Brief
The scale of the buildout is no longer contested even if the return timeline is. America's largest technology companies plan to spend $900 billion on AI infrastructure in 2025, rising to $1.4 trillion by 2027, according to The Economist's analysis of company disclosures. Tom's Hardware separately reports that Amazon, Google, Meta, and Microsoft alone have spent more than $1 trillion on AI infrastructure since 2023, with an additional $745 billion expected in 2026. The Bank for International Settlements' 2026 Annual Report frames this as one of three biggest risks to global financial stability, specifically citing circular financing arrangements between hyperscalers, neoclouds, and construction lessors.
The financing mix has shifted decisively from cash flow to capital markets. Oracle is the clearest case study: it raised $43 billion in debt and $5 billion in equity during fiscal 2026, including a record $25 billion bond issuance, and has since announced another $40 billion in planned financing for fiscal 2027 — bringing two-year financing to roughly $88 billion against FY2027 net capex of approximately $70 billion. Oracle's capex overshot its own guidance, hitting $55.7 billion in fiscal 2026 against a $50 billion target. Its remaining performance obligations reached $638 billion, up 363% year over year, with roughly half of that backlog tied to a single counterparty, OpenAI — itself financed by venture and private capital rather than operating profit. Fitch affirmed Oracle's BBB rating on its financing plan while S&P and Moody's maintained negative outlooks, and Oracle authorized a $20 billion at-the-market equity program specifically because, per one financing analysis, 'debt capacity alone cannot support the buildout.'
Neocloud providers show the same pattern at smaller scale but higher leverage. CoreWeave raised approximately $28 billion in combined equity and debt over the twelve months through March 2026, against FY2026 capex guidance of $30-35 billion; it closed Q1 2026 with roughly $22.7-25 billion in long-term debt and quarterly net interest expense of $536 million. Microsoft has accounted for a majority of CoreWeave's revenue in various disclosed periods (67-70% depending on the period cited), meaning a single hyperscaler customer effectively underwrites a heavily indebted infrastructure vendor whose backlog — $99.4 billion as of March 2026 — is increasingly concentrated in AI-lab counterparties like OpenAI and Anthropic that are themselves not yet cash-flow-positive.
The circularity extends to chip suppliers directly financing their own demand. Nvidia made a $2 billion private placement investment in CoreWeave in January 2026 and committed to a take-or-pay capacity backstop arrangement, while separate reporting indicates Nvidia is in talks for a backstop of up to $250 billion to help OpenAI raise debt against Nvidia's own credit standing — the same period in which Nvidia's credit default swap spread hit a record 82 basis points amid revived circular-financing concerns following roughly $750 billion in AI-related deal announcements. This is the mechanism the CIO.inc analysis calls an 'Ouroboros of financing schemes,' in which hyperscalers and chipmakers invest in AI labs that in turn commit to buy the compute those same investors are selling — allowing a single dollar to be booked as revenue at multiple nodes of the same loop.
On the other side of the ledger, enterprise returns remain unproven at scale. The BIS working paper cited by CIO.inc notes total AI infrastructure investment could reach several trillion dollars over coming years, while a CICC report finds annual AI-application revenue would need to reach $1 trillion to support the financing burden the industry is taking on. Reporting on enterprise AI adoption cites a figure of 56% of CEOs reporting zero measurable return on AI spending so far, and coverage of 'tokenomics' as an emerging measurement discipline reflects that companies still lack a standard way to price what they are getting for the spend. Meanwhile, exposure is migrating into pension and insurance balance sheets: New York and Pennsylvania public pension funds are invested in the same infrastructure fund financing Meta's Louisiana data center project and several Oracle facilities, and Wall Street has begun securitizing this debt into tradable instruments — distributing the risk of a stalled AI-revenue ramp into long-horizon fiduciary capital pools that have no direct visibility into AI product economics.
Money Flows (13)
Public bond and equity markets→Oracle
Investment$43B debt + $5B equity (FY2026), including a $25B bond issuance· AnnualHigh conf.
Funds Oracle Cloud Infrastructure data-center construction and GPU capacity ahead of contracted revenue recognition.
Basis: Global Data Center Hub analysis of Oracle FY2026 results, citing company disclosures (June 19, 2026)
Oracle→Data-center construction, power, and equipment vendors
Contract$55.7B (FY2026 capex, exceeding $50B guidance)· AnnualHigh conf.
