Section 1
Executive Summary
The Trillion-Dollar CapEx Cycle, Shadow Leverage, and Infrastructure Commoditization
The core finding of this analysis is that systemic risk within the artificial intelligence infrastructure buildout is determined at the layer level. The ecosystem exhibits severe structural fragmentation: one tier generates operational liquidity, a second bears contractual liabilities, and a third carries high-depreciation assets. Any market contraction will propagate unevenly across these balance sheets rather than triggering an immediate, system-wide banking collapse.
However, forensic analysis reveals a market entering a phase of hyper-financialization. While base-layer hyperscalers and frontier laboratories absorb historic levels of capital, downstream vulnerabilities are accumulating through shadow leverage, working capital friction, and circular financing. Cash realization is being actively deferred and collateralized. By decoupling GAAP accounting optics from physical cash burn and mapping the counterparty transmission mechanisms across seven distinct layers, this report locates the true fault lines of the 2026 capital cycle.
Section 2
The Ledger
Read every company in the same order: liquidity, fixed exposure, capital source, then the first loss-bearer.
The Ledger: seven layers, twenty-four records
Twenty-one named companies · three aggregate records · hover, focus or select a record
Read every record in the same order: liquidity, fixed exposure, capital source, then the first loss-bearer. An empty field means not disclosed, never zero.
Section 3
Cash-Rich Platforms and the Gap in GAAP
The capital expenditure required to sustain the platform shift has stretched traditional balance sheet optics to their limits.
Companies: Microsoft, Alphabet, Amazon, Meta, Oracle
First-quarter guidance compiled in April 2026 projected combined 2026 capital expenditures for the four largest U.S. platforms at up to $725 billion, representing a 77% increase over the 2025 baseline. Capital expenditure now accounts for approximately 95% of operating cash flow net of dividends and share repurchases.
Depreciation Stacking vs. Cash Outflows
A single $91.1 billion short-lived vintage at Microsoft implies roughly $15.2 billion a year of straight-line replacement under a six-year life, or $30.4 billion under a three-year economic-life stress case. This maintenance expenditure layers directly onto growth CapEx. To shield current-period earnings from this burden, platforms have aggressively extended the estimated useful life of server hardware. Microsoft extended the estimated useful life of its data centers and office buildings from 15 to 25 years effective FY2027. While these extensions artificially reduce GAAP depreciation expense, they do not mechanically reduce the physical cash required to procure hardware.
Off-Balance-Sheet Commitments (Shadow Leverage)
GAAP lease accounting standard: ASC 842.
Under ASC 842 lease accounting, corporations are not required to recognize formal lease liabilities on their balance sheets until physical control of the asset is achieved. Consequently, the massive volume of data centers under construction sits strictly in the footnotes as uncommenced operating leases and unconditional purchase obligations.
- Meta carries $279.0 billion in uncommenced lease obligations and $349.3 billion in non-cancellable commitments.
- Alphabet holds $811.0 billion in purchase commitments and other contractual obligations.
In aggregate, the top platforms hold approximately $1.67 trillion in off-balance-sheet commitments. While this figure may not encapsulate all future variable purchases, it functions as shadow leverage, reserving substantial future cash flows for facilities well before hardware is installed.
The Cash Conversion Constraint
Oracle provides the clearest leading indicator of execution risk. Oracle carries a 90% unconverted share ($638 billion in Remaining Performance Obligations). While RPO represents contracted future revenue, Oracle is financing the infrastructure required to service this RPO through external debt, generating negative ~$23.7 billion in FY26 free cash flow. This maturity mismatch triggered a downgrade by S&P to BBB− and a negative outlook from Moody's.
Fig. 1 · The funding gap
Operating cash flow against capital expenditure · matched or annualised company periods · US$bn
Source: Microsoft FY2026, Amazon trailing twelve months to June 2026, Alphabet H1/Q2 2026 and Meta Q2 2026 filings. Quarter and half-year figures are annualised only for this silhouette.
