# How Technology Roadmaps Rewrite Capital Structure
## AI Chips, Power, and Compute Financing Elasticity

**Publisher: AI Infra Credit**  
**Research type: Flagship report · Technology–finance analysis**  
**Series position: 3/3**  
**Research fact cutoff: August 30, 2026**  
**Editorial and targeted technical-source check: September 5, 2026 (not a comprehensive fact refresh)**  
**Reading time: approximately 35 minutes**  
**Evidence structure: Technical fact + disclosed transaction + structural inference + conditional scenario**  
**Publication status: Free full report · September 5, 2026 revision**  
**Research note: For general research and information only; not individualized professional advice.**  
**Keywords:** training and inference, GPU/ASIC, HBM, advanced packaging, interconnect, liquid cooling, power, cash flow available for debt service (CFADS), borrowing base, residual value, exposure at default (EAD)

> The next AI credit event may not begin with a model company filing for bankruptcy. It may begin with a substation that does not deliver power on schedule, a batch of GPUs that is nearly obsolete when it arrives, an inference contract forced into a price reset, or fixed-rate debt that does not reprice until the day no lender is willing to refinance it.
>
> That is the point at which a technology roadmap becomes a finance variable. A change in chip architecture no longer changes only compute per second. It changes when revenue starts, whether cash flow is bankable, who can buy the collateral, how much a lender will advance, and who ultimately absorbs the tail loss when the system breaks.

## Executive summary

The central claim of this report is simple: **the financing object in AI infrastructure is moving from a GPU count to a capacity–time–credit package that can be energized, accepted by a customer, and operated at qualified-workload levels within the promised service level.** A compute system becomes financeable production capacity only when chips, HBM, packaging, interconnect, software, racks, cooling, power, customer contracts, and acceptance all close at the same time.

The U.S. AI industry is building a layered capital structure. Model companies use equity, convertibles, strategic investments, and long-term cloud commitments to buy runway. Big Tech uses prepayments, leases, minimum purchase obligations, guarantees, and investment-grade credit to turn future demand into contract cash flow. GPU cloud and data-center operators place GPUs, campuses, power, and contracts into delayed-draw term loans (DDTLs), equipment loans, triple-net leases, project notes, private credit, and asset-backed securities. Chip suppliers now participate through strategic equity, warrants, residual-value guarantees, custom-ASIC orders, and financing platforms. BIS analysis of on- and off-balance-sheet AI borrowing and S&P Global Ratings’ work on GPU and data-center securitization point in the same direction: risk is not disappearing; it is being redistributed among companies, special-purpose vehicles, suppliers, customers, lenders, and insurers.[BIS: Financing the AI infrastructure boom](https://www.bis.org/publ/qtrpdf/r_qt2603u.htm) | [S&P: Equipping Data Centers Through Securitization](https://www.spglobal.com/ratings/en/regulatory/article/abs-frontiers-equipping-data-centers-through-securitization-s101645975)

Technology variables are not one-way positives for credit. Same-generation software improvements can raise output from an installed fleet. Lower prices and new products can expand task volume. A new hardware generation can simultaneously compress old-equipment lease rates and residual values. Those channels cannot be collapsed into one “efficiency factor”: **task volume, required compute, equipment-equivalent capacity, MW, and revenue are different units.** Nor does the energy efficiency of a new GPU improve an old GPU fleet at zero cost. New-hardware efficiency reaches cash flow only after new capex, delivery, installation, commissioning, and customer acceptance.

This report advances three core propositions:

1. **The asset being financed is qualified completed workload, not peak FLOPS.** Capital expenditure becomes bankable capacity only when chips, memory, interconnect, software, cooling, power, contracts, and acceptance close together.
2. **Technology progress can split cash flow from collateral value.** Same-generation software optimization can improve existing-asset CFADS. A hardware refresh requires new capital while potentially reducing old-GPU lease rates, liquidation value, and borrowing base.
3. **For most projects under construction, time-to-power is closer to a hard constraint than peak performance.** A late chip delays revenue; late power lets already-delivered chips age while interest accrues.

This report develops a technology–finance elasticity framework around five questions:

1. How do training, batch inference, real-time inference, and agent workloads create different revenue and utilization curves?
2. How do model efficiency, HBM/interconnect, chip generations, liquid cooling, and power bottlenecks flow through to time to revenue and cost per qualified task?
3. How do those variables rewrite CFADS, debt service coverage ratio (DSCR), borrowing base, residual value, refinancing capacity, and guarantee EAD?
4. Why do NVIDIA, Broadcom/AMD, hyperscalers, Neocloud operators, xAI/SpaceX, and power projects carry different capital structures?
5. Across four conditional scenarios—productivity supercycle, disciplined expansion, financing break, and policy/geopolitical restructuring—which thresholds amplify downside and which variables split operating cash from residual value?

The four scenarios are parallel conditional branches; this report assigns no probability to them. Transaction amounts remain in their original economic categories: completed financing, committed capacity, contract value, project cost, valuation, guarantee cap, and non-cash consideration are not interchangeable.

### Evidence boundary

- **[Technical fact]** means a primary paper, standards body, government source, or vendor disclosure directly supports the stated technical status. Vendor performance figures remain vendor claims, not independent cross-workload validation.
- **[Transaction fact]** means that, as of August 30, 2026, the transaction status and amount category can be checked in a company release, regulatory filing, or clearly labeled media report.
- **[Structural inference]** is this report’s interpretation of what a technical or contractual fact means for cash flow, collateral, or capital structure; it is not a conclusion stated by the source itself.
- **[Conditional scenario]** is a parameter set used to test direction and thresholds. It is not a real-world forecast and carries no subjective probability.

The September 5, 2026 work was a targeted check of key technical sources, formula definitions, and Chinese–English consistency. Transaction facts remain frozen at August 30, 2026. This revision should not be described as a comprehensive latest update on every company, transaction, or technology.

At launch, all AI Infra Credit public research, all three flagship reports, and the research newsletter are available in full at no charge.

## I. The asset being financed is qualified workload, not a chip

### 1. From peak FLOPS to bankable capacity

Traditional hardware analysis often treats peak GPU FLOPS, memory capacity, and purchase price as proxies for capacity. Lenders are not buying a specification sheet; they are underwriting future cash recovery. The more useful production chain is:

~~~text
Workload → model/algorithm → accelerator → HBM/packaging → interconnect/software
        → rack/cooling/power → qualified workload within SLO → customer cash
~~~

To prevent tasks, equipment, and utilization from being counted twice, this report uses a bridge with explicit units:

\[
A=\frac{D\times H\times a\times u\times y}{k}
\]

where:

- \(A\) is the number of completed tasks accepted by the customer during a period;
- \(D\) is standardized equipment-equivalent capacity, not MW;
- \(H\) is period hours; \(a\) is technical availability; \(u\) is the share of available device-hours occupied by productive work;
- \(y\) is useful compute per occupied device-hour, including hardware, compiler, batching, and goodput effects;
- \(k\) is useful compute consumed per qualified task. If this convention is used, failed attempts, retries, long context, and in-boundary tool calls enter here once.

