what would weaken the thesis before outcomes arrive.

These tests could strengthen the warning or reduce its estimated scale. The present audit establishes a financing-risk scenario with material qualifications, not a validated crisis forecast.