AI debt: how borrowing by AI companies and data-centre builders could spill into global financial markets

Abstract

1. Introduction

Until 2025 the companies building AI infrastructure paid for it from the profits of their existing businesses. That has changed. From late 2025 the same companies began raising money in the bond markets at a scale the technology sector had never used, and a second tier of borrowers, specialist graphics processing unit (GPU) cloud providers, data-centre developers and the AI laboratories themselves, turned to private credit, leases, vendor financing and special-purpose vehicles.

Does that matter to anyone outside Silicon Valley? A UK accountancy practice with forty staff does not own Oracle bonds. But its pension scheme may; its bank's funding costs are set in the same markets; and its email and files almost certainly sit on infrastructure that two of the borrowers supply. The Bank of England's Financial Policy Committee (FPC) said in July 2026 that a reassessment of AI earnings "could result in spillovers to financial markets that are relevant to UK financial stability and potentially lead to tighter financial conditions for UK borrowers" (Bank of England, 2026b). That sentence is the reason for this paper.

The contribution is to put the official assessments and the primary financing documents side by side, in plain English, and to trace the channels to UK small and medium-sized enterprises (SMEs). Section 2 reviews the literature; Section 3 the method; Section 4 the findings; Section 5 the parallels with the telecoms crash of 2000 to 2002 and with 2008, the counter-argument, and the implications for UK firms; Sections 6 and 7 the limitations and conclusions.

2. Literature review

2.1 From cash flow to debt

At the Bank for International Settlements (BIS), Aldasoro, Doerr and Rees (2026) observe that the large technology firms had relied on "their highly profitable operations to generate the cash flows needed to fund investments", but that "free cash flows have recently lagged capital expenditures in absolute amounts" and the firms are now "increasingly financing AI investment via debt". They put IT manufacturing and data-centre spending at about 1% of United States gross domestic product (GDP) by mid-2025, and note a pricing gap: loans to AI borrowers carry spreads and maturities almost identical to any other sector's while equity markets price the same firms for exceptional growth, so either "lenders may be underestimating the risks" or "equity markets may be overestimating the future cash flows".

The Bank of England (2026a) quantifies the change in the bond market. The five AI hyperscalers, Meta, Alphabet, Amazon, Microsoft and Oracle, "accounted for just 3% of the stock of outstanding US IG debt at the end of 2025, but as of early May accounted over 15% of year-to-date issuance", where IG means investment grade; across currencies that issuance was "broadly comparable in scale to UK gilt issuance over the same period". In the riskier high-yield (HY) market, AI issuers took 41% of non-refinancing issuance in 2026 despite being 1% of the index. S&P Global Ratings (2026) adds that technology-sector bonds reached "about 16.7% of global non-financial corporate bond issuance" in 2025, up from 11.6%, and that data-centre securitisation "topped $30 billion in 2025, nearly tripling".

2.2 Private credit and the second tier

Private credit, lending by funds rather than banks or bond markets, is where the newer borrowers go. The Financial Stability Board (FSB, 2026) sizes the global market at $1.5 to $2 trillion, cites a projection that AI infrastructure capital expenditure of $2.9 trillion between 2025 and 2028 will need $1.5 trillion of external capital "including $800 billion from private credit", and reports that AI's share of private credit deals "reached 34% in 2025, up from a 17% average over the previous five years". The Organisation for Economic Co-operation and Development (OECD, 2026) reports the same jump from a different base, 9% in 2024 to 34% in 2025, and Aldasoro, Doerr and Rees (2026) find AI-related private loans grew "from near zero to over $200 billion".

The FSB (2026) is candid about the weak points: borrowers cluster around single-B ratings with leverage of five to six times earnings before interest, tax, depreciation and amortisation (EBITDA), "true leverage could be closer to 7x", private equity groups increasingly own insurers that buy their funds, and loan-level data is limited. The Federal Reserve Board (2026) records that software had become "the largest sector in private credit portfolios" and that 50% of its market contacts named AI as a salient risk in spring 2026, up from none six months earlier. The International Monetary Fund (IMF, 2026) adds that leveraged private credit vehicles' exposure to software borrowers "may exceed 50 percent of NAV" (net asset value).

