CoreWeave and Nebius lead specialized neocloud shift away from hyperscalers

The emergence of specialized providers means streaming infrastructure is no longer a monolithic purchase from AWS or Google. Media companies must now navigate a federated control plane where hyperscalers manage data and governance while neoclouds handle the concentrated industrial work of AI inference and training. This shift transforms technical operations into a financial discipline, requiring firms to hedge capacity and manage take-or-pay contracts like energy commodities. As the market matures, the competitive advantage will move from simply owning GPUs to mastering the orchestration and power pipelines. Watch for whether CoreWeave can diversify its revenue beyond the three customers that currently drive 72% of its business.

CoreWeave has rapidly expanded its infrastructure footprint through aggressive capital deployment and customer concentration. In March 2026, CoreWeave completed its acquisition of Core Scientific's data center assets for approximately $9 billion, a deal that gave the neocloud direct control over power capacity in key US markets rather than relying on colocation agreements. The company's customer base remains heavily concentrated, with Microsoft accounting for a dominant share of contracted revenue, a dependency that investors have flagged as a structural risk since the company's March 2025 IPO. Nebius, the AI infrastructure spinoff from Yandex, has taken a different approach by building vertically integrated data centers in Europe and the Middle East, announcing a $2 billion expansion of its GPU cluster capacity across Finland and the Netherlands in May 2026, targeting sovereign AI workloads that require data residency guarantees hyperscalers cannot always provide under local regulation. The business model divergence between neoclouds and hyperscalers is becoming clearer as enterprises evaluate total cost of ownership for AI training and inference. AWS responded to competitive pressure by launching its Trainium3 UltraCluster offering in April 2026, which bundles custom silicon with managed networking at prices aimed at undercutting GPU-based neoclouds on per-token inference costs. NVIDIA, whose H100 and B200 GPUs form the backbone of both CoreWeave and Nebius fleets, has maintained allocation leverage by prioritizing its own DGX Cloud and select partners, creating a supply constraint that limits how quickly neoclouds can scale without NVIDIA's cooperation. Applied Digital, a smaller data center developer that has positioned itself as a power-first infrastructure provider, signed a 15-year lease agreement with CoreWeave for a 400-megawatt campus in North Dakota in February 2026, illustrating how the neocloud model depends on securing long-term power commitments that traditional cloud providers historically managed internally. Technical differentiation among neoclouds centers on network topology and GPU interconnect density rather than raw compute alone. CoreWeave's architecture uses NVIDIA's NVLink and InfiniBand fabric to create tightly coupled clusters optimized for large-scale distributed training, a design that delivered 1.8x faster training throughput on Llama-class models compared to equivalent GPU counts on general-purpose cloud instances in benchmarks published by SemiAnalysis in June 2026. Neocloud vertical integration has become a broader industry trend as providers seek to control the full stack from hardware to software. Nebius has emphasized its proprietary scheduling layer, which dynamically reallocates GPU memory across jobs to reduce idle time, claiming utilization rates above 85% versus the 60-70% typical of hyperscaler shared pools. For streaming and media companies evaluating these platforms for video generation, real-time transcoding, and recommendation model training, the choice between neocloud and hyperscaler increasingly hinges on whether workloads require sustained multi-thousand-GPU jobs or bursty inference with tight latency SLAs.