The Neocloud Landscape in 2026: CoreWeave, the Debt Machine, and Who Buys GPU Clouds

"Neocloud" — the GPU-specialist cloud providers that grew up outside the hyperscaler triopoly — went from curiosity to systemically interesting in about eight quarters. CoreWeave, the category's flagship, now reports numbers that read like a utility crossed with a leveraged buyout: over 1 GW of active power, 3.5+ GW contracted, and a $99.4 billion revenue backlog against $2.078 billion in quarterly revenue. Here's how the category actually works, and how to buy from it. The Neocloud Business Model in One Paragraph Sign an anchor tenant to a multi-year committed contract; borrow against that contract to build the data center; deliver capacity; repeat. CoreWeave's $8.5 billion non-recourse delayed-draw facility (SOFR+2.25% floating / ~5.9% fixed) is the template: debt secured by contracted revenue, not the corporate balance sheet. The model works exactly as long as anchor tenants keep paying and keep renewing. The Customer Concentration Question The anchor list is short and heavy. Meta initially committed up to ~$21 billion to CoreWeave through 2032. Jane Street signed a $6 billion AI cloud agreement (and put $1 billion into CoreWeave equity). Perplexity runs its inference on CoreWeave under a multi-year deal. This is the category's core risk and its core strength simultaneously: revenue visibility measured in years, dependency measured in a handful of logos. CoreWeave's balance sheet shows the leverage side — roughly $25 billion of combined current and non-current debt. Why Buyers Choose Neoclouds Over Hyperscalers 1. Availability and time-to-capacity — the founding value proposition still holds for large training and inference blocks. 2. Price per GPU-hour — typically undercuts hyperscaler on like-for-like hardware, especially on committed terms. 3. Specialist stack — CoreWeave has moved up-stack: Sandboxes (an execution layer for RL, agent tool use, and evals, integrated with Weights & Biases) signals the category's shift from "GPUs by the hour" to "AI infrastructure platform." Its embrace of llm-d and portability messaging targets the enterprise fear of single-environment lock-in. Why Buyers Don't Hyperscalers answer with co-location advantages: your data is already in S3/BigQuery/Azure; marketplace procurement is one signature; and their custom silicon changes the price floor — Amazon's chip business passed a $20 billion run rate, and OpenAI itself committed to ~2 GW of Trainium capacity. For inference tightly coupled to existing data estates, the hyperscaler usually wins on gravity, not price. The Second Tier and the Sovereign Angle Below the flagship: regional and sovereign-flavored capacity is proliferating — Oracle's Alloy-based sovereign builds in Japan, Mistral standing up its own 10 MW inference facility near Paris to control capacity directly, and national initiatives like SK's planned 2 GW NVIDIA-based AI factory. The pattern: compute is becoming a jurisdictional asset, and "where does my inference physically run" is now a procurement question with a real answer. Buyer's Checklist - Match commitment length to workload certainty; the discounts are real, and so is the lock-in. - Ask for the provider's debt and anchor-tenant structure — their financing risk is your continuity risk. - Test egress and portability *before* signing; portability marketing ≠ portability. - For data-gravity workloads, price the hyperscaler's committed tiers honestly before assuming the neocloud is cheaper.