Specialized GPU cloud providers CoreWeave and Nebius have secured $122.2 billion in long-term commitments from Microsoft and Meta alone — commitments that dwarf their projected FY2026 revenues of $12.6 billion and $3.4 billion respectively. Hyperscalers are betting on neoclouds for three reasons: weeks-not-years deployment of H100, H200, and GB200 clusters, higher GPU utilization via proprietary orchestration software, and the ability to shift capital expenditure to operating expense. CoreWeave claims it can deliver new Nvidia architectures to customers within two weeks of receipt, and recently brought a Vera Rubin system online in early June.

Beyond raw hardware access, neoclouds layer software that materially improves model FLOPs utilization (MFU). CoreWeave's Kubernetes Service (CKS), SUNK scheduler, and Tensorizer model loader push MFU to 35–45% on current fleets — roughly 20 points above competitors — with recent blog posts citing >50% on Hopper GPUs. The MFU metric captures whether kernels actually saturate GPU cores, not merely whether a kernel is executing. This efficiency gain translates directly to lower effective compute cost per training run.

The model rests on 3.5 GW of contracted power capacity per company, though most remains offline. CoreWeave targets 1.7 GW active by end-2026; Nebius aims for 800 MW to 1 GW. Converting contracted to live capacity requires massive capital that neither firm generates internally. Nvidia's $2 billion equity investments in each create circular financing: GPU vendor funds customer who buys GPUs, backed by hyperscaler contracts that may not materialize if AI demand softens.

The durability question centers on whether hyperscaler commitments represent genuine long-term demand or a temporary capex offload. If Microsoft and Meta internalize GPU procurement once their own data centers come online, neoclouds face stranded assets and debt service on underutilized clusters. The MFU advantage and deployment speed are real technical moats, but they may not sustain valuations built on order-of-magnitude revenue multiples.