Historical precedents for the AI investment boom

Last updated: 2026-10-17 · Author: Claude (claude-fable-5-1 subagent) · Status: reviewed · Reviewer: Codex (gpt-6-astra) cross-check, applied 2026-10-17

Summary

• Seven selected episodes are compared for mechanism: British Railway Mania, electrification and the 1929 utilities/RCA boom, telecom/fiber, dot-com equities, Japan's credit bubble, US housing, and crypto/SPACs. Telecom and dot-com overlap; these are not independent statistical observations.
• Four cases involve technologies that became widely useful despite substantial investment losses. Losses, ownership changes and recovery times differed; the evidence establishes neither a common capacity-absorption period nor a requirement for bankruptcy before useful assets can earn adequate returns (claims 1–30 and Comparison table).
• Financing structure helps determine how losses propagate. Leverage, maturity mismatch, collateral and guarantees can transmit losses beyond equity holders. These cases do not establish a universal recovery ordering or show that equity-led booms cannot coincide with banking stress (claims 13, 16, 25, 48–49).
• Telecom vendor support is a useful comparison: Lucent disclosed $5.391 bn undrawn and $1.263 bn drawn credit at September 2000; Cisco's FY2001 allowance expense was $2.775 bn. Today's connections include Nvidia's phased maximum $105 bn contingent lease guarantee, AMD's conditional purchase-linked warrant and Amazon's $50 bn cumulative OpenAI investment. Loans, guarantees, equity investments and warrants have different loss mechanics (claims 13, 15, 33, 35–36).
• Historical indicators give mixed timing evidence. Annual IPO counts and returns fell, but cannot establish quarterly leads; Lucent's December 2000 downgrades followed the March equity peak. Mortgage and funding-market stress preceded the October 2007 equity peak, so Baa–Aaa alone is insufficient (claims 20, 28, 47, 54).
• External financing has expanded, while warning signals are mixed: September's Baa–Aaa spread was 0.44 pp, but the Fed raised rates and its October minutes describe earlier hyperscaler spread widening. Those minutes also report earnings-supported equity gains and declining multiples. The evidence does not identify a reliable bubble-cycle stage (claims 32, 34, 40–42). Issuance alone does not prove a contemporaneous capex cash deficit (claims 44–46).
• A key uncertainty is the joint effect of demand, margins, financing terms and different asset lives: hardware may need replacement while buildings and power connections remain usable. Profitable reuse depends on acquisition price and continuing costs. The Brookings buildout comparison is a scenario, not committed or realised spending (claims 43, 50).

Key claims

Analysis

The seven episodes

British Railway Mania (1845–50). Steam railways offered a real transport improvement. Parliament sanctioned 4,538 miles in 1846 (claim 3). Financing relied heavily on individuals' partly paid equity, with further capital calls during construction. Cumulative capital rose from £88.5 m to £245.8 m between 1845 and 1850; peak annual investment was approximately 7.3% of GDP (claim 1). Revenue reached £15 m in 1852 against Odlyzko's reconstructed £60 m expectation, not an independently established industry consensus. The major-line share example fell 55%; York & North Midland's £120 to £20 fall illustrates the weaker-line losses (claims 3–4). Odlyzko's aggregate loss estimate is about a third of invested capital, not most or all (claim 2). Mileage grew to approximately 2.7 times its 1845 level by 1850 and revenue reached 6% of GDP by 1905, supporting the usefulness of the technology without proving full utilisation or adequate original-investor returns (claims 5, 53). The 1847 financial crisis and emergency liquidity operations mean this equity-led episode cannot serve as an unqualified no-banking-crisis control; the archival evidence does not establish railway lending as the crisis's cause (claim 48).

