The Hidden Credit Link in the AI Power Boom

Thesis

This report examines the credit channel linking frontier-model economics to data-center and power-development finance. Version 1.1 adds two explicit circuit breakers: proprietary providers can retain disproportionate monetization after losing token share, and stronger technology balance sheets can substitute for weaker standalone lab credit. The likely near-term result is financing bifurcation rather than uniform contraction.

Key chart

The research question

Could frontier-model multiple compression undermine power-development financing?

Yes, but only after two circuit breakers fail.

An equity valuation is the market's estimate of what future residual cash flows are worth to shareholders. Creditworthiness is the probability and severity of failing to meet contractual obligations. The two overlap, but they are not identical.

A frontier lab could fall from a very high private valuation to a lower one because investors apply a smaller terminal multiple, demand a higher risk premium or become less willing to capitalize distant growth. If revenue, margins, liquidity and debt-service capacity continue to improve, the lab's credit quality could remain stable or even strengthen.

The credit concern becomes much more serious when the valuation reset reflects one or more of the following:

• revenue growth that is decelerating faster than contracted compute costs;
• lower model pricing without an offsetting increase in high-margin volume;
• persistent negative free cash flow;
• rising lease, purchase or take-or-pay obligations;
• weaker access to equity capital;
• a shorter runway before another financing round;
• concentration in one cloud, compute provider or infrastructure sponsor;
• reduced willingness by an investment-grade partner to provide a guarantee or backstop.

The correct causal statement is therefore:

Multiple compression is a warning signal only when its cause changes cash-flow resilience, liquidity, funding access or credit support.

FactOpen-weight models represented 72.4% of the top-20 routed token volume on OpenRouter during July 1–27, 2026. The same Mozilla report cites a historical May–September 2025 OpenRouter sample in which closed models generated approximately 96% of platform revenue despite accounting for approximately 80% of usage. Menlo Ventures separately estimated that Anthropic, OpenAI and Google represented 88% of U.S. enterprise LLM API share in December 2025, while open-source models represented 11%. These are platform- and survey-specific observations, not a current global market census. [S22] [S23] FactNVIDIA disclosed guarantees capped at $105 billion supporting approximately 4.25 GW of OpenAI-related leases at SB Energy's Ohio campus. The guarantees phase in as nine data centers commence, decline as payments are made, cover defined portions rather than every obligation and can terminate if OpenAI achieves a satisfactory credit rating. NVIDIA also stated that less-capitalized AI clouds and model makers can lack the long-term contracting and investment-grade financing capacity required for large infrastructure. [S21] FactIREN disclosed approximately $3.6 billion of investment-grade GPU financing for its Microsoft contract at a company-stated blended cost of 6.0%. It separately grouped $2.8 billion of financings with non-investment-grade customer deployments, including a $2.4 billion Mackenzie equipment financing at a fixed 9.0% rate. The structures differ materially, so the comparison supports financing dispersion rather than a universal credit-pricing rule. [S24] [S25] InferenceThe nearer-term risk is therefore better described as credit migration and financing bifurcation. Strong workloads can continue to receive capital when a stronger balance sheet wraps the obligation. Unwrapped or weakly supported projects may require lower leverage, higher coupons, more prepayment, more sponsor equity or smaller phases. Broad project delay or shrinkage remains a forward hypothesis to test, not a completed market-wide fact.

Evidence status

The evidence is strongest at the transaction and legal-obligation level. It is weaker at the private model-economics level and weakest when attempting to infer a market-wide construction response.

Revision finding: two circuit breakers slow the transmission

Circuit breaker one: token share is not revenue or gross profit

Open-weight adoption can be economically important without immediately destroying proprietary monetization.

Mozilla reports that open-weight models dominated the top of OpenRouter's routed-token leaderboard in July 2026. Yet the report's historical May–September 2025 platform sample attributes approximately 96% of revenue to closed models. Menlo's December 2025 U.S. enterprise study also estimates that Anthropic, OpenAI and Google represented 88% of enterprise LLM API share, while open-source models represented 11%. [S22] [S23]

Those observations support a narrower statement than “closed models retain most global revenue”:

Available platform and enterprise samples show that proprietary providers can retain disproportionate monetization even after open models gain substantial usage.

They do not establish current global revenue share. OpenRouter is a routed developer platform rather than the full market. Menlo's estimate is based on 495 U.S. enterprise decision-makers, weighted production usage and public financial information as of December 2025. Neither source provides audited, model-level gross profit.

The model-layer compression thesis therefore remains a risk pathway, not a completed event. The first circuit breaker fails only if lower prices and lost workload reduce gross profit or cash generation faster than volume growth, serving efficiency and premium-workload monetization can offset them.

Circuit breaker two: financeable credit can come from somebody else

NVIDIA's August 2026 filing provides direct evidence of credit substitution. NVIDIA said AI clouds and model makers can lack the ability to secure long-term infrastructure contracts and investment-grade financing. It then disclosed up to $105 billion of guarantees supporting OpenAI-related land, power and shell obligations for approximately 4.25 GW. [S21]

The workload beneficiary, lease tenant, infrastructure provider and financeable credit can therefore be different entities:

Frontier workload → lab or AI-cloud contract → strategic or hyperscaler credit wrapper → project financing.

This can preserve construction even when the lab's standalone credit is insufficient. It also moves contingent risk onto the guarantor and may concentrate infrastructure exposure within a small number of technology balance sheets.

What the six-versus-nine evidence supports

IREN's disclosures provide unusually useful same-issuer evidence. The company describes approximately $3.6 billion of Microsoft-related investment-grade GPU financing at 6.0% and a separate $2.4 billion Mackenzie financing at 9.0% within a group of non-investment-grade customer deployments. [S24] [S25]

That supports the proposition that counterparty quality and contract structure materially affect financing economics. It does not prove that every investment-grade project prices at 6%, that non-investment-grade projects cannot finance or that tenant credit alone explains the spread. The facilities differ in tenor, prepayments, collateral, amortization, parent support, customer concentration and deployment risk.

