Structural movement across the AI value chain, for people who allocate capital and capacity.
The Collateral Is the Contract
Six GPU-backed facilities priced this year separate cleanly into two bands, and the variable is not the hardware. It is who signed the offtake. Power still decides whether a data center can exist; a rated counterparty now decides what it costs to build.
2026-08-22 to 2026-09-05 · September 5, 2026 · Investors · C-suite · REIT leaders · Market analysts

Key takeaways
- GPU-backed debt has split into two visible pricing bands. Facilities behind an investment-grade offtaker cleared at SOFR plus 225 to 300 basis points this year. Facilities behind non-investment-grade customers cleared at SOFR plus 450 to 550.
- The cleanest evidence is same-sponsor. CoreWeave priced three facilities off one platform in five months: 225 over SOFR against Meta take-or-pay contracts, 450 against two non-investment-grade customers, 550 against shorter non-investment-grade contracts.
- NVIDIA's residual value guaranties for OpenAI-leased capacity terminate, by their own terms, upon OpenAI achieving a satisfactory credit rating. The guarantee is not a partnership gesture. It is a substitute for a rating that does not yet exist, written to expire when one does.
- This is not a new financial technology. It is project finance and vendor credit support at unprecedented scale, and the closest historical analogues in vendor-supported demand did not end well for the vendors.
- Power remains the physical constraint. Credit determines what access to power costs, not whether it exists. Treating the two as interchangeable is the most common analytical error in the current cycle.
- Local resistance is concentrating where data centers are arriving rather than where they already are. Virginia and Texas together account for five of 533 tracked local moratorium instruments; Michigan alone accounts for 64.
By the numbers
- 225 vs 550Basis points over SOFR, same sponsor, same asset classCoreWeave DDTL 4.0 against Meta take-or-pay contracts versus DDTL 5.5 against shorter non-investment-grade contracts, both closed in 2026.
- $105.0BCap on NVIDIA's residual value guaranties for OpenAI-leased capacityCovers roughly 4.25 GW of IT load at the Portsmouth site, with credit support for a further 3.8 GW exercisable at NVIDIA's sole discretion.
- 300 bpCurve-adjusted gap between IREN's two 2026 GPU financingsRoughly 183 basis points over Treasuries on the Microsoft-contracted notes against roughly 479 on the non-investment-grade facility.
- 533Local moratorium instruments tracked across 42 states429 in force as of 19 August 2026. Michigan leads with 64; Virginia has two and Texas three.
- 53 of 116 GWShare of GE Vernova's gas commitments that is firm backlogThe remaining 63 GW is slot reservation agreements, which are not equipment orders.
- $89.0BNVIDIA Data Center revenue in the quarter ended 26 July 2026Up 117 percent year over year, with management describing supply as constrained through fiscal 2028.
Audience implications
- Investors
Rebuild neocloud comparisons around weighted average cost of capital and offtaker rating rather than fleet size. A 325 basis point spread differential on multi-billion-dollar facilities dominates most operating differences between otherwise similar operators.
- C-suite
If your AI infrastructure partner's economics depend on vendor guarantees or customer prepayments, understand what happens when that support terminates. In at least one disclosed case it terminates automatically once the tenant earns a credit rating.
- REIT leaders
Report energised, contracted and announced capacity as three separate numbers, and underwrite entitlement risk by market maturity rather than by state incentive regime. Both distinctions are now being priced.
- Market analysts
Stop deriving load growth from interconnection queues. One market is verifying roughly 205 gigawatts against a 474 gigawatt queue, and the correction will look like a demand shock when it is a measurement fix.
Two bands, one variable
Six GPU-backed facilities priced in 2026 fall into two clean bands. Everything physical about them is similar: NVIDIA accelerators, special-purpose borrowers, delayed-draw structures aligned to deployment. What separates them by more than 300 basis points is the credit quality of the counterparty that signed the offtake, and whether that contract could be pledged.
Sources CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · CoreWeave (primary) · IREN Limited via GlobeNewswire (filing)
Where credit enters the buildout
Basis points of financing cost where shown · 2026 financings and guaranties
Method A schematic of the observed financing paths, drawn from the disclosed structures of 2026 GPU facilities and the NVIDIA residual value guaranties. Edge weights show the announced all-in financing cost in basis points where a facility discloses one; unweighted edges are structural rather than priced.
As of 2026-09-05 · Source data
What moved
- closedobserved$6.4 billion across both packages
IREN priced the same asset class 300 basis points apart in one fiscal year
IREN raised $3.6 billion against its Microsoft contract at a blended 6.00 percent, comprising $2.1 billion of 5.96 percent senior notes due 2031 and a $1.5 billion delayed-draw term loan at SOFR plus 225, rated A by Fitch and A(low) by Morningstar DBRS. Separately it raised $2.8 billion for non-investment-grade customer deployments, including $2.4 billion led by Blue Owl and PIMCO at 9.00 percent fixed.
The Microsoft facility is secured by the GPUs and the contracted Microsoft cash flows. The 9.00 percent facility takes first-priority security in the financed equipment rather than customer contracts, but IREN unconditionally guarantees the borrower's payment obligations, so it is not a pure equipment-only risk.
Capital
- Mechanism
- A pledged, non-cancellable payment stream from a rated counterparty lets a lender underwrite the contract rather than the hardware. Fitch's analysis of the Microsoft facility leans on fixed payments, full amortisation inside the contract term, hedged power and operator replacement rights. None of those supports exist when the only pledge is equipment and a parent guarantee.
- Market consequence
- Cost of capital, not accelerator allocation, becomes the differentiator between operators holding otherwise identical hardware. At these spreads a 300 basis point gap on a multi-billion-dollar facility is worth more than most operating efficiencies an operator can engineer.
- Audience implications
- InvestorsPrice neocloud equity off weighted average cost of capital and contract quality, not fleet size. Two operators with the same accelerators and different offtakers are not the same business.
- Market analystsModels that treat GPU debt as a single asset class will misprice it. Split the stack by offtaker rating and by whether the customer contract is pledged.
Evidence and falsifiers
Sources IREN Limited via GlobeNewswire (filing) · Fitch Ratings (analyst) · U.S. Department of the Treasury (regulatory)
What would change the viewThe Mackenzie facility documents expected with IREN's September-quarter filing. If they show assigned customer receivables, materially weaker covenants, or large original issue discount and fees, then 9.00 percent is not a clean comparison to 6.00 percent and the gap narrows.
- signedobserved$105.0 billion cap; $108.5 billion of maximum gross guarantee exposure in total
NVIDIA guaranteed roughly 4.25 GW of OpenAI lease obligations, capped at $105 billion
NVIDIA entered residual value guaranties with SB Energy covering approximately 4.25 gigawatts of IT load at the Portsmouth site in Pike County, Ohio, where an OpenAI affiliate is the tenant. The aggregate payment obligation is capped at $105 billion, with credit support for roughly 3.8 further gigawatts exercisable at NVIDIA's sole discretion. Ready-for-service conditions are expected to begin in 2028.
Trigger events are OpenAI insolvency resulting in a lease default, or OpenAI failure to pay. NVIDIA's obligations terminate on the earliest of the twentieth anniversary of lease commencement, OpenAI terminating the lease, or OpenAI achieving a satisfactory credit rating.
Capital · Power · Land
- Mechanism
- The termination clause is the tell. A guarantee that expires when the tenant earns a rating is, by construction, a rental of the guarantor's balance sheet until the tenant can stand on its own. SB Energy and its project creditors get NVIDIA credit today in place of OpenAI credit that does not yet exist in rateable form.
- Market consequence
- Long-dated data centre leases can now be underwritten against a chip vendor's balance sheet. That widens the set of projects that can reach financial close, and it concentrates tail risk in a single supplier that also recognises revenue on the hardware going into those buildings.
- Audience implications
- InvestorsTrack NVIDIA's maximum gross guarantee exposure as a distinct line from product demand. It is $108.5 billion and it behaves like written credit protection, not like backlog.
- REIT leadersVendor credit support is now a competing form of tenant enhancement. Underwriting teams should ask whether a counterparty's covenant strength is its own or borrowed, and what happens on the day the support terminates.
- C-suiteIf a supplier is willing to guarantee your obligations, price that support explicitly in the negotiation. It is a financing product with a cost, not a favour.
Evidence and falsifiers
Sources NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing)
What would change the viewOpenAI obtaining a public credit rating would terminate the guaranties and convert this from a structural feature into a transitional one. NVIDIA declining to exercise the additional 3.8 GW option while comparable projects still reach close would suggest the support was never load-bearing.
- approvedreported474 GW requested; approximately 205 GW preliminarily eligible
ERCOT began verifying roughly 205 GW of large load against a 474 GW queue
ERCOT reported approximately 474 GW in its large-load queue, of which about 90 percent by its own account is data centres, and identified roughly 205 GW as preliminarily eligible for Batch Zero. The eligibility verification report is scheduled to be filed on 10 December 2026, with a commission meeting the following week.
