MTS

Wiki/Operators

title: Neoclouds description: GPU-cloud companies sitting between Nvidia/hyperscalers and end customers — the leveraged, fragile middle of the AI capex stack. last_updated: 2026-05-10

Neoclouds

"Neocloud" is the catch-all label for the GPU-cloud operators that sit between Nvidia (the chip vendor), the hyperscalers (who lease them capacity), and end-customers (who run training and inference). They are leveraged, concentrated, and strategically dependent — a fragile middle layer that has nonetheless absorbed the largest share of debt issuance in the AI capex boom. This page is the wiki's reference on the layer's structure, the public-disclosure facts on each operator, and the underexplored angles (Nvidia's circular financing, miner conversion economics, GPU-backed debt as a new asset class).

The layer at a glance (May 2026)

Operator Status Contracted backlog Largest customer (% of revenue or backlog) Operating MW Cost of secured debt
CoreWeave (CRWV) Public; IPO'd Mar 2025 $66.8bn (Dec 31 2025) Microsoft ~67% of FY25 revenue; ~35% of backlog ~850 MW (43 sites) SOFR + 2.25% / 5.9% fixed (DDTL 4.0)
Nebius (NBIS) Public (Nasdaq) Multi-year; >2 GW contracted Microsoft ($17.4bn / 5y); Meta added 2026 ~220 MW operating; >3 GW contracted n/a (mostly equity-funded so far)
Lambda Private "Multi-billion" w/ Microsoft (Nov 2025) Microsoft (Nov 2025 deal) mid-hundreds MW mixed; Nvidia GPU lease-back $1.5bn
Crusoe Private ~$998m FY25 revenue; pipeline >45 GW Microsoft, Oracle/OpenAI (Abilene) Abilene 1.2 GW (phased); + Iceland $3.4bn JV w/ Blue Owl; project finance
Applied Digital (APLD) Public ~$11bn contracted lease revenue CoreWeave (effectively 100% of AI segment) 100 MW live (Polaris Forge 1) $5bn+ project finance (Macquarie, SMBC)
IREN (IREN) Public $9.7bn w/ Microsoft (5y, 20% prepay) Microsoft Childress 750 MW pad; 200 MW IT load planned Equity + Dell vendor agreement
Core Scientific (CORZ) Public $10.2bn (12y w/ CoreWeave) CoreWeave 100% of HPC segment 590 MW HPC contracted across 6 sites Emergence-financing + reorg notes
Hut 8 (HUT) Public $7.0bn base / up to $17.7bn River Bend Fluidstack/Anthropic (Google-backstopped) 245 MW (River Bend, Q2 2027) $3.25bn IG project debt (Louisiana)
TeraWulf (WULF) Public ~$13.2bn (Lake Mariner + Abernathy) Fluidstack/Google Lake Mariner; Abernathy TX Project-level secured
Cipher Mining (CIFR) Public ~$9.3bn AWS (300 MW deal); Fluidstack/Google Barber Lake TX Project-level secured
Together AI Private; SaaS layer n/a (consumption) ARR ~$300m (Sep 2025) n/a (rents capacity) Equity ($3.3bn val Feb 2025)
Fireworks AI Private; SaaS layer n/a (consumption) ARR ~$315m (Feb 2026) n/a (rents capacity) Equity ($4bn val Oct 2025)
RunPod / Vast.ai / Voltage Park Private; long-tail n/a / smaller Diffuse SMB & researcher base Voltage Park ~35k GPUs Mixed; smaller credit facilities

Sources cited per-section below. Figures as of public disclosure through May 10 2026 (Q1 2026 earnings season).

1. CoreWeave (NASDAQ: CRWV)

CoreWeave is the bellwether. It is the only neocloud with a full public-disclosure stack (S-1, four 10-Qs, one 10-K, multiple 8-Ks) and the only one whose GPU-backed credit lines have a continuous price-discovery history.

1.1 Customer concentration

  • FY24: Microsoft was ~62% of revenue (S-1).
  • FY25: Microsoft rose to ~67% of FY25 revenue — i.e., concentration increased in the first full public year, despite management's diversification narrative.
  • Backlog mix: No single customer is more than ~35% of the $55.6bn Q3'25 revenue backlog (down from ~50% in Q2'25 and ~85% at start of 2025); >60% of backlog tied to investment-grade counterparties (Q3 earnings call).
  • OpenAI: $11.9bn / 5-year contract announced March 2025; expanded subsequently. OpenAI is the principal driver of the backlog diversification away from Microsoft.
  • Other anchors: Meta and (per management) "investment-grade hyperscalers" account for the bulk of the post-OpenAI additions.