“Strong credit gets built; everything else waits, pays up or shrinks” is therefore a useful monitoring hypothesis. The pays-up branch already has direct support. The waits and shrinks branches require a broader cohort of delayed, resized or cancelled projects.

Key findings

1. Credit, not theoretical demand, converts AI adoption into physical infrastructure.
2. Token share is not revenue share, and revenue is not gross profit. Open-weight usage gains do not by themselves prove proprietary-model economic compression.
3. Model-layer multiple compression is not yet a completed, market-wide event. Available samples show substantial proprietary monetization, but current global share and margins are not publicly observable.
4. Investment-grade credit can substitute for weaker standalone lab credit. NVIDIA's OpenAI-related support makes this mechanism explicit at unprecedented scale. [S21]
5. The risk can migrate upward rather than disappear. Guarantors, hyperscalers and strategic suppliers can warehouse obligations that the model developer could not independently finance.
6. A direct investment-grade tenant and an unrated tenant with limited support are not economically equivalent.
7. Lease term alone is insufficient. Guarantee scope, commencement conditions, termination rights, completion support, power certainty, leverage and amortization determine bankability.
8. Published transactions and IREN's same-issuer disclosures show financing dispersion, not a universal coupon rule. [S10]-[S16] [S24] [S25]
9. The most defensible base pathway is bifurcation. Strongly wrapped projects continue while weakly supported pipelines pay more, use more equity, phase more slowly or remain unfinanced.
10. Valuation models should use linked credit states. Debt share, interest cost, sponsor equity and construction probability should move together when driven by the same contract evidence.

Why equity multiples and credit can diverge

This distinction is especially important for frontier labs because private valuations can be enormous relative to current cash generation. Reuters reported private financing rounds valuing OpenAI at $852 billion and Anthropic at $965 billion in 2026. Those figures demonstrate access to capital, but they are not credit ratings and do not by themselves prove that long-duration infrastructure obligations are self-funding. [S3] [S4]

The conditional credit-transmission mechanism

The revised causal chain is:

1. Open-weight convergence changes usage and pricing. It can expand the market while shifting routine workload toward cheaper systems.
2. Monetization is the first circuit breaker. Token-share loss matters to credit only if proprietary revenue, gross profit and cash generation weaken after accounting for higher volume, lower serving cost and premium workloads.
3. Standalone lab credit is then tested. Cash flow, liquidity and equity access determine how safely the lab can support long-dated compute commitments.
4. Credit substitution is the second circuit breaker. A hyperscaler, chip supplier, cloud provider or strategic investor can directly sign, guarantee, prepay or otherwise support the obligation.
5. The legal package determines bankability. Lenders underwrite the obligated entity, guarantee cap and duration, commencement conditions, remedies, collateral and completion support.
6. Bankability sets the capital stack. Coupon, spread, leverage, tenor, amortization, reserves and sponsor equity determine whether the project earns an acceptable return.
7. Only then does the project build, resize, wait or fail to close.

The two circuit breakers create three distinct outcomes:

NVIDIA's filing demonstrates the middle branch. It does not prove that every lab commitment will receive a wrapper or that the guarantor can absorb unlimited exposure. It proves that the project can be financed against a different credit than the entity generating the underlying AI workload. [S21]

Desired compute versus financed compute

A compact representation is:

Financed MW = minimum of demand-backed MW, credit-supported MW, power-deliverable MW and construction-ready MW.

This is not a forecasting equation. It is a constraint map.

A model could justify another 50 GW of economically useful inference, but that demand will not become physical load unless customers sign contracts, the contracts survive legal and credit review, utilities or behind-the-meter systems can deliver power, permits are obtained, equipment is procured and the construction capital is available.

This distinction resolves an apparent contradiction:

• Open models can increase total desired compute by making inference cheaper and more deployable.
• The same competitive process can reduce the proprietary margins available to support long-dated commitments.
• Total desired demand can rise while the marginal project's financing probability falls.

What lenders actually underwrite

1. Tenant and guarantor credit

The lender asks who is legally obligated to pay rent or capacity charges and whether that entity can withstand a downturn. A direct obligation from an investment-grade hyperscaler is different from an obligation from an unrated compute intermediary. A backstop can narrow the gap, but only to the extent of its legal scope.

2. Lease legal strength

The headline contract value is less important than the payment mechanics. Key questions include:

• when rent begins;
• whether charges are fixed, usage-based or conditional;
• whether payments survive customer underutilization;
• termination rights for delay, outage or performance failure;
• caps on damages and service credits;
• change-of-control and assignment provisions;
• renewal and remarketing risk;
• whether the lease is triple-net, near-triple-net or operating-expense exposed.

3. Completion support

A long-term tenant may have no payment obligation until the facility reaches a defined completion standard. Construction lenders therefore need protection before rent begins. Relevant evidence includes guaranteed-maximum-price contracts, parent completion guarantees, liquidated damages, contingency budgets, long-lead procurement and sponsor equity already invested.

4. Power delivery

A completed building without deliverable power cannot generate contracted rent. Lenders examine interconnection rights, utility agreements, substation scope, generation fuel, curtailment rights, upgrade costs, deposits, timing and who bears cost overruns.

FERC's 2026 large-load proceedings illustrate the issue. The Commission directed six regional grid operators to address interconnection processes, cost shifting, co-location, flexible load and nearby generation. Commissioner Chang separately emphasized cost-recovery structures intended to make the customers serving large loads bear induced system costs. [S17] [S18]

5. Leverage and debt-service coverage

Even a strong lease can be overlevered. Lenders size debt against cash flow after operating costs, reserves, taxes, downtime, credits and required capital expenditure. The same project can be bankable at 55% loan-to-cost and unbankable at 80%.