The roughly 90 percent data-centre share describes the overall queue, not Batch Zero specifically. ERCOT's materials do not support attributing that share to the audited subset.
Power · Land
- Mechanism
- Requiring evidence of site control, financing and equipment procurement converts a free option into a costly one. Speculative and duplicate requests carry a real filing burden for the first time, which should compress apparent pipeline volume without any change in genuine demand.
- Market consequence
- Headline queue figures stop being usable as demand forecasts. Analysts who have been deriving load growth from interconnection requests will need to reset their denominators, and the reset will look like a demand shock when it is a measurement correction.
- Audience implications
- Market analystsStop treating interconnection requests as a demand series. Rebase load-growth models on verified, contracted, or energised capacity and state which of the three you are using.
- REIT leadersReadiness documentation is becoming a competitive asset. Sites with executed site control, reserved long-lead equipment and financing evidence will clear screens that undeveloped acreage will not.
Evidence and falsifiers
Sources Electric Reliability Council of Texas (regulatory) · Electric Reliability Council of Texas (regulatory)
What would change the viewA December verification outcome clearing most of the 205 GW without conditions would suggest the queue was substantially real and that readiness screening is a formality rather than a filter.
- reportedobserved116 GW combined at GE Vernova; 95 GW combined at Siemens Energy
Gas turbine order books are roughly half reservations rather than firm equipment
GE Vernova reported gas equipment backlog growing from 44 to 53 GW while slot reservation agreements grew from 56 to 63 GW, giving a 116 GW headline of which 53 GW is firm. Siemens Energy reported 69 GW of backlog alongside 26 GW of slot reservation agreements.
Siemens Energy describes the combined categories as customer commitments. Reservations remain distinct from backlog at both suppliers and should not be added into a single delivery figure.
Power · Construction
- Mechanism
- A slot reservation secures a manufacturing position, not equipment. It can lapse or be repriced. Counting reservations as backlog inflates the apparent supply of firm generation behind announced campuses.
- Market consequence
- Announced behind-the-meter generation is less contracted than headline gigawatt figures imply, which pushes real energisation dates to the right and raises the value of already-reserved equipment.
- Audience implications
- InvestorsSeparate firm backlog from slot reservations in turbine supplier models. The two convert at different rates and carry different cancellation economics.
Evidence and falsifiers
Sources GE Vernova (earnings) · Siemens Energy (earnings)
What would change the viewA sustained conversion of reservations into firm backlog at either supplier, disclosed as such, would indicate the pipeline is firmer than the current split suggests.
- reportedobserved
Local moratoria reached 533 instruments across 42 states, concentrated outside the core markets
An open dataset of primary ordinances, resolutions and board minutes tracked 533 local moratorium instruments across 42 states as of 19 August 2026, with 429 in force. Michigan leads with 64, followed by Ohio at 53 and Georgia at 47. Virginia records two and Texas three.
The maintainers describe the inventory as a floor rather than a precise census, because moratoria are frequently extended, amended or allowed to lapse without public notice.
Land · Policy
- Mechanism
- Opposition tracks arrival, not presence. Established markets have already absorbed the land-use fight and normalised the asset class; markets receiving their first large campuses are litigating it now, and doing so before local tax bases have come to depend on the revenue.
- Market consequence
- Entitlement risk is inverted relative to where capital has been rotating. The secondary markets pitched as escape valves from Northern Virginia power scarcity carry materially higher political risk per megawatt than the core they were meant to relieve.
- Audience implications
- REIT leadersUnderwrite entitlement risk by market maturity rather than by state incentive regime. A generous tax abatement does not survive a county-level moratorium.
- C-suiteSite selection in emerging markets should budget for a contested public process and a longer entitlement calendar than the core markets required five years ago.
Evidence and falsifiers
Sources Moratorium Data 2026 project (community)
What would change the viewA sustained rise in moratoria within Virginia or Texas would indicate the pattern is a lag rather than a structural difference between arriving and established markets.
- reportedforecast$42.3 billion total end-user spending
Inference is forecast to pass training in AI infrastructure spending for the first time
Gartner forecasts worldwide AI-optimised infrastructure-as-a-service spending of $42.276 billion in 2026, with inference at $23.3 billion surpassing training at $19.0 billion for the first time.
This is a forecast of global end-user spending from a single forecaster, not reported revenue, and rests on Gartner's segment definitions.
Workloads · Capital
- Mechanism
- Training demand is lumpy, project-shaped and can be deferred. Inference demand is recurring and tracks deployed application usage, so it behaves more like utility load than like capital projects.
- Market consequence
- Recurring inference load is exactly the kind of demand that supports long-dated contracts, which is what the financing market has begun to price. The workload mix shift and the credit repricing are the same phenomenon viewed from two angles.
- Audience implications
- InvestorsWeight utilisation, tokens per megawatt and serving efficiency more heavily than peak training capability when comparing operators.
Coverage audit
All eleven pillars, every issue. Thin and degraded pillars are printed rather than hidden.
| Pillar | Status | Evidence | Primary | Gaps |
|---|---|---|---|---|
| Power | met | 5 | 4 | No consistently defined median large-load interconnection study time across RTOs; generation-queue statistics are not a substitute. |
| Land | met | 2 | 1 | Moratorium counts are a floor, not a census; the maintainers note extensions and lapses often go unannounced. |
| Construction | thin | 2 | 1 | Vendor-specific lead times for chillers, CDUs, busway and generators vary too widely by geography to publish. No verified per-megawatt cost disclosure inside the window; the JLL benchmark is a forecast. |
| Workloads | thin | 2 | 0 | No rigorous independent enterprise AI return-on-investment study published inside the window. Inference and training split rests on a single forecaster. |
| Networking | met | 2 | 1 | Dell'Oro reports the crossover but not absolute back-end and front-end revenue, so the magnitude is not publishable. |
| Chips and memory | met | 4 | 4 | No supplier has disclosed completed NVIDIA HBM4 volume qualification; commercial shipment is not the same claim. Custom accelerator unit volumes are not disclosed by any hyperscaler. |
| Models | thin | 2 | 1 | Frontier training compute is no longer disclosed by leading labs; any figure would be a reconstruction. Benchmark index positions move between versions and were not independently reproduced here. |
| Distribution | met | 4 | 4 | Private coding-harness revenue figures circulating in the press are not in audited filings and are excluded. |
| Data | thin | 3 | 1 | No major data-licensing agreement with disclosed economics was signed inside the window. |
| Policy | met | 3 | 3 | Federal preemption of state AI law remains an administrative and litigation posture, not completed statute. |
| Capital | met | 9 | 8 | Original issue discount, fees and hedging costs are not disclosed for most facilities, so coupon comparisons are approximations rather than option-adjusted spreads. |
Proof of readiness replaces queue position
ERCOT is auditing roughly 205 GW of preliminarily eligible large load against a 474 GW queue. PJM qualified 715 of 811 proposals for study. MISO has filed reliability requirements aimed specifically at computational load. The through-line is that grid operators have stopped treating a queue slot as evidence of anything, and local governments have stopped treating data centers as by-right.
Sources Electric Reliability Council of Texas (regulatory) · Electric Reliability Council of Texas (regulatory) · PJM Interconnection (regulatory) · Federal Energy Regulatory Commission (regulatory) · Moratorium Data 2026 project (community)
Local moratoria concentrate outside the core markets
Local moratorium instruments · 2018 to 19 August 2026
Method Counts of local moratorium instruments per state from an open dataset built from primary ordinances, resolutions and board minutes. The maintainers describe the inventory as a floor rather than a precise census, because moratoria are frequently extended, amended or allowed to lapse without public notice. Includes instruments no longer in force.
As of 2026-08-19 · Source data
Firm backlog against slot reservations
Gigawatts · As at 30 June 2026
| Supplier | Firm Backlog | Slot Reservations |
|---|---|---|
| GE Vernova | 53 | 63 |
| Siemens Energy | 69 | 26 |
Method Figures as reported by each supplier. Slot reservation agreements secure a manufacturing position rather than equipment and are reported separately from backlog by both companies. The two categories should not be summed into a single delivery figure.
As of 2026-06-30 · Source data
Moves in this section
- signedobserved$105.0 billion cap; $108.5 billion of maximum gross guarantee exposure in total
NVIDIA guaranteed roughly 4.25 GW of OpenAI lease obligations, capped at $105 billion
NVIDIA entered residual value guaranties with SB Energy covering approximately 4.25 gigawatts of IT load at the Portsmouth site in Pike County, Ohio, where an OpenAI affiliate is the tenant. The aggregate payment obligation is capped at $105 billion, with credit support for roughly 3.8 further gigawatts exercisable at NVIDIA's sole discretion. Ready-for-service conditions are expected to begin in 2028.
Trigger events are OpenAI insolvency resulting in a lease default, or OpenAI failure to pay. NVIDIA's obligations terminate on the earliest of the twentieth anniversary of lease commencement, OpenAI terminating the lease, or OpenAI achieving a satisfactory credit rating.