Fig. 2 · Oracle: the contract book and the cash clock
RPO accumulation, recognition timing and current funding pressure · separately scaled
Source: Oracle FY2026 Form 10-K and dated RPO series in the 22 August evidence base; rating actions as recorded in July 2026.
Fig. 3 · The commitment horizon
Four filed obligation bases, separated rather than summed · US$bn
Source: Microsoft FY2026 contractual obligations; Alphabet, Amazon and Meta Q2 2026 filings. Prior-period comparison appears only where a matched value is available.
Section 4
Leading Hardware Suppliers and Working Capital Friction
At the epicenter of the hardware supply chain, operational mechanics indicate that unconstrained exponential growth is encountering working capital friction.
Companies: Nvidia, TSMC
Inventory Buildup and Concentration
Nvidia is self-funded, generating $102.7 billion in operating cash flow against $215.9 billion in FY2026 revenue. However, working capital is stretching. Nvidia carries $95.2 billion of non-cancellable forward purchase obligations that must be paid regardless of downstream demand. Customer concentration is severe: Nvidia's top two customers account for 36% of revenue, and the top three hold 64% of receivables as of Q1 FY2027.
Vendor Financing and Circularity
To sustain neocloud unit economics and protect its own top-line revenue, the supplier layer has transitioned into providing structural credit enhancement. Nvidia filed an August 2026 Form 8-K confirming a $105 billion capped residual-value guarantee to SB Energy for the initial phase of an OpenAI campus in Ohio. By guaranteeing downstream project debt, Nvidia's net-cash position becomes subordinated to massive contingent liabilities, establishing a circular revenue loop.
Two-layer concentration at the supplier
Live · Q1 FY2027Legal credit and collection risk
Sits with the entity Nvidia invoices
Receivables at risk
Top three customers, Q1 FY2027
The result is two-layer concentration: legal credit risk sits with the billed counterparty, while economic demand risk may sit with a frontier laboratory one step downstream. Nvidia’s liquidity makes a normal concentration shock survivable, but it does not make earnings or valuation insensitive to one or two customers changing their orders.
Section 5
Frontier Laboratories and the Profitability Bifurcation
The frontier laboratory layer no longer exhibits a uniform financial profile of unmitigated cash burn.
Companies: OpenAI, Anthropic, xAI
Disclosures reveal a sharp bifurcation between enterprise software conversion and heavily subsidized consumer deployments.
- Anthropic: Achieving unprecedented revenue velocity, Anthropic’s run-rate revenue moved from $47 billion in May 2026 to $65 billion by the end of July. By strictly focusing on enterprise API distribution, preliminary figures suggest Anthropic achieved adjusted operating profitability, proving enterprise-focused unit economics can offset model training costs.
- OpenAI: Despite an aggressive $852 billion post-money valuation supported by $122 billion in committed capital, the consumer-heavy model dictates profound operating deficits. Generating roughly $13 billion in recognized revenue, the firm reportedly suffered a $39 billion net loss, driven by a $34 billion direct cost base compounded by significant variable and non-cash expenses.
- xAI: Following its merger into SpaceX, xAI's audited financials reveal it lost roughly three dollars for every dollar earned. In Q1 2026, xAI generated $818 million in revenue against a $2.47 billion operating loss—effectively wiping out more than half of the $4.42 billion operating profit generated by SpaceX's Starlink division over the previous year.
Fig. 4 · Three laboratories, three financial languages
Revenue basis on the left · operating result or burn evidence on the right · no invented common denominator
Source: OpenAI and Anthropic reported run-rates; Reuters’ 20 May 2026 report of Anthropic’s investor projection; xAI Q1 2026 revenue and operating loss filed via SpaceX. All values retain their original basis.
Section 6
Neoclouds and the Refinancing Hinge
Specialized infrastructure providers native to graphics processing units carry the ecosystem's highest refinancing risk.