US electrification and 1929. Electrification and radio supplied the technology thesis. RCA rose from 85 to 549; utilities were 18% of NYSE value, while holding companies and investment trusts layered financial claims over operating assets (claims 7–8). Utilities fell 55% between end-September and end-November versus 48% for industrials; these endpoints alone do not identify which turned first. The Dow lost 89% and regained its nominal peak in November 1954 (claim 6). Baa–Aaa rose from 1.01 pp in January 1929 to 1.32 pp in September and 5.59 pp in June 1932 (claim 9); that selected rating spread does not establish the timing of all credit warnings. The useful lesson is how leverage can amplify sector losses, with aggregate overvaluation disputed (see Conflicts).

Telecom and fiber, 1997–2002. This is a useful mechanism comparison because equipment suppliers supported their customers' investment. The demand narrative claimed traffic doubled every 100 days; the cited measurements indicated roughly annual doubling (claim 17). Public service-provider capex rose from $47 bn to $121 bn before falling to $49 bn (claim 10). Financing included CLEC equity, bonds and vendor credit. CLEC market value fell from $86.4 bn to $32.1 bn and then $3.77 bn; WorldCom filed with $41 bn of debt (claims 11, 16). Lucent disclosed $5.391 bn of undrawn and $1.263 bn of drawn customer credit, separate guarantees, and customer-loan sales to a trust (claim 13). Nortel's original filing citation reports up to $4.1 bn unfunded financing, but the independent review could not re-fetch it (claim 14; see Conflicts).

The telecom equity peak came in March 2000; Lucent's December guidance and rating cuts were confirmation after that peak (claims 11, 54). Cisco's $2.775 bn full-year FY2001 allowance expense included a separately identified $2.249 bn additional Q3 charge (claim 15); the annual total is not one sudden quarterly event. The telecom index fell approximately 88.9% through May 2003, and the stated employment endpoints imply an 18.2% decline, despite the paper's printed 22% (claims 11–12). WorldCom's bankruptcy illustrates the destruction of equity alongside a continuing network, but does not establish universal profitable reuse. The note has no verified capacity-absorption end date. Richmond's technology/regulatory explanation and Doms's rational spare-fiber argument qualify a pure speculative-overbuild interpretation (claim 52).

Dot-com equities, 1995–2002. The commercial internet became widely useful while many securities failed. Equity financing through IPOs expanded: 476 offerings with a 71.2% mean first-day return in 1999, 380 in 2000 and 80 in 2001; 1999 was not the count peak, since 1996 had 677 (claim 20). CAPE reached 44.2 in December 1999 (claim 19). Nasdaq lost 77.9% over 31 months and regained its nominal price peak after 15.1 years (claim 18). Annual IPO observations establish a contraction, not a quarterly lead. This case illustrates substantial equity losses without demonstrating that financing type alone explains recession severity or excludes banking exposures.

Japan, 1986–90. Bank credit against land, rather than a technology thesis, is the relevant mechanism. The reported stock-and-land gains were 452% of GDP over 1986–89 (claim 23). BOJ tightening ran from May 1989 to August 1990; equities peaked after the third rate increase, while the six-city commercial-land index peaked about nine months after equities (claims 21–24). The Nikkei fell 81.9% and took 34.1 years to regain its nominal high. The land decline of approximately 80% refers specifically to JREI commercial land in six major cities, not nationwide property (claims 21–22). Banks and public finances were exposed, but ¥60 tn authorised for support is not a measured realised taxpayer loss (claim 25).

US housing, 2003–09. Mortgage debt, securitisation, repo and off-balance-sheet vehicles linked housing prices to funding markets. New Century's April 2007 bankruptcy was a prominent failure, nine months after the July 2006 house-price peak and six before the October 2007 equity peak, not evidence that no lender failed earlier (claim 28). The Fed documents mortgage repricing, securitisation disruption, leveraged-loan stress and ABCP spread widening before the equity peak (claim 47). Baa–Aaa alone concealed important sector and funding stress. The national house-price index lost 27.4%, troughed after 67 months and regained its nominal level after 10.5 years; the S&P 500 fell 57% (claims 26–27). Neither price stabilisation nor index recovery measures the absorption of vacant housing.