Capital · Power · Land
- Mechanism
- The termination clause is the tell. A guarantee that expires when the tenant earns a rating is, by construction, a rental of the guarantor's balance sheet until the tenant can stand on its own. SB Energy and its project creditors get NVIDIA credit today in place of OpenAI credit that does not yet exist in rateable form.
- Market consequence
- Long-dated data centre leases can now be underwritten against a chip vendor's balance sheet. That widens the set of projects that can reach financial close, and it concentrates tail risk in a single supplier that also recognises revenue on the hardware going into those buildings.
- Audience implications
- InvestorsTrack NVIDIA's maximum gross guarantee exposure as a distinct line from product demand. It is $108.5 billion and it behaves like written credit protection, not like backlog.
- REIT leadersVendor credit support is now a competing form of tenant enhancement. Underwriting teams should ask whether a counterparty's covenant strength is its own or borrowed, and what happens on the day the support terminates.
- C-suiteIf a supplier is willing to guarantee your obligations, price that support explicitly in the negotiation. It is a financing product with a cost, not a favour.
Evidence and falsifiers
Sources NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing)
What would change the viewOpenAI obtaining a public credit rating would terminate the guaranties and convert this from a structural feature into a transitional one. NVIDIA declining to exercise the additional 3.8 GW option while comparable projects still reach close would suggest the support was never load-bearing.
- approvedreported474 GW requested; approximately 205 GW preliminarily eligible
ERCOT began verifying roughly 205 GW of large load against a 474 GW queue
ERCOT reported approximately 474 GW in its large-load queue, of which about 90 percent by its own account is data centres, and identified roughly 205 GW as preliminarily eligible for Batch Zero. The eligibility verification report is scheduled to be filed on 10 December 2026, with a commission meeting the following week.
The roughly 90 percent data-centre share describes the overall queue, not Batch Zero specifically. ERCOT's materials do not support attributing that share to the audited subset.
Power · Land
- Mechanism
- Requiring evidence of site control, financing and equipment procurement converts a free option into a costly one. Speculative and duplicate requests carry a real filing burden for the first time, which should compress apparent pipeline volume without any change in genuine demand.
- Market consequence
- Headline queue figures stop being usable as demand forecasts. Analysts who have been deriving load growth from interconnection requests will need to reset their denominators, and the reset will look like a demand shock when it is a measurement correction.
- Audience implications
- Market analystsStop treating interconnection requests as a demand series. Rebase load-growth models on verified, contracted, or energised capacity and state which of the three you are using.
- REIT leadersReadiness documentation is becoming a competitive asset. Sites with executed site control, reserved long-lead equipment and financing evidence will clear screens that undeveloped acreage will not.
Evidence and falsifiers
Sources Electric Reliability Council of Texas (regulatory) · Electric Reliability Council of Texas (regulatory)
What would change the viewA December verification outcome clearing most of the 205 GW without conditions would suggest the queue was substantially real and that readiness screening is a formality rather than a filter.
- reportedobserved116 GW combined at GE Vernova; 95 GW combined at Siemens Energy
Gas turbine order books are roughly half reservations rather than firm equipment
GE Vernova reported gas equipment backlog growing from 44 to 53 GW while slot reservation agreements grew from 56 to 63 GW, giving a 116 GW headline of which 53 GW is firm. Siemens Energy reported 69 GW of backlog alongside 26 GW of slot reservation agreements.
Siemens Energy describes the combined categories as customer commitments. Reservations remain distinct from backlog at both suppliers and should not be added into a single delivery figure.
Power · Construction
- Mechanism
- A slot reservation secures a manufacturing position, not equipment. It can lapse or be repriced. Counting reservations as backlog inflates the apparent supply of firm generation behind announced campuses.
- Market consequence
- Announced behind-the-meter generation is less contracted than headline gigawatt figures imply, which pushes real energisation dates to the right and raises the value of already-reserved equipment.
- Audience implications
- InvestorsSeparate firm backlog from slot reservations in turbine supplier models. The two convert at different rates and carry different cancellation economics.
Evidence and falsifiers
Sources GE Vernova (earnings) · Siemens Energy (earnings)
What would change the viewA sustained conversion of reservations into firm backlog at either supplier, disclosed as such, would indicate the pipeline is firmer than the current split suggests.
- reportedobserved
Local moratoria reached 533 instruments across 42 states, concentrated outside the core markets
An open dataset of primary ordinances, resolutions and board minutes tracked 533 local moratorium instruments across 42 states as of 19 August 2026, with 429 in force. Michigan leads with 64, followed by Ohio at 53 and Georgia at 47. Virginia records two and Texas three.
The maintainers describe the inventory as a floor rather than a precise census, because moratoria are frequently extended, amended or allowed to lapse without public notice.
Land · Policy
- Mechanism
- Opposition tracks arrival, not presence. Established markets have already absorbed the land-use fight and normalised the asset class; markets receiving their first large campuses are litigating it now, and doing so before local tax bases have come to depend on the revenue.
- Market consequence
- Entitlement risk is inverted relative to where capital has been rotating. The secondary markets pitched as escape valves from Northern Virginia power scarcity carry materially higher political risk per megawatt than the core they were meant to relieve.
- Audience implications
- REIT leadersUnderwrite entitlement risk by market maturity rather than by state incentive regime. A generous tax abatement does not survive a county-level moratorium.
- C-suiteSite selection in emerging markets should budget for a contested public process and a longer entitlement calendar than the core markets required five years ago.
Evidence and falsifiers
Sources Moratorium Data 2026 project (community)
What would change the viewA sustained rise in moratoria within Virginia or Texas would indicate the pattern is a lag rather than a structural difference between arriving and established markets.
Coverage audit
| Pillar | Status | Evidence | Primary | Gaps |
|---|---|---|---|---|
| Power | met | 5 | 4 | No consistently defined median large-load interconnection study time across RTOs; generation-queue statistics are not a substitute. |
| Land | met | 2 | 1 | Moratorium counts are a floor, not a census; the maintainers note extensions and lapses often go unannounced. |
| Construction | thin | 2 | 1 | Vendor-specific lead times for chillers, CDUs, busway and generators vary too widely by geography to publish. No verified per-megawatt cost disclosure inside the window; the JLL benchmark is a forecast. |
Supply-constrained by the vendor's own account
NVIDIA reported $89.0 billion of Data Center revenue and guided to $108.0 billion while assuming nothing from China, and described supply as constrained through fiscal 2028. Broadcom's custom silicon reached $16.7 billion in a quarter. AI back-end switching passed front-end switching for the first time. The constraint on near-term revenue is deliverability, not demand.
Sources NVIDIA Corporation (earnings) · Broadcom Inc. via PR Newswire (earnings) · Marvell Technology (earnings) · Advanced Micro Devices (earnings) · Dell'Oro Group (analyst)
Quarterly data centre and AI silicon revenue
Billions of US dollars per quarter · Most recently reported quarter for each company
Method Each company's most recently reported quarter under its own segment definitions, which are not identical. NVIDIA and AMD report Data Center segments, Broadcom reports AI semiconductor revenue, and Marvell reports a Data Center end market. Fiscal calendars differ, so quarters are not perfectly aligned.
As of 2026-09-02 · Source data
Moves in this section
- reportedforecast$42.3 billion total end-user spending
Inference is forecast to pass training in AI infrastructure spending for the first time
Gartner forecasts worldwide AI-optimised infrastructure-as-a-service spending of $42.276 billion in 2026, with inference at $23.3 billion surpassing training at $19.0 billion for the first time.
This is a forecast of global end-user spending from a single forecaster, not reported revenue, and rests on Gartner's segment definitions.
Workloads · Capital
- Mechanism
- Training demand is lumpy, project-shaped and can be deferred. Inference demand is recurring and tracks deployed application usage, so it behaves more like utility load than like capital projects.
- Market consequence
- Recurring inference load is exactly the kind of demand that supports long-dated contracts, which is what the financing market has begun to price. The workload mix shift and the credit repricing are the same phenomenon viewed from two angles.
- Audience implications
- InvestorsWeight utilisation, tokens per megawatt and serving efficiency more heavily than peak training capability when comparing operators.
Coverage audit
| Pillar | Status | Evidence | Primary | Gaps |
|---|---|---|---|---|
| Workloads | thin | 2 | 0 | No rigorous independent enterprise AI return-on-investment study published inside the window. Inference and training split rests on a single forecaster. |
| Networking | met | 2 | 1 | Dell'Oro reports the crossover but not absolute back-end and front-end revenue, so the magnitude is not publishable. |
| Chips and memory | met | 4 | 4 | No supplier has disclosed completed NVIDIA HBM4 volume qualification; commercial shipment is not the same claim. Custom accelerator unit volumes are not disclosed by any hyperscaler. |
The frontier narrows, the harness collects
Open-weight models have closed much of the capability gap while licences have quietly stopped being uniformly open. Meanwhile the layer that actually bills the customer keeps consolidating: 50 million GitHub Copilot users, more than 30 million paid Microsoft 365 Copilot seats, 950 million monthly Gemini app users. Distribution is compounding faster than model differentiation.
