Nvidia's Neocloud Loop: How Circular Financing Props Up CoreWeave and Nebius

Neoclouds like CoreWeave and Nebius have become central to the AI infrastructure race by offering hyperscalers something they struggle to build themselves: fast access to the newest Nvidia GPUs and higher compute utilization. Both firms deploy cutting-edge hardware within weeks—CoreWeave was first to run a Vera Rubin system and first to ship GB200 NVL72 instances—and layer on software (Kubernetes orchestration, combined training/inference clusters, fast model loading) that pushes their model FLOPs utilization past 50% on Hopper, well above the ~30% typical elsewhere. Microsoft and Meta have committed a combined $122 billion-plus to these providers, with OpenAI and Anthropic deals pushing total potential commitments past $145 billion, even though CoreWeave’s 2026 revenue is projected at just $12.6 billion and Nebius at $3.4 billion.

Beyond speed and efficiency, a quieter driver is accounting: leasing neocloud capacity lets hyperscalers shift spending off the balance sheet from capex to opex, avoiding recognizing massive capital outlays directly. That structural incentive helps explain why Big Tech is funneling roughly 90% of AWS’s trailing revenue into long-term capacity deals with a relatively unproven business model.

The bearish case is financial fragility. CoreWeave and Nebius are racing to convert huge contracted power footprints (3.5 GW each, mostly not yet online) into revenue, an extremely capital-intensive process they fund through GPU-backed debt rather than the operating cash flow Big Tech enjoys. The most striking red flag is circularity: Nvidia recently invested $2 billion in each firm and effectively backstops the buildout, meaning the chipmaker is helping finance the customers buying its chips. In a tightening macro environment, soaring debt loads and this self-referential funding structure raise real questions about how durable the current surge actually is.