Neoclouds
What Does Runpod Do? The Per-Second GPU Cloud
Runpod sells GPU hours by the second to AI developers, on machines that sit in partner data centers it does not own.
The Stack
Status Tightening as of 25 Oct 2026: short bridge, shut market. Nebius signed a $3.5B, 364-day Citi bridge whose margin steps up every 90 days, Crusoe borrowed against gas generators, and high-yield desks wait out the midterms. The long refinancing is what binds.
Neoclouds
Runpod sells GPU hours by the second to AI developers, on machines that sit in partner data centers it does not own.
Neoclouds
TensorWave is a GPU cloud that runs only AMD chips. Its $6.5 billion lease at Fermi’s Texas campus leans on AMD’s credit, and the hall still has no construction money, so financing is the step that binds.
Neoclouds
Together AI sells open-model inference and GPU hours; investors capitalize more than 500 MW of compute separately from the $800 million equity — cash sticks when tokens or reserved hours bill, while those megawatts are still a clock.
Neoclouds
Lightning AI sells GPU hours and AI software from a fleet it says it bought outright — cash sticks when those hours bill, not when a loan is drawn.
Neoclouds
Nscale signs take-or-pay GPU contracts, funds the halls with customer prepayments and delayed-draw loans, and gets paid only after clusters are delivered — most of the $103.4bn book is still waiting on live halls.
Neoclouds
Fluidstack leases power-ready halls, gets Google to backstop those leases so landlords can raise project debt, and sells GPU hours to AI labs.
Neoclouds
Lambda buys GPU clusters, funds them with delayed-draw debt that only pays out as each hall switches on, and gets paid when contracted customers run the hours.
Neoclouds
Crusoe sits AI halls on the electrons first: it gets paid as the energy-first developer that energises campuses for hyperscalers, and as Crusoe Cloud for GPU hours, while 6 GW-plus of contracted capacity still waits on 1 GW live.
Neoclouds
Nebius sells specialist AI compute, funds a large share of the halls with customer prepayments, and gets paid as that capacity goes live — connecting contracted megawatts is the stuck step.
Neoclouds
CoreWeave buys accelerators, funds them with debt, and sells AI compute on multi-year contracts — why the model loses money while growing.
Neoclouds are specialist GPU clouds. They buy or finance GPU fleets, rent power-ready halls, and sell cluster time by the hour or on contract. This Stack layer tracks who gets paid in that chain, and which step is holding the cash. New to the term? Start with who the neoclouds are.
Company models: CoreWeave · Nebius · Lambda · Nscale · Crusoe · Fluidstack · Lightning AI · IREN
Status: Tightening · Updated: 2026-10-25
Investor read: On 21 October this page said the buyer sets the rate: a GPU loan backed by Microsoft clears near 6%, and one backed by a lab pays 9–10%. This week showed what happens to the money that no buyer stands behind. On 22 October Nebius disclosed a $3.5 billion loan from Citi that it calls a bridge. It runs 364 days, it is secured on substantially all of the borrowers' assets, and its margin rises every 90 days until it is refinanced. Its chief operating officer named the way out: asset-backed borrowing against more than $40 billion of investment-grade contracts. That longer loan has not been arranged. The market a neocloud would normally use instead is shut for now. Credit desks told CNBC that high-yield issuance is on hold until after the US midterm elections. CoreWeave's default swaps sit at 1,145 basis points, and its unsecured notes yield above 15%. So the layer is borrowing short and against things it can point to: Crusoe against gas generators it has ordered, and QumulusAI against the GPUs themselves. Money can still be found. What binds now is the takeout, the long refinancing that has to replace each short loan before it gets dearer or comes due.
| Figure | Reported | |
|---|---|---|
| Nebius × Citi bridge | $3.5B senior secured · 364 days (signed 17 Oct, due 16 Oct 2027) | Nebius 6-K, 22 Oct 2026 |
| Nebius bridge cost | Term SOFR + 2.75%, +0.25% every 90 days, capped at 3.50% · prepay without penalty | Nebius 6-K, 22 Oct 2026 |
| Nebius bridge security | Substantially all borrower assets and subsidiary shares · minimum-liquidity covenant | Nebius 6-K, 22 Oct 2026 |
| Nebius takeout named | Asset-backed financing on more than $40B of investment-grade contracts · not yet arranged | Nebius COO, 22 Oct 2026 |
| High-yield window | Effectively on hold until after the midterms (3 Nov), per credit desks | CNBC, 21 Oct 2026 |
| CoreWeave default swaps | Five-year CDS 1,145 bp · unsecured notes above 15% | CNBC, 21 and 23 Oct 2026 |
| CoreWeave maturities and cash | $20.7B due through 2030, $14.9B after · $5.5B cash at end-Q2 · Q3 FCF consensus about −$10.1B | CNBC, 21 Oct 2026 |
| CoreWeave funding gap | Goldman about $23B · Bernstein about $76B ($85B less $9B in place) · Baird about $30B in 2027 | The Inference · CNBC, 21–23 Oct 2026 |
| Crusoe × Neuberger | $700M secured on ordered gas generators for its own sites | Crusoe, 22 Oct 2026 |
| QumulusAI × TFC | GPU leases to $123.6M (from $26.0M in May) · lessor funds 85–90% · high single-digit rate | QumulusAI, 22 Oct 2026 |
| Our arithmetic | Bridge margin +0.75 points if held to the cap (about 9 months in) · CoreWeave debt due by 2030 ≈ 3.8× its June cash · Bernstein gap ≈ 3.3× Goldman's | Ours, on the prints above |
Read the table in three blocks. The first four rows are one bridge loan and the exit it depends on. The next four are why that exit matters: the market for company-level debt has paused, and the largest neocloud's bills keep arriving. The last two are how the rest of the layer is borrowing meanwhile, against a named machine rather than the company's own name.