Sources Microsoft Corporation (earnings) · Alphabet (earnings) · Salesforce (earnings) · Z.ai via Hugging Face (primary) · Model Context Protocol project (primary)
Distribution reach at the assistant layer
Millions of users or seats · Most recently reported quarter
Method Company-reported figures that measure different things and must not be treated as equivalent. Gemini app monthly active users are free and paid consumer users, GitHub Copilot users are total users, and Microsoft 365 Copilot seats are paid commercial seats.
As of 2026-07-29 · Source data
Coverage audit
| Pillar | Status | Evidence | Primary | Gaps |
|---|---|---|---|---|
| Models | thin | 2 | 1 | Frontier training compute is no longer disclosed by leading labs; any figure would be a reconstruction. Benchmark index positions move between versions and were not independently reproduced here. |
| Distribution | met | 4 | 4 | Private coding-harness revenue figures circulating in the press are not in audited filings and are excluded. |
Inference outgrows training, and retrieval starts to rot
Gartner puts 2026 AI-optimised infrastructure-as-a-service spending at $42.3 billion, with inference passing training for the first time. At the same time, published work shows retrieval systems degrading under recursive machine-generated content, and small models answering from deliberately misleading context despite explicit instructions to abstain. Provenance is becoming a product requirement, not a compliance checkbox.
Sources Gartner (analyst) · arXiv (analyst) · arXiv (analyst) · United States District Court, via DocumentCloud (regulatory)
Inference passes training
Billions of US dollars, forecast end-user spending · Calendar 2026 forecast
Method Forecast of worldwide AI-optimised infrastructure-as-a-service end-user spending from a single forecaster, using that firm's segment definitions. These are projections rather than reported revenue, and the two segments shown sum to less than the $42.276 billion total.
As of 2026-08-10 · Source data
Coverage audit
| Pillar | Status | Evidence | Primary | Gaps |
|---|---|---|---|---|
| Data | thin | 3 | 1 | No major data-licensing agreement with disclosed economics was signed inside the window. |
The market learned to price payment certainty
Four sponsors priced six GPU-backed facilities this year across a 325 basis point range, and the ratings commentary is explicit about why: contracted, non-cancellable payments from a strong counterparty, fully amortising inside the contract term. Hyperscaler capital expenditure guidance now aggregates into the hundreds of billions, but under at least three different definitions that should not be summed casually.
Sources CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · Nebius Group (filing) · Lambda (primary) · Microsoft Corporation (earnings) · Amazon.com via Yahoo Finance (earnings) · Alphabet (earnings) · Meta Platforms (earnings)
Six GPU-backed facilities, two bands
Basis points over SOFR · Facilities closed March to August 2026
Method Floating-rate spreads as disclosed by each issuer at closing. IREN's Microsoft facility is shown as its delayed-draw term loan tranche only, because the package also contains fixed-rate notes. Spreads exclude original issue discount, arrangement fees and hedging costs, which are not disclosed for most facilities, so these are announced spreads rather than option-adjusted spreads.
As of 2026-09-05 · Source data
Moves in this section
- closedobserved$6.4 billion across both packages
IREN priced the same asset class 300 basis points apart in one fiscal year
IREN raised $3.6 billion against its Microsoft contract at a blended 6.00 percent, comprising $2.1 billion of 5.96 percent senior notes due 2031 and a $1.5 billion delayed-draw term loan at SOFR plus 225, rated A by Fitch and A(low) by Morningstar DBRS. Separately it raised $2.8 billion for non-investment-grade customer deployments, including $2.4 billion led by Blue Owl and PIMCO at 9.00 percent fixed.
The Microsoft facility is secured by the GPUs and the contracted Microsoft cash flows. The 9.00 percent facility takes first-priority security in the financed equipment rather than customer contracts, but IREN unconditionally guarantees the borrower's payment obligations, so it is not a pure equipment-only risk.
Capital
- Mechanism
- A pledged, non-cancellable payment stream from a rated counterparty lets a lender underwrite the contract rather than the hardware. Fitch's analysis of the Microsoft facility leans on fixed payments, full amortisation inside the contract term, hedged power and operator replacement rights. None of those supports exist when the only pledge is equipment and a parent guarantee.
- Market consequence
- Cost of capital, not accelerator allocation, becomes the differentiator between operators holding otherwise identical hardware. At these spreads a 300 basis point gap on a multi-billion-dollar facility is worth more than most operating efficiencies an operator can engineer.
- Audience implications
- InvestorsPrice neocloud equity off weighted average cost of capital and contract quality, not fleet size. Two operators with the same accelerators and different offtakers are not the same business.
- Market analystsModels that treat GPU debt as a single asset class will misprice it. Split the stack by offtaker rating and by whether the customer contract is pledged.
Evidence and falsifiers
Sources IREN Limited via GlobeNewswire (filing) · Fitch Ratings (analyst) · U.S. Department of the Treasury (regulatory)
What would change the viewThe Mackenzie facility documents expected with IREN's September-quarter filing. If they show assigned customer receivables, materially weaker covenants, or large original issue discount and fees, then 9.00 percent is not a clean comparison to 6.00 percent and the gap narrows.
- signedobserved$105.0 billion cap; $108.5 billion of maximum gross guarantee exposure in total
NVIDIA guaranteed roughly 4.25 GW of OpenAI lease obligations, capped at $105 billion
NVIDIA entered residual value guaranties with SB Energy covering approximately 4.25 gigawatts of IT load at the Portsmouth site in Pike County, Ohio, where an OpenAI affiliate is the tenant. The aggregate payment obligation is capped at $105 billion, with credit support for roughly 3.8 further gigawatts exercisable at NVIDIA's sole discretion. Ready-for-service conditions are expected to begin in 2028.
Trigger events are OpenAI insolvency resulting in a lease default, or OpenAI failure to pay. NVIDIA's obligations terminate on the earliest of the twentieth anniversary of lease commencement, OpenAI terminating the lease, or OpenAI achieving a satisfactory credit rating.
Capital · Power · Land
- Mechanism
- The termination clause is the tell. A guarantee that expires when the tenant earns a rating is, by construction, a rental of the guarantor's balance sheet until the tenant can stand on its own. SB Energy and its project creditors get NVIDIA credit today in place of OpenAI credit that does not yet exist in rateable form.
- Market consequence
- Long-dated data centre leases can now be underwritten against a chip vendor's balance sheet. That widens the set of projects that can reach financial close, and it concentrates tail risk in a single supplier that also recognises revenue on the hardware going into those buildings.
- Audience implications
- InvestorsTrack NVIDIA's maximum gross guarantee exposure as a distinct line from product demand. It is $108.5 billion and it behaves like written credit protection, not like backlog.
- REIT leadersVendor credit support is now a competing form of tenant enhancement. Underwriting teams should ask whether a counterparty's covenant strength is its own or borrowed, and what happens on the day the support terminates.
- C-suiteIf a supplier is willing to guarantee your obligations, price that support explicitly in the negotiation. It is a financing product with a cost, not a favour.
Evidence and falsifiers
Sources NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing)
What would change the viewOpenAI obtaining a public credit rating would terminate the guaranties and convert this from a structural feature into a transitional one. NVIDIA declining to exercise the additional 3.8 GW option while comparable projects still reach close would suggest the support was never load-bearing.
- reportedobserved
Local moratoria reached 533 instruments across 42 states, concentrated outside the core markets
An open dataset of primary ordinances, resolutions and board minutes tracked 533 local moratorium instruments across 42 states as of 19 August 2026, with 429 in force. Michigan leads with 64, followed by Ohio at 53 and Georgia at 47. Virginia records two and Texas three.
The maintainers describe the inventory as a floor rather than a precise census, because moratoria are frequently extended, amended or allowed to lapse without public notice.
Land · Policy
- Mechanism
- Opposition tracks arrival, not presence. Established markets have already absorbed the land-use fight and normalised the asset class; markets receiving their first large campuses are litigating it now, and doing so before local tax bases have come to depend on the revenue.
- Market consequence
- Entitlement risk is inverted relative to where capital has been rotating. The secondary markets pitched as escape valves from Northern Virginia power scarcity carry materially higher political risk per megawatt than the core they were meant to relieve.
- Audience implications
- REIT leadersUnderwrite entitlement risk by market maturity rather than by state incentive regime. A generous tax abatement does not survive a county-level moratorium.
- C-suiteSite selection in emerging markets should budget for a contested public process and a longer entitlement calendar than the core markets required five years ago.
Evidence and falsifiers
Sources Moratorium Data 2026 project (community)
What would change the viewA sustained rise in moratoria within Virginia or Texas would indicate the pattern is a lag rather than a structural difference between arriving and established markets.
