Capital gets stranded in neoclouds when a fixed colocation lease and amortizing GPU debt meet falling GPU rental prices. In StackedAI’s model, a GPU-only operator in leased colocation returns 0.30x MOIC while the owner of the building on an NNN lease returns 6.30x on 28% of the equity (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026). Recovery flows to whoever controls the power and the lease, not the silicon.
This page is the public summary of StackedAI’s Stranded Asset Report track. Three supporting pages go deeper: what happens when a neocloud fails, GPU depreciation and secondary value, and the neocloud distress screening framework.
Why do neocloud economics strand capital?
A neocloud is a GPU rental business, and its capital gets trapped because three time horizons do not match. The colocation lease is fixed, often for 10 to 15 years. The GPU loan amortizes over five years. The revenue the GPUs earn reprices continuously, and since 2023 it has repriced downward.
The financing makes the mismatch worse. Lenders who take GPUs as their only collateral advance about 60% of hardware cost at roughly 8% with five-year amortization in StackedAI’s base case, and residual value at the end of that term is modeled at about 15% of cost (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026). The residual is low because the secondary market for rack-scale liquid-cooled systems is thin: a GB200 NVL72 rack draws 120 to 130 kW (StackedAI analysis, density roadmap compilation, Aug 2026) and cannot be moved into a hall that lacks the liquid cooling plant to support it.
Real estate debt shows the contrast. In mid-2026, top-credit data center loans priced at spreads in the low 200 basis points with up to 85% loan-to-cost; non-credit tenants priced 200 to 300 basis points wider at 70 to 80% loan-to-cost (JLL, North America Data Center Report Midyear 2026, Aug 2026). A GPU-only borrower sits below that tier because its collateral loses value faster than the loan amortizes.
At base assumptions the GPU-only operator in leased colocation returns 0.30x MOIC and a negative 28.8% IRR, while the party that owns the building on an NNN lease returns 6.30x MOIC and 26.3% IRR on 28% of the equity, with no technology risk (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026). The GPU-only structure only breaks even with flat-to-rising $/GPU-hour pricing; at a 20% annual price decay the levered structure goes negative and DSCR falls below 1.0x by year five at 64% loan-to-cost (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026). StackedAI’s summary: the building carries the deal; the GPUs consume it.
How large is the exposure?
Neocloud funding tripled to more than $35B in 2025, and S&P Global Market Intelligence describes the sector as in an “unenviable position” with a consolidation wave possible (S&P Global Market Intelligence via DCD, Feb 2026). The largest operators disclose their balance sheets; the table collects the public figures.
| Exposure indicator | Figure | Source |
|---|---|---|
| Neocloud funding raised in 2025 | More than $35B, tripled vs the 2024 record | S&P Global Market Intelligence via DCD, Feb 2026 |
| CoreWeave recourse debt, Q2 2026 | $6,235M current plus $25,170M non-current | CoreWeave, Q2 2026 Results, Aug 2026 |
| CoreWeave quarterly interest expense | $640M, up from $267M a year earlier | CoreWeave, Q2 2026 Results, Aug 2026 |
| CoreWeave scale | Revenue $2,575M (+112%); backlog about $104B; 3.7 GW contracted, 1.5 GW active; H1 capex $14.1B | CoreWeave, Q2 2026 Results, Aug 2026 |
| CoreWeave 2026 capex guidance | Up to $35B; Q1 2026 net loss $740M; Meta commitment $21B | Data Center Knowledge, May 2026 |
| Nebius financing | $4.34B convertible; capex guide $20-25B; Meta up to $27B; Microsoft $17.3B | Reuters via Yahoo Finance, Mar 2026; Data Center Knowledge, May 2026 |
| GPU-backed loans, 2026 | CoreWeave $3.1B loan sale; GMI Cloud $635M; Lambda $917M and $3B pre-IPO talks | Bloomberg, Apr, Jul and Aug 2026 |
| Microsoft neocloud commitments | More than $60B | Bloomberg, Nov 2025 |
| Hyperscaler-to-neocloud compute leases | More than $100B tracked | BloombergNEF, Mar 2026 |
| Credit stress signal | AI credit default swaps surged (CoreWeave, Nebius, Oracle) | Bloomberg, Jul 2026 |
The sector has not yet produced a headline bankruptcy; StackedAI’s source review as of August 2026 found none reported on a major outlet. The stress shows elsewhere. Wholesale colocation deals with neocloud tenants have stalled over credit risk, including one that failed at $155 to 160 per kW on a 15-year term with six months of prepayment (Data Center Knowledge, Apr 2026). The market is pricing neocloud credit as a problem; the legal events lag the financial ones.
