top of page
Glossy 3D 48PEGASUS Logo.png

The AI Capital Loop

Debt, valuations and systemic transmission

The evidence supports interdependence, not a closed economy

RESEARCH NOTE Structural finding | Evidence cutoff: November 19, 2026

A financing edge is not a sale

The strongest documented claim is that some leading technology firms are simultaneously financiers, suppliers, customers and guarantors of counterparties in the AI supply chain. A chip designer can lend or invest in a compute buyer, a hyperscaler can sign a long lease that supports a developer’s debt, and a lab can promise minimum purchases to a shareholder. Financing increases the buyer’s ability to spend; a commitment improves the seller’s planning and may help both procure scarce capacity. None of those facts proves its ultimate users could not support the same demand. [47] [49]

Real assets can be financed in ways that increase risk

A data centre can exist, perform useful work, and generate generally accepted accounting principles revenue while spare capacity, margins or cash collection disappoint. Treat commercial substance, accounting recognition, funding dependence, cash generation, and network risk as separate questions. Evidence of one does not answer the others. In particular, an invoice paid with money from a financing round is still an invoice; the durability of that payment source is the separate research problem. [56] [48]

Where to place the boundary

For this paper, external demand means payments by ultimate users and other customers outside the specified financing network, after excluding directly attributable promotional support when observable. A lab could sell through a cloud marketplace and still reach independent users; calling marketplace revenue internal purely because the cloud firm is an investor would be wrong. A customer outside the named group might itself be financed indirectly by members. Results must identify the network boundary and whether payer funding source is known.

What the public record cannot supply

No public census traces every equity or debt dollar into cloud use or separates network-funded usage from sustainable independent cash collections. BIS relationship matches, IMF disclosed deals, public debt data and national-investment statistics answer different questions. Their overlap prevents assembling one industrywide “loop percentage.” Estimate a firm’s funding dependence only against a declared period, population, weighting method and evidence standard. [47] [48]

BIS counts relationships across the value chain

RESEARCH NOTE Evidence map | Five layers and a five-year matching window

Scope and definition

In the October 2026 Bulletin study, [*REDACTED*] and coauthors start from 1,246 classified AI firms across compute, infrastructure, data tools, models and applications. They pair [*REDACTED*] equity/debt investment or financing data with [*REDACTED*] commercial relationships and manual checks. A relation is circular if two AI firms linked by an investment during 2021–25 also have a supplier-customer relationship at any point during that interval. The authors identify 972 intra-AI investment relationships. [47]

Three directions, one family of arrangements

Supplier financing sends capital and products or services in the same direction initially; the commercial payment may later travel the other way. Customer financing sends capital toward a supplier while the product or service moves back. Reciprocal commercial exchange is another case. Among the relationships classified as circular in the study, the investor also supplies the investee in 64%. The direction finding is not a share of money or a share of all AI investment. [47]

What “circular” means here

A supplier link need not start on the financing date or be written in the same agreement. The method captures an overlapping relationship, not contract tying, negative economic value, or any formal finding of a round trip. In an equity round with several participants one portion could fund product development unrelated to the investor’s sales. Pairwise observation alone cannot trace the money. [47]

Why the sample matters

Coverage of private firms, commercial relationships and transaction values is incomplete. The authors note that database deal values need not match [*REDACTED*] reporting methodology. A stronger test would match counterparties to dated agreements, track actual funding and purchases across those dates, and compare outcomes for firms that did not have that financing support. Public data do not provide the full set required for that causal test. [47]

Four BIS statistics need four labels

RESEARCH NOTE Quantitative evidence | Value and count are not interchangeable

Incoming versus outgoing

By disclosed deal value between 2021 and 2025, 55.2% of investment into classified AI firms came from deals involving other AI firms; 28.7% of investment by classified AI firms went toward deals involving other AI firms. The two directional quantities divide the same broad intra-AI concept by differently composed populations: all observed investments going into AI targets versus all observed investments made by AI investors. An incoming percentage may be materially higher even if the matched numerator is conceptually shared. [47]