2.3 Opacity and circularity

The Bank of England (2026a) describes "securitised data centre and other asset-backed structures, special purpose vehicles, and other bespoke financing arrangements" that increase "the complexity of identifying where risk ultimately sits", and flags "circular financing arrangements" in which technology companies invest in AI companies "which in turn purchase those technology companies' products". Rungcharoenkitkul (2026) models it: hyperscalers take equity in AI laboratories in exchange for compute-purchase commitments, so that "a single lab failure can propagate across the network, affecting even solvent partners through common backers". Aldasoro, Doerr and Rees (2026) put it in six words: "leverage does not disappear by being out of sight".

2.4 Valuations and concentration

The debt sits on top of equity valuations regulators describe as stretched. The Bank of England (2026a) reports that AI companies "now comprise approximately 50% of index value" of the S&P 500, up from about 25% in 2022, and that consensus forecasts of 2028 hyperscaler capital spending rose from under $600 billion in December 2025 to over $1 trillion by mid-2026. The IMF (2026) notes that S&P 500 and Nasdaq earnings would need to grow at "close to 30 percent and 35 percent" a year to justify prices. The European Central Bank (ECB, 2026) finds euro-area portfolios concentrated in "a few large US issuers, especially AI-related firms".

2.5 Historical parallels

On telecoms, testimony to the US House Committee on Financial Services (2002) records that from 1998 the sector "raised over $1 trillion from the capital markets in equity and debt", excluding bank loans and vendor financing; that its market value peaked at $2.7 trillion in March 2000 and lost $1.7 trillion within a year; and that "about 100 banks" were caught in the Global Crossing bankruptcy alone. Starr (2002) quotes the then chairman of the Federal Communications Commission that the industry owed a trillion dollars "much of which will never be repaid", and counts 23 bankruptcies and half a million job losses. On 2008, the Financial Crisis Inquiry Commission (FCIC, 2011) concluded that the crisis "was avoidable", driven by leverage of up to 30 to 1, an "opaque" shadow banking system, off-balance-sheet entities hidden from regulators, and rating agencies that were "key enablers". Rungcharoenkitkul (2026) argues the AI boom has "surpassed every previous comparable episode" and that firms racing for dominance over-invest by roughly 1.4 to 1.5 times the efficient level: "the larger the boom, the deeper the eventual bust".

The literature agrees on the facts and differs on the weight: the BIS and IMF call the current impact modest; the FPC and FSB stress the speed of change and the loss of visibility.

3. Method

This is desk research on secondary data; no interviews or proprietary data were used. Sources were chosen in a fixed order of preference: regulator and central bank reports; primary company documents (Securities and Exchange Commission filings and investor-relations releases); a rating agency's own release; and trade press only where the primary document could not be retrieved (the Alphabet sterling tranche and the Moody's report).

The first ratio is the external funding gap:

Here G is the cash a company must find from outside in period t, K is capital expenditure in that period, C is net cash from operating activities, and s is the share of capital expenditure not covered by operations. A negative G means operations more than covered investment. The measure ignores dividends and buy-backs, which raise the true need, and cash reserves, which reduce it; it is a flow measure, not a solvency test.

The second is an interest coverage ratio with a refinancing stress:

Here R is the coverage ratio, E is quarterly EBITDA, I is quarterly net interest expense, D is total debt outstanding, Δr is an assumed rise in the average annual interest rate on that debt when refinanced, and R with the subscript "stress" is the ratio after that rise; dividing by four converts the annual increase to a quarter. Historical comparisons are qualitative; 2002 dollar figures are not inflation-adjusted because the comparison is about mechanism, not magnitude.