6. Structural protections

Lockboxes, cash waterfalls, reserve accounts, first-lien collateral, restricted payments, covenants, insurance and bankruptcy-remote entities can protect project lenders from sponsor risk. They do not eliminate tenant default, construction failure, power interruption or obsolete-facility risk.

The anatomy of a credit backstop

A public announcement that a major technology company is "backstopping" a lease is not enough to value the contract. Investors should obtain or reconstruct the following:

TeraWulf's SEC filing is a useful example of why precision matters. It states that Google agreed to backstop certain Fluidstack obligations and that the relevant backstop becomes effective at the corresponding lease commencement. The filing also describes recognition agreements, cure or assumption rights and warrants issued in consideration for the support. That is materially more informative than the shorthand phrase "Google guarantee." [S13]

Selected financing case studies

The following transactions disclose coupons and at least part of the tenant, support or rating framework. They should not be treated as a controlled experiment.

What the cases support

The cases are consistent with the proposition that stronger contractual and credit packages can widen the pool of capital and reduce financing cost. Hut 8's River Bend notes reached investment grade with long amortization, a 245 MW project, a Fluidstack lease and Google support. Fleet disclosed a fully leased project with an AA- tenant and a 6.500% coupon. [S10]-[S12]

Galaxy's 9.875% notes provide a useful contrast. S&P cited predictable cash flow under a 15-year CoreWeave lease, but also cited a speculative-grade offtaker, high leverage, construction risk and concentration. [S15] [S16]

A same-issuer 6% versus 9% comparison

IREN's August 2026 disclosures reduce—but do not eliminate—the comparability problem.

InferenceThe three-percentage-point headline difference is consistent with the market charging more for a weaker credit and contract package. It cannot be assigned entirely to customer rating because maturity, collateral, prepayment, amortization, parent support, timing and asset risk differ.

What the evidence does not prove

Coupon differences cannot be attributed to tenant credit alone. The transactions differ in:

• issue date and underlying rates;
• maturity and amortization;
• loan-to-cost and leverage;
• customer prepayments;
• construction and delivery stage;
• collateral and recovery;
• completion and parent support;
• power status;
• project geography;
• tenant concentration;
• call protection and covenants;
• tax, green-bond and investor-base considerations.

The evidence supports credit differentiation and capital-stack bifurcation, not a mechanical rule such as “investment-grade tenant equals 6%.” It also does not yet establish that weakly supported projects are broadly waiting or shrinking. That claim requires observed amendments, deferred notices, reduced phase sizes or failed financings across a wider cohort.

Illustrative capital-stack sensitivity

To show how credit conditions can affect project feasibility, consider a purely illustrative $10 billion project financed with 20-year level-amortizing debt.

This is not a market quote, forecast or representation of any cited transaction. Actual data-center debt is often sculpted, delayed during construction, refinanced, partially amortizing or combined with tax, equipment, utility and sponsor financing.

At 70% loan-to-cost, the sponsor supplies $3 billion. At 60%, it supplies $4 billion. A ten-percentage-point reduction in leverage therefore requires another $1 billion of equity before considering fees, reserves or cost overruns.

Two points illustrate the combined effect:

• 70% loan-to-cost at 6.5%: approximately $3.0 billion of sponsor equity and $635 million of annual level debt service.
• 60% loan-to-cost at 10.0%: approximately $4.0 billion of sponsor equity and $705 million of annual level debt service.

The second structure uses $1 billion less debt but requires roughly $70 million more annual debt service and $1 billion more sponsor equity. That is how a deterioration in perceived counterparty or project risk can make a campus uneconomic without any change in its nameplate power capacity.

Open weights create a volume-margin paradox

Stanford and Epoch AI document declining capability costs and a narrowing open-versus-closed performance gap on selected benchmarks. Mozilla's July 2026 report adds routed-token evidence: the seven highest-volume models on OpenRouter were open weight, and open models represented 72.4% of top-20 token volume during July 1–27. [S1] [S2] [S22]

That is important usage evidence. It is not proof that proprietary revenue, gross profit or durable enterprise demand have fallen at the same rate.

Token share, revenue share and profit are separate

A token can differ materially in price, serving cost, context length, workload value and required reliability. Market share therefore needs at least four separate measures:

1. routed token volume;
2. request or task volume;
3. revenue;
4. gross profit or operating cash flow.

Mozilla's historical May–September 2025 OpenRouter sample attributes approximately 80% of usage and 96% of revenue to closed models. Menlo's December 2025 U.S. enterprise estimate assigns 88% of enterprise LLM API share to Anthropic, OpenAI and Google and 11% to open-source models. [S22] [S23]

The samples have material limitations:

• OpenRouter is one routed developer platform, not the global AI market;
• the 96% revenue observation covers May–September 2025, before open weights' large July 2026 token gains;
• Menlo surveys U.S. enterprise buyers and estimates dollars from weighted production API usage;
• neither source publishes audited model-level gross margins;
• both sources classify heterogeneous products and deployment models.

The defensible conclusion is not that closed models currently own a known global percentage of revenue. It is that proprietary monetization has historically been much more resilient than token-share charts alone imply, and a completed market-wide margin collapse has not been demonstrated.

Volume channel

Lower inference prices and deployable open weights can increase:

• the number of users;
• tasks per user;
• agent duration;
• background automation;
• on-premise and sovereign deployments;
• batch workloads located near cheap energy;
• experimentation that would be uneconomic at proprietary API prices.

This channel is bullish for total compute consumption and potentially for aggregate electricity demand.

Margin-risk channel

Open alternatives can still:

• cap proprietary API pricing;
• increase workload routing across providers;
• weaken customer lock-in at the model layer;
• shorten the economic life of a frontier release;
• force continuous research and infrastructure spending;
• shift routine work toward lower-margin or self-hosted systems.

This is a forward risk to model-layer scarcity rents, not evidence that the compression has already occurred across the market.

The credit-relevant variable

The relevant question is not whether open tokens increase. It is:

Does proprietary gross profit and durable operating cash flow grow faster than the fixed and contingent obligations required to produce that growth?