How far have GPU rental prices fallen, and are they recovering?
H100 rental pricing fell from above $8 per GPU-hour in 2023 to under $3 on the open market (StackedAI analysis, Stranded Asset Report, Aug 2026). That decline is the largest single driver of stranding: a fleet financed at $8 must now service the same debt schedule at under $3.
Some recovery is visible. One-year H100 contract pricing rose from a $1.70 low in October 2025 to $2.35 by March 2026, roughly 40% (Akash Network, H100 Rental Price 2026, Aug 2026). On-demand H100 pricing on dedicated GPU clouds has a median of $4.19 per GPU-hour against about $7.89 at hyperscalers, up 4.3% over 12 months (GetDeploying, GPU Price Index, Aug 2026). B200 capacity rents at $3.70 to $7 per GPU-hour, with a reserved-versus-on-demand spread of 40 to 70% (StackedAI analysis, Stranded Asset Report, Aug 2026).
The recovery is from a floor. An operator whose take-or-pay contracts were written at 2023 rates is fine while those contracts hold. An operator with short-tenor or on-demand revenue is exposed to a spot price far below the level that underwrote most 2023 and 2024 fleets.
What does “stranded” mean in practice?
“Stranded” is not one thing. StackedAI separates a neocloud deployment into four asset layers, each with a different buyer pool and recovery path.
| Asset layer | Mobility | Who wants it | Recovery logic (StackedAI framework) |
|---|---|---|---|
| 1. GPUs and servers | High for air-cooled PCIe; low for liquid-cooled rack-scale | Other compute operators; GPU lenders enforcing collateral | Set by generation and spot rental pricing; residual modeled at about 15% of cost after five years (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026) |
| 2. Fabric (InfiniBand or RoCE switches, optics, cabling) | Medium; topology-specific | Operators building the same cluster size | Recovers only as a matched cluster; component stripping destroys most value |
| 3. Power and cooling fit-out (CDUs, manifolds, busway) | Near zero; bolted to the hall | The landlord and the next tenant | Liquid-cooling fit-out costs $1.5 to 1.6M per MW, 20 to 25% above air-cooled (Data Center Knowledge, Apr 2026); value survives only in place |
| 4. Lease and power position | Zero; a contract and a grid position | Strategic buyers, hyperscalers, the landlord | The scarce asset: North American vacancy is 1% for the third consecutive year (JLL, North America Data Center Report Midyear 2026, Aug 2026) |
The lower the layer, the less mobile it is and the more its value depends on the market for powered, cooled capacity rather than on the chip cycle. StackedAI’s underwriting position follows: own layer 4 and, where possible, layer 3; hold layers 1 and 2 only as debt or preferred. The NNN lease scarcity pillar covers why layer 4 is so hard to source.
How does StackedAI screen a neocloud for distress?
StackedAI maintains a screen of 31 neocloud operators built on eight dimensions (StackedAI analysis, Neocloud Distress Screen, Aug 2026). The dimensions are public; scores on named companies are not.
- Customer concentration and contract tenor: how much revenue sits with one or two customers, and how long it runs relative to the debt.