The split between current revenue concentration (Microsoft at 67%) and backlog concentration (largest customer ~35%) is the single most load-bearing reconciliation in the CRWV credit story. Near-term cash flows are bet on a single counterparty; the forward book is genuinely more diversified — but the forward book is also priced as if every counterparty pays.

1.2 Debt structure

CoreWeave has pioneered the "Delayed Draw Term Loan" (DDTL) structure that has become the template for GPU-backed lending. Each tranche is secured by specific GPU inventory plus the customer contract it serves — not by the corporate parent.

Facility Size Year Rate Notes
DDTL 1.0 $2.3bn Aug 2023 ~15% floating Magnetar/Blackstone-led; covenant-heavy
DDTL 2.0 $7.6bn (accordion to $7.5bn announced) May 2024 ~11% variable Blackstone/Magnetar/Coatue/Carlyle/CDPQ/DigitalBridge/BlackRock
Revolver / unsecured $650m Oct 2024 n/a JPM / GS / MS / syndicate
DDTL 3.0 $2.6bn 2025 SOFR + 400 bps Tighter pricing
DDTL 4.0 $8.5bn Mar 2026 SOFR + 225 bps floating / 5.9% fixed First investment-grade GPU-backed loan (A3 Moody's / A(low) DBRS)

Total funded debt as of FY25 disclosures sits in the ~$11–13bn range (the S-1 disclosed $7.93bn at IPO; FY25 reporting and the Mar 2026 DDTL 4.0 take-out push it materially higher). Approximately 300 bps of cost-of-debt compression in under two years is the headline lender story; the underwriting question is whether SOFR + 225 bps is a fair price for a 6-year asset depreciating against a 2-year semiconductor refresh cycle.

Underexplored angle — GPU-backed debt as a new asset class. CoreWeave's DDTL 4.0 is the first IG-rated facility secured by GPUs and a single customer contract. The rating agencies effectively underwrote it as a securitisation of the underlying contract cash flows (Microsoft / OpenAI), not as a credit on CoreWeave the operating company. Recovery assumptions in a stress scenario depend critically on (a) the secondary market value of Hopper- and Blackwell-class GPUs in a downside and (b) whether the customer contract can be assigned to a new operator if CRWV defaults. Neither assumption has been tested by a cycle. Compare with the Lucent vendor-financing collapse cited by several short sellers.

1.3 GPU accounting

  • Useful life: 6 years, straight-line (S-1, unchanged through FY25 10-K).
  • Comparison: Nebius uses 4 years; AWS extended to 6 years in 2023; Meta extended to 5.5 years in 2024; Google extended to 6 years; Microsoft uses 6 years (with reductions for Hopper specifically).
  • Sensitivity: A switch from 6 → 4 years would raise Q3'25 quarterly D&A from ~$630m to ~$945m, a ~$315m hit per quarter — roughly the entire reported operating profit line. This is the single most-discussed accounting choice in the sector (Michael Burry's late-2025 short thesis is built on it).

1.4 Unit economics

  • Q3'25 revenue: $1.36bn (+134% YoY).
  • GAAP gross margin: ~74% (but cost of revenue excludes GPU depreciation).
  • Operating margin: ~4% reported; contract-level EBIT margin estimated at ~35%; contract-level unlevered IRR estimated at ~20%.
  • Active power: 590 MW (Q3'25) → ~850 MW (year-end 2025) across 43 data centers; 2.9 GW contracted.
  • Revenue density: roughly $1.36bn / 700 MW average active = **$2m / MW / quarter**, or ~$8m / MW / yr at current utilisation — useful as a back-of-envelope and a yardstick against Applied Digital's $11bn / 400 MW / 15-yr lease revenue ($1.8m / MW / yr to the landlord).

2. Nebius (NASDAQ: NBIS)

Nebius is the Yandex N.V. remnant after the 2024 sanctions-driven split: Yandex's Russian assets were sold; Nebius retained the data-centre operations, the autonomous-vehicle business (Avride), the Toloka data labelling business, and the ClickHouse-related stake. It re-listed on Nasdaq in October 2024.