Relationship overlap among AI-only deals

Among AI-to-AI deals, 16.1% by deal count and 46.4% by disclosed deal value were associated with a commercial connection between investor and target. The higher value share is consistent with large deals carrying more weight, not proof that every dollar in a large round is tied to goods or services. Graph 3 counts the full disclosed round when an eligible pair qualifies, and counts a round once if several AI investors participate. Individual ticket values are not observable in the chart. [47]

Use the bars in two panels

A reproducible chart should show the 55.2% and 28.7% in a directional-value panel and the 46.4% and 16.1% in a matched-deal panel, each labeled 2021–25 with no cash-flow arrows scaled by those percentages. For the matched-deal panel, display one denominator scope with two weighting methods, rather than a time trend. Keep 2026 commitments separate from the 2021–25 population.

What not to infer

Multiplying 55.2% by 46.4% cannot identify the share of funded purchases; it splices different units and denominators. Nor do these figures show that 55.2% came directly from suppliers, 28.7% headed back to original investors, or 46.4% of later revenue is unsupported by independent users. Those conclusions would require contract data and tracing beyond the BIS study. [47]

Exhibit 06 · AI investment shares use different denominators

The 55.2% is the share of incoming value to AI firms from other AI firms. The 28.7% is the share of AI firms’ outgoing investment value going to AI firms. These are not portions of the same whole, estimates of recycled cash or sector revenue.

BIS Bulletin 137, October 2026: primary source

Exhibit 07 · Commercial links are larger by value than by count

BIS counts the entire round when at least one AI investor has a supply-chain relationship with the target in the sample window. Investor-level contributions are not separately identified. These findings describe a screened sample, not the global percentage of funds round-tripped.

BIS Bulletin 137, October 2026: primary source

IMF tracks a disclosed transaction network

RESEARCH NOTE Selected deals | Fourteen transactions since 2023

Three magnitudes, not one total

Figure 1.2 in the IMF’s October 2026 financial stability report summarizes fourteen publicly disclosed transactions since 2023. Staff estimate about $300 billion of aggregate equity commitments, about $650 billion in debt, and $1.4 trillion in future revenue commitments through orders or services. The technical note describes debt financing following nine of the fourteen deals and includes financing backed by investor guarantees, contingent financing and other off-balance-sheet debt. Do not add the bars: a debt flow can help finance a purchase commitment already in the revenue category. [48] [49]

A purposive sample, not a denominator

Researchers selected major U.S. deals with public disclosures and coverage. Figure 5 in the technical note numbers the firms; roles overlap, with nodes 1–10 as investors and 9–14 as investees. This is not fourteen companies accounting for all AI finance. An industry percentage cannot be back-solved from a sample built to illustrate significant interdependencies. The total also does not provide a precise tally of capital already transferred or funding still uncommitted. [49]

Concentration inside the sample

Industry-scale firms and a handful of weaker borrowers recur across major arrangements. IMF’s note estimates the publicly listed participants represent nearly half of revenues and about 40% of market cap in its defined AI stack. This is a defined set, not a share of the S&P 500, all U.S. business, or the global tech sector. The note’s combined public/private valuation near $20 trillion is a snapshot of market-implied values, not liquid cash. [49]

Timing matters more than gross order value

A multi-year nominal sales promise becomes more economically meaningful as service is made available, billed, collected, and renewed at a viable cost. Build, power, counterparty performance and amendments can change that path. The stock of future promises is backward-looking evidence of commitment and forward-looking in cash realization; it is not booked revenue. [48] [56]

Exhibit 08 · The IMF tracks three different financial quantities

IMF staff analyzed 14 large publicly disclosed transactions since 2023. This figure compares reported categories, not cumulative realized demand. The three amounts are not additive and are not a census or a representation of funded cash.