Crypto and SPACs, 2020–22, with banking spillovers observed through 2023. Tokens, retail equity and SPAC trust cash were prominent financing channels, with different mechanisms grouped in this case. SPAC IPOs fell from 613 in 2021 to 86 in 2022; redemptions on completed deals rose from under 10% in Q1 2021 to 60% in Q4, a different population from new listings (claim 30). Operating-company IPO counts also fell (claim 20). Bitcoin lost 76.7% over approximately 12.4 months and regained its nominal high in 2.3 years. FTX filed ten days before the November 21, 2022 trough (claims 29–30). Price recovery alone does not demonstrate crypto's payments thesis or identify its cause. Extending observation beyond 2022 reveals Signature's March 2023 failure: its crypto-sector ties and concentrated uninsured deposits invalidate an unqualified claim of no banking spillover, without establishing direct crypto lending as the cause (claim 49).

Comparison table

Four selected cases combine useful technologies with substantial investment losses. This does not establish a historical failure frequency: telecom and dot-com overlap, asset classes differ, and railway recovery is unmeasured. Leverage and funding structure provide plausible loss-transmission mechanisms, not a demonstrated causal ordering of drawdowns or recovery lengths. Nominal price recovery, original-investor total return and physical-capacity utilisation are separate outcomes (claim 56). The evidence does not show that every asset was eventually used, changed owner or became profitable after a distressed sale.

What the precedents imply for AI in 2026

Which precedent fits which layer. These are mechanism hypotheses, not calibrated impairment forecasts.

• Chip equipment (ASML, Applied, Lam, KLA): railway contractors and telecom equipment makers suggest examining cancellations, inventories, receivables and financial exposure. The comparison does not establish a short earnings downturn or an absence of balance-sheet damage.
• Foundry (TSMC): incumbent carriers are a possible comparison, but BellSouth and Qwest's relative resilience also reflected an access-pricing ruling (claims 11, 52). It cannot establish that TSMC is the safest layer or eliminate solvency risk; utilisation, customers and financing need separate assessment.
• GPUs (Nvidia, AMD): vendor support creates links resembling telecom's, with different instruments. Nvidia's $105 bn phased contingent guarantee and separate $36 bn capacity backstops are not present drawn credit; AMD's warrant is conditional dilution (claims 35–36, 51). Lucent's undrawn/drawn credit ratio was 4.27× (claim 13). Relative exposure requires common revenue periods and instrument-specific valuation; the figures do not establish a tenfold increase. Cisco's annual allowance expense and Q3 charge illustrate inventory risk (claim 15), but no measured customer-financing peak establishes a universal lag.
• Hyperscalers (Microsoft, Alphabet, Amazon, Meta): utilities and established railways suggest studying large committed projects and their financing. The supplied Microsoft, Amazon and Meta capex figures are approximately 1.0–1.4 times the $121 bn telecom peak in unadjusted dollars, mixing periods and definitions (claims 10, 31, 33–34, 55). The Brookings scenario instead supplies a GDP-normalised comparison: $10.3 tn over 2025–32, averaging 3.63% of GDP against 1.10% for its selected telecom/fiber period. It is neither realised spending nor a commitment, and separates hardware from durable infrastructure (claim 43).
• Neoclouds (CoreWeave, Nebius, Lambda): CLECs and long-haul wholesalers suggest a stress scenario involving debt, concentrated customers and contract renegotiation. The debt amount and 7–15% facilities cited here are CoreWeave-specific (claim 38), not established for all three. The CLEC equity collapse is a stress path, not an evidenced base case for neoclouds (claim 11). Continued asset operation and returns to new owners would depend on residual values, power, maintenance and replacement costs.
• Labs (OpenAI, Anthropic, xAI): WorldCom illustrates demand concentration interacting with financing, without making the companies equivalent. OpenAI connects Nvidia guarantees, AMD's warrant, Amazon's cumulative $50 bn investment and Oracle's RPO increase (claims 33, 35–37). These disclosures warrant counterparty analysis; they do not establish that no historical customer connected several vendors, or that all Oracle RPO belongs to one lab. Do not sum investments, guarantees and backstops as financed GPU revenue.

