Neocloud AI infrastructure to capture 20% of cloud market by 2030

The rise of neoclouds forces a strategic decoupling of raw compute from the broader enterprise software stack. For streaming and AI-native firms, this provides a high-performance alternative for model training where hyperscaler margins often become prohibitive. However, the streaming ecosystem must weigh these cost savings against the complexity of fragmented security, data egress fees, and the lack of managed service guardrails found in AWS or Azure. As hyperscalers increasingly act as both customers and competitors to these firms, the market will likely bifurcate. Watch for whether neoclouds can successfully transition from simple GPU rental services into mature platforms offering global governance and security certifications to retain enterprise-grade workloads.

CoreWeave has emerged as the most prominent neocloud, reaching a $35 billion valuation after its March 2025 IPO and subsequently expanding its footprint through aggressive infrastructure deals. In May 2025, CoreWeave announced a $1.1 billion acquisition of Core Scientific's data center assets to secure additional power capacity for GPU cluster expansion, a move that underscored how neoclouds are vertically integrating to control their own supply chain rather than relying on colocation partners. Lambda, meanwhile, has pursued a different growth vector. In early 2025, Lambda raised $800 million in a Series D round led by TWG Global Partners at a $2.5 billion valuation, explicitly targeting AI training workloads for research labs and startups that need large-scale GPU access without hyperscaler lock-in. Nebius, spun out of Yandex's international cloud business, launched a dedicated AI cloud platform in late 2024 with NVIDIA H100 and H200 clusters across European data centers, positioning itself as a sovereignty-friendly alternative for EU-based AI developers.

The competitive economics between neoclouds and hyperscalers are being shaped by custom silicon programs that threaten to erode GPU rental margins. AWS has deployed Trainium2 chips across its UltraClusters, claiming up to 40% lower training costs compared to equivalent NVIDIA-based instances, while Google Cloud began shipping Ironwood TPU pods in 2025 designed for trillion-parameter model training at scale. Microsoft's Maia 100 accelerator entered limited production in late 2024, with the company confirming that Maia chips are being used internally for Azure OpenAI inference workloads. These custom silicon efforts give hyperscalers a structural cost advantage on their own platforms, compressing the price gap that neoclouds currently exploit. For streaming companies evaluating where to run recommendation models or generative AI pipelines, the calculus increasingly depends on whether multi-cluster AI infrastructure across these divergent architectures is feasible.

On the infrastructure layer supporting these GPU clusters, storage and networking vendors are positioning themselves as critical enablers of neocloud performance. VAST Data reported in Q1 2025 that its AI-optimized storage platform had been adopted by three of the top five neocloud providers for training data pipelines requiring sustained multi-terabyte-per-second throughput. WEKA has similarly targeted the neocloud segment, with the company announcing in April 2025 that its parallel file system was certified for NVIDIA DGX SuperPOD reference architectures, a designation that signals compatibility with the largest GPU cluster deployments. Backblaze, traditionally a consumer and SMB storage provider, launched a dedicated AI data lake service in March 2025 priced at $6 per terabyte per month, undercutting hyperscaler object storage pricing by roughly 70% and appealing to neocloud operators who need cost-effective cold storage for training datasets. These infrastructure plays illustrate how the neocloud ecosystem is maturing beyond raw GPU rental into a full-stack alternative, though and breadth remain gaps relative to AWS, Azure, and Google Cloud.