Companies: CoreWeave, Nebius, IREN, Lambda, Crusoe
This layer carries fixed maturities against assets that may reprice before those maturities arrive.
Debt Stacking and Coverage Compression
CoreWeave illustrates the sharpest mismatch between growth and funding dependence. At June 30, 2026, CoreWeave reported $35.1 billion of funded debt against $104 billion of contracted backlog. While backlog mitigates spot-pricing exposure, cash coverage is thinning. Cash interest nearly doubled year-over-year in Q2 2026 (from $267 million to $640 million), and Q3 guidance implies the Adjusted EBITDA-to-cash-interest coverage ratio is narrowing from 2.3× toward 1.6–1.7×.
Adjusted DSCR & The Convertible Pivot
Evaluating neocloud solvency requires adjusting traditional coverage ratios to account for hardware obsolescence:
To avoid breaching this ratio via high cash-interest burdens, neoclouds are bridging deficits by issuing low-coupon convertible debt. Nebius guided 2026 CapEx to $20–25 billion, turning to near-zero coupon convertibles to shield cash flow. While this delays immediate cash-interest insolvency, it converts equity dilution into a severe maturity cliff if underlying collateral values compress. Furthermore, Lambda Labs pricing a term loan B facility at SOFR + 300 bps alongside a Baa2 rating highlights a highly unusual pricing anomaly in the broadly syndicated loan market, reflecting immense, potentially mispriced collateral demand.
Fig. 5 · CoreWeave’s refinancing runway
Scheduled principal in payment order · segment width = US$bn · coverage on a separate scale
Source: CoreWeave Q2 2026 filing and guidance in the 22 August evidence base.
Fig. 6 · One asset, six repricings
Six market readings · original units retained · physical, quoted and derived states
Source: CoreWeave financing and market readings; private-credit PIK and BDC income figures in the 22 August 2026 evidence base.
Section 7
Data-Centre and Power Projects: The Duration Mismatch
The physical constraints of power generation and site deployment serve as the ultimate bottleneck.
Companies: Applied Digital
Companies operating at this nexus command extreme valuation premiums based entirely on future capacity.
Applied Digital generated $611.3 million in FY2026 revenue (up 167%) but reported a GAAP net loss of $249.2 million. The enterprise is anchored entirely by $36.2 billion in base-term lease commitments. This layer relies heavily on 15-year non-cancellable leases underwritten by CoreWeave and hyperscalers. The structural risk here is a pure duration mismatch: multi-decade physical buildouts are funded with high-yield short-term developer bonds. If a neocloud defaults, the resulting impairment bypasses the hyperscaler and lands directly on the physical site developer.
Fig. 7 · Asset life and the contract clock
Seven asset classes · Applied Digital’s site, lease and debt · one cell = five years
Source: asset-life ladder and Applied Digital financing disclosures in the 22 August 2026 evidence base.
Section 8
Applications and the Monetization Chasm
The capital accumulation in Layers 1 through 5 assumes the application layer will generate high-margin software revenue sufficient to cover foundational CapEx.
Companies: Perplexity, Harvey, C3.ai
Disclosures indicate this transition is stalling.
- Public Market Compression: C3.ai's FY2026 revenue fell to $250.3 million, with subscription revenue down 31% year-over-year, against a $470.4 million net loss. The expansion of operating losses alongside revenue contraction at the peak of global infrastructure spending points to a systemic failure in enterprise deployment.
- Private Market Exuberance: Conversely, specialized private applications continue to absorb capital at historic multiples. Harvey is in talks for a $15.5 billion valuation on reported revenues surging past $350 million. This dichotomy highlights a fundamental dislocation regarding the ultimate monetization timeline of generative software. Because this layer is entirely equity-funded, failure here is dilutive and equity-absorbed, not credit-transmitting.
Section 9
Credit Intermediaries and Systemic Transmission
As commercial banks rapidly originate loans to fund data centers, they risk breaching internal sectoral concentration limits.