IMF October 2026 Global Financial Stability Report, chapter 1: primary source

Why a lender can prefer a strong tenant

RESEARCH NOTE Project finance | Strong tenant credit is not unconditional cover

The credit transformation

A site developer or special purpose entity supplies a specialized building and underlying equipment. Its debt repayment is often arranged around contracted cash from a tenant, including base rent or a take-or-pay offtake payment. A large parent company may back all or part of those payments, give a residual-value guarantee, arrange a capacity substitute clause, or contribute equity. Credit investors can then have exposure partly to the performance of the large company or to the collateral, rather than exclusively to a thinly capitalized developer. [50] [51]

Transfer is not elimination

Compare a guarantee of residual asset value, a guarantee of a borrower’s loan, a fixed capacity reservation, and a construction completion guarantee. In a downturn, each may respond at a different time and pay for a different shortfall. Financial covenants, notice and cure provisions, termination rights, credit support expiration and collateral seniority matter. A property developer and lender cannot reliably count on the strongest party’s full balance sheet unless the relevant legal obligation exists and is enforceable. [52]

Disclosure is not consolidation

Recognized operating and finance lease liabilities, future commitments for facilities not yet delivered, and third-party debt of an unconsolidated vehicle are generally different disclosure categories; project obligations may nevertheless be visible in footnotes and agreements. Read the lease starts, debt borrower, related guarantees and contractual minimums; do not treat all obligations as hidden or double count a lease reported both as a tabular disclosure and as part of a broader headline. [52] [61]

Direct and indirect risk channels

BIS’s October study finds small and short-lived effects in Treasury and investment-grade bond markets on its identified 2026 financing-news days. This is an observed study-window result, not evidence that a future shock under heavier issuance is limited. A shared project sponsor, lender, tenant or insurer can still concentrate risk even when each project has separate legal entities. [51]

The latest Fed tracking shows a broader funding perimeter

RESEARCH NOTE November 2026 snapshot | Outstanding stock is not gross issuance

Debt issuance accelerated

The Federal Reserve’s November financial stability report estimates advanced computing-related debt issuance at over $680 billion through October 2026 versus roughly $130 billion in all of 2023. Tracked issuers’ advanced-computing borrowing rose from less than 6% of nonfinancial investment-grade corporate issuance in 2023 to about 31% in 2026. The Fed estimates outstanding public debt related to advanced computing above $1.1 trillion as of October. The issuance measure captures public issuance plus identifiable Rule 144A securities and excludes Section 4(a)(2) private placements. [52]

Do not add financing estimates across studies

The BIS March Quarterly Review reports that gross bond issuance of selected hyperscalers exceeded $100 billion in 2025; this is neither a net-debt measure nor a statement that every dollar financed AI. The IMF’s $650 billion tracks additional debt associated with selected transactions, not the Fed’s debt universe. Different issuers, windows, instruments, obligations and degrees of contingent exposure prevent direct substitution or addition. [50] [48] [52]

Beyond recognized debt

The Fed’s Box 2.2 describes over $1 trillion of undiscounted not-yet-started leases and over $2 trillion of other undiscounted commitments in a seven-company analysis. Governor [*REDACTED*] characterized the total as a potential maximum nominal value, and separately put expected on-balance-sheet firm debt above $800 billion before year-end. One total is not a market value, expected loss, or debt payable today. A fair comparison models expected timing, conditions, collateral and overlapping guarantees. [52] [53]

A channel for the wider economy

Large bank holding companies had close to $1 trillion in exposure across a broadly defined set including large technology firms, developers, related special purpose vehicles, private credit and utilities. This set isn’t an AI-only loan balance. Losses may first be absorbed by well-capitalized sponsors or subordinate investors, or circulate through refinancing if several counterparties depend on the same source of credit. [52]

Exhibit 09 · Gross issuance grew in the Fed’s tracked universe

Federal Reserve staff report more than $680 billion through October 2026, and about $130 billion in full-year 2023. Bars use the rounded reference amounts; periods differ. Gross issuance is not outstanding debt, net issuance, or proof all proceeds were used for AI.