4. Findings

4.1 Capital expenditure has overtaken operating cash flow at three of the five hyperscalers

Figure 1 compares, for the quarter to 30 June 2026, cash generated from operations with cash spent on property and equipment.

Applying Equation 1: Alphabet spent $44.9 billion against $39.1 billion of operating cash, a gap of $5.8 billion, so 13% of capital expenditure was externally funded; its free cash flow was negative $5.9 billion (Alphabet, 2026b). Amazon spent $54.2 billion against $45.4 billion, a gap of $8.8 billion or 16%; its trailing-twelve-month free cash flow was an outflow of $7.6 billion (Amazon, 2026b). Meta's $31.1 billion, including finance-lease principal, was just covered by $31.9 billion of operating cash, leaving free cash flow of $784 million against $8.5 billion a year earlier; its long-term debt rose from $58.7 billion at the end of 2025 to $83.7 billion (Meta, 2026). Microsoft is the exception: $55.4 billion of operating cash against $35.8 billion of additions to property and equipment (Microsoft, 2026).

Oracle reports on a May year-end. For fiscal 2026 it generated $32.0 billion of operating cash flow but free cash flow of negative $23.7 billion, implying capital expenditure of about $55.7 billion, 43% of it externally funded. It raised $43 billion of debt and $5 billion of equity in the year and expects about $40 billion more in fiscal 2027 (Oracle, 2026b).

The Bank of England (2026a) notes that "the declining free cash flows of the AI hyperscalers increases their dependence on accommodative future refinancing conditions". None is short of money: Alphabet held $242.5 billion of cash and marketable securities and Meta $90.3 billion (Alphabet, 2026b; Meta, 2026). But the marginal pound is now borrowed or raised, and Moody's expects hyperscaler capital expenditure of $700 billion in 2026 and $870 billion in 2027, warning of a possible "reassessment of creditworthiness" if profit growth disappoints (Data Center Dynamics, 2026).

4.2 The largest verified financings

Table 1 lists transactions for which a primary document was retrieved; widely reported financings by private AI laboratories could not be verified and are excluded.

Three features stand out. Maturity: Alphabet and Amazon both sold 40-year paper and Alphabet a century bond, against assets, GPUs especially, whose useful life is measured in single-digit years. Cost: CoreWeave, which resells GPU capacity, pays 9.625% while Alphabet pays 3.7% to 5.75%; the two ends of the supply chain are financed at very different prices although they depend on each other. Form: Hyperion is not a bond but an 80%-owned vehicle that Meta leases from, with a guarantee that crystallises only if leases lapse (Meta, 2025), exactly the "bespoke financing arrangement" the Bank of England (2026a) says obscures where risk sits.

Equity is also being raised at scale: Alphabet issued $49.6 billion of common and mandatory convertible preferred stock in June 2026 "to scale AI infrastructure and global compute" (Alphabet, 2026b), and Oracle's plan is half equity by design (Oracle, 2026a).

4.3 Vendor and circular financing

The clearest documented example is NVIDIA's September 2025 letter of intent to "invest up to $100 billion in OpenAI progressively as each gigawatt is deployed" of at least 10 gigawatts of NVIDIA systems (NVIDIA, 2025): the chip supplier funds the customer, which buys the chips. Oracle lists OpenAI and NVIDIA among the customers whose "contracted demand" justifies its raise (Oracle, 2026a), and its remaining performance obligations reached $638 billion at 31 May 2026, up 363% in a year (Oracle, 2026b). A backlog is only as good as the customer's ability to pay, and several of those customers are financed by their suppliers. This is the network Rungcharoenkitkul (2026) models.