- GPU generation mix versus debt maturity: whether the fleet is Hopper or Blackwell, and whether loans mature before or after the generation ages out.
- Utilization floor and take-or-pay coverage: what share of fixed obligations is covered by contracted minimums.
- Colocation lease fixed-cost ratio: the lease as a share of contracted revenue.
- Refresh obligations: whether contracts require hardware upgrades on generation change, and who funds them.
- Change-of-control and termination triggers: whether key contracts survive a sale, restructuring or lender step-in.
- SPV and bankruptcy-remote structure: whether GPUs, lease and customer contracts sit in one entity or in separate vehicles.
- Power position: owned, leased from a colocation provider, or behind-the-meter.
The red, amber and green criteria for each and the five diligence questions asked before any GPU term sheet are on the screening framework page.
Who buys stranded GPU assets, and who runs the sale?
The sale is rarely run by the operator. Four fiduciary channels handle distressed compute assets: court-appointed receivers, assignment-for-the-benefit-of-creditors (ABC) firms, Chapter 11 sale processes, and liquidation advisors engaged by lenders or estates (StackedAI analysis, Stranded Asset Report, Aug 2026). Infrastructure vendors typically see the distress six to eighteen months before it is public, through deferred invoices and skipped refreshes (StackedAI analysis, Stranded Asset Report, Aug 2026).
Buyers fall into three groups. Compute operators buy silicon at a discount and pay for layers 1 and 2. GPU lenders enforce collateral and re-market it, with recovery capped by the residual of a depreciating fleet. Strategic buyers, meaning data center operators, hyperscalers and infrastructure funds, want the lease, the utility capacity and the liquid-cooled hall; they may take the GPUs but price the deal on layers 3 and 4.
With primary-market asking rents at $195.94 per kW per month for 250 to 500 kW requirements (CBRE, North America Data Center Trends H2 2025, Feb 2026) and vacancy at 1% (JLL, North America Data Center Report Midyear 2026, Aug 2026), the third group increasingly sets the clearing price. The what happens when a neocloud fails page walks through the sequence and what each stakeholder recovers.
How do hyperscaler and Nvidia backstops change the picture?
Backstops are why the sector carries so much debt without, so far, a bankruptcy. Three forms exist.
Hyperscaler lease guarantees. Google backs about $1.3B of TeraWulf’s lease obligations to Fluidstack on a 168 MW deployment with about $9.5B of contracted revenue (TeraWulf, press release, Oct 2025). Hut 8 signed a 245 MW, 15-year deal with Fluidstack worth $7.0B and up to $17.7B with options, part of Fluidstack’s partnership with Anthropic (Data Center Dynamics, Dec 2025). In both cases the credit that makes the lease financeable belongs to the hyperscaler or the AI lab, not the neocloud.
Chip-vendor utilization floors. Nvidia has entered floor-utilization commitments under which it rents idle capacity at a predetermined rate; StackedAI’s review is that these exist and are obtainable in negotiation (StackedAI analysis, Stranded Asset Report, Aug 2026).
Vendor and customer financing pools. Nvidia has assembled a $500B financing commitment from six financial firms and has discussed backstopping up to $250B for OpenAI (Bloomberg via Fortune, Aug 2026). Meta has committed $21B to CoreWeave (Data Center Knowledge, May 2026) and up to $27B to Nebius, with Microsoft committing $17.3B to Nebius (Reuters via Yahoo Finance, Mar 2026; Data Center Knowledge, May 2026).
A backstop moves risk; it does not remove it. The diligence question is whether the guarantee covers the lease, the compute contract or the debt, and what happens to parties outside it. A landlord whose lease is guaranteed by a hyperscaler holds a very different asset from a GPU lender whose collateral is subject to a vendor’s option to rent idle capacity below debt service. The PE power-risk underwriting page covers the parallel question for power.
Key terms
- Neocloud: a specialist cloud provider that rents GPU compute, usually from leased colocation.