Federal Reserve November 2026 Financial Stability Report, box 2.2: primary source

Time to cash differs from time to build

RESEARCH NOTE Operations anchor | Physical bottlenecks affect the financial calendar

Permits, equipment, grid, commissioning

Power, interconnection service, transmission and local permit approvals are milestone risks separate from equipment orders. A site with a press release or a queue position is not equivalent to one with a binding interconnection agreement, energized racks or contracted operations. The IEA’s updated Key Questions on Energy and AI central case puts global data-center electricity use at 485 TWh in 2025 and 950 TWh in 2030. That covers all data centers, not solely AI. It is a central scenario, not materialized consumption or a capacity forecast in GW. [54]

Changes in deployment matter to liquidity

A delay could shift revenue right while capitalized construction or equipment deposits continue. If the tenant’s obligation starts at acceptance or energization, the delay can also withhold rent for the project vehicle while construction debt accrues. A completion guarantee, stop-work clause or changes to lease commencement can redistribute that gap. Without the relevant term, the final bearer of delay cost is unknown. [51] [61]

Power purchases and geography

Alternative sites can mitigate local grid constraints if technical and contractual integration permits. But redundancy, cooling, water, reserve power, user latency and equipment delivery all constrain substitution. A firm may need a new contract, rather than moving an existing revenue commitment intact. Local approvals, signed grid service agreements and reported energized load are better milestones than press releases. [54] [55]

Monitor mismatch, not only aggregate use

Estimate the timing of contractual minimums and customer billings under a delay range; list conditions precedent, liquidity buffers and alternative suppliers. An increase in projected global electricity use does not resolve one firm’s financial gap. A site can be delayed during strong market demand or empty during a period of weak conversion. The relevant comparison is contractually reachable and billable capacity against capital and financing costs at the same project and time.

Revenue recognition cannot settle payer independence

RESEARCH NOTE Financial statement reading | Separate recognized sales from recurring cash

Recognition is a distinct gate

FASB Topic 606 ties revenue to the satisfaction of performance obligations for a qualifying customer contract. Contract criteria include commercial substance and probable collection of substantially all consideration to which the issuer will be entitled for goods or services transferred. A framed service reservation can produce real revenue over time if the accounting requirements are met. A deposit, undelivered order or future cloud obligation is a different kind of evidence. [56] [57]

Payments to customers need scrutiny

Payments, credits or similar amounts payable to a customer generally reduce transaction price unless paid in exchange for a distinct good or service, with further limits when payments exceed fair value. Scope and facts control. A strategically related equity stake is not automatically a revenue offset or a revenue violation. Look for the issuer’s explanation of why arrangements are separable, with independent commercial obligations, relevant accounting policies and transaction specifics. [56]

Historical comparison without guilt by analogy

In 2002, SEC staff testimony examined telecom capacity swaps, highlighting questions when cash moves between counterparties concurrently with reciprocal asset exchanges and emphasizing economic purpose. Its historical guidance framework predates today’s standards. The relevance is the discipline of asking what changed economically, not assuming modern financing is equivalent to the accounting issues of that cycle. Separate audited findings regarding a specific party from historical comparisons and current hypothesis. [58]

Cash context

An investor-supported buyer may exhibit real growth with some externally funded consumption. To evaluate durability, compare operating cash conversion, customer concentration, emergency financing, promotional or usage credits, payment terms and like-for-like renewal. Capital inflows from newly issued stock are not cash receipts from providing compute. A commercially rational capacity purchase can become fragile when the financing source and primary purchaser depend on the same group. [47] [49]