Companies: Apollo Global Management, Blue Owl Capital, Commercial Banks
To manage regulatory capital, risk is being quietly offloaded into the private markets, setting the stage for opaque systemic vulnerability. The Financial Stability Board estimates the entire private-credit market at $1.5–2.0 trillion.
PIK Swelling & Cash Realization Failure
The private credit transmission channel is visibly shifting from cash collection to non-cash accruals. Payment-In-Kind (PIK) interest—where borrowers pay interest by issuing more debt rather than remitting cash—has swelled to approximately 11% of the aggregate private credit market. Funds are booking non-cash PIK to avoid default declarations, masking cash-flow insolvency at the infrastructure layer.
SRT Risk Transfer and Ownership Overlap
To free up balance sheet capacity, Tier-1 investment banks are structuring Significant Risk Transfers (SRTs). The originating bank retains the underlying assets but offloads the first-loss tranche (0% to 8%) to a private credit firm in exchange for high yields.
The embedded moral hazard comes from ownership overlap: the private equity entities purchasing these SRT tranches are frequently the exact same firms that hold equity stakes in the underlying data center assets. If a wave of neocloud defaults triggers widespread data center lease breaks, the resulting losses will bypass the heavily regulated banking sector and detonate directly within highly levered private credit portfolios.
Section 10
The Emerging “Cash Delay & Circularity Loop”
The private credit transmission channel is visibly shifting from cash collection to non-cash accruals.
Accounting Life Smoothing vs. Hard Cash Outflows
Useful life extensions across hyperscalers (Microsoft 4→6 years, Meta to 5.5 years, Alphabet 6 years) have mechanically shielded GAAP operating income by billions ($3.7B at MSFT, $2.9B at Meta) while masking underlying cash burn. Amazon's reversal (shortening from 6 to 5 years, adding $889M in depreciation) is the sole honest outlier reflecting actual hardware obsolescence.
Formalized Vendor Financing & Utilization Floors
Nvidia is no longer merely an arms-length supplier; it has evolved into a structural credit enhancer and buyer-of-last-resort. Between the $105B OpenAI Ohio backstop, a $1.5B direct equity stake in the project, and a programmatic model renting back unused GPU capacity from neoclouds, supplier revenue is directly backstopping customer utilization.
Neocloud Debt Stacking via Low-Coupon Convertibles
Neoclouds are bridging massive cash deficits by issuing convertible debt (Nebius $5B at 0.5%–4.5%, IREN $3B at 1.00%). This temporarily lowers current cash interest burdens relative to high-yield bank debt, but converts equity dilution into a future debt maturity cliff if underlying share prices compress.
Private Credit PIK Swelling & Bank Risk Shedding (SRTs)
Payment-In-Kind (PIK) interest—where borrowers pay interest by issuing more debt rather than remitting cash—has expanded from 5–7% to 10.6–11.0% of the private credit market. Public BDCs are recognizing tens of millions in non-cash PIK income ($54M at Ares, $38M at FS KKR, $31.5M at Blue Owl). Meanwhile, Tier-1 banks (Morgan Stanley, JPMorgan, Citi, Goldman Sachs) are actively marketing Significant Risk Transfers (SRTs) to offload data-center loan risk to private funds.
Fig. 8 · The cash delay and circularity loop
Five stations · cash and credit descend · risk and non-cash claims move upward
Source: “The Emerging Cash Delay & Circularity Loop,” Research and Analysis 1. Station and link text reproduced verbatim.
Section 11
The Dot-Com Comparison
The dot-com label compresses four different events, with four different loss-bearers.
Which 2000?
The dot-com label compresses four different events. Equity-funded applications failed when new capital disappeared. Cisco remained solvent while its share price collapsed. Debt-funded telecom carriers entered bankruptcy. Fibre and physical network assets survived and were reused at lower values.