A lab can process far more tokens and become less creditworthy if revenue per unit falls faster than serving cost, capital expenditure, leases, R&D and financing needs. It can also lose token share while improving credit if premium pricing, enterprise demand and unit-cost reductions preserve cash generation.

This is the first circuit breaker in the revised framework. Model competition reaches infrastructure credit only after it impairs realized economics.

Why frontier valuations matter even without public debt ratings

Private frontier valuations perform a practical financing function. They determine how much dilution is required to raise the next dollar of equity and influence the confidence of strategic partners, lenders and counterparties.

If a lab needs $50 billion of additional equity:

• at a $1 trillion valuation, the amount is equivalent to 5% of that valuation;
• at a $250 billion valuation, it is equivalent to 20%.

That arithmetic is illustrative, but the mechanism is real. A lower valuation raises the economic cost of equity funding and can reduce the willingness of existing investors to continue financing commitments whose returns are increasingly competed away.

Rating-agency commentary shows that credit markets are already focused on funding structures rather than only headline demand. S&P expects more than $1.3 trillion of combined hyperscaler capital expenditure by 2027 and describes negative free operating cash flow across six large hyperscalers in 2026 and 2027. Moody's notes that developers require multiple forms of debt and equity and that startup AI firms depend heavily on large-technology-company credit support. [S7]-[S9]

Oracle provides a public-company example of the broader mechanism. Reuters reported negative free cash flow, heavy capital spending and roughly $260 billion of signed data-center leases, alongside heightened scrutiny of the company's credit profile. Oracle is not a pure frontier lab, but the case illustrates how long-term AI infrastructure obligations can become a corporate-credit issue. [S20]

Credit support can decouple the project from the lab

The second circuit breaker is the availability of stronger credit.

A project can remain bankable if:

• an investment-grade hyperscaler directly signs the lease;
• an investment-grade entity provides a sufficiently broad, durable and enforceable guarantee;
• the tenant prepays material amounts;
• a strategic supplier commits to buy unused capacity or supply residual-value support;
• debt is sized conservatively;
• sponsor completion support covers the construction period;
• the project is ring-fenced with reserves, collateral and a lender-controlled cash waterfall;
• the facility has credible alternative-use or remarketing value.

NVIDIA makes the substitution mechanism explicit

NVIDIA's Form 10-Q states that less-capitalized AI clouds and model makers can lack the ability to secure long-term infrastructure contracts and investment-grade financing. NVIDIA says it has responded through guarantees, capacity commitments and financing initiatives. [S21]

The most consequential disclosed example is the SB Energy PORTS campus:

• approximately 4.25 GW of IT load under OpenAI-related leases;
• up to $105 billion of NVIDIA guarantees;
• nine phases, with support generally becoming effective as each lease commences;
• 20-year lease terms;
• obligations triggered by specified tenant defaults;
• exposure declining as OpenAI pays;
• coverage limited to defined portions of lease and power payments rather than every tenant or project obligation;
• termination upon specified events, including OpenAI achieving a satisfactory credit rating;
• an NVIDIA option to support approximately 3.8 GW of additional phases.

This is direct evidence that the entity creating AI demand and the entity supplying financeable credit need not be the same.

The risk is transferred, not erased

Credit substitution can lower project-level financing risk while increasing concentration and contingent exposure at the guarantor. The wrapper must therefore be analyzed on two levels:

NVIDIA also disclosed $36 billion of AI-cloud commitments and preliminary memorandums intended to mobilize more than $500 billion of third-party infrastructure capital. The memorandums are not definitive financing commitments, and any residual-value support would be project-specific. [S21]

The practical implication is credit migration. Strong technology balance sheets can keep selected infrastructure moving, but the system becomes more dependent on their willingness and capacity to warehouse AI risk.

Project finance also limits, but does not eliminate, contagion. Nonrecourse debt can protect the sponsor's broader balance sheet. It cannot make an insolvent tenant pay, expand a capped guarantee or create power that was never delivered.

Three forward scenarios

No probabilities are assigned because current evidence does not support precise weights.

The revised base pathway is bifurcation, not an imminent uniform financing stop. The bear case requires both circuit breakers to fail: proprietary economics deteriorate and stronger credit support becomes unavailable or inadequate.

Strongest counterargument

The strongest counterargument is no longer a footnote to the framework; it is an explicit part of the revised base case.

Open-weight progress may expand the market while proprietary providers continue monetizing the hardest and most valuable work. Closed providers can also defend economics through enterprise distribution, workflow integration, security, memory, identity, agent orchestration and lower serving costs. Token volume can migrate without an equivalent loss of dollars or gross profit. [S22] [S23]

Even if standalone lab economics weaken, hyperscalers, NVIDIA and other strategic counterparties may rationally support the infrastructure because they capture value elsewhere—from cloud consumption, chip sales, software, equity ownership, distribution or ecosystem control. NVIDIA's disclosed guarantees show that this is not merely hypothetical. [S21]

There is also a stabilizing supply feedback:

1. weaker economics reduce unwrapped financed capacity;
2. reduced capacity tightens future compute supply;
3. tighter supply supports pricing and utilization;
4. improved economics restore some financing capacity.

The remaining counterargument to the counterargument is concentration. Wrappers have caps, conditions and opportunity costs. A small number of strong balance sheets may be able to preserve the buildout for a long period, but doing so transfers more correlated AI exposure onto those entities.

For these reasons, the report does not conclude that open weights will cause a frontier-lab credit crisis or stop the power buildout. It concludes that investors should track where the credit migrates, what the recipient charges for it and which projects remain unsupported.