- Stranded asset: an asset whose contracted revenue no longer covers its fixed obligations and which cannot be redeployed or sold at a price that repays its financing.
- Take-or-pay: a clause under which the customer pays for a minimum volume of capacity whether or not it uses it.
- Bankruptcy-remote SPV: a special purpose vehicle structured so its assets are not swept into the parent’s bankruptcy estate.
- Step-in rights: a right allowing a landlord, lender or customer to take over an operator’s obligations or assets on default.
- Residual value: the estimated value of an asset at the end of its financing term.
- MOIC: multiple on invested capital, total cash returned divided by equity invested.
How StackedAI applies this
StackedAI advises sponsors, lenders and restructuring professionals on the real estate and power layers of AI infrastructure. In stranded-asset engagements it applies the 31-company distress screen to find where fixed lease obligations, debt maturities and GPU generation risk converge, and it values the four asset layers separately so each fiduciary knows what is recoverable in place versus at auction. Its underwriting position is consistent: own the building and the power position, and hold GPU exposure as debt or preferred, never as common equity (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026). Engagement models are on the advisory services page.
Frequently asked questions
What is a stranded GPU asset?
A GPU fleet, network fabric or fitted-out hall whose contracted revenue no longer covers its fixed obligations and which cannot be redeployed or sold at a price that repays its financing. StackedAI models GPU residual at about 15% of cost after five years, so the stranding sits mainly in the debt and the lease (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026).
Why do neoclouds get into trouble when demand for AI compute is high?
Because their costs are fixed and their revenue is variable. A 10 to 15 year colocation lease and five-year GPU amortization were financed against H100 rates that fell from above $8 per hour in 2023 to under $3 (StackedAI analysis, Stranded Asset Report, Aug 2026). High demand at lower prices does not fix a debt schedule set at higher prices.
How much capital is exposed to neocloud credit?
Neoclouds raised more than $35B in 2025, tripling the 2024 record (S&P Global Market Intelligence via DCD, Feb 2026). BloombergNEF tracks more than $100B of hyperscaler-to-neocloud compute leases (BloombergNEF, Mar 2026) and Microsoft alone has committed more than $60B (Bloomberg, Nov 2025). CoreWeave reported $6,235M of current and $25,170M of non-current recourse debt at Q2 2026 (CoreWeave, Q2 2026 Results, Aug 2026).
Have GPU rental prices recovered?
Partly. One-year H100 contract pricing rose from a $1.70 low in October 2025 to $2.35 by March 2026, about 40% (Akash Network, Aug 2026), and on-demand H100 pricing on dedicated clouds is a median $4.19 per GPU-hour, up 4.3% over 12 months (GetDeploying, Aug 2026). The recovery is from a floor well below the rates that financed 2023 and 2024 fleets.
Who buys GPU clusters out of a failed neocloud?
Three buyer types: compute operators buying silicon at a discount, GPU lenders enforcing on collateral, and strategic buyers who want the powered, cooled hall and the lease rather than the chips. With North American vacancy at 1% (JLL, North America Data Center Report Midyear 2026, Aug 2026), the third group increasingly sets the price.
Does a hyperscaler backstop remove the risk?
It moves it. Google backing about $1.3B of TeraWulf’s lease obligations to Fluidstack (TeraWulf, Oct 2025) shifts risk from the neocloud to Google’s credit for those obligations. Parties whose claims sit outside the backstop are not protected, which is why diligence must confirm exactly what is covered: the lease, the compute contract or the debt.