Backlog, performance obligations and bookings can diverge

RESEARCH NOTE Operational milestone | Named definition and time window required

Backlog is issuer-defined

[*REDACTED*]’s November 10, 2026 results reported approximately $132.5 billion of revenue backlog at September 30. Its reporting definition includes remaining performance obligations plus other amounts it expects to recognize if delivery and availability conditions are met. The Form 10-Q presented approximately $132.3 billion of unsatisfied remaining performance obligations (RPO), including billed and unbilled amounts from committed contracts. They are close at that date but not interchangeable and should not be added. [59] [60]

Timing of conversion

The filing reported expected recognition of 36% of unsatisfied RPO within the first 24 months ending September 30, 2028, 38% during months 25–48 and the balance across months 49–86. The 36% applies to the $132.3 billion RPO denominator, not to the annual sales base or the $132.5 billion backlog measure. Expectations may change. If service is made ready in stages, the shape of recognition and collection matters more than a headline sum. [60]

Diligence questions

Ask whether the disclosed contract entitles the seller to fixed payment if capacity stays unused, what milestones trigger acceptance, who finances deployment, and whether the buyer has rights to terminate, substitute or renegotiate. Where definitions change, bridge opening to closing backlog with new commitments, delivery, cancellation, financial reclassification and revisions. If the issuer has not disclosed those components, mark each as unknown rather than using all change as new demand.

Do not confuse the related measures

Bookings may be signed deal value, an annualized measure, or estimates based on commercial channels; no single universal accounting convention defines every issuer’s use. RPO is a disclosure on unsatisfied obligations as specified by rules, not banked cash. Billed receivables may become cash sooner than an unbilled order, but the payment period, credit quality and collection record matter separately. [56] [60]

The commitment and collection calendars can disagree

RESEARCH NOTE Prospectus case | Nine-month results and multi-year future obligations

[*REDACTED*]’s disclosed commitments

[*REDACTED*]’s November 2026 final prospectus described approximately $900.9 billion of committed spend as of and subsequent to September 30, substantially all tied to compute and infrastructure, including short-term commitments and leases not yet commenced. About $451.6 billion was expected prior to the end of 2030; about $70.7 billion was due within twelve months of September 30. Do not add $70.7 billion to $451.6 billion. About $186.2 billion of aggregate contractual commitments involved exercisable termination rights or conditions before commencement; the two smaller totals are not disjoint slices unless the document shows so. [61]

Other potential spend is excluded

The same filing separately estimated a $150–$175 billion range under arrangements without fixed dollar commitments. These payments depend on deployed capacity, pricing and delivery. They should not be presented as firm incremental liabilities with guaranteed execution. The filing also sets out the timing mismatch: much revenue is primarily consumption-based and generally not subject to commensurate long-term commitments, while many infrastructure contracts span multiple years. [61]

Evidence customers are using services

For nine months to September 30, 2026 the prospectus indicates $24.5 billion of revenue from customers generating $100,000 or more on a trailing twelve-month basis, about 74% of company revenue for that nine-month period. That is evidence of recognized usage and concentration by size cohort, not verification that each customer independently funded demand or will renew. Some usage is directly contracted; some goes through marketplaces and partners. [61]

Avoid an obligation-to-revenue trap

Quoting a ratio of gross commitments to recent revenue without aligning maturities, recognition, adjusted margin or available cash can suggest a false solvency conclusion. The better comparison estimates year-by-year cash payments, utilization, payment concentration and capex/offtake flexibility against operating cash plus available funding.