Pays for data-center buildout across 71+ sites to serve contracted hyperscaler and AI-lab demand.
Basis: The Next Web reporting on Oracle FY2026 earnings (June 11, 2026)
OpenAI (financed by venture/private capital)→Oracle
Contract~$319B of Oracle's $638B RPO (roughly half)· RecurringMedium conf.
Contracted future compute purchases that Oracle books as backlog but must fund construction for upfront, before cash is collected.
Basis: Sunday Guardian Live financial-engineering analysis citing Oracle's disclosed remaining performance obligations (July 18, 2026)
Microsoft→CoreWeave
Revenue~67-70% of CoreWeave's revenue in disclosed periods· RecurringMedium conf.
Compensates CoreWeave for GPU cloud capacity Microsoft resells or uses internally for Azure AI workloads.
Basis: Sacra CoreWeave company profile (2026) and Kerrisdale Capital research report; ranges cited across sources
Bond and private-credit markets→CoreWeave
Loan~$28B combined equity and debt raised over 12 months through March 2026; ~$22.7-25B long-term debt outstanding· RecurringHigh conf.
Funds GPU purchases and data-center leases needed to meet a $99.4B revenue backlog before that backlog converts to cash.
Basis: Sacra CoreWeave profile and InsiderFinance Q1 2026 earnings coverage (May 7, 2026)
CoreWeave→Bondholders (senior notes, term loans)
Fee$536M net interest expense (Q1 2026)· RecurringHigh conf.
Services debt raised to fund GPU and data-center capex ahead of full contract monetization.
Basis: InsiderFinance and CNBC coverage of CoreWeave Q1 2026 earnings (May 2026)
Nvidia→CoreWeave
Investment$2.0B private placement (January 2026)· One-timeHigh conf.
Nvidia takes an equity stake in a major GPU-cloud customer as part of an expanded collaboration targeting over 5GW of AI-factory capacity by 2030.
Basis: Sacra CoreWeave profile citing the January 2026 private placement at $87.20/share
CoreWeave / Oracle / other neoclouds→Nvidia
Revenue$6.3B take-or-pay order form (CoreWeave-Nvidia specific)· RecurringHigh conf.
Locks in guaranteed GPU purchase volumes regardless of whether end-customer AI demand materializes at the contracted level.
Basis: Sacra CoreWeave profile citing the Nvidia take-or-pay capacity backstop arrangement through April 2032
Nvidia (reported)→OpenAI
OtherReportedly in talks for a backstop of up to $250B· UnknownLow conf.
Would let OpenAI raise debt against Nvidia's own credit standing rather than OpenAI's unproven cash flows.
Basis: Prior scan finding (July 30, 2026), described as reported talks, not a confirmed closed transaction
Public pension funds (New York, Pennsylvania)→Infrastructure funds financing Meta and Oracle data centers
InvestmentUndisclosed· One-timeMedium conf.
Provides long-horizon fiduciary capital that is being distributed into securitized AI data-center debt, per Wall Street bundling described in the same reporting.
Basis: Trellis Group reporting confirms the pension funds hold stakes in the same infrastructure fund but does not disclose a specific dollar allocation (July 27, 2026)
Alphabet→Chipmakers, data-center contractors, and power suppliers
Contract$195B-$205B (2026 capex guidance); $44.9B spent in Q2 2026 alone· AnnualHigh conf.
Funds Google Cloud and AI infrastructure buildout, one of the five hyperscalers driving the combined $1T+ AI capex figure.
Basis: Prior scan finding (July 30, 2026) grounded in Alphabet's own capex guidance disclosure
Amazon, Alphabet, Microsoft, Meta (aggregate)→Chip, construction, and power vendors (aggregate)
ContractCombined sequential quarterly capex comparable across the four in H1 2026, per Tomasz Tunguz's tracking (e.g., Amazon $53.1B, Alphabet $44.9B, Microsoft $41.0B, Meta $31.1B in a single quarter)· RecurringHigh conf.
Direct hyperscaler capex funds physical AI infrastructure construction and chip procurement independent of AI product revenue realized to date.
Basis: tomtunguz.com hyperscaler capex tracker (August 5, 2026)
Enterprises deploying AI tools→Hyperscalers / AI vendors (indirect, via subscriptions and API spend)
FeeUndisclosed aggregate; enterprise implementation and integration cited as the largest cost category at an estimated 41% of AI spend· RecurringLow conf.