Disconfirming evidence

The financing-bifurcation thesis would weaken if:

• proprietary providers deliver sustained positive free cash flow despite continued open-weight convergence;
• gross profit grows materially faster than compute, lease and R&D obligations;
• current, broad market data show open models taking revenue and gross-profit share much more slowly than assumed;
• unwrapped non-investment-grade projects repeatedly obtain leverage, tenor and spreads comparable with investment-grade-supported projects;
• direct investment-grade tenants replace startup intermediaries across most new developments;
• guarantees become broader, longer, standardized and less conditional without materially burdening guarantors;
• project-level defaults and losses remain low through a demand or pricing slowdown;
• weakly supported projects continue to close at full planned scale rather than being deferred or resized.

The thesis would strengthen if:

• proprietary pricing or gross profit falls despite continuing token growth;
• equity rounds become more dilutive, conditional or concentrated among strategic backers;
• guarantee caps tighten, commencement is delayed or wrappers are withdrawn;
• project coupons widen and loan-to-cost falls after controlling for base rates, tenor and structure;
• developers increase prepayments, sponsor equity, warrants or parent guarantees to close;
• phase sizes are reduced, delivery dates are extended or projects are cancelled for financing reasons;
• ratings, bond spreads or contingent-obligation disclosures deteriorate at the guarantors absorbing the risk.

Risks to the framework

Measurement risk

Model benchmarks do not directly measure willingness to pay, enterprise reliability, security, agent performance or integration costs. A four-month capability lag is not equivalent to a four-month economic lag.

Private-information risk

Frontier-lab unit economics, side letters, strategic guarantees and capacity commitments are often confidential. Public analysis may miss both protections and liabilities.

Causality risk

Interest rates, tenor, security, leverage, construction stage and market timing affect coupons. The selected financing chart is descriptive.

Structural-change risk

Custom silicon, edge inference, new cooling systems, software efficiency, sovereign funding or regulation could alter the relationship between tokens, data centers and power.

Policy risk

FERC and state regulators may require deposits, cost-recovery agreements, curtailment rights or dedicated generation. These can improve ratepayer protection while raising the capital required from large-load customers. [S17] [S18]

Feedback risk

Credit tightening can reduce supply and support compute prices. Static models may overstate downside if they ignore this feedback.

Investment implications

1. Gross pipeline MW should not be valued as financed MW

A development pipeline should be segmented into:

• land or concept MW;
• power-requested MW;
• power-approved MW;
• contracted MW;
• credit-supported contracted MW;
• financed MW;
• under-construction MW;
• energized and rent-paying MW.

The value difference between those states can be larger than the difference between two terminal capitalization rates.

2. Credit tightening can hurt pipeline assets and help de-risked assets

If lenders reduce leverage and require more equity, weaker projects are cancelled. That is bearish for developers whose value depends on uncontracted future capacity. It can be relatively bullish for completed, powered and strongly contracted assets because competing supply falls.

3. The guarantee should be modeled as a structure, not a checkbox

A model should record:

• direct tenant rating or implied credit;
• guarantor identity;
• guarantee scope;
• cap;
• duration;
• commencement date;
• triggers;
• cure and assumption rights;
• termination-payment coverage;
• construction-period coverage.

4. Financing assumptions should be linked

A stronger contract can simultaneously increase:

• financing probability;
• debt share;
• tenor;
• amortization quality;
• interest rate;
• completion probability;
• stabilized cap rate.

These effects should not be adjusted independently without checking for double counting.

Implications for Ephesus Research models

This publication proposes a methodology for future model revisions. It does not change any live model input, scenario probability, or public valuation output.

A credit-adjusted capacity framework

A more defensible development model can use:

Credit-adjusted MW = gross MW x probability of power delivery x probability of financing given the contract x probability of completion given financing x economic ownership.

The probabilities are conditional and must not be multiplied mechanically if they already capture the same risk. For example, a lower construction probability and a higher discount rate may both reflect weak tenant credit; applying both without a documented distinction can double count risk.

Suggested financing states for a development campus

For a TCDC-style development model, this matrix should be applied phase by phase. Management statements about discussions with investment-grade tenants are commercially relevant, but they should not receive the same treatment as an executed lease, an enforceable guarantee and closed construction financing.

For an IREN-style operating-company model, obligations should be separated into corporate debt, project-level nonrecourse debt, leases, GPU or equipment financing, customer prepayments and contingent commitments. The model should identify which cash flows and assets support each claim.

Monitoring dashboard

The next 12–18 months should be monitored as a financing cohort, not as isolated headlines.

Methodology

This report uses a causal-chain framework rather than a valuation target.

1. It identifies evidence that open-weight models are narrowing parts of the capability and cost gap.
2. It separates token share, request share, revenue, gross profit and operating cash flow.
3. It separates equity valuation from standalone credit quality.
4. It adds two explicit circuit breakers: monetization resilience and investment-grade credit substitution.
5. It maps guarantee scope and counterparty quality into project bankability and financed MW.
6. It reviews selected public financing cases with disclosed coupons, ratings and support structures.
7. It adds a same-issuer IREN comparison while preserving structural qualifications.
8. It applies a transparent, illustrative $10 billion project-finance sensitivity.
9. It develops bear, base and bull pathways without assigning unsupported probabilities.
10. It identifies disconfirming evidence and a 12–18 month monitoring cohort.

Primary filings, company transaction releases, rating-agency rationales and government publications were prioritized. Mozilla and Menlo Ventures were used for original platform and survey evidence, with their scope, time period, methodology and incentives explicitly qualified. Reuters remains limited to private-company valuations, reported commitments and market context where complete primary documentation was unavailable.

