Neocloud Capital Raise (2026): SPV Off-Take Structures and the GPU-Backed Capital Stack

Neocloud Capital Raise (2026): SPV Off-Take Structures and the GPU-Backed Capital Stack

Last updated: August 2026

I'm Sean Yu, co-founder of Peony, and I spend most of my days around the deal teams and lenders who finance hard assets. GPU clusters are the strangest asset I watch get financed, and the reason is a single tension that sits underneath every structure choice. When a neocloud raises against its fleet, it is asking a lender to lend against something that starts losing value the day it is racked, on a schedule nobody can agree on. The whole art of the raise is arranging the capital so that the debt gets repaid before the collateral decays and before the customer contract runs out. Get that arrangement right and a 40-person operator raises hundreds of millions on terms that used to be reserved for utilities. Get it wrong and the facility that funded your growth becomes the thing that ends you.

Quick answer: A neocloud capital raise is how a GPU-cloud operator funds its fleet and the capacity that runs it, climbing a ladder from equity to venture debt to a GPU-backed private-credit facility to an investment-grade or ABS takeout. Every layer is governed by one constraint I'll call the amortization-depreciation race: a GPU-backed facility only works if the debt amortizes faster than the collateral loses value and the off-take outlives the loan. Every structure choice — SPV ring-fence, advance rate, tenor, cash sweep, off-take coverage — is downstream of that race. The chips are the collateral on paper; the contracted revenue is the collateral that actually gets underwritten, because GPUs depreciate on a contested curve (four-to-six years per operators, two-to-three per skeptics) while a creditworthy customer contract does not. This is the RAISE — capital-stack design, SPV mechanics, off-take-backed debt sizing, the lender universe, the process, and what kills these deals. The room a lender reads it in is the companion GPU cluster financing data room guide.

Where this post sits. This is the deliberate sibling of the GPU cluster financing data room guide, and the two split the job cleanly. That post is the room — the folder tree a credit committee reads, staged disclosure, redaction, watermarking, visitor groups, and the lender Q&A workflow. This post is the raise — how you design the capital stack, why the SPV exists, how off-take sizes the debt, who the lenders are, how the process runs, and what goes wrong. Wherever this guide brushes against room mechanics, it hands off with a link rather than repeating; the GPU cluster financing data room guide covers the room itself, and this one covers the transaction the room exists to serve.

What is the amortization-depreciation race, and why does it govern everything?

Because a GPU-backed loan is a bet that two clocks run in a particular order, and every term in the facility exists to enforce that order. The first clock is amortization: how fast the debt gets paid down. The second is depreciation: how fast the GPUs lose value. A GPU-backed facility only works if the first clock beats the second — if the loan amortizes faster than the collateral decays — and if the customer contract outlives the loan so the cash to pay it down is actually there. Name that race and every otherwise-arbitrary structure choice becomes legible: the SPV ring-fence protects the collateral pool so the race is run on a clean track; the advance rate sets how much debt you start with relative to the asset; the tenor decides how long you have to win; the cash sweep accelerates amortization when cash is available; and the off-take coverage is what guarantees the contract outlasts the debt.

The reason the race is even close — the reason this is hard and a building is easy — is that GPUs depreciate on a genuinely contested curve. Operators and Nvidia point to four-to-six-year depreciable lives. Skeptics argue real economic life is closer to two to three years given Nvidia's roughly annual cadence; in November 2025 the short-seller Michael Burry made the loudest version of this case, accusing hyperscalers of understating depreciation by extending useful lives — "one of the more common frauds of the modern era," he wrote — and estimating the industry would understate depreciation by about $176 billion from 2026 through 2028. Under GAAP there is no single correct number: useful life is a company-specific estimate, and a change to it is treated as a change in accounting estimate under ASC 250, applied prospectively rather than as an error correction. So the depreciation clock is not just fast; it is uncertain, and a lender cannot assume it away.

That uncertainty is why the winning move in the race is never to argue the chips will last longer. It is to make the amortization clock so much faster that it wins even on the pessimistic depreciation case. A facility structured to repay over a three-year customer contract survives a two-to-three-year useful life; a facility structured to repay over the hardware's optimistic six-year tail does not. Hold this frame through the rest of the post: every good structure choice speeds up the amortization clock or protects the collateral the depreciation clock is eating, and the single most important lever on both is the off-take contract that carries the credit.