4.4 The second tier under stress: a worked coverage ratio

CoreWeave publishes enough to apply Equation 2. In the quarter to 30 June 2026 it reported adjusted EBITDA of $1,510 million and net interest expense of $640 million, a coverage ratio of 2.4. It carried $31.4 billion of recourse and $3.7 billion of non-recourse debt, about $35.1 billion, against $5.5 billion of cash; the quarter's net loss was $626 million (CoreWeave, 2026c). Refinanced at an average rate two percentage points higher, quarterly interest would rise by roughly $175 million to about $815 million and the stressed ratio would fall to 1.9; at four points higher, 1.5. Coverage of around two on a measure outside generally accepted accounting principles (non-GAAP), before depreciation of hardware that ages quickly, is thin. Against that, its revenue backlog was about $104 billion. The whole AI debt question is in that pair of numbers: enormous contracted demand, financed at high cost, resting on customers who are themselves raising money.

4.5 Transmission channels

Figure 2 draws the routes from the borrowers to a UK business, assembled from the FPC's description of the channels (Bank of England, 2026b), the FSB (2026) on private credit interconnections, the ECB (2026) on cross-border spillovers and the Competition and Markets Authority (CMA, 2025) on UK cloud supply.

The middle columns show who holds the paper. The FSB (2026) records about $220 billion of bank credit lines to private credit funds and notes commercial estimates "more than twice as large". Open-ended funds, which the ECB (2026) reports have already faced "sizeable redemption requests" where heavily exposed to software and AI, are the fast channel.

The right-hand column is what arrives in the UK. The FPC's chain runs from an AI earnings disappointment to "a broader negative repricing in sovereign debt markets", to doubts about "the sustainability of current debt trajectories", to "tighter financial conditions for UK borrowers" (Bank of England, 2026b). The second route is wealth: UK funded occupational pension schemes held £1,120 billion in private-sector defined-benefit (DB) and hybrid schemes and £945 billion in defined-contribution (DC) and public-sector schemes at 30 September 2025, with DC growth "driven primarily by mixed assets and equities" (Office for National Statistics, 2026), while AI companies are now about half of the S&P 500 by value (Bank of England, 2026a). The third is operational: the CMA (2025) found Microsoft and Amazon Web Services each held 30% to 40% of UK infrastructure-as-a-service supply in 2024, in a £10.5 billion market, and that "competition is not working well". Two of the five borrowers are, for most UK firms, the cloud.

5. Discussion

5.1 What the evidence does and does not show

The evidence supports three claims. AI investment at the largest firms is no longer self-funding at the margin. The financing is spreading from the most transparent market, investment-grade bonds, to the least transparent, private credit, leases, joint ventures and vendor equity, and regulators say openly they cannot fully see it. And the exposures are concentrated: five borrowers, two of whom supply most of the UK's cloud, and a private credit industry that took a third of its 2025 business from one theme. The evidence does not show that a crash is imminent: the IMF (2026) says the impact "appears modest currently", and Aldasoro, Doerr and Rees (2026) size the boom at about 1% of GDP, "half as large as the rise in IT investment during the dot-com boom".

5.2 The telecoms parallel, and why it is imperfect

The telecoms crash rhymes in mechanism. A technology with real long-run value attracted a race to build capacity ahead of demand; capital markets supplied "over $1 trillion" in a few years; vendor financing hid the leverage; and writing off debt that "will never be repaid" took a hundred banks into one bankruptcy court (US House Committee on Financial Services, 2002; Starr, 2002). The fibre was eventually used, as the data centres will be, but not by the firms that borrowed to build it.

The differences are substantial. The telecoms borrowers were mostly new companies without profits; the hyperscalers are among the most profitable enterprises ever, with Azure revenue above $100 billion a year (Microsoft, 2026) and Amazon Web Services growing 36.7% (Amazon, 2026b). Demand is contracted rather than hoped for: Oracle's $638 billion of performance obligations and CoreWeave's $104 billion backlog have no 1999 equivalent. And the depreciating asset is a GPU, not a fibre strand, which shortens the time available to earn a return.

