The Neocloud Arbitrage: Inside the Infrastructure Economics of CoreWeave and Nebius
Hyperscalers are offloading capital expenditure to specialized neoclouds through over $120 billion in long-term compute commitments, shifting balance-sheet capex to operating expenses. To support this demand, these unprofitable neoclouds rely on circular financing backed by Nvidia equity and debt to fund rapid power and hardware deployments.
The Capex-to-Opex Offload
Hyperscalers are utilizing specialized neocloud providers to restructure their balance sheets and defer massive upfront capital expenditures. Meta, facing estimated 2026 capital expenditures of $125 billion to $145 billion against $136 billion in projected operating cash flow, risks operating at a negative free cash flow. By entering into long-term compute agreements with neoclouds, Meta and Microsoft shift these expenditures from balance-sheet capex to long-term operating expenses.
Microsoft has committed approximately $60 billion to neoclouds including CoreWeave, Nebius, and Nscale. Meta has structured up to $62.2 billion in agreements, comprising $35.2 billion to CoreWeave and up to $27 billion to Nebius. Combined with deals from OpenAI and Anthropic, total hyperscaler commitments to these specialized providers surpass $145 billion. This total commitment is equivalent to roughly 90% of AWS's trailing twelve-month revenue, highlighting the scale of the outsourcing.
Infrastructure Scaling and Power Allocations
The primary constraint on scaling these commitments is physical infrastructure, specifically power delivery. CoreWeave and Nebius have each secured 3.5 GW of contracted power capacity, though the majority of this footprint has yet to come online. CoreWeave is targeting 1.7 GW of active power by the end of 2026, while Nebius aims to bring 800 MW to 1 GW of connected power online in the same timeframe.
To convert these backlogs into active revenue, neoclouds must minimize deployment latency. While hyperscalers require years for greenfield data center builds, neoclouds deploy high-density GPU clusters within months. CoreWeave claims deployment times as short as two weeks from hardware receipt, having rapidly fielded clusters using Nvidia H100, H200, GH200, and GB200 NVL72 architectures, as well as the Vera Rubin system.
Software-Enabled Compute Optimization
To differentiate from general-purpose public clouds, neoclouds leverage specialized software layers to maximize Model FLOPs Utilization (MFU) and bridge the 30% to 40% efficiency gap common in large-scale AI workloads. CoreWeave employs CoreWeave Kubernetes Service (CKS) to coordinate workload distribution across multi-thousand GPU clusters.
To mitigate idle time and optimize resource allocation, CoreWeave uses SUNK to co-schedule training and inference workloads on shared clusters, alongside CoreWeave Tensorizer for high-speed model loading. These software optimizations allowed CoreWeave to report an initial MFU of 35% to 45%—approximately 20% higher than average competitor baselines of 30%—and subsequently achieve over 50% MFU on Hopper GPU architectures.
The Circular Capital Loop
The rapid expansion of neocloud infrastructure is highly leveraged and relies on circular capital flows. Neither CoreWeave nor Nebius has achieved profitability, operating with constrained cash flows and escalating debt. To sustain their capital expenditure, both companies rely on financial backing from Nvidia, which has invested $2 billion in each provider.
Under these strategic partnerships, both neoclouds plan to deploy over 5 GW of data center capacity by 2030. This financing structure creates a closed loop: Nvidia provides equity investments and backstops to the neoclouds, who secure GPU-backed debt to purchase Nvidia hardware, which is then leased to hyperscalers to generate the revenue necessary to service the debt.