How does the neocloud capital stack actually work, layer by layer?

The capital stack is a ladder that climbs as the business de-risks, and each layer prices off a different thing relative to the depreciation curve. Early on you have no contracted revenue and nothing to secure, so you raise equity; as you sign customers and build an asset base, cheaper and more senior capital becomes available, culminating — for the largest names — in an investment-grade or securitized takeout that would have been unimaginable for the asset class three years ago. The critical insight is that you do not choose one layer; you ascend them, and where you can reach on the ladder is set by the strength of your off-take, not the size of your fleet.

Here is the stack, from the equity floor to the investment-grade ceiling:

The ladder is not theoretical; the market has now demonstrated every rung. Lambda's first GPU-backed facility was a US$500 million special-purpose GPU financing vehicle led by Macquarie in April 2024, described as a first-of-its-kind GPU-backed structure placed into a dedicated SPV; by May 2026 Lambda had climbed to a $1 billion syndicated senior secured facility led by J.P. Morgan, upsized from a $275 million facility established in August 2025. At the top of the ladder, CoreWeave closed the first investment-grade-rated GPU-backed financing — an $8.5 billion delayed-draw term loan rated A3 by Moody's and A (low) by DBRS, closed March 31, 2026, with the floating tranche priced at SOFR plus 2.25% and maturing March 2032. The lesson for your own raise is that the stack is a progression: you enter where your off-take lets you, and each contract you sign moves you up a rung toward cheaper, more senior capital.

Why do lenders demand a bankruptcy-remote SPV?

Because a lender lending against a fast-depreciating asset needs the collateral legally walled off from the operating business, so that if the operating company fails, the chips and the contracts are still there and still reachable. This is the structural spine of a GPU-backed raise, and it is not optional above the smallest deals. The mechanics have three moving parts, and each one is a piece of the ring-fence.

True sale of the GPUs. The neocloud sells a defined pool of GPUs into a bankruptcy-remote special-purpose entity — a true sale, transferring ownership, not merely a pledge from the operating company. The distinction matters because a true sale is what keeps the chips out of the operating company's bankruptcy estate. The SPV owns the fleet; the lender lends to the SPV.

Assignment of the off-take contracts. The associated customer contracts and their receivables are assigned into the same entity, so the revenue that repays the loan flows to the SPV, not through the operating company where other creditors could reach it. This is the step that makes the off-take actually collateral rather than just a nice story — the cash is legally directed to the entity the lender has security over.

The security package. The lender takes a first-priority security interest over both the chips and the assigned contracts and receivables, plus account control agreements over the SPV's accounts. The convention is one-cluster-one-SPV: each facility gets its own bankruptcy-remote entity so a default on one does not cross-default into another, and each lender has a bounded, clean collateral pool.

The landmark deals make the pattern unmistakable. CoreWeave's facilities are secured on dedicated single-purpose entities — the publicly disclosed CoreWeave Compute Acquisition Co. VII and VIII, LLC borrowers — into which a defined pool of GPUs and the associated customer contract are ring-fenced. The Meta-Blue Owl Hyperion structure runs the same logic at hyperscale, using an SPV to keep tens of billions of debt off the parent's balance sheet. Your raise applies the identical mechanic at your scale: which chips and which contracts sit inside the entity, that it is separate and not commingled with operating cash, and that the lender's security attaches to assets the SPV actually owns. That is the capital-raise reason the SPV exists. Proving the ring-fence to a credit committee — the formation documents, the non-consolidation opinion, the org chart that shows where the entity sits — is a room job, and the GPU cluster financing data room guide covers exactly how to evidence it.

How does off-take coverage size the debt?

Contracted revenue sizes and repays the debt, and the hardware is the fallback — so the first number in the raise is not the value of your chips but the strength of your off-take. This is the single most important thing to internalize about a GPU-backed facility, and it is the direct consequence of the amortization-depreciation race: because the collateral is decaying on a contested curve, a lender cannot rely on the chips, so they lend against the contracted cash flow the chips produce and underwrite your customer's credit, not your enthusiasm.