External capital supplements hyperscaler funding, but issuance alone does not establish a contemporaneous capex cash deficit. Alphabet's H1 operating cash exceeded rounded capex by about $4.26 bn, Meta's by $13.17 bn, while Amazon's TTM comparison was −$7.604 bn; Alphabet's Q2 alone showed capex above operating cash (claims 44–46). These different periods and definitions cannot be summed. Meta's Hyperion venture adds a separate ownership and loss-allocation channel: Blue Owl-managed funds own 80%, Meta 20%, with a four-year initial lease and a conditional capped residual guarantee across the first 16 operating years (claim 50). The guarantee's contract-level exposure remains to be assessed.

Which indicators to watch now, from history.

1. Vendor commitments and their effectiveness. Track drawn credit, contingent guarantees, capacity commitments, warrant vesting and receivable sales separately. Lucent's cited September snapshots show both drawn and undrawn credit lower than a year earlier, not an established shift toward drawdown (claim 13).
2. IPO window. Annual counts and first-day returns document contractions, not reliable quarterly leads. The 2025 observations are historical context, not a measurement of the 2026 window; a large first-day pop is not a validated crash threshold (claim 20).
3. Sector equities versus broad equities. Telecom divergence motivates monitoring customer-exposed firms, but regulation confounds the historical incumbent comparison (claims 11, 52). A future split would need interpretation, not automatically date the first warning.
4. Guidance and ratings. Lucent's December 2000 cuts followed the March equity peak; they confirmed deterioration rather than predicting that peak (claim 54). Monitor both vendors' cancellations and buyers' funding.
5. Policy and all-in funding costs. Japan's equities peaked after the third rate rise, before tightening ended (claims 21, 24). On September 16, 2026 the Fed raised its range 25 bp to 3.75–4.00%; tightening is an active risk. October minutes report an approximately 35 bp intermeeting Treasury-yield rise and earlier hyperscaler spread widening, despite some subsequent narrowing (claims 41–42). Those minutes also report earnings-supported equity gains and falling multiples, counterevidence to a simple euphoria account. These are mixed signals, not a reliable cycle-stage classification.
6. Instrument-specific credit and funding availability. Baa–Aaa alone missed important mortgage and money-market stress before the October 2007 equity peak (claim 47). Watch sector spreads, spreads over government benchmarks, lending terms and funding availability alongside Baa–Aaa and broad HY OAS (claims 9, 28, 40). Narrow rating spreads do not negate sector repricing or higher all-in costs.
7. Redemption-style exits. SPAC redemptions suggest watching reductions in contracted volumes and renegotiations while deals continue to be announced (claim 30). Oracle RPO revisions are one possible signal, but are not mechanically equivalent to cash redemptions (claim 37).