- reportedforecast$42.3 billion total end-user spending
Inference is forecast to pass training in AI infrastructure spending for the first time
Gartner forecasts worldwide AI-optimised infrastructure-as-a-service spending of $42.276 billion in 2026, with inference at $23.3 billion surpassing training at $19.0 billion for the first time.
This is a forecast of global end-user spending from a single forecaster, not reported revenue, and rests on Gartner's segment definitions.
Workloads · Capital
- Mechanism
- Training demand is lumpy, project-shaped and can be deferred. Inference demand is recurring and tracks deployed application usage, so it behaves more like utility load than like capital projects.
- Market consequence
- Recurring inference load is exactly the kind of demand that supports long-dated contracts, which is what the financing market has begun to price. The workload mix shift and the credit repricing are the same phenomenon viewed from two angles.
- Audience implications
- InvestorsWeight utilisation, tokens per megawatt and serving efficiency more heavily than peak training capability when comparing operators.
Coverage audit
| Pillar | Status | Evidence | Primary | Gaps |
|---|---|---|---|---|
| Policy | met | 3 | 3 | Federal preemption of state AI law remains an administrative and litigation posture, not completed statute. |
| Capital | met | 9 | 8 | Original issue discount, fees and hedging costs are not disclosed for most facilities, so coupon comparisons are approximations rather than option-adjusted spreads. |
Who is underwriting whom
NVIDIA now carries $108.5 billion of maximum gross guarantee exposure alongside $366 billion of future commitments. CoreWeave carries roughly $35.1 billion of debt with three customers at 36, 26 and 10 percent of quarterly revenue. The interesting question is no longer who has the best accelerator. It is whose balance sheet absorbs the shortfall when a tenant cannot pay.
Sources NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · Equinix (earnings)
Strategy dossiers
Each inferred strategy carries the alternative reading that would explain the same moves. A coherent story is not a confirmed one.
NVIDIA NVDA
high confidence- Reported $89.0 billion of Data Center revenue for the quarter ended 26 July 2026, up 117 percent year over year at a 75.0 percent gross margin.
- Guided the October quarter to $108.0 billion plus or minus 2 percent while assuming no Data Center compute revenue from China.
- Entered residual value guaranties capped at $105.0 billion covering roughly 4.25 GW of OpenAI-tenanted capacity, with $108.5 billion of maximum gross guarantee exposure in total.
- Disclosed $366 billion of future commitments, including $279 billion of supply and capacity and $29 billion of cloud service agreements.
- Inferred strategy
- Underwrite the demand side of its own market. By making non-rated tenants financeable, NVIDIA converts a customer credit problem into a deployment schedule it can supply, keeping the buildout on a cadence that matches its overlapping Blackwell and Rubin ramps.
- Alternative explanation
- The guaranties may be narrowly opportunistic rather than strategic: a way to secure land, power and shell capacity at a specific attractive site on favourable terms, with the OpenAI tenancy an incidental structuring detail rather than the point.
- Moat
- Control of the full rack-scale platform combined with a balance sheet large enough to act as lender of last resort to its own customers. No competitor can currently offer both.
- Pressure point
- The OpenAI indemnity is least collectible in precisely the scenario that triggers the guarantee. NVIDIA's own filings acknowledge it may not recover amounts promptly or in full, which means the protection works for technical defaults and thins out against genuine insolvency.
Sources NVIDIA Corporation (earnings) · NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing)
CoreWeave CRWV
high confidence- Closed an $8.5 billion facility in March 2026 rated A3 and A(low), priced at SOFR plus 225 on the floating tranche, secured against Meta take-or-pay contracts.
- Closed a $3.1 billion publicly syndicated facility in May 2026 rated Ba2 and BB+ at SOFR plus 450, behind two large non-investment-grade customers.
- Closed a $2.6 billion facility at SOFR plus 550 behind shorter non-investment-grade contracts.
- Carried roughly $35.1 billion of net debt at 30 June 2026 with three customers at 36, 26 and 10 percent of quarterly revenue.
- Inferred strategy
- Build a repeatable financing assembly line by segmenting facilities by customer credit, so that the strongest contracts subsidise access to capital markets and the weaker ones are priced rather than refused.
- Alternative explanation
- The tiering may be a consequence rather than a design: the company may simply be financing whatever contracts it can win, with lenders imposing the segmentation from outside.
- Moat
- Demonstrated repeat access to rated, syndicated GPU-backed debt across three structures in five months, which is a capability most competitors have not shown.
- Pressure point
- Customer concentration above 70 percent across three counterparties, against roughly $6.2 billion of 2027 contractual principal repayments. Concentration is tolerable while contracts perform and becomes the whole story if one does not.
Sources CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · CoreWeave (primary) · CoreWeave, filed with the U.S. Securities and Exchange Commission (filing)
Equinix EQIX
medium confidence- Delivered $424 million of annualised gross bookings in the June quarter, up 23 percent year over year.
- Raised 2026 revenue growth guidance to approximately 11 to 12 percent.
- Inferred strategy
- Compete on diversified, interconnection-linked demand rather than on single-tenant hyperscale capacity, which keeps revenue visibility less dependent on any one frontier lab's funding cycle.
- Alternative explanation
- The bookings breadth may reflect enterprise migration cycles largely unrelated to AI, in which case it is a legacy strength rather than a deliberate hedge against concentration.
- Moat
- A diversified tenant base and interconnection density that does not reprice with any single counterparty's credit.
- Pressure point
- Development yields must clear a rising cost of capital while competing for the same constrained power, equipment and construction labour as better-capitalised single-tenant developers.
Sources Equinix (earnings)
Control points on the move
What the lender actually underwrites in a GPU financinghigh confidence
Value moves from operators who can secure allocation to operators who can secure signatures. CoreWeave's own description of its first investment-grade facility names the collateral as high-performance computing infrastructure and an associated customer contract, in that order. The hardware alone does not reach investment grade.
Sources CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · Fitch Ratings (analyst) · Kroll Bond Rating Agency (analyst)
What a grid interconnection queue position signifieshigh confidence
Value moves from early filers to prepared developers. ERCOT's verification of roughly 205 GW and PJM's qualification of 715 proposals from 811 submissions both convert queue position from an option into an obligation to demonstrate readiness.
Sources Electric Reliability Council of Texas (regulatory) · PJM Interconnection (regulatory) · Federal Energy Regulatory Commission (regulatory)
Three hops out
Every chain names its mechanism at each hop and the evidence that would break it.
An investment-grade counterparty signs a long-dated, non-cancellable offtake→The facility financing that deployment prices 225 to 325 basis points tighter
The contract itself becomes pledgeable collateral, letting rating agencies underwrite a fixed payment stream that fully amortises inside the contract term rather than a hardware residual with no observable secondary market.
Sources CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · Fitch Ratings (analyst)
Cheaper capital lowers the return threshold a site must clear→Projects that failed underwriting at wider spreads reach financial close
Data centre development is capital-intensity dominated, so the discount rate does more work than operating margin. A few hundred basis points moves the marginal project from rejected to funded.
Sources Lambda (primary) · Nebius Group (filing)
More projects reach close than the physical supply chain can serve→Competition shifts to reserved turbines, transformers and interconnection position
Financing capacity is elastic and can be manufactured through structure and guarantees. Long-lead electrical equipment and grid access are not, so the queue re-forms at the physical layer regardless of how much capital clears.
Sources GE Vernova (earnings) · Siemens Energy (earnings) · Electric Reliability Council of Texas (regulatory)
Cheap credit does not create megawatts. It moves the bottleneck downstream and raises the scarcity value of whatever remains physically constrained, which is why the same cycle can show abundant capital and lengthening delivery dates at once.
What would prove this wrong
- Equipment lead times and interconnection timelines shorten materially while financing spreads stay tight, indicating physical supply is keeping pace with capital.
- A majority of delayed or cancelled large AI projects through 2027 cite financing or offtake rather than interconnection, generation, transmission or permitting.
A chip vendor guarantees a non-rated tenant's lease obligations→The landlord and its project creditors underwrite the vendor instead of the tenant
Residual value guaranties transfer the shortfall between guaranteed minimum lease value and amounts recovered through reletting or sale onto the guarantor's balance sheet, which is genuine risk transfer from the landlord's perspective.
Sources NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing)
The guaranteed capacity is built and filled with the guarantor's own hardware→Hardware revenue is recognised against demand the guarantor itself made financeable
The tenant deploys the vendor's full-stack platform at the guaranteed site, so the same balance sheet commitment that unlocked the building also determines what goes inside it.
Sources NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing)
A trigger event requires the guarantor to pay the landlord→The guarantor's recovery depends on indemnity from the party that just defaulted
The tenant has agreed to reimburse the guarantor, but that indemnity is least collectible in exactly the insolvency scenario that triggers payment. The vendor's filings acknowledge it may not recover promptly or in full.
Sources NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing)
Vendor credit support protects the landlord well and the vendor poorly. It works for technical or temporary defaults and thins out precisely where the tail risk lives, which is the same asymmetry that made customer financing dangerous for network equipment suppliers a generation ago.