Limitations

• This is not a credit rating of OpenAI, Anthropic, NVIDIA, Microsoft, CoreWeave, Fluidstack, Oracle or any project issuer.
• Private frontier-lab financial statements, model-level gross margins and complete contract schedules were not available.
• Current global closed-versus-open revenue share is not publicly observable. OpenRouter and Menlo provide useful but non-comprehensive samples.
• Mozilla's July 2026 token data and May–September 2025 revenue data cover different periods; they cannot be combined into a current market-share estimate.
• Menlo's study is a December 2025 estimate based on 495 U.S. enterprise decision-makers and a market-sizing model, not audited global revenue.
• NVIDIA's $105 billion disclosure is a capped, conditional guarantee framework, not proof that every dollar will be drawn or that every project cost is covered.
• IREN's 6.0% and 9.0% financing examples differ in tenor, collateral, prepayments, amortization, parent support and deployment risk; they are not a matched credit experiment.
• The broader transaction sample is selected, small and not statistically controlled.
• Coupon comparisons do not fully adjust for the yield curve, original issue discount, fees, call protection, hedging, tenor or security.
• The assertion that weaker projects will wait or shrink is a forward hypothesis. A broad cohort of observed deferrals, resizing or failed closings is not yet available.
• The $10 billion sensitivity assumes level annual debt service over 20 years and excludes construction-period interest, fees, taxes, reserves, working capital and refinancing.
• Power availability, permitting and utility cost allocation can bind before tenant credit does.
• Open-model benchmark convergence may not translate into enterprise substitution or price sensitivity.
• No Ephesus model input, scenario probability or valuation output was changed as part of this revision.

What would change the conclusion

The conclusion would become more bearish if:

• proprietary prices and gross profit fall faster than serving efficiency and workload volume improve;
• frontier funding rounds become smaller, more dilutive or contingent;
• strategic guarantors reduce support or approach concentration limits;
• project coupons widen and loan-to-cost falls despite stable benchmark rates;
• lease amendments shift more construction, utilization or termination risk back to developers;
• weakly supported campuses are delayed, resized or cancelled because financing cannot close;
• contingent AI obligations begin to impair ratings or funding costs at the guarantors.

The conclusion would become more bullish if:

• frontier operating cash flow turns sustainably positive;
• premium model workloads retain pricing power despite open alternatives;
• current market data continue to show proprietary revenue resilience after open token-share gains;
• investment-grade counterparties directly sign or broadly guarantee a larger share of leases;
• standardized guarantees, securitization and third-party capital deepen the investor base;
• non-investment-grade projects demonstrate durable financeability without excessive sponsor support;
• project-level default and loss experience remains low through a demand slowdown;
• power and interconnection reforms reduce construction uncertainty without shifting unfinanceable costs to large-load customers.

Detailed conclusion

The original thesis survives, but the likely transmission path is slower and more concentrated than a direct model-to-power contagion story implies.

Open-weight models can win tokens without immediately winning the same dollars or gross profit. Available platform and U.S. enterprise evidence shows that proprietary monetization has remained materially more resilient than usage share alone would suggest. That means model-layer multiple and margin compression should be treated as a contingent forward risk, not a completed market-wide fact. [S22] [S23]

Standalone lab credit is also not the only credit available to the infrastructure stack. NVIDIA's OpenAI-related guarantees show that a strategic balance sheet can substitute for a weaker customer's financing capacity at very large scale. This can preserve construction while moving contingent exposure upward to the guarantor. [S21]

The nearer-term implication is therefore financing bifurcation:

• strongly contracted and investment-grade-supported capacity can access deeper, cheaper capital;
• weaker but financeable deployments may use higher coupons, shorter tenor, more prepayment, more collateral and parent support;
• unwrapped speculative capacity remains most exposed to delay, resizing or failed financial close.

IREN's disclosed 6.0% Microsoft-related financing and 9.0% Mackenzie financing provide direct evidence of the pays-more branch. They do not establish a universal 6% threshold, and the market-wide waits-or-shrinks branch remains to be demonstrated through actual project cohorts. [S24] [S25]

The most important question is no longer simply whether a frontier lab is creditworthy. It is:

Who ultimately warehouses the lab's infrastructure obligations, on what terms, and how much correlated AI exposure can that balance sheet absorb?

For infrastructure investors, the scarce asset remains the megawatt attached to deliverable power, completed construction and durable credit. The revision adds one further distinction: durable credit may be substituted, but it is never free and it never eliminates the underlying risk.

---

Complete sources

The complete claim-level source register is also available in CSV and JSON formats.

S1 - The 2025 AI Index Report

Publisher: Stanford Institute for Human-Centered Artificial Intelligence
Publication date: 2025-04-07
Source type: Independent research report
Used for: Reports that the inference cost of GPT-3.5-level performance fell more than 280-fold from November 2022 to October 2024, hardware costs declined about 30% annually, hardware energy efficiency improved about 40% annually, and the open-versus-closed performance gap narrowed materially on some benchmarks.
Qualification: Original Stanford AI Index synthesis. Benchmark-specific convergence does not prove parity across all economically important workloads.

S1 - Direct source

S2 - Open models lag state-of-the-art closed models by 4 months

Publisher: Epoch AI
Publication date: 2026-05-29
Source type: Independent model-capability analysis
Used for: Estimates that since January 2026 the leading open-weight models lagged the leading closed models by an average of four months and eight points on the Epoch Capabilities Index.
Qualification: The authors disclose sensitivity to benchmark coverage and methodology. A stricter interpretation can produce a longer lag, and public benchmarks may not capture private or deployment-specific advantages.

S2 - Direct source

S3 - Artificial Intelligencer: OpenAI's $852 billion problem: finding focus

Publisher: Reuters
Publication date: 2026-04-01
Source type: News analysis
Used for: Reports that OpenAI raised $122 billion at an $852 billion valuation.
Qualification: Private-market valuation and financing figures are reported by Reuters; they are not audited credit metrics and should not be treated as a rating.

S3 - Direct source

S4 - Anthropic's valuation surges to $965 billion, surpassing OpenAI

Publisher: Reuters
Publication date: 2026-05-28
Source type: News report
Used for: Reports that Anthropic raised $65 billion at a $965 billion post-money valuation.
Qualification: Private-market valuation and financing figures are reported by Reuters; they are not audited credit metrics and should not be treated as a rating.

S4 - Direct source

S5 - What happens if OpenAI or Anthropic fail?