Four things about the off-take determine how much debt it supports.

Contracted versus merchant revenue. Term debt is sized against contracted revenue — a signed, long-term customer commitment — not merchant (spot) revenue, because spot revenue can evaporate. The evaporation is not hypothetical: H100 rental rates fell from roughly $8/hour at their 2023-24 peak to the $2.85-3.50/hour range by late 2025 as supply caught up with demand. A facility modeled on spot rates that then halve is a facility in breach. This is why lenders discount or ignore merchant revenue in sizing and lean almost entirely on the contracted book.

Counterparty credit quality of the off-taker. The lender is effectively underwriting your customer's ability to pay. A contract with an investment-grade counterparty is worth far more, and supports far more debt, than the same contract with a shaky one — which is why the strongest deals lead with named, creditworthy off-takers. Analysts describe the rating agencies on the CoreWeave deals as underwriting the hyperscaler customer, not the GPU hardware, with long-term take-or-pay agreements from a Meta or a Microsoft functioning as near-sovereign credit support.

Tenor matching — the off-take must outlive the loan. This is the coverage rule the race demands: the debt has to amortize inside the contracted-revenue window, so the loan matures in step with the contract and the fleet's useful life rather than relying on the hardware's optimistic tail. It is why the investment-grade deals mature five to six years out — CoreWeave's A3 facility matures in 2032 — instead of the ten-to-twenty-year horizon a building supports. If your off-take runs three years, your term debt should be structured to repay inside three years.

Coverage as lenders frame it. Lenders size the facility to a cushion over the contracted cash flows, so that the contract more than covers debt service and the deployment cost. The cleanest public statement of coverage is Nebius's $775 million senior secured facility, backed by deployed GPUs and an investment-grade customer's contracted cash flows and structured so that, in Nebius's own words, "together with cash flows under the customer agreement, the facility covers more than 100% of the capital expenditure required to deploy the underlying GPU infrastructure." Nebius framed the raise around more than $40 billion of additional contracted revenue from investment-grade customers such as Microsoft and Meta — the contracted book is the thing that sizes the debt.

A word on advance rates, because the question — "what advance rate can I get on my H100 fleet?" — comes up constantly and is really a covenant-quality question wearing a pricing costume. The advance rate (the share of appraised fleet value a lender will lend against) is set deal by deal, driven by off-take durability, counterparty credit, the useful-life assumption, and the residual-value view — and it is not reliably public. Rather than chase a number that circulates without a source, make the inputs legible and let the lender size it: a stronger contract and a more defensible depreciation case support a higher advance. The advance follows the evidence, not the other way around.

What is the 2025-26 market read a neocloud raises into?

The market read is that GPU-backed debt de-risked in public, ratable steps — from a private-credit experiment to an investment-grade, securitizable asset class in under three years — and your raise is benchmarked against that precedent ladder. This is liberating and demanding at once: the template exists, so you are not inventing the argument, but your file gets compared to the deals that set it. The ladder is worth knowing because it is the narrative your raise is measured against.

• CoreWeave is the reference borrower. Its debt escalated in public steps to the $8.5 billion investment-grade facility (Moody's A3, DBRS A low), closed March 31, 2026 at SOFR plus 2.25%, secured on CoreWeave Compute Acquisition Co. VIII, LLC plus a customer contract. Before that first investment-grade deal, this risk was the domain of private credit accepting risk premiums reported in the 10-15% range — the spread compression from there to SOFR plus 2.25% is the asset class de-risking in real time.
• The SPV structure went hyperscale. In October 2025, Meta announced a joint venture with funds managed by Blue Owl Capital to develop the Hyperion data center, an SPV structure arranged by Morgan Stanley carrying roughly $27 billion in total development costs plus $2.5 billion in equity, with Blue Owl funds owning 80% and Meta 20%, and PIMCO anchoring the debt. It marked an inflection in Big Tech's use of SPVs to fund AI infrastructure with debt that stays off the parent's balance sheet.
• The buildout context is enormous. Crusoe secured $11.6 billion in debt and equity in May 2025 for the Abilene, Texas campus anchored to OpenAI's Stargate program, designed around Nvidia GB200 (Blackwell) systems. Morgan Stanley has estimated that between 2025 and 2028, roughly $800 billion of private-credit capital will be required to finance AI data centers, power, and fiber.
• Securitization arrived. DataBank raised $1.1 billion in September 2025 in the industry's first data-center ABS dual-rated by S&P and Moody's — its first hyperscale master trust, taking its securitized portfolio to $3.23 billion of investment-grade bonds — and rating agencies formalized their approach as KBRA published a Data Center ABS Global Rating Methodology on January 9, 2026, applying multiple rounds of tenant-default and revenue-haircut stresses in its scenarios.

