In March 2025, CoreWeave listed on Nasdaq positioning itself as a dedicated AI hyperscaler. Regulatory disclosures in its IPO registration highlighted a business model anchored in multi-year take-or-pay customer contracts, with Microsoft accounting for 62% of fiscal 2024 revenue.
Unlike legacy hyperscalers with diversified product catalogs and millions of enterprise accounts, specialized GPU clouds operate as capital-intensive infrastructure providers underpinned by asset-backed borrowing and long-term counterparty commitments.
Tap to enlargeDefining the neocloud architecture
The term "neocloud" designates specialized infrastructure providers focused primarily on high-density GPU compute and high-bandwidth networking (such as CoreWeave, Lambda, Crusoe Cloud, and Nebius), distinguishing them from general-purpose public cloud platforms:
- Hardware Specialization: Rather than offering broad SaaS suites, object storage ecosystems, and managed databases, neoclouds optimize specifically for large-scale model training and high-throughput inference clusters.
- Take-or-Pay Long-Term Contracts: Customers reserve dedicated GPU clusters for multi-year horizons (typically 3 to 5 years), committing to recurring payments regardless of runtime cluster utilization.
- Counterparty-Backed Financing: Debt facilities for hardware procurement are secured directly against these multi-year customer contracts, allowing lenders to underwrite loans based on the investment-grade credit ratings of the anchor enterprise clients.
Regulatory filings and revenue concentration
Public filings from CoreWeave provide detailed insight into the unit economics of the sector:
- Revenue Trajectory: Annual revenue grew from $16 million in 2022 to $229 million in 2023, $1.915 billion in 2024, and $5.131 billion in 2025.
- Operating Losses & Interest Expenses: CoreWeave reported net losses of $863 million in 2024 and $1.167 billion in 2025, with net interest expenses reaching $1.229 billion in 2025 alongside adjusted EBITDA of $3.093 billion.
- Client Concentration: In 2024, the top two customers generated 77% of total revenue. Committed long-term contracts represented 98% of revenue in early 2026 filings, with on-demand consumption accounting for the remaining 2%.
Tap to enlargeCredit structures and delayed-draw facilities
To fund extensive capital expenditure without excessive equity dilution, neoclouds utilize Delayed-Draw Term Loans (DDTLs) and special purpose vehicles (SPVs):
- In March 2026, CoreWeave closed an $8.5 billion financing facility (DDTL 4.0) through an isolated subsidiary, achieving investment-grade credit ratings (Moody's A3, DBRS Morningstar A-low) secured directly by high-performance computing assets and associated long-term customer contracts.
- Similarly, Nebius disclosed a multi-year infrastructure agreement with Microsoft valued between $17.4 billion and $19.4 billion through 2031, using the counterparty's credit profile to secure low-cost debt for cluster buildouts. Nebius also entered into an infrastructure agreement with Meta featuring a $12 billion base commitment and up to $15 billion in optional expansion tranches.
Structural comparison: Data center REITs vs. neoclouds
| Dimension | Powered Shell & Real Estate (e.g. Applied Digital, Hut 8) | Specialized Neocloud (e.g. CoreWeave, Nebius) |
|---|---|---|
| Primary Asset | Land, high-voltage substations, building shells | GPU server racks, InfiniBand/RoCE networking |
| Lease Duration | 10 to 15 years | 3 to 5 years |
| Depreciation Horizon | 20 to 30 years (structural assets) | 3 to 5 years (rapid silicon obsolescence) |
| Residual Value Risk | Low (real estate and power capacity retain value) | High (older GPU generations reprice downward) |
| Customer Obligation | Power availability and facility cooling | Cluster uptime, interconnect bandwidth, SLA |
Capital intensity and infrastructure limits
By early 2026, CoreWeave reported operating over 1 GW of active power with several gigawatts under long-term utility reservation, while IREN disclosed over 800 MW of operational capacity alongside a planned 480 MW AI cloud expansion.
The primary operational constraint across the ecosystem remains physical grid connection lead times, transformer manufacturing schedules, and high-density liquid cooling integration.
The long-term economic sustainability of the neocloud model depends on counterparty contract fulfillment, capital cost management across interest cycles, and the residual market value of hardware assets as new accelerator generations enter the market.