A neocloud turns money and megawatts into billed GPU hours, in this order:
In the AI build chain this layer sits downstream of Semiconductors, beside Hyperscalers, and next to AI Data Centers (landlords, megawatt halls, power-to-hall). Capex and cloud backlog stay on Hyperscalers. This page stays on GPU renters and the money between a signed contract and a billed hour. Background: neocloud vs hyperscaler, take-or-pay GPU contracts and GPU depreciation.
Who gets paid first right now: banks that write short secured bridges, because they hold a claim on the whole company and a margin that rises if the takeout is late; asset lenders and lessors who fund a named machine — GPUs, gas generators — with QumulusAI's lessor at an implied high single-digit rate; and Nvidia, paid on delivery once a cluster is funded. Who waits: neoclouds whose next phase depends on a long loan or a bond that cannot be placed until the high-yield market reopens, and IPO candidates waiting for buyers. Who carries the risk: unsecured bondholders and shareholders, who sit behind every secured layer, and who are the funders a refinancing would turn to if asset-backed money falls short.
Roles only — not a leaderboard, not a buy list.
| Role | Examples | What they do in this layer |
|---|---|---|
| Specialist GPU clouds | CoreWeave, Nebius, Lambda, Nscale, IREN, Crusoe, Fluidstack, QumulusAI | Sign buyers, raise GPU money, switch clusters on, bill hours, and now refinance short money into long money |
| Bridge banks | Citi and other arranging banks | Lend short, secured on the whole company, with margins that rise the longer the loan stays out |
| Asset lenders and lessors | Neuberger Specialty Finance, Tech Finance Co., Blue Owl, PIMCO-advised funds | Lend against a named object — GPUs, gas generators — and the contract that pays for it |
| Investment-grade buyers | Microsoft, Meta and other hyperscalers | Their contracts are the collateral a long-term takeout loan is built on |
| Company-level funders | Unsecured bondholders, convertible holders, IPO and pre-IPO investors | Fund what secured lenders do not; paid last; their market is the one that has paused |
| Chip and system sellers | Nvidia | Paid on delivery once a cluster is funded; also an investor and backstop in parts of the layer |
Always verify listing, ADR, and tax treatment for your own jurisdiction. Names here map roles. They are not recommendations.
The 13 October Stack said signed is not funded. On 17 October the funding split in two: cheap secured loans and expensive company bonds. On 21 October the secured price itself turned out to depend on the buyer. This week the question moved from price to time. The money arriving is short, and the long money that should replace it is waiting on a market that has paused.
1. A bridge with a clock on it — PRIMARY (new 2026-10-25). On 22 October Nebius filed a 6-K describing a Bridge Facility Agreement signed with Citibank on 17 October: a $3.5 billion senior secured term loan to build out GPU capacity, maturing on 16 October 2027. Interest starts at Term SOFR plus 2.75% and the margin rises by 0.25% for each 90-day period, up to 3.50%. Nebius can repay early without penalty. The loan is secured on substantially all of each borrower's assets, including the shares of its US and Dutch borrowing subsidiaries, and carries a minimum-liquidity covenant. On our arithmetic the margin reaches its cap about nine months in, 0.75 points above where it started. A loan built that way is meant to be replaced. Nebius says so: chief operating officer Ophir Nave called sequencing the financing the way to optimise long-term cost of capital, and named the next source as asset-backed financing on more than $40 billion of long-term contracts with investment-grade customers. It has not said when that will be arranged. See the Wire on the Citi bridge and what Nebius does.
2. The usual exit is shut for now — SECOND PRINT. On 21 October CNBC reported that, according to sources on credit desks, the high-yield debt market is effectively on hold until after the midterm elections, about two weeks away, because fixed-income investors want permitting and regulatory questions around data centers substantially resolved first. CoreWeave's five-year default swaps stood at 1,145 basis points, its unsecured debt yielded about 15%, and by 23 October its 2031 and 2032 notes were above 15%. Shares of CoreWeave, Nebius and IREN had each fallen well over 30% in October. Nvidia's Jensen Huang put it plainly in Berlin: the demand for capital is much greater than its availability. For a neocloud with a bridge outstanding, the company-bond market is the fallback takeout if asset-backed money falls short, and it is the one that has paused. See the Analysis on neocloud debt near 1,000 bp.