Read the ladder and the practical consequence for your raise is that the categories now exist. A lender can slot a GPU-backed SPV facility into a shape they understand, which turns the question from "can this even be underwritten?" into "does this look like the deals that already cleared?" — a far easier question to pass if your structure is built to the precedent.

Who actually lends against GPUs — the lender universe?

Four lender types finance GPUs, and they line up roughly by the maturity of the borrower and the strength of the off-take. Knowing which one fits your raise is half the work of running it, because approaching the wrong lender universe for your stage wastes the scarcest thing you have — time on the clock.

Private-credit and infrastructure-credit funds are the workhorse for the $20M-$500M SPV facility this post is written for. They underwrite the ring-fenced, off-take-backed structure directly, move faster than a syndicated bank process, and are comfortable with the asset class — Blackstone, PIMCO, BlackRock, Carlyle, JP Morgan, and Macquarie are all active lenders to GPU-backed structures, and Fluidstack has secured up to $10 billion in borrowing capacity from Macquarie and other lenders. This is the same fund universe that underwrites the broader direct-lending market; the fund-side architecture behind these facilities is covered in the private credit data room guide, and the recurring-reporting cadence a credit fund will expect from you post-close is in the borrower reporting data room guide.

Banks arrange and syndicate the larger facilities. Once a deal is big enough to syndicate and to pursue a rating, a bank-arranged structure is cheaper and deeper — CoreWeave's $8.5 billion investment-grade loan was arranged by Morgan Stanley, MUFG, Goldman Sachs, and JPMorgan. This is the altitude where procurement expects an enterprise process, and where the raise starts to look like a capital-markets transaction rather than a bilateral loan.

Data-center and GPU ABS is the securitized takeout for the largest, most contracted books. DataBank's $1.1 billion hyperscale securitization in September 2025 — the industry's first data-center ABS dual-rated by S&P and Moody's, its fifth securitization since 2021, taking the securitized portfolio to $3.23 billion of investment-grade bonds — is the template: a proven operator recycling capital against a seasoned, diversified contract book. This is not a first-raise instrument; it is where the ladder ends once the book is big and boring enough to securitize.

Vendor financing is the newest and most debated layer. Nvidia has moved from chip supplier to financier, agreeing to backstop up to $105 billion for an OpenAI data center in Ohio in August 2026 — a figure cut from a reported $250 billion after investors worried about the chipmaker's own risk exposure. Vendor backstops can guarantee a demand floor beneath a customer's collateral, but they sit at the center of the circularity debate covered below. For most neoclouds, vendor financing is context, not a layer you can count on.

How does the raise process run, from teaser to close?

The raise runs teaser to term sheet to confirmatory diligence to documentation to close, and the pace is set by how legible your off-take and SPV structure are to a credit committee. The workstreams here are the deal process, not the room mechanics — where this touches folder structure, staged disclosure, or lender Q&A, the GPU cluster financing data room guide is the reference and this post does not repeat it.