The cleanest present pairing is between leveraged carriers and neoclouds. Both fund rapid capacity growth with debt against demand that may weaken before the assets or obligations expire. Cash-rich platforms and Nvidia resemble Cisco more closely: corporate survival can coexist with a severe valuation loss and large impairments.
At the March 2000 peak, Cisco held no material funded long-term debt and no customer supplied ten per cent of sales. Nvidia is also liquid, but its top three customers now hold 64 per cent of receivables and it carries of forward supply obligations. The analogy holds on solvency. It weakens on concentration.
Fig. 9 · Four events called 2000
Four separate failures · paired with the corresponding 2026 layer · outcome shown inside each case
Source: NTIA telecom-bust record, Cisco FY2001, and the settled lifeline register in the 22 August evidence base.
The macro setting, 2000 against 2026
Mixed · history settledmore loss-absorbing similar reading less loss-absorbing / more opaque
Source: Federal Reserve H.15/FRED; BLS CPI and productivity; BEA GDP and corporate profits; CBO budget outlook; FINRA margin statistics; FSB private-credit vulnerability review.
Section 12
The 2008 Comparison
The comparison turns on whether multiple claims reference the same underlying exposure.
Layered claims are not yet a synthetic CDO
The 2008 comparison turns on one mechanical property: synthetic instruments referenced the same mortgage pools repeatedly, allowing notional exposure to exceed the face value of the loans. AI infrastructure already has layered claims on the same asset, including equipment finance, secured loans, assigned customer contracts, SPV equity, securitisation and risk transfer.
What remains unproven is the closed loop. No public disclosure shows overlapping reference portfolios, notional above the underlying exposure, or bank capital impaired by an AI data-centre chain. The Financial Stability Board identifies roughly of bank credit lines to private-credit funds, but that is private-credit-wide and below 0.5 per cent of bank assets in reporting jurisdictions.
Rung four in the claims ladder is the hinge: the customer contract assigned to the lender. It is the first point where cash from an unaffiliated payer enters. Every claim below it depends on that same customer cash from farther away. By the last rung, the paying party can no longer be identified from public disclosure.
Fig. 10 · Nine claims on one chip
Nine claim types · payer at each rung · evidence partition after rungs five and eight
Source: nine-rung claims register in the 22 August 2026 evidence base; FSB and FCIC references as recorded there.
Section 13
Futures, Falsification and Gaps
Scenarios identify what must change before a loss moves from equity into credit.
Scenarios are not forecasts. They identify what must change before a loss moves from equity into credit.
A · Soft landing. Usage and paid revenue grow fast enough to absorb falling unit prices. Backlog converts, renewals hold, and CapEx moderates without impairments.
B · Infrastructure bust. Demand persists but prices and margins fall. Applications and weaker laboratories lose equity funding first. Neocloud refinancing spreads widen, collateral values compress, and projects restructure.
C · Credit transmission. The infrastructure bust becomes correlated across funds, insurers and bank counterparties. This branch requires loan-level exposure and SRT data that nobody publishes.
D · Depreciation mirage. Reported CapEx or earnings improve because useful-life assumptions change while the physical build continues. The first loss is a pricing error in public equity. A second comparable useful-life extension at another platform would make this scenario live rather than a one-company observation.
The most useful near-term signal is not a headline default. It is the spread and covenant package on the next neocloud refinancing, read beside backlog conversion and customer concentration.
Fig. 11 · The verdict switchboard
Four paths · decisive confirm and falsify conditions · current reading stated without a score
Open the 25-indicator evidence register 15 linked to Scenario B · open to inspect
Source: scenario definitions, falsification tests and 25-row leading-indicator register in the 22 August 2026 evidence base.
Twelve circulating claims, against what the primary source says
Corrections · settledWhat nobody publishes
Corporate and laboratory economics. No company publishes a complete AI-specific bridge from cash revenue to depreciation, power, inference cost and replacement CapEx.
Credit and counterparty exposure. Loan-level AI exposure, SRT reference portfolios, protection sellers, attachment points and bank residual exposure remain unavailable. This is the gap that could move the conclusion from selective loss to systemic stress.