If task volume \(A\) has already been inferred from realized utilization, \(u\) cannot be multiplied again. If \(y\) is already a goodput measure, goodput cannot be multiplied again. If \(k\) includes token length and retries, neither tokens nor a retry multiplier can be applied again to equipment demand. MW is also not a task unit: for a specified system power, \(MW_{IT}=D\times kW_{system}/1000\), and facility load then applies the stated PUE convention. This report converts equipment-equivalent capacity to MW only when equipment type, average power, and PUE are all specified. EPRI likewise distinguishes nominal IT capacity, facility load, and annual energy use; announced MW is not completed workload.[EPRI: Understanding Key Data Center Power Metrics](https://powering-intelligence.epri.com/understanding-metrics.html)

**[Structural inference]** If any physical or contractual link approaches zero, incremental supply elsewhere is absorbed by the bottleneck. More GPUs cannot repair a grid-interconnection delay. A completed building cannot repair an HBM shortage. A contract cannot substitute for acceptance.

### 2. Training and inference are now different capital goods

**[Technical fact]** Google’s April 2026 architecture description positions TPU 8t for large-scale pretraining and TPU 8i for sampling, serving, and reasoning. The same-day launch material said they would be available to Cloud customers “soon”; that did not establish general industry availability at that date.[Google TPU 8t/8i technical deep dive](https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive) | [Google Cloud Next ’26 infrastructure announcement](https://cloud.google.com/blog/products/compute/ai-infrastructure-at-next26)

**[Structural inference]** Frontier pretraining, continued training, reinforcement learning, synthetic data, and safety evaluation are capability-production capital: high-lag, high-variance, and dependent on model milestones. Ordinary inference, retrieval-augmented generation, tool use, video generation, and agent services are capability-consumption capital: closer to metered usage, SLOs, and customer revenue. “AI chip demand” is no longer one category, and vendor performance claims still must be discounted to the actual workload, software release, and service target.

A shared GPU pool that can move between training and multiple inference workloads generally has higher utilization, stronger substitute-customer value, and better residual value. A dedicated inference ASIC with one anchor customer looks more like a project asset backed by long-term offtake. Training contracts should not be financed solely on GPU-hours; they should be tied to model milestones, quality acceptance, downstream inference revenue, and continuing purchase capacity.

### 3. Agents turn inference demand into an option—and failure into a debt variable

The measurement unit for an agent should not stop at the token. For an observed operating period, a cleaner accounting identity is:

\[
Cost^{accepted\ task}=\frac{Model+Retrieval+Tools+FailedAttempts+Human+AllocatedCapex}{N_{accepted}}
\]

If a forward model instead uses the shortcut \(Cost_{attempt}/P(task\ accepted)\), its numerator cannot also include separately modeled retry cost. Doing both would count failure twice.

Even with a 99% per-step success rate, the independent illustration for 100 consecutive steps yields \(0.99^{100}\approx36.6\%\) end-to-end success. That is arithmetic under an independence assumption, not an empirical success rate for deployed agents. **[Structural inference]** The financial value of an agent workload depends on whether tasks are verifiable, failures can be rolled back, and human intervention can be priced—not merely on how many hours the agent runs autonomously. Higher success improves revenue quality; higher retry rates act like hidden excess consumption of power and chips and erode CFADS.

## II. Five channels through which technology roadmaps enter capital structure

### Channel 1: Model efficiency → task cost → price and rebound

**[Technical fact]** Quantization, speculative decoding, continuous batching, PagedAttention, KV-cache management, and compiler optimization can raise throughput or reduce latency for particular models, hardware, and service targets. Benchmark results from a paper are not a permanent efficiency coefficient for every production cluster.[vLLM/PagedAttention paper](https://arxiv.org/abs/2309.06180) | [Speculative Sampling paper](https://arxiv.org/abs/2302.01318)

To keep demand rebound separate from equipment demand, define the price ratio \(r_p=P_1/P_0\), the absolute price elasticity of demand \(\eta>0\), a new-task/quality task-volume multiplier \(M_q\), compute intensity per qualified task \(m_k=k_1/k_0\), and useful compute per device-hour improvement \(g_y=y_1/y_0\). Under a constant-elasticity demand assumption:

\[
\frac{A_1}{A_0}=r_p^{-\eta}\times M_q
\]

\[
\frac{D_1}{D_0}=\frac{A_1}{A_0}\times m_k\times\frac{a_0u_0}{a_1u_1}\times\frac{1}{g_y}
\]

The first equation gives **task volume**. Only the second gives **equipment-equivalent capacity demand**. If price pass-through is written as \(r_p=g_y^{-\phi}\), then, with other variables unchanged, the equipment ratio is \(g_y^{\phi\eta-1}M_qm_k\). This is a constant-elasticity working model, not a universal law. For a material price move, the report does not use the local linear approximation \(\eta\times\Delta P\).

The anti-double-counting rule is straightforward. If rebound already enters \(A_1/A_0\), it is not multiplied into \(D_1/D_0\) again. If agent steps, long context, or video workloads already enter \(m_k\), they do not also enter \(M_q\) or \(g_y\). If utilization is inferred from queue data, it cannot reappear as an independent demand multiplier. Only with those denominators fixed can an analyst decide whether efficiency reduces equipment, increases equipment, or merely changes idle time and price.