The sequence, and what actually gates each step:

1. Teaser and lender outreach. You (or an advisor) circulate a short teaser that frames the fleet, the contracted off-take, and the proposed structure — without exposing customer names, contract pricing, or the SPV specifics. This is the document that gets a fund to engage, and its whole job is to make the off-take and the structure legible in a page.
2. NDA and room access. Interested lenders sign an NDA and enter the data room, where the off-take contracts stay staged behind redaction until a counterparty is genuinely progressing. Because your customers are frequently your lenders' competitors or their other borrowers, the named off-take contract is your most sensitive document — the staging mechanics are a room job.
3. Term sheet. The lead lender issues a term sheet sizing the facility and setting the advance rate, spread, tenor, covenants, and cash-sweep triggers. This is where the amortization-depreciation race gets priced: the tenor and amortization schedule are negotiated against your off-take term and useful-life case.
4. Confirmatory diligence. The lender and its technical advisors work the fleet file, the off-take contracts, the SPV formation, the power and colo arrangements, and the model. A lender ties draws to chip-delivery milestones, because the collateral does not exist until the GPUs are delivered and racked.
5. Documentation and close. Legal papers the credit agreement, the security package (the true-sale documents, the assignment of contracts, account control agreements), and any intercreditor terms; then close and initial draw.

A focused SPV raise with a single strong off-take and a lender you already know can move in weeks; a syndicated or rating-agency-path deal takes materially longer because syndication and a rating add process. The single biggest accelerant on the diligence phase is a room a credit committee can underwrite fast — which is precisely the job the companion GPU cluster financing data room guide is written to do. Run the raise from this post; build the room from that one.

What kills these raises?

They die when the amortization clock loses the race — when the debt cannot get repaid before the collateral decays or the contract runs out — and the failure shows up in one of six recognizable forms. A serious lender will raise each of these whether or not you do, so the honest move is to address them on your terms, in the model and the room, rather than hope nobody asks.

Depreciation and obsolescence. This is the master risk the whole post is organized around. If the GPUs lose value faster than the debt amortizes, the collateral cushion evaporates and a stressed refinancing has nothing to lean on. The depreciation debate — four-to-six-year operator lives versus the two-to-three-year skeptic case, framed as a company-specific estimate under ASC 250 — is the modeled assumption a committee probes hardest. The mitigation is structural, not rhetorical: size the debt to amortize inside the contract even on the pessimistic life, and show it.

Off-taker concentration. These structures live or die on the creditworthiness and stickiness of a few customers, and reliance on a single off-taker is the risk a committee flags first. The disclosure that works is what share of contracted revenue each off-taker represents, alongside your mitigants — contract term, minimum commitments, counterparty credit, and any diversification in the pipeline. Concentration does not shrink because you hide it; disclosing it lets you frame the mitigation.

Merchant-revenue assumptions. A model that leans on spot revenue is a model exposed to the spot market, and the spot market moved violently — H100 rates fell from roughly $8/hour at peak to the $3/hour range into late 2025. The discipline is to size term debt against contracted, not merchant, revenue and to treat merchant upside as upside, never as debt-service coverage.

Residual value. The secondary market for used GPUs is thin and volatile, so recovery-value assumptions are a weak foundation for a loan. This is exactly why the contract carries the credit and the chips are the fallback — and why a facility that depends on a rich residual to repay is fragile. Some structures transfer this risk explicitly: the Meta-Blue Owl Hyperion deal reportedly carries a residual-value guarantee that moves technology-obsolescence risk to Meta, a luxury most raises do not have.

Refinancing risk. A facility that must be refinanced into a worse market — higher rates, a soured sentiment on the asset class, a weakened off-take — is exposed at exactly the wrong moment. The defense is a tenor and amortization profile that minimizes the balance needing refinancing and, ideally, an off-take that extends past the refinancing point.

Chip-generation transitions. Nvidia's roughly annual cadence — Hopper, then Blackwell, then Rubin — means the generation you finance today can be superseded while your loan is outstanding, denting secondary value and competitive utilization. The room addresses it by showing the generation you are buying, the refresh plan, and — critically — that the off-take commits the customer across the loan term regardless of what ships next. A contract that survives the next generation is what converts "your chips will be obsolete before you repay us" from a deal-killer into a modeled, mitigated risk.