Conflicts and open questions

• 1929 valuation. Bierman and McGrattan–Prescott dispute aggregate overvaluation; Bierman singles out utilities. The original note favoured sector leverage over aggregate valuation. The cross-check retains leverage as a useful mechanism but finds that the selected sector-loss endpoints do not establish which sector turned first. Both qualifications matter.
• Universal historical lessons and recovery. The original note claimed six technology successes, universal eventual reuse after ownership transfer, and longer recoveries under credit finance. The cross-check counts four successes in the note's own table and rejects those universal and causal claims: railway aggregate losses were estimated at about a third, recoveries mix assets and omit railway recovery, and no common utilisation series exists. The revised Analysis treats these as selected mechanisms. Nominal index recovery, with changing constituents, does not restore failed firms' equity or measure total return or utilisation.
• Railway loss size and banking links. Odlyzko's estimate remains C; Campbell and Turner's differing assessment of investor sophistication has not been verified here. The original equity-finance/no-banking-crisis framing differs from the cross-check's reading of the 1847 archival evidence (claim 48). A concurrent financial crisis is established; railway-loan causation is not.
• Telecom explanation and utilisation. The original note emphasised speculative overbuild and universal distressed reuse. Richmond stresses technology/regulatory uncertainty, including access pricing; Doms argues spare fiber can be rational when reopening trenches is expensive (claim 52). The cross-check regards these as material contrary explanations. Low utilisation alone does not identify irrational investment; no verified absorption interval or full-reuse series is available. The repeated few-percent-lit figure remains unused.
• Telecom employment arithmetic. Richmond prints 22%, but its 1.59 m and 1.30 m endpoints imply approximately 18.2%. The review's arithmetic is used; the inconsistent source statement is retained explicitly in claim 12 rather than silently discarded.
• Financing and cycle stage. The original note interpreted issuance as cash-flow funding exhaustion and placed AI before indicator reversal. The cross-check instead finds mixed warnings: rate tightening and sector-credit repricing coexist with narrow broad spreads, earnings-supported equities and positive H1 cash-minus-capex at Alphabet and Meta (claims 40–46). The revised note does not assign a calibrated stage or infer cash deficits from issuance alone.
• Credit timing. The original note treated spreads as lagging confirmation only; the cross-check points to summer 2007 mortgage and funding-market warnings (claim 47). Baa–Aaa is a different instrument from sector spreads or spreads over government debt; timing depends on both instrument and reference peak.
• Analog strength and loss allocation. The original note made the CLEC path a neocloud base case and treated vendor exposures as economically comparable. The cross-check supports a contingent-credit stress scenario, not that base case or the tenfold size comparisons. Loans, equity, conditional dilution, backstops and guarantees require separate exposure and maturity analysis. Meta's primary Hyperion announcement resolves the original filing-text-search omission, but the residual guarantee still needs contract-level treatment (claim 50).
• Crypto horizon. The original no-banking-loss assertion stopped at 2022; the cross-check extends spillover observation to Signature in 2023 (claim 49). Crypto-linked deposits are a transmission channel, not proof that direct crypto loans caused the failure. Bitcoin's price recovery does not establish a payments-use thesis or its cause.
• Brookings source date. The cross-check labels the draft September 23. The fetched PDF identifies a September 4 draft for the September 24–25 conference; this note records those source dates rather than inferring a September 23 publication date. The scenario values agree.
• Asset lives and the new scenario. The original fiber-versus-PC framing treated a universal short hardware life as decisive and assumed profitable second ownership. The cross-check finds no established universal GPU economic life; hardware, buildings and power differ. Brookings distinguishes those components and cautions against prematurely declaring housing-style systemic risk (claim 43). Demand, margins, refinancing and operating costs jointly determine losses and reuse; a low purchase price alone is insufficient.
• Unverified leads. Nortel's legacy $4.1 bn filing could not be independently re-fetched in the cross-check; claim 14 retains the original direct-source grade and citation, not a new verification. Its separate reported quarterly-loss figure is omitted here rather than treated as a filing fact. Global Crossing transaction values, insider selling, quarterly IPO leads and the historical vendor-counterparty census remain unestablished. The review's September SB Energy prospectus lead was not verified and its reported prepayment/share-purchase details are not added.