What would prove this wrong
- OpenAI obtains a satisfactory credit rating, terminating the guaranties and converting the arrangement into a transitional bridge rather than a structural feature.
- The guarantor discloses collateral, escrow or third-party reinsurance that makes recovery independent of the tenant's solvency.
What the spread will and will not tell you
The pricing evidence is strong enough to act on and weak enough to require care. Six facilities is a selected sample, structures differ in tenor and recourse, and at least one investment-grade-tenanted project still rated BB because completion risk dominated. The tripwires below are the observable thresholds that would confirm or break the read over the next four quarters.
Sources CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · Fitch Ratings (analyst) · Fitch Ratings (analyst) · Kroll Bond Rating Agency (analyst)
One issuer, one year, two costs of capital
Per cent, all-in as disclosed · Facilities signed May and August 2026
Method All-in rates as disclosed by the issuer. The Microsoft figure is the blended package rate across $2.1 billion of 5.96 per cent senior notes and a $1.5 billion delayed-draw term loan at SOFR plus 225. The non-investment-grade figure is a fixed rate. The comparison holds sponsor, asset class and vendor constant but not tenor, seniority or collateral structure.
As of 2026-08-27 · Source data
Cases
Public scenarios with named tripwires. Not recommendations, and no price targets.
GPU-backed structured credit
12 to 24 months · as of 2026-09-05 · current
bull
Contracted payment certainty becomes a standardised, ratable input, and GPU-backed debt matures into a recognised asset class with deep secondary liquidity and progressively tighter spreads for well-structured investment-grade paper.
- Catalysts
- Further publicly syndicated facilities that trade in secondary markets; Additional investment-grade ratings on contract-backed GPU structures; Rating agency criteria published specifically for accelerator-intensive assets
- Risks
- Spread compression that stops discriminating between offtaker quality; Structures that migrate risk to equity while preserving the rating
- Tripwires
- A non-investment-grade offtaker facility pricing inside SOFR plus 350 without vendor guarantees, prepayments or material equity subordination
- Investment-grade GPU facilities repeatedly clearing above SOFR plus 450 absent identifiable construction defects
Evidence
Sources CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · Kroll Bond Rating Agency (analyst)
base
The two-band structure persists. Investment-grade-contracted paper prices in the 225 to 325 range over SOFR, non-investment-grade paper in the 450 to 550 range, and the spread between them stays wide enough to determine which operators can grow.
- Catalysts
- Continued segmentation of facilities by customer credit quality; Disclosure of Mackenzie-style facility documents allowing cleaner comparison
- Risks
- Original issue discount, fees and hedging costs proving large enough to distort coupon comparisons; Structural differences in tenor and recourse explaining more of the gap than counterparty credit does
- Tripwires
- IREN's September-quarter filing revealing assigned customer receivables in the 9.00 percent facility
- A same-sponsor pair pricing within 100 basis points despite differing offtaker ratings
Evidence
Sources IREN Limited via GlobeNewswire (filing) · CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · Fitch Ratings (analyst)
bear
The spread reflects structure, tenor and sponsor risk more than counterparty credit, and the apparent pattern dissolves once fees, amortisation and recourse are controlled for. Six selected facilities were never a sample.
- Catalysts
- Disclosure showing comparable economics once original issue discount and fees are included; Investment-grade-tenanted projects rating below investment grade on completion risk
- Risks
- Over-reading a small, self-selected set of publicly announced transactions; Ignoring that announced deals are those that closed, which is a survivorship filter
- Tripwires
- A controlled study across twenty or more disclosed facilities showing an investment-grade offtake effect below 50 basis points
- Further investment-grade-tenanted projects rated below investment grade for reasons unrelated to the offtaker
Evidence
Sources Fitch Ratings (analyst) · Lambda (primary)
Long-lead electrical equipment and grid access
18 to 36 months · as of 2026-09-05 · current
bull
Firm equipment positions and verified interconnection rights appreciate as abundant capital chases a physically constrained supply of energisable capacity, rewarding suppliers and developers holding real delivery slots.
- Catalysts
- Verification processes disqualifying speculative queue positions; Continued conversion of slot reservations into firm backlog at turbine suppliers
- Risks
- Demand normalising before reserved capacity is delivered; Reservation cancellations revealing the order book was softer than reported
- Tripwires
- GE Vernova firm gas equipment backlog exceeding 65 GW as distinct from reservations
- ERCOT verification clearing materially less than half of the roughly 205 GW under review
Evidence
Sources GE Vernova (earnings) · Electric Reliability Council of Texas (regulatory)
base
Power and long-lead equipment remain the binding physical constraint through the horizon. Financing capacity expands faster than energisable capacity, so delivery dates rather than capital availability set the pace of the buildout.
- Catalysts
- Grid operators formalising readiness screening across additional markets; Rack densities continuing to rise and pulling cooling and electrical scope forward
- Risks
- Behind-the-meter generation scaling faster than expected and bypassing the grid constraint; Load flexibility commitments materially reducing firm capacity requirements
- Tripwires
- A majority of delayed or cancelled large projects citing financing rather than interconnection or permitting
- Critical equipment lead times shortening below pre-2024 norms
Evidence
Sources JLL (analyst) · Federal Energy Regulatory Commission (regulatory) · Vertiv (primary)
bear
Flexible load commitments, behind-the-meter generation and customer-funded infrastructure relieve the physical constraint faster than expected, and the scarcity premium on equipment and interconnection compresses.
- Catalysts
- Large flexibility commitments becoming standard in utility agreements; Islanded and on-site generation projects reaching commercial operation on schedule
- Risks
- Islanded projects internalising utility obligations they are not equipped to carry; Air permitting and fuel supply becoming the replacement constraint
- Tripwires
- Multiple utility agreements adopting flexibility commitments above 25 percent of contracted load
- Turbine slot reservations declining without offsetting firm backlog growth
Evidence
Sources Georgia Power (primary) · Siemens Energy (earnings)
Possible mispricings
Neocloud equity
Operators are frequently compared on fleet size and contracted backlog in aggregate, without adjusting for the credit quality of the counterparties behind that backlog.
The financing market prices those same contracts across a 325 basis point range. If debt investors distinguish sharply between offtakers and equity investors do not, one of the two is wrong about the risk.
Counter Equity may be correctly looking through to recontracting optionality and utilisation upside that debt structurally ignores, since lenders are paid par at best and care only about downside.
Sources CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · CoreWeave, filed with the U.S. Securities and Exchange Commission (filing) · CoreWeave, filed with the U.S. Securities and Exchange Commission (filing)
Announced generation capacity
Behind-the-meter generation announcements are widely reported in gigawatts without distinguishing firm equipment backlog from slot reservations.
At GE Vernova only 53 of 116 headline gigawatts is firm backlog. Treating reservations as secured equipment overstates how much announced capacity has a delivery date attached.
Counter Reservations convert at a high rate in a supply-constrained market, and holding a manufacturing slot may be nearly as valuable as an order when the alternative is a multi-year queue.
Sources GE Vernova (earnings) · Siemens Energy (earnings)
Bubble indicators
Vendor credit support as a share of demand · $108.5 billion of maximum gross guarantee exposure at a single semiconductor supplier, against $366 billion of future commitments
When a supplier underwrites its customers' obligations, reported demand stops being fully independent of the supplier's own balance sheet. Lucent carried $6.7 billion of customer credit commitments and $1.4 billion of customer debt guarantees in 2000, then recorded $2.2 billion of provisions for uncollectibles and customer financing the following fiscal year.
Offsetting The guaranties are conditional on ready-for-service milestones beginning in 2028, capped, indemnified by the tenant, and terminate if the tenant earns a rating. That is materially more disciplined than the uncapped vendor lending of the telecom cycle, and the underlying demand is currently supply-constrained rather than inventory-financed.
Sources NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission (filing) · Lucent Technologies, filed with the U.S. Securities and Exchange Commission (filing)
Gap between requested and verified capacity · 474 GW of large-load interconnection requests in one market against roughly 205 GW preliminarily eligible for verification
A queue more than twice the size of the subset that can document readiness suggests substantial duplicate, speculative or unfunded requests, which have been feeding public load-growth forecasts.
Offsetting Filing in multiple locations is rational under uncertainty and does not imply bad faith. The verification process is designed to resolve exactly this, and the correction is a measurement improvement rather than evidence of demand collapse.
Sources Electric Reliability Council of Texas (regulatory) · Electric Reliability Council of Texas (regulatory)
What we could not answer
How much of the observed spread differential survives once original issue discount, arrangement fees, hedging costs and amortisation schedules are known?
Coupon-to-benchmark comparisons are approximations, not option-adjusted spreads. If undisclosed fees are large, the headline gap between investment-grade and non-investment-grade GPU paper could narrow materially and the central read of this issue would weaken.