Valuation and finance transmit different signals

RESEARCH NOTE Market reading | An investment markup is not operating cash generation

Separate prices, marks and revenues

A financing-round valuation reflects what an investor paid for securities under particular rights and dates, not necessarily one price at which every outstanding claim is equivalent. Public equity pricing reflects expectations about margins, growth, reinvestment and terminal profits. It can move even if current recognized revenue does not change. An investor’s reported gain can be noncash, and neither gain nor valuation funds construction without a financing transaction or liquidity event. [49] [61]

Realized growth and high expectations coexist

IMF’s October technical note judges that the boom reflects both strong earnings by incumbents and elevated valuations, with concentration greater than earlier technology cycles in its comparison. The Fed’s November report also described S&P 500 prices relative to earnings as historically elevated. These observations do not identify fair prices for any stock or predict a correction date. They show that similar near-term operations can accompany materially different expectations about longer-run gains. [49] [52]

One deal can move several economic channels

A major purchase commitment can support a supplier’s forecast, the investee’s claim to scarce capacity, and a credit investor’s confidence in repayment. Equity financing may then appear less costly if market expectations improve. The same set of promises becomes a vulnerability if expected monetization slows and a funding repricing reduces both cash availability and valuation. That feedback is an analytical mechanism, not evidence of a coordinated attempt to misprice securities. [48]

Model terms before valuation

Use scenarios varying utilization, unit revenue, prices, operating margins, replacement cycles and financing costs. Test robustness if the firm can refinance at a higher cost or must absorb more depreciation. Mark announced capacity separately from operational assets. Do not multiply forward revenue commitment by a public multiple without establishing timing, margins, contractual rights and how much revenue is already included in the market estimate.

Wrong-way risk is about timing and dependence

RESEARCH NOTE Transmission mechanism | The honorability of support can weaken when needed most

Two exposures, one shock

A compute customer’s slowdown may reduce the supplier’s projected sales and lower the value of its existing investment in that customer. A contractual backstop to support that same business could become payable at the time the supplier’s business forecast deteriorates. If both parties are financed with the same securities-market optimism, a concurrent mark-down in collateral and lower appetite for refinancing can intensify the cash deficit. This is the mechanism described as wrong-way risk in the IMF financial stability assessment. [48]

Who absorbs first losses

Strong upstream sponsors can absorb losses or guarantee replacement financing, which can protect other investors in ordinary stress. Under narrower contractual support, thinly capitalized developers or junior debt holders may absorb losses first. For the chain to become systemic, correlated defaults, redemptions, forced deleveraging, margin pressure or tighter new credit would have to transmit losses beyond the original group. No published case identification by itself establishes impending systemwide distress. [52] [51]

A practical shock path

Suppose a service provider depends on a major purchaser for contracted utilization. If that purchaser delays accepted capacity, operating cash of the project declines even while interest and salaries continue. Equipment resale values may fall if multiple projects revise demand simultaneously. The sponsor’s claim on limited remaining cash has to be mapped against lien priority and trigger conditions. Draw a separate node for each economic claim and do not count each manifestation as independent loss.

What to watch

Condition the assessment on payment, refinancing and capacity timing, not a market price alone. Evidence of a structural shock includes multiple verified amendments or delayed deliveries, tightening in both credit and equity funding, and lower renewals or collections among independent customers. Correlation in listed returns can be an early signal of shared risk, but cannot prove which contracts or firms transmitted stress. [49]

Correlated portfolios do not prove coordinated decisions

RESEARCH NOTE Cross-market stress | Exposure, leverage, and behavior are separate

Three channels of equity risk

One investor can own equities in several links of the same value chain, increasing sensitivity to the same changes in demand or financing. A capitalization-weighted market index can also become concentrated as constituent values rise. Neither common ownership nor index weight mechanically proves that asset managers directed individual companies’ AI strategies. A third channel is investor leverage through margin, derivatives or levered funds; leverage can produce involuntary selling and liquidity stress even without any change in shared managers’ governance. [48] [52]

Model scenarios are not market predictions

The IMF technical note models the behavior of selected circular-financing stocks under changed correlation and shock assumptions. It estimates an approximately 8–10% weaker cumulative return relative to baseline under stressed correlation assumptions in that exercise, not a forecast for market indexes or a probability of loss. The study uses market-adjusted returns and model simulations. Event-window co-movement can reflect macro news, shared end demand or overlapping financing, without distinguishing all three. [49]