The circularity question sits over all of it. There is a live 2025-26 debate about vendor-financing loops — Nvidia invests in and backstops neoclouds and AI labs, those parties buy Nvidia chips and capacity, and value loops back to Nvidia, with the Nvidia-OpenAI Ohio backstop, cut from a reported $250 billion to up to $105 billion on investor risk concerns, the most-cited example, and comparisons drawn to Lucent's late-1990s vendor financing. A serious lender knows this discourse cold. If your structure touches any of these loops, name it and explain why your contracted revenue is genuine end-demand rather than a circular flow. The strength of your actual off-take is the answer to the skepticism.

How Peony fits the raise

I run Peony, a data room company, and I will be precise about where it fits: Peony is the room you run the raise from, not the thing that makes your raise bankable. What makes a GPU-backed raise bankable is your SPV structure, your off-take contracts, and a model that amortizes inside the contract — the substance of everything above. What Peony does is give a small neocloud deal team the controlled room a credit committee's process expects, without a bank-scale procurement, so a 40-person operator's raise reads as institutional. That is the leverage the room gives, and it is exactly the surface the companion GPU cluster financing data room guide details.

The fit is structural, and it starts with pricing. A GPU raise is document-heavy by nature — hardware invoices, colo leases, off-take contracts, a large model, technical reports — and it runs several lenders and their advisors through one room in parallel. Peony is priced per admin seat with every recipient free: the Business plan is $30 per admin per month, the Data Room plan is $52 per admin per month (adding dynamic watermarking and unlimited rooms), and the Deal Team plan is $64 per admin per month (adding advanced redaction and an advanced Q&A module), all on annual billing. Every lender, every advisor, and every vendor you invite is a free viewer, so a raise run with two to four admins over a few months lands in the low hundreds of dollars total regardless of how many funds are reading — the bill is bounded by your small deal team, not by document count or counterparty list. Against a facility measured in tens or hundreds of millions, the room cost is a rounding error, which is one reason 6,800+ customers run controlled disclosure this way rather than on a shared drive.

Where Peony earns its place in the raise specifically: NDA gating so a lender signs before they see a named off-take contract; visitor groups so competing funds, their technical advisors, and vendors each get an isolated view and never see each other; dynamic watermarks that burn each viewer's identity into the named contracts and the model so a leak is traceable; and page-level analytics that show which fund actually opened the off-take contracts and how long they spent — your clearest read on who is moving toward a term sheet. On the claims a diligence-minded lender checks, Peony is SOC 2 Type II, with custom data residency, BYOK, and self-hosted deployment available on the Enterprise plan. Consider the anonymized archetype this post is built around — a Texas neocloud raising a nine-figure SPV facility against a single contracted off-take, whose entire case is that the ring-fence, the contract, and the depreciation model hold together. Peony makes the room that carries that case look like it was run by a team three times the size, which on a first institutional raise is exactly the credibility the room exists to lend — and why 6,800+ customers run their most sensitive raises on flat per-admin pricing rather than a metered enterprise bill.

Related resources

• GPU Cluster Financing Data Room — the companion post: the room a lender reads, folder structure, staged disclosure, and credit-committee Q&A. This post is the raise; that one is the room.
• Private Credit Data Room — the fund-side architecture behind the private-credit lenders who underwrite these facilities: the three-room model a direct-lending fund runs.
• Borrower Reporting Data Room — the recurring-reporting cadence a credit fund expects from you post-close: no-login uploads, reminders, and the audit trail.
• Data Center Data Room — the facility acquisition and buildout raise: powered land, Will-Serve letters, and energization risk, for when you own rather than rent the site.
• Data Room for Infrastructure Projects — the horizontal project-finance raise room across asset classes, the cross-asset cousin of a compute raise.
• UK Data Centre Due Diligence — UK development-stage diligence detail: the grid queue, planning, and energization timeline behind a UK compute site.
• Data Room for Investors — the general fundraise-room structure behind the equity layer of the neocloud capital stack.
• Data Centers solution — the digital-infrastructure overview for operators financing compute and the facilities that house it.

About the author: Sean Yu is co-founder of Peony, the data room used by 6,800+ M&A, private equity, private credit, and infrastructure-finance teams. He was previously a venture investor at Backed and Target Global and has evaluated hundreds of deals from the capital-provider side.

Try Peony free for 14 days — no credit card required.