- Full credit agreements or offering memoranda for the compared facilities
- Disclosed original issue discount and arrangement fees
- Weighted average life and prepayment terms for each tranche
What is OpenAI's complete schedule of compute and lease obligations across all counterparties?
A $105 billion guarantee cap is only interpretable against total obligations. Individual relationships are disclosed piecemeal, so no one outside the company can currently assess how much of the buildout depends on a single unrated balance sheet.
- A consolidated obligation schedule from a primary source
- Any credit rating assigned to OpenAI or its financing entities
- Counterparty disclosure of contracted minimums by year
Does any non-investment-grade offtaker financing exist that priced at investment-grade-like spreads without vendor guarantees or prepayments?
A clean counterexample would show that structure and sponsor strength can substitute for counterparty credit, which would falsify the central claim rather than merely qualify it. None was located for this issue, and absence of evidence in a market this opaque is weak evidence of absence.
- A disclosed facility above $100 million behind a non-rated offtaker inside SOFR plus 350
- Rating agency commentary explaining how it achieved that pricing
Which memory suppliers have actually completed NVIDIA HBM4 volume qualification, as distinct from announcing commercial shipment?
All three major suppliers claim HBM4 activity, but qualification determines who captures Rubin-generation volume and at what margin. Conflating shipment with qualification would overstate competitive parity in the highest-value memory tier.
- Supplier or customer confirmation of completed qualification by part number
- Disclosed HBM4 revenue split by supplier
Provenance
- Source policy
- Public sources only. Filings, regulatory dockets, earnings releases and rating agency commentary are preferred over journalism, and journalism is preferred over analyst estimates. Every quantitative claim in this issue resolves to a numbered evidence reference. Where a figure is a forecast, an estimate, or an aggregate of non-comparable definitions, the surrounding text says so rather than presenting it as observed fact.
- Methodology
- Eleven pillars are swept every cycle. Financing comparisons use announced spreads and coupons, curve-adjusted against Treasury yields at each signing date where tenor differs materially; they are not option-adjusted spreads, because original issue discount, arrangement fees and hedging costs are undisclosed for most facilities. The central claim of this issue was drafted, then independently attacked before publication, and narrowed in response: the evidence supports contracted payment certainty as a major determinant of financing cost, and does not support the stronger claim that credit has displaced power as the binding physical constraint.
- Degraded coverage
- datacenter-construction: no verified per-megawatt cost disclosure inside the window, workloads: no rigorous independent enterprise return-on-investment study inside the window, models: frontier training compute is no longer disclosed by leading laboratories, data: no major data-licensing agreement with disclosed economics inside the window
- Generated
- 2026-09-05
Evidence index
47 references. Every claim above resolves to one of these.
- CoreWeave closes $8.5 billion delayed draw term loan facility, first investment-grade rated GPU-backed financingCoreWeave, filed with the U.S. Securities and Exchange Commission · filing · published 2026-03-31 · accessed 2026-09-05
The DDTL 4.0 Facility received ratings of A3 by Moody's and A (low) by DBRS, respectively, representing the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract. The facility includes a floating rate tranche financed at SOFR + 2.25% and a fixed rate tranche financed at approximately 5.9%.
The issuer names the customer contract as part of the collateral, which is the clearest available statement of the mechanism described in this issue. - CoreWeave closes $3.1 billion loan facility, expanding access to public markets for GPU-backed financingCoreWeave, filed with the U.S. Securities and Exchange Commission · filing · published 2026-05-18 · accessed 2026-09-05
Proceeds from the DDTL 5.0 Facility will support the purchase and deployment of infrastructure dedicated to customer contracts with two large, non-investment grade customers... Strong market interest enabled pricing to tighten by 50 basis points from initial discussions, resulting in final pricing at SOFR + 4.50%.
Rated Ba2 by Moody's and BB+ by Fitch. Same sponsor and asset class as DDTL 4.0, priced 225 basis points wider. - CoreWeave closes $2.6 billion loan facility, expanding financing flexibility for AI infrastructureCoreWeave · primary · published 2026-08-01 · accessed 2026-09-05Priced at SOFR plus 550 and rated BB+ with a recovery rating of RR2 by Fitch, behind shorter non-investment-grade customer contracts.
- CoreWeave quarterly report on Form 10-Q, debt and concentration disclosuresCoreWeave, filed with the U.S. Securities and Exchange Commission · filing · published 2026-08-11 · accessed 2026-09-05Approximately $35.068 billion of net carrying debt against $35.551 billion of gross principal. Three customers represented 36, 26 and 10 per cent of quarterly revenue. Contractual 2027 principal repayments of approximately $6.184 billion.
- IREN reports FY26 resultsIREN Limited via GlobeNewswire · filing · published 2026-08-27 · accessed 2026-09-05
$3.6bn investment grade GPU financing for Microsoft contract at 6.0%... New $2.8bn GPU financings supporting non-investment grade customer deployments. Includes $2.4bn financing led by Blue Owl and Pacific Investment Management Company LLC, as adviser to certain investors, at a 9.0% fixed rate for Mackenzie air-cooled expansion.
Also discloses $638.8 million of FY26 non-cash impairments, primarily from decommissioning Bitcoin mining hardware as sites converted to AI cloud use. - Fitch publishes IE US Hardware 3 LLC delayed draw term loan and notes ratingFitch Ratings · analyst · published 2026-06-02 · accessed 2026-09-05Rating analysis emphasises fixed contractual payments, full amortisation inside the contract term, hedged power and operator replacement rights rather than hardware residual value.
- NVIDIA current report on Form 8-K, residual value guaranties with SB EnergyNVIDIA Corporation, filed with the U.S. Securities and Exchange Commission · filing · published 2026-08-17 · accessed 2026-09-05
NVIDIA's aggregate payment obligation is cumulatively capped at $105 billion for its initial commitment under the Agreements... NVIDIA's obligations under an Agreement will terminate upon the earliest to occur of (i) the 20th anniversary of the commencement of the applicable lease, (ii) the termination of the applicable lease by OpenAI in accordance with its terms, (iii) OpenAI achieving a satisfactory credit rating.
Covers approximately 4.25 GW of IT load, with credit support for approximately 3.8 further gigawatts exercisable at NVIDIA's sole discretion. OpenAI has agreed to reimburse and indemnify NVIDIA for amounts paid. - NVIDIA quarterly report on Form 10-Q for the quarter ended 26 July 2026NVIDIA Corporation, filed with the U.S. Securities and Exchange Commission · filing · published 2026-08-26 · accessed 2026-09-05Discloses $366 billion of total future commitments including $279 billion of supply and capacity and $29 billion of cloud service agreements, plus $105.0 billion of Portsmouth guarantees and $3.5 billion of other land, power and shell guarantees. The commitment total combines unlike obligations and should not be read as a single exposure.
- NVIDIA announces financial results for the second quarter of fiscal 2027NVIDIA Corporation · earnings · published 2026-08-26 · accessed 2026-09-05
Revenue of $96.2 billion... Data Center revenue of $89.0 billion, up 117%... Revenue is expected to be $108.0 billion, plus or minus 2%. NVIDIA is not assuming any Data Center compute revenue from China.
- Nebius second quarter 2026 shareholder letterNebius Group · filing · published 2026-08-12 · accessed 2026-09-05Approximately $775 million secured facility priced at SOFR plus 250, backed by deployed GPUs and contracted cash flows from an investment-grade customer. Also reports more than $9 billion of expected 2026 customer prepayments.
- Lambda closes $926 million senior secured term loan B facilityLambda · primary · published 2026-08-27 · accessed 2026-09-05Rated Baa2 by Moody's, priced at SOFR plus 300 and issued at 99.5, fully amortising to 31 December 2030, behind a committed investment-grade offtaker.
- Fitch expects to rate SV RNO proposed senior secured notes BB(EXP)Fitch Ratings · analyst · published 2026-02-13 · accessed 2026-09-05A project fully preleased to an investment-grade tenant under a long triple-net lease still rated below investment grade because construction and completion risk dominated. The counterexample to any claim that investment-grade offtake is sufficient on its own.
- KBRA research on cross-sector credit considerations for GPU, TPU and accelerator-intensive assetsKroll Bond Rating Agency · analyst · published 2026-07-27 · accessed 2026-09-05Identifies cash-flow durability, contract structure, counterparty concentration, hardware refresh and residual value among the principal credit considerations, confirming that offtaker credit is a major but not sole determinant.
- Daily Treasury par yield curve rates, August 2026U.S. Department of the Treasury · regulatory · accessed 2026-09-05Used to curve-adjust the comparison between IREN's May and August financings, which removes the possibility that the observed gap is an artefact of benchmark rates moving between signing dates.
- Lucent Technologies annual report on Form 10-K for the fiscal year ended 30 September 2000Lucent Technologies, filed with the U.S. Securities and Exchange Commission · filing · published 2000-12-15 · accessed 2026-09-05Historical precedent for vendor-supported demand: $6.7 billion of customer credit commitments and $1.4 billion of guarantees of customer debt, followed by $2.2 billion of provisions for uncollectibles and customer financing in the next fiscal year.