Cross-border exposure

Foreign holders of U.S. equities and credit can be exposed even when their domestic stock market has relatively low AI beta. Semiconductor, hardware and electricity investments affect suppliers around the world. When interpreting a cross-border balance table, distinguish issuer residence, asset currency, ultimate operating exposures, and fund domicile; a fund’s incorporation jurisdiction does not necessarily identify where its final beneficiaries live. [48]

No single portfolio beta proves systemic risk

Monitor flows as well as prices: collateral haircuts, trading liquidity, funding terms, duration and redemption terms can alter whether investor mark-to-market losses transmit into forced sales or reduced new credit. Portfolio concentration is an exposure descriptor. What determines amplification is whether holders have overlapping exposures, leverage and common funding dependencies and are forced to act together.

Macro output is neither ecosystem revenue nor bookings

RESEARCH NOTE National accounts | Recognized private transactions do not all enter GDP as final output

Recent investment signal

BEA’s October 29 advance estimate put U.S. real GDP growth in 2026 Q3 at a 3.4% annual rate versus 2.2% in Q2. Investment led part of the acceleration; information-processing equipment contributed to equipment growth. This category covers computers and peripheral equipment, not AI-only purchases. Imports are separately deducted in GDP. A domestic procurement using imported hardware does not add its full invoice to domestic production. The estimate is also an advance vintage subject to revision. [62]

Why summing private revenue overstates the macro analogy

A chip sale, a lease payment and a downstream AI service each matter to a firm’s statements. As intermediate activity inside a value chain, they cannot all be counted as final consumption or added directly to national GDP. An investment effect on construction and installed equipment is different from realized household consumption, and financial transfers do not measure current production. As an empirical proxy for disruption, tracked equipment orders may reveal a change in expectations rather than a current change in final output. [62] [63]

Cross-sector channels and buffers

A slowdown can affect construction hours, equipment manufacturing and related utilities with different lags. An expansion can be meaningful without being wholly funded by external customer demand in the early years. Large diversified firms may finance the adjustment from non-AI cash flows while thinly capitalized counterparties cannot. A region can still face changes in tax revenue and utility requirements even when aggregate national effects are modest. [48] [54]

Macro-monitoring rule

Use official investment and output data with dates, revisions and sector definitions. Pair them with energized capacity, delivery acceptance, funds raised, credit availability and actual customer collections where observable. Never use whole-ecosystem contracted future dollars as today’s GDP contribution.

Three testable futures, not point forecasts

RESEARCH NOTE Scenario framework | Monitor which observable would overturn the working case

Independent demand catches up

If diversified customers renew with stable pricing, their receipts cover variable service costs and increasingly support fixed compute and finance costs. Net usage and cash collection grow in cohorts that rely less on introductory support. Deposits, leases and capacity options let firms secure inputs during the transition. Track evidence of cash collection, planned-versus-actual deployment, contract performance and refinance terms rather than treating a larger order book as proof. [47] [61]

Installation and demand proceed unevenly

Power and equipment availability, financing terms or permits slow projects, while some demand still grows. Delays push the conversion of nominal commitments into actual revenue and can shift the segment or region where projects land. Legacy sponsors remain liquid, but some individual developers need schedule changes or renegotiation. Watch utilization at accepted sites, not just total operating megawatts or data-center announcements. [54] [51]

Financing and end-market weakness compound

Under weaker monetization, high fixed commitments persist and short-run price pressure reduces margin. Credit spreads rise, guarantees conditionally activate, and development slows. Revised forecasts can be reflected in equity prices before a default occurs. Downstream users may benefit from cheaper capacity even as providers and capital owners lose value. Systemwide consequences depend on the scale and distribution of leverage, collateral haircuts and redemption sensitivity. [48] [52]

Decision trigger

Specify actions contingent on checks. A temporary decline in public market prices or one amended agreement is not confirmation. Simultaneous and persistent changes in cash collection, secured capacity schedule and ability to raise capital warrant a fuller review. Evidence of improving revenue diversification and collections, delivery and lower funding sensitivity would weaken the severe scenario. Do not attach unsupported probabilities or dates.