3. The bill keeps coming — THIRD PRINT. CNBC also set out CoreWeave's calendar: about $20.7 billion of debt due through 2030 and $14.9 billion after, against $5.5 billion of cash at the end of June. Analysts expect third-quarter free cash flow of about minus $10.1 billion, and Baird's Rob Oliver estimates roughly $30 billion of new capital needed in 2027. The Inference set two estimates side by side: Goldman Sachs puts the remaining funding need near $23 billion, while Bernstein starts from about $85 billion to fund the committed backlog, subtracts $9 billion already in place and gets about $76 billion. On our arithmetic, debt due by 2030 is about 3.8 times June's cash, and the two estimates differ by about 3.3 times. Chief executive Mike Intrator told staff the company combines infrastructure-backed borrowing, corporate debt, equity and customer prepayments. Every one of those is a takeout, and each has to be arranged in this market. See what CoreWeave does and the Analysis on CoreWeave's signed-versus-live gap.
4. Borrow against the machine, not the name — FOURTH PRINT. Also on 22 October, Crusoe and Neuberger Specialty Finance announced a $700 million facility secured on gas generation equipment Crusoe has recently ordered, to power its own sites behind the meter. The lender said it sees significant value in the power assets themselves. The same day QumulusAI said its equipment leases with Tech Finance Co. had grown to about $123.6 million from $26.0 million in May. The lessor funds 85–90% of each Blackwell GPU schedule, customer deposits fund the rest, and the implied rate is in the high single digits, with an option to hand the GPUs back after three years. These loans do not need the company-bond market. They need an object a lender is willing to repossess. Crusoe's is power equipment it has already ordered, and QumulusAI's is a GPU whose value falls each year. See what Crusoe does and why GPU depreciation matters to the lender.
5. Late sites make every bridge longer — FIFTH PRINT. Bloomberg reported from Nvidia's Berlin conference on 22 October that Huang told Nscale to "work around the clock" on problems bringing some data centers online, and that Nscale has faced delays, equipment failures and power constraints at some sites, on top of pushback on its roughly $30 billion IPO valuation target. A long-term loan secured on a contract can only be sized once the capacity bills. Every month a site runs late is a month the short money stays out at a higher margin. See what Nscale does and the Wire on Nscale's second Anthropic talks.
Demand is not the bind. Nebius's revenue chief Marc Boroditsky said on 21 October that compute prices are rising faster than Nebius expected, that an auction it ran that week cleared higher again, and that its 2027 capacity would sell out if it were online. Nebius also began publishing live spot GPU prices. Buyers are there. What is missing is time on the funding: long money arranged before the short money gets dear. Our read, labelled as ours: neoclouds that can fund the next phase against investment-grade contracts or a physical asset keep building; those relying on the high-yield window wait for it to reopen.
Explained
Company models
Wire, Catalyst & Analysis
Markets
Sibling Stack layers
No. This is a map of roles and constraints. Not investment advice. Do your own research.
Because the largest new neocloud loan this week is a 364-day bridge with a rising margin, high-yield issuance is reported on hold until after the midterms, CoreWeave's default swaps are above 1,100 basis points, and estimates of its funding gap run as high as $76 billion. We would ease if Nebius closes an asset-backed takeout on its investment-grade contracts, or if a neocloud bond prices well below today's yields once the market reopens.
The 21 October desk led with The buyer sets the rate / Lab credit binds: the same GPU loan costs about 6% with an investment-grade buyer and 9–10% with a lab. This update asks how long the money lasts as well as what it costs. New money is short (Nebius's bridge) or tied to a named asset (Crusoe's generators, QumulusAI's GPUs), and the long market that would refinance it has paused. The headline becomes Short bridge, shut market / The takeout binds.
A short loan that pays for GPUs and sites now, on the plan that longer-term money will repay it later. Nebius's is 364 days, secured on substantially all borrower assets, with a margin that rises every 90 days. The design pushes the borrower to refinance early.
The long-term financing that replaces a bridge or construction loan. For a neocloud it is usually a multi-year loan secured on GPUs and the customer contract, a bond, a convertible or new equity.
Credit-desk sources told CNBC on 21 October that fixed-income investors want permitting and regulatory questions around data centers substantially resolved before committing, and that issuance is effectively paused until after the midterm elections. That is a report from sources, not an official closure.
No. It is the price of insurance against default, and CNBC notes that a level above 11% is generally seen as distressed. CoreWeave has not defaulted and says it has a range of funding options.
Hyperscalers covers the cloud companies' spending and the power they buy. AI Data Centers covers landlords, megawatt halls and power-to-hall. Neoclouds covers the GPU renters and the money between a signed contract and a billed hour.
Not investment advice. Do your own research. The 0.75-point step-up, 3.8× and 3.3× figures are our arithmetic on the filed and reported figures. Goldman's, Bernstein's and Baird's figures are analyst estimates as reported by The Inference and CNBC.
Last updated: 2026-10-25 · Second Order by TKN