Contract durability. Cancellation, assignment, price-reset, availability and cross-default terms are not disclosed consistently. Backlog is only as strong as its weakest enforceability clause.
Collateral and recovery. Forced-sale prices by accelerator generation and redeployment costs after tenant failure are not public.
Historical comparability. No matched market-cap basket applies consistent constituents across 2000 and 2026. Every aggregate comparison must therefore remain approximate.
Section 14
Source appendix
Disclosed facts remain separate from estimates, commitments, run-rates and company-wide totals.
Every load-bearing figure resolves to a dated source or is marked as derived, estimate, forecast or gap. The bibliography below is the audit trail; these rules explain how to read it.
Source classes
Company disclosure Filings, earnings materials, investor announcements and product documentation. The primary record.
Regulator & official SEC, FSB, central-bank, statistical-agency, court and tribunal records.
Institutional research Disclosed market datasets and research with an identifiable method. Treated as estimate unless independently filed.
Lama Research Aggregations and arithmetic produced here from cited inputs. Reproducible, but not a company disclosure.
Disclaimer
This report is for information only and is not investment advice, an offer, or a solicitation. Estimates, forecasts and derived figures may prove wrong; figures marked gap are known to be missing. Historical comparisons describe a record and do not guarantee repetition.
State taxonomy
Actual is printed in a filing or official release. Estimate is an outside number where no filing exists. Forecast is company guidance or another stated claim about the future. Derived is reproducible arithmetic on cited rows. Gap is a number that does not exist in auditable form and is never interpolated into silence.
Counting caveats
Company-wide totals are not relabelled AI-specific. Run-rate revenue is not audited annual revenue. Backlog and RPO are not cash. Commitments are not assumed funded. Ranges remain ranges, and mixed accounting bases are never added without a stated reconciliation.
BibliographyPrimary, official, institutional and corroborative recordsOpen
Company filings and announcements — Tier 1
- Nvidia FY2026 Form 10-K — revenue, earnings, liquidity, customer concentration, supply commitments.
- Nvidia Q1 FY2027 Form 10-Q — three-customer revenue and receivables concentration, inventory, indirect demand.
- Microsoft FY2026 Form 10-K and Q4 materials — CapEx, useful lives, RPO, obligations, lease-classification revision.
- Microsoft quarterly earnings: Q1, Q2, Q3 — CapEx composition and the OpenAI-linked RPO share.
- Alphabet Q2 2026 Form 10-Q, Q4 2025 call and useful-life FAQ — H1 PP&E, cloud RPO, the 60/40 mix, six-year server life.
- Amazon Q2 2026 Form 10-Q, Q2 results and 2025 shareholder letter — AWS additions, trailing cash CapEx, lab investments, free cash flow.
- Meta Q2 2026 Form 10-Q — uncommenced leases, non-cancellable commitments, restricted cash, residual-value guarantees.
- Meta project and supplier announcements: Hyperion / Blue Owl, El Paso / BlackRock, AMD, Nebius, Corning — named transaction values and structures, not a complete allocation of Meta's aggregate notes.
- CoreWeave Q2 2026 earnings exhibit and CoreWeave–Meta Form 8-K — revenue, backlog, debt, CapEx, interest, and the ~$21B commitment through 2032.
- Nebius Q1 2026 filing and Microsoft agreement; IREN–Microsoft, IREN–Nvidia; Applied Digital 8-K; Lambda; Crusoe.
- Oracle FY2026 Form 10-K — $638B RPO, recognition schedule, customer prepayments.
- OpenAI funding announcement and Anthropic Series H — committed capital, valuations and run-rate revenue. Neither is a financial statement.
- Reuters, 20 May 2026 — Anthropic’s investor projection of $559M Q2 operating profit on $10.9B projected revenue. Carried as forecast, not audited actual.