  • Microsoft deal (8 Sept 2025): Up to $17.39bn over 5 years, including ~$6.96bn upfront, with options to $19.4bn. Delivered in 9 tranches across 2025–26 from a new Vineland, NJ data centre. The upfront prepayment is unusually large and economically resembles vendor financing of Nebius's GPU buy.
  • Meta deal (early 2026): Combined with Microsoft, Nebius now has ~$46bn in disclosed multi-year hyperscaler commitments.
  • Capacity: >2 GW contracted as of Feb 2026 reporting; guidance raised to >3 GW contracted by year-end 2026; 800 MW – 1 GW expected connected during 2026.
  • Nvidia equity: NVDA invested ~$2bn in Nebius (Jan 2025) — one of the canonical "circular financing" data points.
  • Financials: Q4'25 revenue $228m (+547% YoY); annualised run rate $1.2bn exiting 2025; 2026 guidance $3.0–3.4bn revenue, ~40% EBITDA margin, $16–20bn capex.
  • GPU useful life: 4 years (vs. CoreWeave's 6).

The Nebius and CoreWeave depreciation gap is the cleanest natural experiment in neocloud accounting: same chips, same vendors, different policy. At equivalent fleet sizes, Nebius's reported D&A burden will be ~50% higher per dollar of GPU on the balance sheet.

3. Lambda

Private. The most secretive of the top-tier neoclouds.

  • Series D (Feb 2025): $480m at ~$2.5bn valuation; Nvidia participated.
  • Series E (Nov 2025): ~$1.5bn led by TWG Global and USIT at ~$5.9bn valuation.
  • Microsoft deal (Nov 2025): Multi-billion, multi-year contract for tens of thousands of GPUs (size undisclosed but reported as "multibillion-dollar").
  • Nvidia lease-back: Nvidia leases back ~18,000 GPUs from Lambda for $1.5bn — Nvidia is simultaneously equity investor, GPU vendor, and lease customer. This is the structural pattern that has driven the "circular financing" critique.
  • Run rate: ~$760m ARR exiting 2025 (per Sacra).
  • Customer base: 5,000+ customers, more diversified than CoreWeave; manufacturing, financial services, federal.

4. Crusoe

Originated as a stranded-gas / flared-gas bitcoin miner; pivoted aggressively to AI hosting from 2023 onward and is now a top-five neocloud by megawatt pipeline.

  • Stargate Abilene: Crusoe-developed, on Lancium's Clean Campus. Initial phase ~200 MW live by Q3 2025; full design 1.2 GW. Oracle is the lease counterparty; OpenAI is the workload customer. JPMorgan led a $2.3bn loan (May 2025) for the Abilene buildout.
  • Abilene plans (post-October 2025): Oracle/OpenAI cancelled a planned 600 MW expansion at Abilene; the broader 4.5 GW Oracle-OpenAI commitment was redistributed to other sites (Detroit, Wisconsin). Meta is reportedly in talks (with Nvidia's help) to absorb some of the freed capacity.
  • Microsoft (Dec 2025): 900 MW additional AI factory at Abilene specifically for Microsoft workloads. Total Abilene planned capacity ~2.1 GW.
  • Revenue: ~$276m (2024) → ~$998m (2025E); ~17x YoY growth in added contract value; cloud ARR up ~150% YoY.
  • Series E (Oct 2025): $1.38bn at >$10bn valuation. Lead investors: Valor Equity Partners, Mubadala. Others: Nvidia, Altimeter, BAM Elevate, Founders Fund, Fidelity, Salesforce Ventures, Super Micro.
  • Project debt: $3.4bn JV with Blue Owl Capital + Primary Digital Infrastructure (Oct 2024). Project-level financing rather than corporate.
  • Power pipeline: >45 GW as of Q1 2026 disclosure.

Crusoe's Abilene cancellation is the single most concrete data point that the contracted pipeline is not immutable. A hyperscaler that wants to walk a 600 MW expansion can walk it; the operator is left holding the site.

5. Applied Digital (NASDAQ: APLD)

Former crypto miner that converted its North Dakota footprint into AI hosting. The cleanest single-customer-concentration case study in the sector.

  • CoreWeave leases at Polaris Forge 1 (Ellendale, ND):
    • May 2025: two ~15-year leases totalling 250 MW; ~$7bn contracted revenue.
    • Aug 2025: additional 150 MW lease ("Building 4"); brings total to 400 MW / $11bn contracted.
  • Operational status: First 100 MW building energised in Q1 2026 (Phase II "Ready for Service"). Second 150 MW building ready mid-2026. Third 150 MW ready 2027.
  • Customer concentration: Effectively 100% of the AI segment is leased to CoreWeave. APLD is therefore a de facto downstream credit on CoreWeave, which is in turn a downstream credit on Microsoft and OpenAI. The chain is APLD → CoreWeave → Microsoft/OpenAI.