### Channel 2: HBM, advanced packaging, and interconnect → delivery and utilization

**[Technical fact]** An AI system is not the sum of advanced logic dies. HBM stacks, 2.5D/3D packaging, substrates, network interfaces, switching ASICs, optical modules, power-management ICs, and memory controllers jointly determine whether the system can go live. Micron disclosed that its 36GB 12H HBM4 designed for Vera Rubin began volume shipments in the first quarter of 2026. SK hynix disclosed that HBM4 mass shipments began in the second quarter, with a production ramp planned for the second half. Those are vendor supply-status disclosures; they do not establish that every customer, package, or complete system receives qualified supply on schedule.[Micron HBM4](https://investors.micron.com/news/press-release/2026/Micron-in-High-Volume-Production-of-HBM4-Designed-for-NVIDIA-Vera-Rubin-PCIe-Gen6-SSD-and-SOCAMM2-03-16-2026/default.aspx) | [SK hynix Q2 2026](https://news.skhynix.com/en/q2-2026-business-results/)

**[Technical fact]** UCIe, CXL 4.0, UALink 1.0, and Ultra Ethernet 1.0 specifications have been released. They address distinct layers: in-package chiplets, host–device coherent interconnect, accelerator scale-up, and Ethernet scale-out.[UCIe specifications](https://www.uciexpress.org/specifications) | [CXL 4.0](https://computeexpresslink.org/wp-content/uploads/2025/11/CXL_4.0-Specification-Release_FINAL_Website-Copy.pdf) | [UALink 1.0](https://ualinkconsortium.org/blog/ualink-200g-1-0-specification-overview-802/) | [Ultra Ethernet 1.0](https://ultraethernet.org/ultra-ethernet-consortium-uec-launches-specification-1-0-transforming-ethernet-for-ai-and-hpc-at-scale/)

**[Structural inference]** Publishing a standard does not increase loan recovery by itself. Openness may raise portability and borrowing base only after multi-vendor interoperability, drivers, spares, operating responsibility, and substitute customers reach commercial maturity. During transition, openness can instead increase software adaptation, integration responsibility, and the size of the failure domain.

### Channel 3: GPU/ASIC and generational change → residual value and borrowing base

General-purpose GPUs have mature software, a broader secondary market, and cross-workload portability. They fit equipment debt, offtake-backed loans, leases, and ABS. Their residual values are nevertheless exposed to new-architecture performance per dollar, memory capacity, interconnect, export controls, rack compatibility, and secondary-market depth. Cloud-provider ASICs can achieve lower TCO under stable, large-scale internal workloads with mature compilers and high utilization, but external resale should not be the primary recovery assumption.

A lender’s borrowing base can be expressed as the following simplified minimum; the actual definition remains governed by the loan documents, valuation policy, and eligible-asset rules:

\[
BorrowingBase_t=\min\left(AdvanceRate_t\times OLV_t,\frac{PV(Firm\ CFADS_t)}{Required\ DSCR},Eligible\ Contract\ Capacity_t\right)
\]

OLV means orderly liquidation value, not carrying value. `Eligible Contract Capacity` is a monetary amount after contract tail, concentration, cancellation rights, and customer credit are applied. **[Structural inference]** A new chip generation can improve cash flow on new assets while reducing the lease rate, OLV, and advance rate on the old GPU pool. That is a **cash-flow/collateral-value split** and can create downside nonlinearity at LTV or borrowing-base triggers. This report does not call an adverse effect “negative convexity” without defining and estimating the curvature of a value function.

### Channel 4: Power density, liquid cooling, and the grid → time-to-revenue

**[Technical fact]** NVIDIA’s DGX GB system guide gives an approximately 120kW power figure for GB200/GB300 NVL72 racks. Its August 2026 800VDC article describes a transition path from hybrid use in existing AC facilities toward native 800VDC facilities. Together they establish high-density power distribution and liquid cooling as real design constraints; they do not establish that the installed fleet already has those capabilities.[NVIDIA DGX GB hardware guide](https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html) | [NVIDIA 800VDC](https://blogs.nvidia.com/blog/800-vdc-power-architecture-ai-factory/)

Project time to revenue is determined by the slowest link:

\[
TTR=\max(T_{grid},T_{transformer},T_{civil},T_{cooling},T_{chip},T_{network})+T_{commissioning}
\]

**[Conditional scenario]** If chips arrive first and power is delayed by 12–18 months, the project incurs capitalized interest, inventory and custody cost, a rent-start mismatch, old-generation depreciation, and customer delay. This is not a factual claim about the schedule of any named campus.

**[Technical fact]** DOE’s data-center resource hub lists generation, transmission upgrades, backup resources, and large-load flexibility among the tools for expansion. EPRI’s related work likewise frames data-center load and grid coordination as a system problem.[DOE Data Center Resource Hub](https://www.energy.gov/powering-americas-ai-future-data-center-resource-hub) | [EPRI Powering Intelligence](https://powering-intelligence.epri.com/executive-summary.html)

**[Structural inference]** Power risk cannot be reduced to an average electricity price. Hourly firm power, interconnection queues, transmission, backup, load flexibility, demand charges, and curtailment rules often matter more to project NPV.

### Channel 5: Supply bottlenecks → capex and the financing clock

Deliverable capacity is not the sum of every component. It is:

\[
Q^{deliverable}=\min\left(\frac{Logic}{n_L},\frac{HBM}{n_H},Package,Substrate,Network,Rack,Power\right)
\]

When the bottleneck shifts from leading-edge logic to HBM, advanced packaging, networking, liquid cooling, transformers, or power, a 10% increase in a non-constrained component may produce almost no additional output. Yet financing may already have been drawn against total project cost and interest may already accrue against committed capacity. The result is a three-way squeeze: assets remain on the balance sheet, revenue remains in the future, and the technology’s competitive life is running down.

## III. The elasticity matrix: thresholds, divergence, and downside nonlinearity

| Technical variable | First operating transmission | Capital-structure transmission | Threshold, divergence, and downside nonlinearity |
|---|---|---|---|
| Model efficiency and quantization | Lower task cost, higher throughput | Existing-asset CFADS may rise; equipment demand depends on rebound | Task volume and equipment demand can move in opposite directions |
| Agent steps and retry rate | Changes cost per qualified task and peak inference demand | Revenue, power, and labor cost move together | Unit-task margin deteriorates quickly beyond failure or human-takeover thresholds |
| MoE and long context | Changes the mix of active FLOPs, weights, HBM, network, and KV-cache load | Chip count may not fall; memory/interconnect capex may rise | A memory or network constraint can force discrete derating |
| GPU generational refresh | Better TCO for new systems | Lower old-asset OLV, lease rate, and advance rate; higher new capex | New-asset cash flow diverges from old-asset recovery |
| ASIC share | Lower unit cost on stable workloads | Better cloud balance-sheet efficiency, weaker external residual value | High utilization supports cash flow; customer mismatch magnifies recovery loss |
| Open-interconnect maturity | More migration and substitute-customer options | Recovery, insurance acceptance, and tenor may improve | Adaptation cost arrives first; portability is realized only after maturity |
| HBM/packaging yield | Changes delivery and system throughput | WIP, prepayment, and rent-start mismatch | A small bottleneck shortfall can block acceptance of the full system |
| Power density and liquid cooling | More output per MW | Cooling, distribution, and residual-value stratification | Capex and downtime jump when a facility-retrofit threshold is crossed |
| Firm power/time-to-power | Determines available capacity and rent start | CFADS, construction debt, and DSCR move together | Downside is strongest when delay crosses liquidity or completion-test dates |
| Contract portability | Substitute workloads after customer exit | Contract tail, step-in, and recovery improve | Recovery range can narrow abruptly when substitute customers are scarce |
| Utilization and peak reservation | Revenue and energy cost | Debt capacity and refresh reserve | Tail latency and service penalties can rise nonlinearly near queue saturation |

“Downside nonlinearity” here means that the marginal change in cash flow or loss can jump after an acceptance, memory, cooling, DSCR, or refinancing threshold is crossed. Strict convexity or negative-convexity language requires a defined state variable \(x\), value function \(V(x)\), and curvature \(V''(x)\). This report does not estimate that curvature.