Sources
- StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026 (internal, illustrative)
- StackedAI analysis, Stranded Asset Report, Aug 2026 (internal)
- StackedAI analysis, Neocloud Distress Screen (31 companies), Aug 2026 (internal, dimensions only)
- StackedAI analysis, density roadmap compilation, Aug 2026 (internal, compiled from Goldman Sachs and Dell’Oro estimates)
- S&P Global Market Intelligence via Data Center Dynamics, Data center M&A topped $69bn in 2025; neoclouds in “unenviable position”, Feb 2026, https://www.datacenterdynamics.com/en/news/sp-global-data-center-ma-topped-69bn-in-2025-neoclouds-in-unenviable-position/
- CoreWeave, Q2 2026 Results, Aug 2026, https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx
- Data Center Knowledge, Earnings Roundup: Neoclouds Shift From GPU Race to Power Wars, May 2026, https://www.datacenterknowledge.com/cloud/earnings-roundup-neoclouds-shift-from-gpu-race-to-power-wars
- Reuters via Yahoo Finance, Nebius says well funded, Mar 2026, https://finance.yahoo.com/sectors/technology/articles/nebius-says-well-funded-ai-125730023.html
- Bloomberg, Banks kick off $3.1 billion loan sale for CoreWeave, Apr 2026, https://www.bloomberg.com/news/articles/2026-04-30/banks-kick-off-3-1-billion-loan-sale-for-coreweave-amid-ai-boom
- Bloomberg, Nvidia partner GMI Cloud seeks $635 million GPU-backed bank loan, Jul 2026, https://www.bloomberg.com/news/articles/2026-07-14/nvidia-partner-gmi-cloud-seeks-635-million-gpu-backed-bank-loan
- Bloomberg, AI cloud provider Lambda taps loans for Nvidia-tied chip deal, Aug 2026, https://www.bloomberg.com/news/articles/2026-08-10/ai-cloud-provider-lambda-taps-loans-for-nvidia-tied-chip-deal
- Bloomberg, Wall Street picks AI winners and losers as credit swaps surge, Jul 2026, https://www.bloomberg.com/news/articles/2026-07-29/wall-street-picks-ai-winners-and-losers-as-credit-swaps-surge
- Bloomberg, Microsoft neocloud deals cross $60 billion, Nov 2025, https://www.bloomberg.com/news/articles/2025-11-04/microsoft-neocloud-deals-cross-60-billion-in-ai-spending-frenzy
- BloombergNEF, AI Data Center Build Advances at Full Speed: Five Things to Know, Mar 2026, https://about.bnef.com/insights/commodities/ai-data-center-build-advances-at-full-speed-five-things-to-know/
- GetDeploying, GPU Price Index, Aug 2026, https://getdeploying.com/gpu-price-index
- Akash Network, H100 Rental Price 2026: Cost Per Hour, Aug 2026, https://akash.network/the-bid/h100-rental-price-2026-cost-per-hour/
- Data Center Knowledge, Neocloud Storm Gathers as Data Center Deals Stall Over Credit Risk, Apr 2026, https://www.datacenterknowledge.com/cloud/neocloud-storm-gathers-as-data-center-deals-stall-over-credit-risk
- TeraWulf, press release: TeraWulf Expands Strategic Partnership with Fluidstack Through New 168 MW AI Compute Joint Venture, Oct 2025, https://investors.terawulf.com/news-events/press-releases/detail/121/terawulf-expands-strategic-partnership-with-fluidstack-through-new-168-mw-ai-compute-joint-venture
- Data Center Dynamics, Hut 8 signs 245MW capacity deal with Fluidstack as part of multi-gigawatt partnership with Anthropic, Dec 2025, https://www.datacenterdynamics.com/en/news/hut-8-signs-245mw-capacity-deal-with-fluidstack-as-part-of-multi-gigawatt-partnership-with-anthropic/
- Bloomberg via Fortune, Nvidia $500 billion AI financing, Aug 2026, https://fortune.com/2026/08/11/nvidia-500-billion-ai-financing/
- JLL, North America Data Center Report Midyear 2026, Aug 2026, https://www.jll.com/en-us/insights/market-dynamics/north-america-data-centers
- CBRE, North America Data Center Trends H2 2025, Feb 2026, https://www.cbre.com/insights/books/north-america-data-center-trends-h2-2025