Make the unknown external demand share measurable

RESEARCH NOTE Analytical template | Use intervals, not a precise estimate from public scraps

Build the audited base

Start with specified period and unit. List all cash receipts from customers for delivered AI services, where disclosed, and map known related organizations, distinct end users, consumption intermediaries and financing sponsors. Separate a parent or investor from a customer using the provider as a reseller or marketplace. Third-party demand should be tested at the ultimate payer rather than inferred from a company name. Reconcile to receivables and customer credits. [56] [47]

Transparent range

For an illustrative firm, receipts equal $100. If $40 is identified as independent, $20 associated with a clearly documented funding tie, and $40 indeterminate, the bounded independent share is 40% to 80% under that classification. The lower bound is confirmed independent receipts; the upper bound adds unknown receipts, not assumed self-funded demand. The range is a toy method, not a sector estimate. Capacity usage and cash conversion are separate weights.

Contract and performance overlays

Report how many transactions you classified and how much value was known, who sponsored or guaranteed which payments, whether incentives were market-standard, and the minimum and optional quantities. Account for overlap and the time between equity cash and consumption; money is generally fungible. An agreement or relation can establish a link without tracing any particular dollar. Large future commitments can matter even if today’s customer receipts are independent. [49] [61]

Comparability and challenge

Use consistent rules, payer identities at the entity level, dates, currencies and thresholds across companies. Avoid claiming a round-trip return by dividing total orders by initial equity. When private links prevent reliable classification, publish the unclassified share along with sensitivity analysis and the highest-value data gap. The point of the range is to show where evidence changes the view, not to substitute precision for unavailable records.

A review protocol for correlated buildout risks

RESEARCH NOTE Decision checklist | Monitor each type of claim with its correct unit

Capture agreements and cash separately

For each counterpart list authorizing entity, legal status, execution or funding date, minimum commitment, options and amendments, term and maturity, venue, trigger, guarantor, collateral and evidence source. Create separate fields for contractual maximum, recognized obligations, commitments excluded from recognition, cash received and expected residual value. A financing announcement can reserve future capital without transferring cash, and a companywide debt figure may finance multiple activities. [52] [61]

Map bottlenecks to counterparties

Record who is responsible for access to power, equipment, build, acceptance and operation, and when debt service begins relative to earnings. Group exposures sharing the same tenant, guarantor, collateral class or refinancing condition. Record geographic concentration of utility, equipment and public approval dependencies. Maintain both gross and net exposure maps and compare secured and under-construction assets against delivery assumptions. [50] [54]

Read the indicators by scenario

Rising utilization and independently funded renewals support a stronger commercial case. Accelerating obligations without delivered capacity or a disproportionate shift toward weaker issuers warrant review. A mark-to-market loss on related equity does not equal a debt write-down; document which accounting and credit indicators changed and why their timing differs. An issuer’s business that is outside the AI network can cushion earnings, though it does not erase commitments. [48] [51]

Tier the conclusion

Use grades such as documented, commitment subject to conditions, issuer estimate, secondary claim not verified, calculation, and scenario. Establish named follow-up metrics: leverage against stressed cash, committed versus energized capacity, cash conversion, maturity clustering, credit support and how much ultimate demand is not publicly discoverable. This is a diagnostic checklist, not investment, legal or accounting advice.

Exhibit 10 · One counterparty can carry several exposures

Equity value, guarantee exposure, trade receivables and underused capacity can all respond to distress at the same counterparty. These are distinct claims with distinct timing and protection. This diagram supplies no aggregate loss estimate.

48PEGASUS Research conceptual exhibit. Not empirical data.

Previous chapter

Next chapter

Back to the report

Browse: White Papers · Reports · Downloads

bottom of page