Underexplored angle — single-customer credit chains. APLD's $11bn of contracted lease revenue is only as good as CoreWeave's ability to pay rent, which is only as good as CoreWeave's collections from Microsoft and OpenAI, which is only as good as those firms' continued willingness to deploy compute through CoreWeave specifically (rather than insource). A default at any layer is a default at every layer below it. This is the structure regulators, rating agencies and short sellers will eventually stress-test.

6. IREN (NASDAQ: IREN)

Australia-listed → Nasdaq-listed, formerly Iris Energy. Bitcoin miner that has reframed as an AI infrastructure operator (the rebrand to "IREN" was explicit signalling).

  • Microsoft deal (3 Nov 2025): $9.7bn over 5 years, 20% upfront prepayment.
  • GPU: Nvidia GB300 capacity, deployed in 4 × 50 MW phases through 2026 at Childress, TX.
  • Site: Childress 750 MW site, of which 200 MW IT load is dedicated to the Microsoft contract.
  • Dell vendor agreement: ~$5.8bn for GPUs and ancillary equipment — a notable shift in the vendor stack (Dell rather than Supermicro).
  • Funding: Largely equity-funded to date; the Microsoft prepayment is doing a meaningful share of working capital duty.

7. Core Scientific (NASDAQ: CORZ)

Post-Chapter-11 (emerged Jan 2024) bitcoin miner. Pivoted hard into hosting.

  • CoreWeave master deal (Feb 2025 expansion): ~590 MW HPC critical IT load across six sites; $10.2bn contracted revenue over 12-year terms.
  • Capital: $1.5m/MW shell/powered-core capex funded by CORZ; balance funded by CoreWeave.
  • Bitcoin exit: Sold $208m of BTC in Q1 2026 to fund AI capex.
  • Customer concentration: 100% of HPC contract pipeline = CoreWeave.

The CORZ-CRWV pair therefore form a closed economic loop nearly identical to APLD-CRWV, layered on top of the CRWV-MSFT loop.

8. Bitcoin miners converting to AI hosting

The bitcoin-miner-to-AI-hoster pivot is the most distinctive feature of the 2025 neocloud cohort. The basic trade: miners have stranded power (often on the order of 200–700 MW per site, secured years ago for ASIC mining), and AI compute desperately needs power. Match them, and the miners' equity re-rates.

Hut 8 (NASDAQ: HUT) — River Bend, LA

  • 15-year, 245 MW lease w/ Fluidstack (announced Dec 2025).
  • Base contract value $7.0bn; renewal options to $17.7bn. 3.0% annual base rent escalator; triple-net.
  • Fluidstack ROFO on up to 1,000 MW additional capacity. 330 MW utility capacity already secured (Entergy), expandable to 1.33 GW.
  • Anthropic is the workload customer; Google backstops Fluidstack's payment obligations.
  • $3.25bn investment-grade project debt closed for the Louisiana site.
  • First data hall: Q2 2027 commissioning.

TeraWulf (NASDAQ: WULF)

  • Abernathy, TX: $9.5 billion, Google-backed Fluidstack agreement.
  • Lake Mariner, NY: $3.7bn in earlier deals.
  • Combined ~$13.2bn contracted backlog locking up otherwise modest-scale megawatts at hyperscaler-credit pricing. WULF's stock outperformed BTC by ~70% in 2025 on contract announcements.

Cipher Mining (NASDAQ: CIFR)

  • ~$9.3bn HPC backlog. 300 MW AWS deal (direct, not through Fluidstack) plus a Google-backstopped Fluidstack agreement at Barber Lake TX. CIFR has exited most of its remaining bitcoin operations.

Underexplored angle — are mining sites actually fit-for-purpose for AI? The miner-pivot thesis assumes that "power + land + fibre" is the binding constraint and the buildings themselves are commodity. In practice: ASIC mining halls run very high PUE (1.3+ for forced-air), have minimal redundancy (single feed common, no Tier-III concurrent-maintainability), and were sited for cheap power rather than fibre routes. Liquid cooling for Blackwell-class hardware needs 60–80 kW/rack and chilled-water loops the original mine hall never had. The new hosting deals are almost universally describing new buildings on the site, not the retrofit of existing mining halls. Net: the AI revenue is on the power interconnect, not on the mining infrastructure. The historical mining capex is, to first order, stranded.