- C3.ai FY2026 Form 10-K — revenue, subscription revenue, net loss and accumulated deficit.
- TSMC 2026 capital-budget record — $60–64B company guidance and the separate $100B Arizona commitment, carried from the verified 22 August evidence file. The ledger displays the $62B midpoint only as a range marker.
Historical record — Tier 1
- Cisco 2001 Annual Report — FY2000–01 liquidity, sales, inventory provision, debt position, customer concentration.
- Financial Crisis Inquiry Commission, Final Report — synthetic CDOs referencing the same mortgage bonds repeatedly, and the resulting multiplication of exposure beyond the underlying loans. The single citation added for this edition; the mechanism is not in the source report's own evidence base.
- Genuity 2001 Form 10-K — 2000 gross PP&E categories and useful lives, used as a representative rather than industry-wide proxy.
- FCC dark-fibre order, SEC telecom-accounting testimony and SEC Qwest order — end-2000 dark fibre, IRU accounting, reciprocal capacity transactions.
- NTIA remarks, OECD, After the Telecommunications Bubble and Federal Reserve retrospective — bankruptcies, layoffs, CapEx contraction, overcapacity.
- FHWA Interstate FAQ and FHWA cost history — 46,876 miles, 1956–92, $128.9B cumulative then-year cost with a $114.3B federal share.
Regulators and statistical agencies — Tier 1
- FSB, Vulnerabilities in Private Credit — market size, bank links, leverage, data gaps.
- Federal Reserve Financial Stability Report and Bank of England Financial Stability Report — near-term risks, AI debt, private credit, SRT.
- BEA Q2 2026 GDP advance estimate; BLS Q2 2026 Productivity and historical table; CBO 2026–2036 Outlook; OMB historical tables; FINRA margin statistics.
- FRED series: NASDAQCOM, FEDFUNDS, GDP, CPIAUCSL, CPROFIT — comparison inputs, retrieval cutoff 11 August 2026.
Institutional research and market data — Tier 2
- Goldman Sachs, Tracking Trillions — the $7.6T 2026–31 baseline and useful-life sensitivities. This is the source of that figure; it is not a J.P. Morgan estimate.
- J.P. Morgan Asset Management, How is AI being monetized? — the ~$650B annual revenue hurdle, the figure most often confused with Goldman's capital baseline.
- J.P. Morgan Asset Management, Smothering Heights and Silicon Data, Silicon Index — comparable H100 hourly rental series and generation-specific rental indices.
- BofA hyperscaler cash-flow chart — consensus CapEx as a share of operating cash flow after dividends and repurchases. An aggregate analyst measure, not audited AI-only CapEx.
- NVCA 2026 Yearbook and S&P DJI, In the Shadows of Giants — venture concentration and index concentration.
- McKinsey, State of AI and Gartner agentic-AI forecast — adoption and EBIT attribution; a cancellation forecast, not a count.
- DeepSeek-V3 technical report — 2.664M H800 GPU-hours and an estimated $5.576M official training cost, excluding prior research, ablations, data and broader organisational cost.
Corroborative aggregation — Tier 3
- 2026 four-company CapEx compilation — based on first-quarter guidance compiled by the Financial Times; up to $725B for 2026 against $410B in 2025. Not relabelled as audited AI-only spending.
- BofA bond-issuance reporting — $121B of 2025 issuance across five firms against a $28B average. Used with company filings, not as an AI-debt total.
- 2026 Interstate retrospective — the ~$634B current-purchasing-power estimate. The official FHWA nominal series remains the base fact, and the conversion is labelled approximate.
How to update this page
Every time-varying figure lives in a single data file rather than in the prose. A quarterly refresh means editing roughly forty-five numbers in that file and rebuilding; nothing numeric is typed into a sentence. Seven of those figures carry an explicit guard — a threshold beyond which the sentence around the number stops being true — so a refresh surfaces the contradiction instead of quietly publishing a stale claim. Panels marked settled draw on the historical record and do not update.