## IV. Rewriting capital structure: CFADS, residual value, and EAD move separately

### 1. Keep three funding gaps separate

This research separates three clocks:

- **Construction funding gap:** insufficient equity, undrawn construction debt, customer prepayment, or completion support before completion;
- **Operating debt-service gap:** CFADS and reserves cannot cover cash interest and scheduled principal;
- **Maturity refinancing gap:** maturity debt exceeds refinancing capacity calculated from CFADS, DSCR, rate, tenor, contract tail, and market availability.

A construction funding gap is not automatically a default. A passing operating DSCR does not make a balloon debt safe. **[Conditional model]** In this research program’s existing conditional model, DDTL5’s minimum forward-four-quarter DSCR is approximately 1.5525, yet a refinancing gap of about $620 million still appears in the second quarter of 2031. The model probe date is not legal maturity; the disclosed legal maturity is November 15, 2031. River Bend’s project notes are fully amortizing and have no base-case balloon, but move risk earlier into construction, acceptance, and rent start during 2027–2029. For Lake Mariner, under fixed-rate project debt, a rate shock primarily reduces maturity refinancing capacity rather than current interest expense. These are conditional model results, not forecasts of real-world default probability.

A borrowing-base decline also does not automatically create an equal refinancing gap. The fuller bridge is:

\[
RefiGap_t=\max\left(0,DebtDue_t+Fees_t-NewDebtCapacity_t-DedicatedCash_t-CommittedEquity_t\right)
\]

Debt balance depends on interim amortization, prepayment, and capitalized interest. Dedicated cash and committed equity reduce the gap. New refresh capex is a separate use of funds and must be modeled separately; it cannot be included in `DebtDue` and then added again as new-project cost.

### 2. How technology changes CFADS

Project CFADS should not be written as “capacity × lease rate.” A more realistic bridge is:

\[
CFADS=(A\times Price)-Power-Opex-Maintenance-RefreshReserve-Tax
\]

Here \(A\) is the qualified-task measure defined above; tokens, retries, and utilization are not multiplied again. Power depends on PUE, power density, firm power, gas/grid arrangements, and pass-through terms. Maintenance/RefreshReserve must include equipment refresh, liquid-cooling maintenance, network spares, and contractually required reserves.

Two kinds of efficiency must remain separate:

1. **Same-generation software efficiency.** Compilers, quantization, batching, and cache management can improve output on existing hardware, but still require engineering, validation, possible downtime, quality-regression control, and price pass-through. Only the net benefit reaches existing-asset CFADS.
2. **New-generation hardware efficiency.** A new GPU’s energy per task does not benefit an old GPU at zero cost. It enters new-asset or consolidated-project CFADS only after new capex, delivery, power/cooling work, installation, commissioning, and acceptance. Old-asset lease rates and recovery may fall before that happens.

### 3. Residual value is a path variable, not a depreciation rate

GPU liquidation value depends on workload retention, software compatibility, memory capacity, networking, rack and cooling compatibility, export controls, secondary-market depth, dismantling, and redeployment cost. A new generation is a state transition: it raises output per MW for the new system while potentially reducing lease rates and sale value for the old system. Treating “technology progress” as a uniform positive residual-value factor is one of the most dangerous assumptions for a lender.

New capex and old recovery require two accounts. Old-asset recovery is a source for repayment of old debt. New-asset capex and its borrowing base are sources and uses in the refresh financing. Unless the documents create cross-collateral, a sale-and-leaseback, or an explicit cash transfer, the two do not net automatically.

### 4. Guarantee EAD is exposure at default, not final loss

The guarantee cap is not debt or expected loss. For a shortfall guarantee that pays after a trigger, contractual exposure can be expressed as:

\[
EAD^g_t=I^{active}_t\times I^{triggered}_t\times\min\left(Cap_t,\max(0,CoveredMinimum_t-EligibleRecovery_t)\right)
\]

EAD means **exposure at default**: the contractual payment exposure that can arise when activation and trigger conditions are met. It is not ultimate risk, expected loss, or system loss. A security holder’s ultimate loss still depends on asset recovery, enforceability, the guarantor’s own credit, payment timing, and other support. The guarantor’s economic loss also depends on subrogation and subsequent recoveries.

When the same Big Tech company or chip supplier supports multiple projects, exposures must be aggregated across common demand, common GPU generations, and common power nodes. A credit guarantee has two possible channels. One merely transfers loss from the security holder to the guarantor. The other—if support is timely, enforceable, funded, and paired with effective governance—can prevent a liquidity-driven shutdown, preserve maintenance, power, and customer service, and buy time for re-leasing or equipment migration. That can raise going-concern value and reduce real economic loss. It is not a general conservation law; the analyst must test shutdown probability, recovery time, incremental operating cost, and recovery value.

## V. Live cases: chip companies are becoming credit organizers

### 1. NVIDIA–SB Energy–OpenAI: residual-value guarantees put the supplier inside the project capital stack

**[Transaction fact]** On August 17, 2026, NVIDIA and SB Energy disclosed the PORTS-Pike arrangement: approximately 4.25GW of initial IT load, an OpenAI-affiliated tenant, a 20-year lease chain, and an initial cumulative NVIDIA payment obligation capped at $105 billion. A further approximately 3.8GW option had not been exercised. Guarantees generally activate building by building after the building is ready for service and the lease begins, with the first expected activations from 2028. NVIDIA also announced a $1.5 billion investment in SB Energy. SB Energy and SoftBank plan at least 10GW of new generation and at least $4.2 billion of regional-grid investment through AEP Ohio.[NVIDIA project announcement](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Guarantees-SB-Energys-PORTS-Pike-Technology-Campus-in-Ohio-to-Exclusively-Host-NVIDIA-AI-Compute/default.aspx) | [NVIDIA 8-K](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm)

$105 billion is not a loan received on signing and is not a cash investment by NVIDIA in OpenAI. Following a trigger, NVIDIA generally pays the gap between minimum lease value and recovery from a substitute lease or asset disposal; OpenAI’s reimbursement obligation must be discounted if OpenAI itself defaults. S&P’s rating adjustment treats the arrangement as economic leverage based on building-by-building activation and assessed recovery; its public analysis shows an adjusted-debt measure rising from approximately $4.2 billion in 2028 to $37.7 billion in 2031 before declining.[S&P: NVIDIA/SB Energy rating action](https://www.spglobal.com/ratings/en/regulatory/article/-/view/type/HTML/id/3613248)

**[Structural inference]** This is supplier-led project-finance credit enhancement. NVIDIA supplies credit and possible ecosystem economics; the developer carries land, power, and construction risk; the tenant supplies future cash flow; and lenders receive residual and support layers. If support arrives at an early shortfall, it may also preserve project value by avoiding shutdown. The wrong-way risk is equally clear: if OpenAI demand, NVIDIA equipment sales, and campus re-leasing value deteriorate together, the guarantor’s earnings, collateral, and capacity to honor support weaken in the same state.