Signals to watch

Sources

1. Andrew Odlyzko, "Collective hallucinations and inefficient markets: The British Railway Mania of the 1840s", preliminary version 2010-01-15. https://www-users.cse.umn.edu/~odlyzko/doc/hallucinations.pdf. Accessed 2026-10-17.
2. Gary Richardson, Alejandro Komai, Michael Gou, Daniel Park, "Stock Market Crash of 1929", Federal Reserve History, 2013-11-22. https://www.federalreservehistory.org/essays/stock-market-crash-of-1929. Accessed 2026-10-17.
3. William H. Donaldson, SEC Chairman, speech to the Detroit Economic Club, 2004-10-29. https://www.sec.gov/news/speech/spch102904whd.htm. Accessed 2026-10-17.
4. Harold Bierman Jr., "The 1929 Stock Market Crash", EH.Net Encyclopedia, 2008-03-26. https://eh.net/encyclopedia/the-1929-stock-market-crash/. Accessed 2026-10-17.
5. Ellen McGrattan and Edward Prescott, "The 1929 Stock Market: Irving Fisher Was Right", Minneapolis Fed Staff Report 294 (published in International Economic Review, Nov 2004). https://www.minneapolisfed.org/research/staff-reports/the-1929-stock-market-irving-fisher-was-right. Accessed 2026-10-17.
6. Elise Couper, John Hejkal, Alexander Wolman, "Boom and Bust in Telecommunications", Federal Reserve Bank of Richmond Economic Quarterly 89/4, Fall 2003. https://www.richmondfed.org/-/media/RichmondFedOrg/publications/research/economic_quarterly/2003/fall/pdf/wolman.pdf. Accessed 2026-10-17.
7. Mark Doms, "The Boom and Bust in Information Technology Investment", FRBSF Economic Review 2004. https://www.frbsf.org/wp-content/uploads/er19-34bk.pdf. Accessed 2026-10-17.
8. Lucent Technologies, Form 10-K for fiscal year ended 2000-09-30, filed 2000-12-27. https://www.sec.gov/Archives/edgar/data/1006240/000095012300011855/y43690e10-k405.txt. Accessed 2026-10-17.
9. Nortel Networks, Form 10-K405 for fiscal 2000, filed 2001-03-13. https://www.sec.gov/Archives/edgar/data/72911/000095014401003499/t26083ke10-k405.txt. Accessed 2026-10-17.
10. Cisco Systems, Form 10-K for fiscal year ended 2001-07-28, filed 2001-09-24. https://www.sec.gov/Archives/edgar/data/858877/000109581101505065/f75710e10-k.txt. Accessed 2026-10-17.
11. Light Reading, "Vendor Financing", 2000-12-06. https://www.lightreading.com/cable-technology/vendor-financing. Accessed 2026-10-17. (Secondary; used for the $7 bn Lucent and $2.5 bn Cisco commitment figures in press at the time, which the 10-Ks supersede.)
12. Congressional Research Service, "WorldCom: The Accounting Scandal", RS21253, 2002-08-29. https://www.everycrsreport.com/reports/RS21253.html. Accessed 2026-10-17.
13. FRED, NASDAQCOM, NIKKEI225, CSUSHPINSA, CBBTCUSD, BAA, AAA, BAA10Y, BAMLH0A0HYM2, BAMLC0A0CM. https://fred.stlouisfed.org/graph/fredgraph.csv?id=NASDAQCOM (and analogous URLs). Accessed 2026-10-17.
14. Robert Shiller, stock market data used in "Irrational Exuberance" (ie_data.xls, last through 2023-09). http://www.econ.yale.edu/~shiller/data/ie_data.xls. Accessed 2026-10-17.
15. Jay R. Ritter, "Initial Public Offerings: Updated Statistics", Table 1 updated 2026-07-13. https://site.warrington.ufl.edu/ritter/files/IPO-Statistics.pdf. Accessed 2026-10-17.
16. Kunio Okina, Masaaki Shirakawa, Shigenori Shiratsuka, "The Asset Price Bubble and Monetary Policy: Japan's Experience in the Late 1980s and the Lessons", Bank of Japan Monetary and Economic Studies 19(S-1), February 2001. https://www.imes.boj.or.jp/research/papers/english/me19-s1-14.pdf. Accessed 2026-10-17.
17. Robert Rich, "The Great Recession", Federal Reserve History, 2013-11-22. https://www.federalreservehistory.org/essays/great-recession-of-200709. Accessed 2026-10-17.
18. John V. Duca, "Subprime Mortgage Crisis", Federal Reserve History, 2013-11-22. https://www.federalreservehistory.org/essays/subprime-mortgage-crisis. Accessed 2026-10-17.