5.3 The 2008 parallel, and why it is the more useful one

What matters for global markets is less the overbuild than the opacity. The FCIC (2011) found that 2008 was made by leverage regulators could not see, held outside the regulated perimeter, rated by agencies paid by the issuer, and funded short. Read against that, the FSB's list of what it does not know about private credit (no common definitions, no loan-level data, bank exposures possibly twice the recorded figure, insurers owned by the fund managers) is uncomfortable. It is not 2008: the hyperscaler bonds are plain, senior and public. But the marginal dollar is increasingly raised in the part of the system the FSB describes, and the BIS warning that "leverage does not disappear by being out of sight" is aimed there.

5.4 The strongest counter-argument

The best case against this paper's concern is this. The borrowers hold hundreds of billions in cash; their bonds are investment grade; they are issuing equity alongside debt; their AI revenues are growing at 30% to 40% a year; their customers have signed contracts worth hundreds of billions; and the boom is, on the BIS's measure, half the size of the dot-com surge. Private credit lending to AI, at perhaps $200 billion, is small against the $1.4 trillion of US private credit recorded by the Federal Reserve Board (2026). On this view the debt is the rational way to fund an asset with contracted revenue, like a utility building a power station.

The reply is that the argument is right about the centre and silent about the edges. The centre, Microsoft above all, is fine. The risk sits where coverage is 2.4 and the coupon 9.625%, where the customer is funded by the supplier, and where the lender's exposure is not reported. That is where a shock would start, and Rungcharoenkitkul (2026) shows that in a network of circular holdings a failure there reaches "even solvent partners". The counter-argument shows the system is unlikely to break, not that prices and credit conditions would stay still while it bent.

5.5 Implications for UK small and medium businesses

For an SME the exposures are indirect and, in order of importance:

Cost and availability of credit. The FPC's transmission runs through gilt yields and bank funding to "tighter financial conditions for UK borrowers" (Bank of England, 2026b). A firm expecting to refinance within two years should not assume today's conditions persist, and should fix what it can.

Pensions. Global index funds track a US market in which AI companies are about half the S&P 500 by value (Bank of England, 2026a); an owner-manager should look at what the default fund actually holds.

Cloud pricing and continuity. Two of the five borrowers supply 60% to 80% of UK infrastructure-as-a-service, in a market where the CMA (2025) has already found switching barriers, and an indebted supplier under stress has every incentive to raise prices to captive customers. Contracts running more than three years without price caps, and architectures that cannot be moved, are worth reviewing now; our cloud services page describes the portable, backed-up approach we recommend.

Supplier risk in the second tier. Any SME buying AI services from a start-up should ask who funds it, at what cost, and for how long.

None of this argues against using AI tools or cloud services; it argues for the ordinary disciplines of contracting, backup and diversification, applied to a risk now large enough to appear in the Bank of England's report.

6. Limitations

This is a desk study of published documents; it cannot see the private loan books the FSB says are unreported, so private credit figures are lower bounds. Figure 1 mixes capital-expenditure definitions (Meta includes finance-lease principal) and year-ends (hence Oracle's exclusion from the chart). Equation 1 ignores dividends, buy-backs and cash reserves; Equation 2 uses a non-GAAP earnings measure and an instantaneous refinancing assumption that overstates the first-year effect. Two data points rest on secondary reporting: the size of Alphabet's sterling century bond and Moody's capital-expenditure projections. Financings by private AI laboratories are excluded for lack of a primary source, so Table 1 understates the second tier. The latest official assessment used is dated July 2026.

7. Conclusion

The AI build-out has crossed from being paid for out of profits to being paid for, at the margin, with borrowed and raised money: capital expenditure exceeded operating cash flow at Alphabet, Amazon and Oracle in their latest reported periods, hyperscaler bond issuance went from 3% of the investment-grade stock to over 15% of new supply, and private credit's share of AI financing rose from 9% to 34% in a year. Regulators do not call this a crisis. They call it a fast-growing, concentrated and increasingly opaque set of exposures, linked by circular financing, and they have named the channels by which a disappointment would reach the UK: sovereign yields, bank funding, pension values and cloud supply. The AI case differs from the telecoms crash and from 2008 in the strength of the borrowers at the centre, and resembles both at the edges. For a UK business the sensible response is not to predict the outcome but to reduce dependence on any single outcome: fixed-price, portable cloud arrangements; credit secured before it is needed; and a pension default fund that someone has actually looked at.