9. Smaller / second-tier

Operator Niche Notable 2025 data point
Together AI Inference-as-a-service / fine-tune $533.5m raised; ~$3.3bn val (Feb 2025); $300m ARR (Sep 2025)
Fireworks AI Inference platform $250m Series C at $4bn val (Oct 2025); $315m ARR (Feb 2026)
RunPod Bare-metal GPU rental, low-cost spot Private; no public valuation disclosed in 2025 cycle
Vast.ai Decentralised GPU marketplace SMB / researcher-facing; long-tail of supply
Voltage Park Mid-market AI cloud ~24k H100s; merged with Lightning AI (mid-2025) → 35k+ GPUs combined
Fluidstack Broker / operating layer Counterparty for HUT, WULF, CIFR deals; effectively a Google front-end for capacity acquisition

Together and Fireworks sit one layer up the stack — they are consumption companies that rent capacity from the layer below. They are exposed to the same demand cycle but with negligible capex and no GPU-backed debt.

Fluidstack is worth flagging specifically: it is doing the broker / operator function for a sequence of miner-AI deals (Hut 8, TeraWulf, Cipher) with Google as the ultimate paymaster. Fluidstack's emergence as the standard miner-side counterparty, with Google standing behind, is one of the more under-reported structural facts of the 2025 cycle — Google has effectively built a third-party AI hosting capacity stack that competes with CoreWeave but routes through publicly traded miner balance sheets.

10. Nvidia's circular financing

The taxonomy of NVDA-funded GPU demand:

Mechanism Example Magnitude
Direct equity into customer NVDA → CoreWeave ($2bn, Jan 2025); NVDA → Nebius ($2bn, Jan 2025); NVDA → Lambda (Series D, Feb 2025); NVDA → Crusoe (Series E, Oct 2025) >$40bn NVDA AI equity commitments in first 4 months of 2026 (incl. $30bn into OpenAI)
Capacity-purchase agreements NVDA → CoreWeave $6.3bn CPA: NVDA itself commits to buy CRWV's capacity through 2032 Single largest data point; most-cited "Lucent comparison"
Lease-back NVDA leases ~18,000 GPUs back from Lambda $1.5bn
Lease guarantees NVDA backstops $860m of a partner data centre's lease obligations (reported Q1 2026) $0.86bn

The accounting question. None of this is fraud. The disclosure question is: of NVDA's revenue from these neoclouds, how much represents genuinely third-party demand vs. demand that NVDA itself financed (via equity, capacity commitments, or guarantees)? On the most aggressive reading, multi-billion dollars of NVDA's reported quarterly data-centre revenue is currently being recognised against counterparties that NVDA has invested into within the past 12 months. On the most conservative reading, NVDA's equity stakes are small relative to the contracts and the counterparties have independent investment-grade demand drivers (Microsoft, Meta, OpenAI).

The shape of the question — what fraction of NVDA's revenue is downstream-financed by NVDA itself — is the right one for the cycle. The number is not yet observable, because (a) NVDA does not separately disclose revenue per neocloud and (b) the neoclouds disclose customer concentration but not vendor concentration. The first rigorous attempt at quantifying this will be a watershed for the sector. Until then, this is the single biggest unresolved question on the demand side of the AI capex cycle.

11. Underexplored angles — summary

  1. Forward-book vs. current-revenue concentration gap (CoreWeave). Microsoft is 67% of FY25 revenue but only ~35% of the backlog. The cash collection risk of the next four quarters is concentrated; the underwriting case for the next ten years is diversified. The two are not the same risk.
  2. Credit chains. APLD → CoreWeave → Microsoft. CORZ → CoreWeave → Microsoft. Hut 8 → Fluidstack → Anthropic → Google (backstop). Default propagates downward through these chains; recovery analysis at each link assumes the chain holds.
  3. GPU-backed debt as a new asset class. DDTL 4.0 is the first investment-grade rating. Recovery assumptions are untested. The asset is depreciating at 4–6 year accounting life against a 2-year semiconductor refresh cycle.
  4. Mining-site retrofit economics. Most disclosed miner-AI deals are new buildings on existing sites, not retrofits of mining halls. The historical mining capex is functionally stranded; the real asset is the power interconnect.
  5. Nvidia's circular financing. Equity investment + capacity commitment + lease guarantees, all flowing into the same counterparty whose GPU purchases are NVDA's reported revenue. Quantification is the open research problem.
  6. Depreciation policy as the single biggest accounting choice. CoreWeave 6yr vs. Nebius 4yr is the cleanest comparison; the difference is ~50% of D&A. The right answer depends on whether Blackwell + successor generations command pricing parity with Hopper at year 5–6 — for which there is no historical evidence either way, because no GPU generation has ever been operated to year 6 in a production AI workload.

Sources

CoreWeave

Nebius

Lambda

Crusoe

Applied Digital

IREN

Core Scientific

Bitcoin-miner conversions

Smaller operators

Nvidia circular financing

Stargate / OpenAI