### 2. Broadcom: custom-chip collateral is narrower, and the stack is deeper

**[Transaction fact]** In June 2026, Broadcom, Apollo, and Blackstone announced an approximately $35 billion initial capital solution under an AI XPV platform supporting more than 1GW of Anthropic-related compute. Reports in August said Broadcom was discussing a larger, layered debt package of approximately $70–80 billion. As of this report’s fact cutoff, no final Broadcom SEC filing or company closing document had been found. The $70–80 billion range is therefore a reported negotiation range, not completed financing, and cannot be added to the $35 billion.[Apollo $35 billion capital solution](https://ir.apollo.com/news-events/press-releases/detail/629/apollo-leads-35-billion-capital-solution-for-broadcom-ai) | [CNBC report](https://www.cnbc.com/2026/08/21/broadcom-debt-deal-expected-to-reach-upwards-of-70-billion-sources.html)

**[Structural inference]** If the reported structure closes, Broadcom would provide a useful contrast to NVIDIA’s campus residual-value guarantee. The Broadcom model is closer to layered senior/subordinated financing against custom XPU equipment, purchase commitments, customer contracts, and project assets, potentially with supplier support. Recovery depends more heavily on continued customer use of a specialized chip, software-stack migration cost, and the subordinated layer’s loss absorption. The more efficient the custom ASIC and the more stable the customer, the stronger the operating cash-flow support. If the customer shifts to GPUs or an internally designed chip, equipment specificity can magnify residual-value loss nonlinearly.

### 3. AMD, Marvell, and Google: warrants turn purchase commitments into a cost of capital

**[Transaction fact]** AMD and Anthropic announced deployment of up to 2GW of MI450-series systems, together with up to $5 billion of future equity investment. This is a hybrid of supplier equipment, customer purchase, and milestone-based equity; future investment must not be booked as current cash.[AMD–Anthropic](https://ir.amd.com/news-events/press-releases/detail/1292/amd-and-anthropic-announce-strategic-partnership-to-deploy-up-to-2-gigawatts-of-amd-instinct-mi450-series-gpus)

Marvell issued Google up to approximately 58.97 million warrants, with an initial exercise price of approximately $206.58 per share. Some warrants vest in tranches based on qualified revenue, and the performance warrants require approximately $120 billion of cumulative qualified revenue to vest fully. Mechanical exercise proceeds are not a current investment, and warrant fair value is not financing proceeds. The economic substance is a supplier exchanging future dilution for long-term custom-chip demand, customer lock-in, and ecosystem position.[Marvell 8-K](https://investor.marvell.com/sec-filings/all-sec-filings/content/0001193125-26-356217/d412696d8k.htm) | [Marvell warrant exhibit](https://investor.marvell.com/sec-filings/all-sec-filings/content/0001193125-26-356217/d412696dex41.htm)

**[Structural inference]** These warrants are a financial form of the technology roadmap. As customers move from GPUs toward custom ASICs, purchase volume, chip revenue, supplier valuation, and dilution obligations become linked. Technical success can raise the supplier’s equity value while increasing dilution for existing holders. Technical failure may prevent vesting, but leaves the supplier carrying custom-development cost.

## VI. Hyperscalers and the power layer: separating demand credit from power credit

### 1. AWS, Google, Microsoft, Meta, and Oracle occupy different positions

**[Transaction fact]** The Amazon–Anthropic arrangement includes a current $5 billion investment, up to $20 billion of future milestone-based investment, and an Anthropic commitment to spend more than $100 billion on AWS technology over ten years. The economic loop is clear: Amazon supplies equity and Trainium/Graviton capacity; Anthropic buys AWS; Claude is distributed through AWS. The $100 billion-plus figure is future spending, not financing received by Anthropic.[Amazon announcement](https://www.aboutamazon.com/news/company-news/amazon-invests-additional-5-billion-anthropic-ai) | [Anthropic–Amazon expanded collaboration](https://www.anthropic.com/news/anthropic-amazon-compute?invite=1)

The multi-GW next-generation TPU partnership among Google, Broadcom, and Anthropic, together with Google’s TPU 8t/8i split, shows how a cloud provider can optimize chips, compilers, networking, and data centers as one system. But ASIC external residual value is weaker. The financing fit is therefore the cloud provider’s balance sheet, stable internal load, or strong offtake—not an equipment loan that depends on secondary-market resale.

The Microsoft–Constellation–DOE Crane case combines a long-term power purchase agreement, government financing, and a nuclear restart. Microsoft has signed a 20-year PPA; DOE has closed a $1 billion loan; and the 835MW unit still depends on permits and engineering, with a projected 2028 in-service date. The PPA capitalizes future power demand. The DOE loan supplies tenor. Neither is a grant, and neither eliminates restart or cost-overrun risk.[Constellation Crane](https://www.constellationenergy.com/newsroom/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html) | [DOE Crane Restart](https://www.energy.gov/edf/crane-restart)

Meta, Amazon, and Google’s investments in nuclear, advanced nuclear, and clean power make the same point: power is becoming a distinct financing layer. Power-ready rights, interconnection queues, PPAs, backup generation, and grid upgrades can attract capital before GPUs arrive, but distant MW cannot offset near-term time-to-power.