References

• Aldasoro, I., Doerr, S. and Rees, D. (2026) Financing the AI boom: from cash flows to debt. BIS Bulletin No 120, 7 January. Basel: Bank for International Settlements. https://www.bis.org/publ/bisbull120.pdf (accessed 25 August 2026).
• Alphabet Inc. (2026a) Prospectus supplement, Form 424B2, 9 February 2026. Washington DC: US Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1652044/000119312526046846/d108033d424b2.htm (accessed 25 August 2026).
• Alphabet Inc. (2026b) Alphabet announces second quarter 2026 results, 22 July. Exhibit 99.1 to Form 8-K. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000066/googexhibit991q22026.htm (accessed 25 August 2026).
• Amazon.com, Inc. (2026a) Prospectus supplement, Form 424B5, 7 July 2026. Washington DC: US Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1018724/000110465926081786/tm2619352-2_424b5.htm (accessed 25 August 2026).
• Amazon.com, Inc. (2026b) Amazon.com announces second quarter results, 30 July. Exhibit 99.1 to Form 8-K. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000024/amzn-20260630xex991.htm (accessed 25 August 2026).
• Bank of England (2026a) Financial Stability Report – July 2026. London: Bank of England. https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026 (accessed 25 August 2026).
• Bank of England (2026b) Financial Policy Committee Record – July 2026. London: Bank of England. https://www.bankofengland.co.uk/financial-policy-committee-record/2026/july-2026 (accessed 25 August 2026).
• Competition and Markets Authority (2025) Cloud services market investigation: summary of final decision, 31 July. London: CMA. https://assets.publishing.service.gov.uk/media/688b20e6ff8c05468cb7b120/summary_of_final_decision.pdf (accessed 25 August 2026).
• CoreWeave, Inc. (2026a) CoreWeave prices upsized $3.5 billion convertible senior notes offering, 10 April. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Prices-Upsized-3-5-Billion-Convertible-Senior-Notes-Offering/default.aspx (accessed 25 August 2026).
• CoreWeave, Inc. (2026b) CoreWeave announces pricing of $1.25 billion of senior notes and €2 billion of senior notes, 11 June. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Announces-Pricing-of-1-25-Billion-of-Senior-Notes-and-2-Billion-of-Senior-Notes/default.aspx (accessed 25 August 2026).
• CoreWeave, Inc. (2026c) CoreWeave reports strong second quarter 2026 results, 11 August. Exhibit 99.1 to Form 8-K. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm (accessed 25 August 2026).
• Data Center Dynamics (2026) US hyperscaler capex to top $700bn in 2026, investors fear overbuild and weak returns – Moody's. https://www.datacenterdynamics.com/en/news/us-hyperscaler-capex-to-top-700bn-in-2026-investors-fear-overbuild-and-weak-returns-moodys/ (accessed 25 August 2026).
• European Central Bank (2026) Financial Stability Review, May 2026. Frankfurt: ECB. https://www.ecb.europa.eu/press/financial-stability-publications/fsr/html/ecb.fsr202605~50566915a7.en.html (accessed 25 August 2026).
• Federal Reserve Board (2026) Financial Stability Report, May 2026. Washington DC: Board of Governors of the Federal Reserve System. https://www.federalreserve.gov/publications/files/financial-stability-report-20260508.pdf (accessed 25 August 2026).
• Financial Crisis Inquiry Commission (2011) The Financial Crisis Inquiry Report. Washington DC: US Government Printing Office. https://www.govinfo.gov/content/pkg/GPO-FCIC/pdf/GPO-FCIC.pdf (accessed 25 August 2026).
• Financial Stability Board (2026) Report on vulnerabilities in private credit, 6 May. Basel: FSB. https://www.fsb.org/uploads/P060526.pdf (accessed 25 August 2026).