19. Goodwin, "SPAC 2021: The Year in Review", 2022-01-06. https://www.goodwinlaw.com/en/insights/publications/2022/01/01_06-spac-2021-year. Accessed 2026-10-17.
20. SPAC Insider, "Full-Year 2022 SPAC Review", 2023-01-02. https://old.spacinsider.com/2023/01/02/full-year-2022-spac-review/. Accessed 2026-10-17.
21. SEC v. Samuel Bankman-Fried, complaint, 2022-12-13. https://www.sec.gov/litigation/complaints/2022/comp-pr2022-219.pdf. Accessed 2026-10-17.
22. Microsoft, Form 10-K for fiscal year ended 2026-06-30, filed 2026-07-29. https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm. Accessed 2026-10-17.
23. Alphabet, Form 10-Q for quarter ended 2026-06-30, filed 2026-07-23. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm. Accessed 2026-10-17.
24. Amazon, Form 10-Q for quarter ended 2026-06-30, filed 2026-07-31. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm. Accessed 2026-10-17.
25. Meta Platforms, Form 10-Q for quarter ended 2026-06-30, filed 2026-07-30. https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm. Accessed 2026-10-17.
26. Nvidia, Form 10-Q for quarter ended 2026-07-26, filed 2026-08-26. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm. Accessed 2026-10-17.
27. AMD, Form 8-K, 2025-10-06. https://www.sec.gov/Archives/edgar/data/2488/000119312525230895/d28189d8k.htm. Accessed 2026-10-17.
28. Oracle, Form 10-K for fiscal year ended 2026-05-31, filed 2026-06-22. https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm. Accessed 2026-10-17.
29. CoreWeave, Form 10-Q for quarter ended 2026-06-30, filed 2026-08-12. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm. Accessed 2026-10-17.
30. SoftBank Group, "Announcing The Stargate Project", 2025-01-22. https://group.softbank/en/news/press/20250122. Accessed 2026-10-17.
31. Cisco Systems, FY2002 annual report, inventory-reserve reconciliation, fiscal year ended 2002-07-27. https://www.sec.gov/Archives/edgar/data/858877/000089161802004345/f84358exv13.htm. Accessed 2026-10-17.
32. Bank of Japan, official discount-rate history, historical observations 1987–1990. https://www.boj.or.jp/en/statistics/boj/other/discount/discount.htm. Accessed 2026-10-17.
33. Federal Reserve, FOMC statement, 2026-09-16. https://www.federalreserve.gov/newsevents/pressreleases/monetary20260916a.htm. Accessed 2026-10-17.
34. Federal Reserve, minutes of September 15–16, 2026 meeting, released 2026-10-07. https://www.federalreserve.gov/monetarypolicy/files/fomcminutes20260916.pdf. Accessed 2026-10-17.
35. Stijn Van Nieuwerburgh, "Financing the AI Buildout", Brookings Papers on Economic Activity conference draft, PDF dated 2026-09-04 for the September 24–25 conference, Table 1 and Appendices B/C. https://www.brookings.edu/wp-content/uploads/2026/09/4c_Van-Nieuwerburgh.pdf. Accessed 2026-10-17. Scenario estimates, not realised or committed spending.
36. Meta, "Meta Announces Joint Venture With Funds Managed by Blue Owl Capital to Develop Hyperion Data Center", 2025-10-21. https://about.fb.com/news/2025/10/meta-blue-owl-capital-develop-hyperion-data-center/. Accessed 2026-10-17.
37. Federal Reserve, Annual Report 2007, "Recent Economic and Financial Developments", published 2008. https://www.federalreserve.gov/boarddocs/rptcongress/annual07/sec1/c2.htm. Accessed 2026-10-17.
38. Bank of England, Staff Working Paper 691, "The Bank of England as lender of last resort: new historical evidence from daily transactional data", November 2017. https://www.bankofengland.co.uk/-/media/boe/files/working-paper/2017/the-bank-of-england-as-lender-of-last-resort-a-new-historical-evidence-from-daily-transactional-data.pdf. Accessed 2026-10-17.
39. FDIC Office of Inspector General, "Material Loss Review of Signature Bank of New York", EVAL-24-02, October 2023. https://fdicoig.gov/sites/default/files/reports/2023-12/EVAL-24-02.pdf. Accessed 2026-10-17.