Oracle’s public results illustrate a different combination: customer-funded GPUs alongside corporate debt. For fiscal 2026, Oracle disclosed approximately $43 billion of debt, $5 billion of equity financing, and approximately $75 billion of customer prepayments or customer-supplied GPU support. The point is not to add these numbers into “Oracle financing.” It is to separate corporate debt, customer contract liabilities, asset ownership, and delivery obligations.[Oracle FY2026 results](https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Record-Q4-and-FY-2026-Results-Driven-by-Cloud-Infrastructure--Cloud-Applications/)

### 2. Neocloud: investment grade comes from structure, not a label

CoreWeave has completed an $8.5 billion investment-grade GPU-backed DDTL and a $3.1 billion publicly syndicated HPC-backed facility. Lambda priced a $926 million senior secured TLB at Baa2, SOFR+300bp, maturing at the end of 2030 with full amortization. The credit lesson is structural: investment-grade offtake, clear draw conditions, direct step-in rights, rapid amortization, and both asset and contract recovery matter more than the operator’s name.[CoreWeave $8.5 billion financing](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx) | [Lambda TLB](https://lambda.ai/blog/lambda-prices-926-million-senior-secured-term-loan-b-facility)

CoreWeave also has higher-spread DDTL structures in which customer contracts are shorter than the loan tenor. The lender is underwriting re-contracting after contract expiry, GPU migration, and future lease rates. “GPU-backed” has therefore become a credit spectrum: an investment-grade project facility, a non-investment-grade DDTL, a HoldCo mezzanine loan, and a short bridge loan are not the same asset.

### 3. A unit-consistent elasticity pass-through: why efficiency does not automatically reduce borrowing

**[Conditional scenario]** Consider a project with 100MW of energized IT load, one generation of general-purpose GPUs, paid occupancy \(u_0=65\%\), 100 qualified task units per occupied MW-year, a task price of 1, annual power and operating costs equal to 45% of revenue, and original equipment-plus-supporting capex of 100. Baseline task volume is \(100\times65\%\times100=6,500\) task units. These are research illustrations, not projections for any real project. Availability and utilization remain constant in the ratio calculations below, so 65% is not multiplied a second time.

**Step one: isolate a software improvement on the installed generation.** If compiler, quantization, and batching improvements raise useful compute per occupied device-hour by 30%, \(g_y=1.30\). The equipment-equivalent capacity needed for the same task volume is \(1/1.30=76.92\%\), a theoretical decline of 23.08%. If the project keeps the full fleet, task capacity rises. But the CFADS effect still must deduct engineering, validation, downtime, and price pass-through; software efficiency is not a zero-cost gain.

**Step two: use exact constant-elasticity rebound.** If unit price falls from 1 to 0.8, the absolute demand elasticity is \(\eta=0.8\), and there is no separate new-task multiplier, then:

\[
\frac{A_1}{A_0}=0.8^{-0.8}=1.1954
\]

Task volume rises approximately **19.54%**. `20% × 0.8 = 16%` is a local linear approximation for a small price move; this example does not use it. Dividing task volume by the 30% efficiency improvement gives:

\[
\frac{D_1}{D_0}=\frac{1.1954}{1.30}=0.9195
\]

The approximately **8.05%** decline is in equipment-equivalent capacity demand, not task volume. If agent steps, long context, and video raise independent compute intensity per qualified task by 25%, and that complexity has not already entered task volume or efficiency, then:

\[
\frac{D_1}{D_0}=\frac{1.1954\times1.25}{1.30}=1.1494
\]

Equipment-equivalent capacity demand instead rises approximately **14.94%**. Rebound, complexity, and efficiency each enter once. If observed \(k\) already includes retries or long context, the 1.25 multiplier is not applied again.

**Step three: separate old-asset operations from a new-hardware refresh.** Assume a new GPU generation improves energy per task by 40% while old-GPU lease rates fall 35%. In the no-refresh branch, the old fleet receives none of the 40% energy improvement but bears lower rent and OLV; old-asset CFADS may deteriorate. In the refresh branch, the 40% improvement enters new-asset CFADS only after new equipment capex, power/cooling work, installation, commissioning, and acceptance. Refresh downtime, financing cost, and old-equipment disposal must also be deducted. New-equipment efficiency cannot be booked as a zero-cost operating benefit for old collateral.

**Step four: state debt and funding sources before calling a refinancing gap.** To isolate the borrowing-base effect, assume the refresh date is the end of year two: original capex is 100; outstanding old debt is still 70 because the first two years are interest-only with no scheduled principal amortization; dedicated cash is 0; committed new equity is 0; transaction fees are ignored; and new debt capacity against the old assets is 50. Only under those assumptions does:

\[
RefiGap=70-50-0-0=20
\]

The 20 is an old-debt refinancing gap under a specified case. It is not an accounting identity mechanically created when the borrowing base falls from 70 to 50. If debt had amortized to 60 by the end of year two and the borrower had 5 of dedicated cash, the gap would be 5. New-generation equipment capex, its new borrowing base, and new equity must be shown separately; they are not part of the 20.

**Step five: put facility retrofit and time back into the model.** If high-density racks add 15% to cooling and distribution capex but power arrives two quarters late, revenue start is delayed, capitalized interest rises, and old equipment depreciates while idle. The binding constraint is time-to-power. Performance per dollar cannot offset two quarters of zero revenue. A lender should track five dates—equipment arrival, energization, acceptance, rent start, and amortization—not just the equipment invoice.

The point is not that a 30% efficiency improvement saves a predetermined amount of capex. At least five quantities move together: task volume, compute per task, useful output per device or MW, old-equipment OLV, and the date of first positive CFADS. Their signs can differ, and they can produce downside nonlinearity at DSCR, cooling, contract-tail, and refinancing thresholds.

## VII. xAI/SpaceX: group equity can replace stand-alone credit, but does not remove risk

**[Transaction fact]** SpaceX’s all-stock combination with xAI, the $20 billion bridge, and the $25 billion long-term bond issue show how a group can use a larger equity and debt pool to replace the financing constraint of a stand-alone model company.[SpaceX–xAI merger agreement](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/exhibit21-sx1.htm) | [SpaceX bridge-loan disclosure](https://www.sec.gov/Archives/edgar/data/1181412/000162828026040364/spaceexplorationtechnologib.htm) | [SpaceX $25 billion bonds](https://ir.spacex.com/updates/releases-details/2026/SpaceX-Announces-Pricing-of-25-Billion-Inaugural-Bond-Issuance-2026-33VwNgsx3O/default.aspx)

On August 14, 2026, SpaceX also completed its stock acquisition of Cursor, at an implied equity value of approximately $60 billion. That is not new cash financing. It transfers Cursor’s growth, employee-retention, and valuation risk to SpaceX shareholders through parent-company equity.[SpaceX–Cursor 8-K](https://www.sec.gov/Archives/edgar/data/1181412/000162828026056945/spcx-20260814.htm)

More important, SpaceX disclosed that certain Valor-related AI-hardware sale-and-leaseback transactions did not qualify as sales and were accounted for as financing. Related financing debt was approximately $13.329 billion. A transaction called a sale-and-leaseback has not necessarily sold the asset economically or removed it from the balance sheet. For an AI group, the real questions are which layer owns the hardware, which layer pays the rent, which layer has legal recourse, and which layer relies on public equity—not the headline valuation of an acquisition.