• Global Finance (2026) Alphabet taps debt markets with 100-year issuance. https://gfmag.com/capital-raising-corporate-finance/alphabet-taps-debt-markets-with-100-year-issuance/ (accessed 25 August 2026).
• International Monetary Fund (2026) Global Financial Stability Report, April 2026: Chapter 1. Washington DC: IMF. https://www.imf.org/-/media/files/publications/gfsr/2026/april/english/ch1.pdf (accessed 25 August 2026).
• Meta Platforms, Inc. (2025) Meta announces joint venture with funds managed by Blue Owl Capital to develop Hyperion data center, 21 October. https://investor.atmeta.com/investor-news/press-release-details/2025/Meta-Announces-Joint-Venture-with-Funds-Managed-by-Blue-Owl-Capital-to-Develop-Hyperion-Data-Center/default.aspx (accessed 25 August 2026).
• Meta Platforms, Inc. (2026) Meta reports second quarter 2026 results, 29 July. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/default.aspx (accessed 25 August 2026).
• Microsoft Corporation (2026) Earnings release FY26 Q4, 29 July. https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast (accessed 25 August 2026).
• NVIDIA Corporation (2025) OpenAI and NVIDIA announce strategic partnership to deploy 10 gigawatts of NVIDIA systems, 22 September. https://investor.nvidia.com/news/press-release-details/2025/OpenAI-and-NVIDIA-Announce-Strategic-Partnership-to-Deploy-10-Gigawatts-of-NVIDIA-Systems/default.aspx (accessed 25 August 2026).
• Office for National Statistics (2026) Funded occupational pension schemes in the UK: April to September 2025, 2 April. https://www.ons.gov.uk/economy/investmentspensionsandtrusts/bulletins/fundedoccupationalpensionschemesintheuk/apriltoseptember2025 (accessed 25 August 2026).
• Organisation for Economic Co-operation and Development (2026) AI firms capture 61% of global venture capital in 2025, February. Paris: OECD. https://www.oecd.org/en/about/news/announcements/2026/02/ai-firms-capture-61-percent-of-global-venture-capital-in-2025.html (accessed 25 August 2026).
• Oracle Corporation (2026a) Oracle announces equity and debt financing plan for calendar year 2026, 1 February. https://investor.oracle.com/investor-news/news-details/2026/Oracle-announces-Equity-and-Debt-Financing-Plan-for-Calendar-Year-2026/default.aspx (accessed 25 August 2026).
• Oracle Corporation (2026b) Oracle announces record Q4 and FY 2026 results driven by cloud infrastructure and cloud applications, 10 June. https://www.oracle.com/news/announcement/q4fy26-earnings-release-2026-06-10/ (accessed 25 August 2026).
• Rungcharoenkitkul, P. (2026) The AI investment race. BIS Working Papers No 1367, July. Basel: Bank for International Settlements. https://www.bis.org/publ/work1367.pdf (accessed 25 August 2026).
• S&P Global Ratings (2026) Private credit, tech issuance fuelled by AI, and increasing leverage among key driving factors impacting credit market liquidity in 2026, 17 February. https://press.spglobal.com/2026-02-17-Private-Credit,-Tech-Issuance-fuelled-by-AI,-and-Increasing-Leverage-Among-Key-Driving-Factors-Impacting-Credit-Market-Liquidity-in-2026-according-to-S-P-Global-Ratings (accessed 25 August 2026).
• Starr, P. (2002) 'The great telecom implosion', The American Prospect, 8 September. https://www.princeton.edu/~starr/articles/articles02/Starr-TelecomImplosion-9-02.htm (accessed 25 August 2026).
• US House Committee on Financial Services (2002) Testimony on market conditions of the US telecom and energy sectors, March. Washington DC. https://financialservices.house.gov/media/pdf/032102wm.pdf (accessed 25 August 2026).