Enterprises pay for AI tool subscriptions, integration, and data-pipeline work without yet generating offsetting AI-driven revenue at most firms.
Basis: aibusinessweekly.net AI spending statistics compilation (July 25, 2026); labeled as an industry estimate, not a filed figure
Who Profits (3)
Nvidia
How They Gain: Sells GPUs into take-or-pay contracts while simultaneously taking equity stakes in and extending credit backstops to its own customers, capturing both product revenue and investment upside from the same capital cycle it helps finance.
Reportedly in talks for a backstop of up to $250B for OpenAI and made a $2.0B direct equity investment in CoreWeave in January 2026, per sourced reporting.
Oracle bondholders and equity underwriters (e.g., Citigroup as lead-left bookrunner)
How They Gain: Earn fees and yield (the mandatory convertible priced at a 6.5% dividend) for underwriting Oracle's capex financing, paid regardless of whether Oracle's AI-driven cloud revenue ultimately covers the debt service.
Oracle's FY2026-27 financing plan totals roughly $88B in debt and equity issuance, per Global Data Center Hub's analysis of company disclosures.
Data-center construction, power, and equipment contractors
How They Gain: Are paid upfront in cash for construction and equipment as hyperscalers deploy capex immediately, ahead of any AI-revenue realization.
Capacity Media reporting frames contractor capacity as one of three binding constraints (alongside power and capital) on the pace of the buildout.
Who Pays (4)
Oracle shareholders and fixed-income holders
How They Pay: Bear dilution risk from the $20B at-the-market equity program and negative free cash flow, while credit-rating agencies S&P and Moody's maintain negative outlooks even as Fitch affirms the rating.
Oracle's remaining performance obligations reached $638B, up 363% year over year, with roughly half tied to a single counterparty (OpenAI), concentrating repayment risk.
CoreWeave equity and debt holders
How They Pay: Fund a business paying roughly a quarter of its top-line revenue in net interest expense, with revenue concentrated in a small number of hyperscaler and AI-lab customers.
CoreWeave's net interest expense reached $536M in Q1 2026 against long-term debt of roughly $22.7-25B.
Public pension beneficiaries (New York, Pennsylvania)
How They Pay: Are exposed, through fiduciary infrastructure-fund holdings, to AI data-center debt whose repayment depends on AI demand that has not yet been proven at the revenue scale needed to service it.
Exposure is described qualitatively in sourcing as distributed into 'long-horizon, fiduciary pools of capital'; no specific dollar allocation is disclosed.
Enterprises deploying production AI without offsetting revenue
How They Pay: Continue paying subscription, integration, and infrastructure costs while a reported majority of CEOs see no measurable return, per industry surveys cited in trade coverage.
A cited estimate of 56% of CEOs reporting zero measurable AI return; this figure originates from industry/trade reporting, not a regulatory filing, and should be treated as directional rather than precise.
Where the Trail Goes Dark
⚠The exact size and terms of Nvidia's reported backstop for OpenAI debt (up to $250B) have not been confirmed in a filing; the reporting describes talks, not a closed transaction.
If this backstop closes at scale, it would materially increase Nvidia's contingent liability exposure to a single customer's demand shortfall, but the opacity around terms means the true risk transfer cannot yet be assessed.
⚠CoreWeave's exact customer concentration percentages (Microsoft's share of revenue) vary across sources from 67% to 70% depending on the reporting period and source, and the company itself has not disclosed exact percentages in some periods.
Without a single disclosed figure, outside analysts cannot precisely quantify how exposed CoreWeave's bondholders are to a single counterparty's AI spending decisions.
⚠The specific dollar allocation of New York and Pennsylvania pension funds to the infrastructure fund financing Meta's and Oracle's data centers is not disclosed in available reporting.
This obscures how much retirement-system capital is actually at risk if AI data-center assets underperform, making it impossible for beneficiaries or outside observers to size the exposure.
⚠The degree to which vendor-to-customer circular financing (chipmakers investing in cloud providers who then buy chips from the same chipmaker) inflates reported revenue at each node is acknowledged qualitatively by multiple sources but not quantified by any single disclosure.
Without a standardized accounting treatment or disclosure requirement, the true organic revenue growth in the AI infrastructure chain cannot be separated from revenue generated by circular investment arrangements.