Publisher: Reuters Breakingviews
Publication date: 2026-03-11
Source type: Financial commentary
Used for: Cites a Morgan Stanley forecast of $2.9 trillion of global data-center investment from 2025 through 2028, including roughly $900 billion from private credit and asset-backed lending.
Qualification: Commentary and third-party forecast. Used to frame financing scale, not as a verified base-case forecast.

S5 - Direct source

S6 - Anthropic was cautious on mega infra deals. Then demand surged

Publisher: Reuters
Publication date: 2026-09-02
Source type: News analysis
Used for: Reports a $30 billion Azure commitment, an approximately $45 billion Nscale infrastructure agreement, and a $1.25 billion monthly SpaceX compute agreement for Anthropic; it also reports that OpenAI represents about 8.753 GW of an 8.8 GW contracted SB Energy portfolio.
Qualification: Figures are reported by Reuters and illustrate scale and concentration. Underlying contracts were not all publicly available for independent review.

S6 - Direct source

S7 - AI Infrastructure Investment To Exceed $1.3 Trillion By 2027

Publisher: S&P Global Ratings
Publication date: 2026-08-27
Source type: Credit research press release
Used for: Projects combined capital expenditure by six large hyperscalers to exceed $1.3 trillion by 2027, expects negative free operating cash flow across the group in 2026 and 2027, and highlights debt, leases, guarantees, SPVs, joint ventures and residual-value guarantees as increasingly important.
Qualification: Original rating-agency summary. Forecasts are S&P views, not realized outcomes.

S7 - Direct source

S8 - Credit Risk Insights for Global Data Centers

Publisher: Moody's
Publication date: 2026-09-02
Source type: Credit risk topic hub
Used for: States that data-center developers and landlords require substantial equity, bank debt, bonds, securitization and project-finance capital, and explains that data-center ABS repayment is primarily sourced from tenant lease payments.
Qualification: Continuously updated Moody's topic hub; publication date records the access and verification date used for this report.

S8 - Direct source

S9 - How will the 2026 digital economy be impacted by changes in artificial intelligence, digital finance, cyber risk, and data centers?

Publisher: Moody's
Publication date: 2026-01-15
Source type: Credit outlook summary
Used for: Highlights narrowing model-performance gaps, questions about proprietary AI monetization, concentrated preleasing, power and construction constraints, and dependence by startup AI firms on large-technology-company credit support.
Qualification: Original Moody's outlook summary. Forward-looking statements are analytical views, not guaranteed outcomes.

S9 - Direct source

S10 - Hut 8 Closes $3.25 Billion of Investment-Grade Senior Secured Notes in Landmark Financing for River Bend Data Center Project

Publisher: Hut 8
Publication date: 2026-04-30
Source type: Company press release
Used for: Discloses $3.25 billion of 6.192% senior secured notes due 2042, BBB- ratings, 245 MW of critical IT capacity, a fully amortizing 16.5-year tenor and project-level financing without recourse to Hut 8.
Qualification: Company-furnished source. Coupon, size, capacity and stated structure were cross-checked against S&P's transaction summary.

S10 - Direct source

S11 - Hut 8 DC LLC Senior Secured Notes Assigned 'BBB-'

Publisher: S&P Global Ratings
Publication date: 2026-04-22
Source type: Credit rating rationale
Used for: Describes the River Bend project, the Fluidstack lease, Google's guarantee of tenant payment obligations, and construction protections supporting the investment-grade notes.
Qualification: Original S&P rating rationale. Ratings are opinions of credit risk, not investment recommendations or guarantees.

S11 - Direct source

S12 - Fleet Data Centers Announces Closing of Upsized $4.6 Billion 6.500% Senior Secured Notes Offering

Publisher: Fleet Data Centers
Publication date: 2026-05-01
Source type: Company press release
Used for: Discloses $4.6 billion of 6.500% senior secured notes due 2031 to finance a 230 MW utility / 200 MW critical IT facility that is 100% leased for 197 months to an AA- investment-grade tenant.
Qualification: Company-furnished source. This report does not infer an issue rating where the source does not provide one.

S12 - Direct source

S13 - TeraWulf Form 8-K - Fluidstack Leases and Google Recognition Agreements

Publisher: U.S. Securities and Exchange Commission / TeraWulf
Publication date: 2025-08-14
Source type: SEC filing
Used for: Discloses leases for more than 200 MW of critical IT load, ten-year rent terms, and recognition agreements under which Google agreed to backstop certain Fluidstack obligations, effective at the relevant lease commencement.
Qualification: Primary legal filing. The wording 'certain obligations' and the effective-date conditions matter; it should not be summarized as an unlimited unconditional guarantee.

S13 - Direct source

S14 - TeraWulf Q3 2025 Investor Update

Publisher: U.S. Securities and Exchange Commission / TeraWulf
Publication date: 2025-11-10
Source type: SEC-filed company presentation
Used for: Discloses a $3.2 billion offering of 7.750% senior secured notes due 2030, ratings of Ba2/BB/BB, and a stated use of proceeds to finance 522.5 MW of AI / high-powered-compute data centers.
Qualification: SEC-filed presentation. Company labels and forward-looking construction statements remain company representations.

S14 - Direct source

S15 - Galaxy Digital Form 8-K - Helios II Senior Secured Notes

Publisher: U.S. Securities and Exchange Commission / Galaxy Digital
Publication date: 2026-07-28
Source type: SEC filing
Used for: Discloses $3.507 billion of 9.875% senior secured notes due 2031 issued by Galaxy Helios Data Centers II, with semiannual interest and scheduled amortization.
Qualification: Primary legal filing. Capacity, tenant quality and rating rationale are cross-referenced to S&P's transaction analysis.