- ERCOT presentation to the Texas Senate, assessing the gridElectric Reliability Council of Texas · regulatory · published 2026-07-29 · accessed 2026-09-05Approximately 474 GW in the large-load queue, of which roughly 90 per cent are data centres. The 90 per cent share describes the overall queue rather than the Batch Zero subset.
- Batch Zero verification and audit presentation to the Public Utility Commission of TexasElectric Reliability Council of Texas · regulatory · published 2026-08-20 · accessed 2026-09-05Approximately 205 GW preliminarily eligible for Batch Zero, with the eligibility verification report scheduled for 10 December 2026.
- Over 700 new generation projects accepted into first cycle of reformed interconnection processPJM Interconnection · regulatory · published 2026-08-03 · accessed 2026-09-05715 projects totalling 201.5 GW qualified for study from 811 proposals, led by 99.8 GW of natural gas. Projects that did not qualify are not described as rejected.
- MISO large load interconnection reliability requirements filing, docket ER26-3650Federal Energy Regulatory Commission · regulatory · published 2026-08-28 · accessed 2026-09-05Applies telemetry, modelling, ramp-rate and ride-through requirements to loads of at least 50 MW, with a requested effective date of 4 December 2026. Not yet accepted by the Commission.
- GE Vernova second quarter 2026 resultsGE Vernova · earnings · published 2026-07-22 · accessed 2026-09-05
Resulting in backlog growth from 44 to 53 GW and an increase in slot reservation agreements from 56 to 63 GW.
- Siemens Energy third quarter fiscal 2026 analyst presentationSiemens Energy · earnings · published 2026-08-05 · accessed 2026-09-05
High level of firm customer commitments (69 GW backlog / 26 GW SRAs).
- Contract with OpenAI approved as part of portfolio delivering $950 million in annual savingsGeorgia Power · primary · published 2026-08-26 · accessed 2026-09-053,200 MW phased from 2028 to 2032 under a 25-year agreement, with OpenAI paying full cost and committing up to 1,000 MW of flexible demand response.
- Open data on local-government moratoria targeting data centers and related infrastructureMoratorium Data 2026 project · community · published 2026-08-19 · accessed 2026-09-05
As of August 19, 2026, 533 local moratorium instruments are tracked across 42 states.
Built from primary ordinances, resolutions and board minutes. The maintainers state the figures should be treated as a floor rather than a precise census. - Broadcom announces third quarter fiscal year 2026 financial resultsBroadcom Inc. via PR Newswire · earnings · published 2026-09-02 · accessed 2026-09-05
Q3 AI semiconductor revenue of $16.7 billion grew 221% year-over-year, and 54% quarter-over-quarter.
- Marvell Technology reports second quarter of fiscal year 2027 financial resultsMarvell Technology · earnings · published 2026-08-27 · accessed 2026-09-05Total revenue of $2.7393 billion with Data Center revenue of $2.1715 billion.
- AMD reports second quarter 2026 financial resultsAdvanced Micro Devices · earnings · published 2026-08-04 · accessed 2026-09-05Data Center segment revenue of $6.718 billion, up 107 per cent year over year.
- SK hynix second quarter 2026 business resultsSK hynix · earnings · published 2026-07-29 · accessed 2026-09-05Began mass HBM4 shipments in the second quarter and reported that customer demand exceeded available supply capability.
- AI back-end network switch sales surpass front-end networks for the first time in 2Q 2026Dell'Oro Group · analyst · published 2026-09-03 · accessed 2026-09-05
2Q 2026 marked a major inflection point as switch sales in AI back-end networks surpassed those in front-end networks for the first time.
- Microsoft fiscal year 2026 fourth quarter earnings callMicrosoft Corporation · earnings · published 2026-07-29 · accessed 2026-09-05
GitHub Copilot now has 50 million users... over 30 million paid Microsoft 365 Copilot seats... Capital expenditures were $41 billion.
Also announced extending estimated useful lives of data centres and office buildings from 15 to 25 years effective fiscal 2027. - Amazon second quarter 2026 earnings call remarks on capital expenditureAmazon.com via Yahoo Finance · earnings · published 2026-07-30 · accessed 2026-09-05
We now believe we will spend approximately $220 billion in cash CapEx in 2026. The higher cost of memory pushing this number up from our prior estimate of about $200 billion.
- Alphabet second quarter 2026 earnings callAlphabet · earnings · published 2026-07-22 · accessed 2026-09-05
CapEx was $44.9 billion... updating our full year 2026 CapEx guidance range to $195-205 billion... The Gemini App now has 950 million monthly active users.
- Meta reports second quarter 2026 resultsMeta Platforms · earnings · published 2026-07-29 · accessed 2026-09-05Capital expenditures of $31.08 billion in the quarter with full-year guidance of $130 billion to $145 billion, both including principal payments on finance leases.
- Salesforce delivers record second quarter fiscal 2027 resultsSalesforce · earnings · published 2026-08-26 · accessed 2026-09-05
Agentforce ARR exceeded $1.5 billion, up over 240% Y/Y.
From this quarter Agentforce annual recurring revenue includes Slackbot and Headless 360, so the growth rate is not like for like. - Equinix reports second quarter 2026 results and raises 2026 guidanceEquinix · earnings · published 2026-07-29 · accessed 2026-09-05$424 million of annualised gross bookings, up 23 per cent year over year, with 2026 revenue growth guidance raised to approximately 11 to 12 per cent.
- Gartner forecasts worldwide AI-optimized IaaS spending to grow 96 per cent in 2026Gartner · analyst · published 2026-08-10 · accessed 2026-09-05
Inference ($23.3 billion) will surpass spending on training ($19 billion).
Total AI-optimised infrastructure-as-a-service spending forecast at $42.276 billion. - The yield imperative: turning AI infrastructure into useful intelligenceMicrosoft · primary · published 2026-09-01 · accessed 2026-09-05States that power availability is setting limits on what AI infrastructure can be built and when, as rack loads move from tens to hundreds of kilowatts.
- JLL 2026 global data center outlookJLL · analyst · accessed 2026-09-05Forecasts average 2026 shell-and-core construction cost of $11.3 million per megawatt and average critical-equipment lead times of approximately 33 weeks. Forecast figures, not reported transactions.
- Vertiv reference architecture for the NVIDIA GB300 NVL72Vertiv · primary · accessed 2026-09-05Validated architecture supporting up to 142 kW per rack using direct-to-chip liquid cooling.
- GLM-5.3 model licenceZ.ai via Hugging Face · primary · accessed 2026-09-05Requires model-as-a-service operators above a revenue threshold to pass a security review before commercial use, so downloadable weights are not equivalent to unrestricted open source.
- Quality-adjusted price measurement for AI inferencearXiv · analyst · published 2026-08-31 · accessed 2026-09-05Finds inference prices declining far faster on a quality-adjusted basis than on a matched-model basis, while cost per completed task is not clearly declining because reasoning models consume more tokens. A preprint, not peer-reviewed.
- Model Context Protocol roadmapModel Context Protocol project · primary · published 2026-08-22 · accessed 2026-09-05Prioritises proof-of-possession, workload identity federation, token exchange and enterprise-managed authorisation, indicating identity rather than connectivity is the current blocker to enterprise agent deployment.
- Recursive retrieval degradation in retrieval-augmented generation systemsarXiv · analyst · published 2026-08-22 · accessed 2026-09-05Simulation study reporting retrieval collapse in the large majority of runs as machine-generated content re-enters the corpus. A preprint using simulated rather than production corpora.
- Abstention failure under misleading retrieved contextarXiv · analyst · published 2026-08-23 · accessed 2026-09-05Finds small models answering a substantial share of questions from deliberately misleading context despite explicit instructions to abstain. A preprint.
- Order granting final approval of the Anthropic copyright class settlementUnited States District Court, via DocumentCloud · regulatory · published 2026-07-20 · accessed 2026-09-05A $1.5 billion settlement fund covering a works list of 482,460 titles. The release is past-only and does not cover claims concerning model outputs or future conduct.
- Commerce revises license review policy for semiconductors exported to ChinaU.S. Bureau of Industry and Security · regulatory · published 2026-01-13 · accessed 2026-09-05Moves specified accelerator exports to case-by-case review subject to security, customer screening and testing conditions. Permits limited revenue without restoring unrestricted access.
- Regulation (EU) 2026/1744 amending the Artificial Intelligence ActOfficial Journal of the European Union · regulatory · published 2026-07-24 · accessed 2026-09-05Moves Annex III high-risk obligations to December 2027 and product-embedded high-risk obligations to August 2028, while most transparency duties remain effective from August 2026.
- GAO report on CHIPS Act award disbursementU.S. Government Accountability Office · regulatory · published 2026-08-06 · accessed 2026-09-05Reports $13.1 billion disbursed of $31.5 billion in direct awards as of April 2026, so cash deployment materially trails announced awards.