**[Structural inference]** The group structure offers more financing options and can put aerospace, satellites, cloud, models, and developer tools into one equity market. Its weakness is that common stock, common liquidity, and related-party transactions can re-aggregate risk. Group credit substitutes for stand-alone credit; it does not create additional end-customer CFADS.

## VIII. Four forward scenarios: scenarios are not probabilities

This report uses four parallel conditional scenarios—productivity supercycle, disciplined expansion, financing break and asset restructuring, and policy/geopolitical restructuring. It assigns no probability to them. The question is when a technology–finance combination switches state and which balance sheet shows stress first.

### Scenario A: Productivity supercycle

Agents, coding, science, advertising, robotics, and enterprise workflows produce measurable productivity, and end customers pay for inference. Same-generation software efficiency and new-generation hardware investment separately convert into higher paid utilization and lower unit cost; model-company gross margin grows faster than compute prices fall. Big Tech guarantees are called less often, projects move from short-term warehouse funding to operating cash and long-duration capital, and GPUs, TPUs, Trainium, AMD platforms, and custom ASICs all participate in growth.

The first assets to re-rate are operating power, standardized data centers, portable GPUs, networking/storage, and model-optimization software. The underappreciated negative is that generational refresh simultaneously improves new-asset operating efficiency and lowers old-GPU residual value. The stronger the boom, the less defensible it is to treat long-dated residual value as permanent capital. Equipment debt should amortize rapidly while utilization is high.

### Scenario B: Disciplined expansion

End-market AI revenue keeps growing but trails the most aggressive capex plans. Big Tech continues to invest, but concentrates capital in power-secure, well-contracted, multi-purpose projects. Financing remains open but sharply tiered: investment-grade offtake projects enter banks, insurers, and ABS; weaker customers face thicker equity, wider spreads, and shorter amortization.

In this scenario, the valuable asset is not the largest backlog. It is the ability to turn announced GW into energized, accepted, CFADS-producing MW on schedule. Neocloud consolidation accelerates as strong platforms absorb projects that have power or contracts but lack capital. Fixed leases, minimum purchase obligations, and guarantees are re-priced for economic leverage.

### Scenario C: Financing break and asset restructuring

End demand still grows but falls short of the capital plan. Inference prices decline faster than cost, old-GPU lease rates fall, multiple projects reach refinancing in the same window, and power or construction delays push revenue further out. Customers cut optional capacity. Neocloud equity and convertibles weaken. Private-credit warehouse lines tighten and the ABS market temporarily closes.

The usual sequence is: utilization falls → lease rates fall → convertibles stop converting → old-GPU OLV is marked down → LTV/advance-rate triggers fire → assets are sold and projects are restructured. Developers’ common equity, warrants, high-valuation private equity, and equipment debt that relies on terminal value take the first losses. Senior debt with strong customers, rapid amortization, explicit step-in rights, and real power recovers better. A timely, enforceable guarantee may also preserve going-concern recovery by preventing shutdown, but it cannot be assumed to offset common demand, common technology generations, and guarantor-credit correlation.

### Scenario D: Policy and geopolitical restructuring

Export controls, data sovereignty, energy regulation, CHIPS funding, foreign-investment review, and utility rate design change both the available technology set and the cost of capital. Government loans, equity, and milestone grants may lower the early capital cost of advanced logic, packaging, nuclear, and grid projects. Domestic sourcing, permits, data isolation, and reliability requirements also raise construction cost.

The key outcome is not simply “policy is positive” or “policy is negative.” Capital structures split by jurisdiction. Certain ASICs, GPUs, HBM supplies, and cloud capacity receive a policy premium while other equipment is discounted for export and secondary-market liquidity. Projects with domestic power, manufacturing, and government offtake attract longer-term capital more easily; cross-border, single-customer, and specialized-architecture projects face narrower recovery ranges.

## IX. Monitoring signals: connect the technology roadmap to market pricing

Quarterly updates should track more than financing headlines:

1. **Task denominator:** the definition of a qualified task, task count, and token/step/tool use per task, so token growth is not counted in both volume and complexity;
2. **Training/inference mix:** pretraining, inference-time search, agent, and video workloads as a share of total device-hours;
3. **Cost per qualified task:** model, retrieval, tool, failed-attempt, and human-takeover cost;
4. **Effective utilization:** separate training, batch-inference, and real-time-inference assumptions, with the denominator identified as installed, available, energized, or customer-reserved capacity;
5. **SLO goodput:** the gap between peak throughput and qualified work plus tail latency;
6. **Generational chip spread:** qualified work per dollar, per watt, per GB of HBM, per device-hour, and per MW across new and old systems;
7. **HBM/packaging lead time:** qualification, yield, WIP, and system delivery, not wafer capacity alone;
8. **Energized MW:** separate land, announced capacity, interconnection capacity, available IT load, facility load, and customer-accepted capacity;
9. **Liquid cooling and transformers:** whether CDU, pump, busway, switchgear, and transformer lead times exceed chip lead times;
10. **Contract tail:** minimum payment, rent start, cancellation, portability, substitute customer, and maturity;
11. **Borrowing base:** which of OLV, advance rate, contract PV, and DSCR is the minimum;
12. **Refinancing window:** comparable issuance volume, spreads, rating migration, and capital availability in 2030–2032;
13. **Guarantee and cross-exposure:** how many projects reuse the same Big Tech company, chip supplier, power node, or private-credit fund, and whether support only transfers loss or actually shortens shutdown and recovery time.

The most informative leading indicator is not announced GW. It is the number of quarters between announced capacity and first revenue. When time to revenue stretches, projects often show a liquidity gap first and DSCR, residual-value, and exposure-at-default deterioration later.

## X. Conclusion: technology becomes finance only after it passes through cash flow

AI capital structure is being rewritten in three ways.

First, **the asset unit is moving from a single card to an integrated system.** HBM, packaging, interconnect, networking, racks, liquid cooling, and power are complementary. The old assumption that a GPU can be financed in isolation is breaking down. Lenders care whether the complete system can be seized, migrated, and re-leased.

Second, **the cash-flow unit is moving from the token to the qualified task.** Model efficiency, agent retries, SLOs, customer acceptance, and human intervention determine revenue quality. More tasks do not imply the same percentage increase in equipment demand; lower equipment demand does not imply lower revenue. CFADS becomes interpretable only when tasks, compute per task, useful throughput, utilization, and price sit on one unit-consistent bridge.