S15 - Direct source

S16 - Galaxy Helios Assigned 'B+' Issuer Credit Rating; Outlook Stable; Debt Rated 'BB-'

Publisher: S&P Global Ratings
Publication date: 2026-07-23
Source type: Credit rating rationale
Used for: Describes $3.5 billion of notes funding two facilities with 400 MW of utility capacity and 260 MW of critical IT load, a 15-year CoreWeave lease, B+/BB- ratings, predictable lease cash flow, speculative-grade tenant concentration, high leverage and construction protections.
Qualification: Original S&P rating rationale. Ratings are opinions of credit risk, not investment recommendations or guarantees.

S16 - Direct source

S17 - FERC Launches Aggressive Targeted Action to Speed Large Load Integration

Publisher: Federal Energy Regulatory Commission
Publication date: 2026-06-18
Source type: Government order summary
Used for: States that FERC ordered six regional grid operators to justify or reform large-load tariffs, including transmission studies, cost-shift protections, co-location, flexible-load service and proximate generation processes.
Qualification: Official government source. Proceedings were ongoing as of publication and did not establish one uniform national final rule.

S17 - Direct source

S18 - Commissioner Chang's Concurrence in California Independent System Operator Corporation - Docket No. EL26-71-000

Publisher: Federal Energy Regulatory Commission
Publication date: 2026-06-18
Source type: Government commissioner statement
Used for: Explains concerns that rapid large-load growth strains planning and resource procurement and supports cost-recovery agreements intended to make customers serving large loads bear the costs they induce.
Qualification: Official commissioner statement, not a final commission-wide rule. Used to identify the credit and cost-allocation questions facing power development.

S18 - Direct source

S19 - EIA forecasts strongest four-year growth in U.S. electricity demand since 2000, fueled by data centers

Publisher: U.S. Energy Information Administration
Publication date: 2026-01-13
Source type: Government forecast
Used for: Forecasts U.S. electricity use growth of 1% in 2026 and 3% in 2027 and attributes the strongest four-year demand-growth period since 2000 primarily to increasing demand from large computing centers.
Qualification: Official short-term forecast. Forecasts can change with economic activity, project delays, weather, efficiency, pricing and supply constraints.

S19 - Direct source

S20 - Oracle Corp goes for high-stakes ratings gamble in AI strategy

Publisher: Reuters
Publication date: 2026-08-04
Source type: Credit-market news analysis
Used for: Reports negative free cash flow, heavy capital spending and roughly $260 billion of signed data-center leases at Oracle, alongside heightened rating-agency scrutiny.
Qualification: Reuters analysis based on company disclosures and credit-market commentary. Used as an example of how infrastructure commitments can affect corporate credit.

S20 - Direct source

S21 - NVIDIA Form 10-Q for the quarter ended July 26, 2026

Publisher: U.S. Securities and Exchange Commission / NVIDIA
Publication date: 2026-08-26
Source type: SEC filing
Used for: Discloses guarantees capped at $105 billion supporting approximately 4.25 GW of OpenAI-related leases at SB Energy's PORTS campus; nine-phase commencement mechanics; 20-year lease exposure; defined payment coverage; termination conditions; broader AI-cloud guarantees, commitments and financing initiatives.
Qualification: Primary legal filing. The guarantees are capped, conditional and limited to defined portions of lease and power payments; the maximum exposure is not the same as an expected cash payment or full project cost.

S21 - Direct source

S22 - The State of Open Source AI — v1.0.1

Publisher: Mozilla
Publication date: 2026-07-14
Source type: Original industry report and developer survey
Used for: Reports that open-weight models represented 72.4% of top-20 OpenRouter routed-token volume during July 1–27, 2026; cites a historical May–September 2025 OpenRouter sample in which closed models generated approximately 96% of revenue; and reports developer adoption and production-deployment differences.
Qualification: Platform- and survey-specific evidence from an organization that advocates for open ecosystems. Token and revenue observations cover different periods, OpenRouter is not the global market, and the report does not provide audited model-level economics.

S22 - Direct source

S23 - 2025: The State of Generative AI in the Enterprise

Publisher: Menlo Ventures
Publication date: 2025-12-09
Source type: Enterprise survey and market-sizing report
Used for: Estimates that Anthropic, OpenAI and Google represented 88% of U.S. enterprise LLM API share in December 2025, while open-source models represented 11%; documents the report's 495-person U.S. enterprise survey and weighted production-usage methodology.
Qualification: Original study, but not an audited market census. Results are limited to U.S. enterprises, use modeled dollar estimates and represent the authors' best assessment as of December 2025. Menlo is an AI investor and discloses portfolio relationships.

S23 - Direct source

S24 - IREN Form 10-K for the year ended June 30, 2026

Publisher: U.S. Securities and Exchange Commission / IREN
Publication date: 2026-08-27
Source type: SEC filing
Used for: Discloses the Microsoft-related delayed-draw term loan and 5.96% private-placement notes; customer prepayment mechanics; the $2.4 billion Mackenzie equipment financing at 9.0%; maturity, amortization, collateral and IREN guarantee terms.
Qualification: Primary legal filing. The financing structures differ in tenor, collateral, customer support, draw mechanics and parent exposure, so their pricing cannot be attributed solely to customer credit.

S24 - Direct source

S25 - IREN FY26 and Q4 FY26 Results

Publisher: U.S. Securities and Exchange Commission / IREN
Publication date: 2026-08-27
Source type: SEC-filed company results presentation
Used for: Company classification of approximately $3.6 billion of Microsoft-related investment-grade GPU financing at a blended 6.0% and $2.8 billion of financings supporting non-investment-grade customer deployments, including $2.4 billion at 9.0% for Mackenzie.
Qualification: Company-furnished presentation. “Investment grade,” “non-investment grade,” blended cost and capex-coverage percentages are IREN's descriptions and should be read with the underlying 10-K terms.

S25 - Direct source

Article revision summary

Disclosure

Ephesus Research publishes research and educational material for self-directed investors and researchers. This report is not personalized investment advice, a credit rating, a legal opinion or a promise of future results. Private-company information is incomplete, forecasts can be wrong, ratings can change, and contractual terms can differ from public summaries. Readers should review the underlying sources and obtain professional advice where appropriate.