Third, **the credit unit is moving from the company to a network of common nodes.** NVIDIA’s equipment, equity, and residual support; Broadcom and AMD’s custom chips and warrants; cloud providers’ purchases and guarantees; Neocloud project debt; and developers’ PPAs and grid capital may all depend on a small set of model companies and financing windows. Legal SPV isolation does not automatically produce economic independence. A credit guarantee can transfer loss or preserve value by sustaining operations; those two effects must be tested separately.

The right diligence questions are therefore:

> When do the chips arrive? When does power arrive? When does the customer accept the system? Is the task qualified? What does same-generation software improvement cost? Who funds the new hardware? Does CFADS cover debt and refresh? Is debt amortized within the GPU’s competitive life? Who can buy the old collateral after a technology change? Who takes the first loss, and who has the capital and control to step in?

If the answers are verified power, portable systems, minimum cash obligations, rapid amortization, independent first-loss capital, real end-customer revenue, and separately funded new capex and old-asset recovery, GPU loans, PPAs, project debt, ABS, warrants, and supplier guarantees can form an efficient industrial-capital machine.

If the answers are only a large contract, another financing round, implicit Big Tech support, and a rising valuation, the technology roadmap is not reducing risk. It is distributing future risk across more balance sheets that have not yet seen the same exposure at the same time.

By 2030, the winners will not necessarily be the companies with the most announced capital, GPUs, or planned MW. They will be the platforms that close model efficiency, chip generations, interconnect, cooling, power, contracts, and capital tenor on one timeline. **Technology determines what a system can do. Cash flow determines how much the system can borrow. Capital structure determines whether the system survives its next refresh cycle.**

## Continue the series

- **The upper-level framework:** [The AI Capital Stack: When Compute Becomes a Credit System](/en/research/ai-capital-stack)—how technology variables enter the broader capital-formation system.
- **Transaction deep dive:** [How Contracts Become Credit: CoreWeave, IREN, and the Neocloud Financing Machine](/en/research/how-contracts-become-credit)—how technology residual value enters the borrowing base and loss waterfall.
- **This report:** How Technology Roadmaps Rewrite Capital Structure—connecting qualified workload, time-to-power, CFADS, residual value, and exposure at default.

> Keep following the research: subscribe to the free AI Infra Credit research newsletter or submit a company, project, or capital structure for future coverage. All public launch research is available in full at no charge.

## Primary sources and further reading

- [BIS: Financing the AI infrastructure boom](https://www.bis.org/publ/qtrpdf/r_qt2603u.htm)
- [IMF: Global Financial Stability Report, April 2026](https://www.imf.org/-/media/files/publications/gfsr/2026/april/english/text.pdf)
- [Google: TPU 8t/8i technical deep dive](https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive)
- [vLLM: Efficient Memory Management for LLM Serving with PagedAttention](https://arxiv.org/abs/2309.06180)
- [DeepMind: Accelerating Large Language Model Decoding with Speculative Sampling](https://arxiv.org/abs/2302.01318)
- [Micron: HBM4 high-volume production](https://investors.micron.com/news/press-release/2026/Micron-in-High-Volume-Production-of-HBM4-Designed-for-NVIDIA-Vera-Rubin-PCIe-Gen6-SSD-and-SOCAMM2-03-16-2026/default.aspx)
- [SK hynix: Q2 2026 HBM4 update](https://news.skhynix.com/en/q2-2026-business-results/)
- [NVIDIA: DGX GB rack-scale hardware guide](https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html)
- [NVIDIA: 800VDC power architecture](https://blogs.nvidia.com/blog/800-vdc-power-architecture-ai-factory/)
- [NVIDIA: Vera Rubin full production](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Vera-Rubin-Ramps-Into-Full-Production-to-Power-Agentic-AI-Factories-Worldwide/default.aspx)
- [AMD: MI400/Helios full-stack compute](https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era)
- [TSMC: 2026 Technology Symposium](https://pr.tsmc.com/system/files/newspdf/attachment/49337b40ff139d51d533076cf7a945b30e107e07/2026%20Tech%20Symposium%20%28E%29_Final_wmn.pdf)
- [UCIe specifications](https://www.uciexpress.org/specifications)
- [CXL 4.0 specification](https://computeexpresslink.org/wp-content/uploads/2025/11/CXL_4.0-Specification-Release_FINAL_Website-Copy.pdf)
- [UALink 1.0 specification overview](https://ualinkconsortium.org/blog/ualink-200g-1-0-specification-overview-802/)
- [Ultra Ethernet Specification 1.0](https://ultraethernet.org/ultra-ethernet-consortium-uec-launches-specification-1-0-transforming-ethernet-for-ai-and-hpc-at-scale/)
- [DOE: Data Center Resource Hub](https://www.energy.gov/powering-americas-ai-future-data-center-resource-hub)
- [EPRI: Powering Intelligence](https://powering-intelligence.epri.com/executive-summary.html)
- [EPRI: Understanding Key Data Center Power Metrics](https://powering-intelligence.epri.com/understanding-metrics.html)
- [CoreWeave: $8.5 billion investment-grade GPU-backed financing](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx)
- [DataBank: $1.1 billion hyperscale asset securitization](https://www.databank.com/resources/press-releases/databank-raises-1-1-billion-in-hyperscale-asset-securitization/)
- [IREN: $3.65 billion investment-grade GPU financing](https://irisenergy.gcs-web.com/news-releases/news-release-details/iren-closes-365bn-investment-grade-gpu-financing)
- [Amazon: additional Anthropic investment](https://www.aboutamazon.com/news/company-news/amazon-invests-additional-5-billion-anthropic-ai)
- [Anthropic: Google/Broadcom compute partnership](https://www.anthropic.com/news/google-broadcom-partnership-compute?gsid=44d84525-2cee-4a37-a6d7-ff2c2aa95f78)
- [Meta: nuclear energy projects](https://about.fb.com/news/2026/01/meta-nuclear-energy-projects-power-american-ai-leadership/)
- [SpaceX: xAI merger agreement](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/exhibit21-sx1.htm)
- [SpaceX: Cursor acquisition 8-K](https://www.sec.gov/Archives/edgar/data/1181412/000162828026056945/spcx-20260814.htm)
- [CFTC: request for comment on compute derivatives](https://www.cftc.gov/PressRoom/PressReleases/9286-26)

**Reading note:** The scenarios, elasticities, and loss figures in this report are conditional analysis. Reported media figures, undrawn commitments, future options, total contract value, project cost, guarantee caps, valuations, and non-cash consideration are not treated as funded financing. The September 5, 2026 check covers the key technical sources and formulas identified in this report; it is not a comprehensive market refresh.