CoreWeave

CoreWeave started in 2017 as a cryptocurrency mining operation that accumulated a substantial fleet of NVIDIA GPUs, then pivoted in 2019 to repurpose that hardware as cloud GPU compute for machine learning workloads as crypto profitability declined. The timing proved prescient: demand for GPU compute for AI training and inference exploded in 2022 and 2023, and CoreWeave — with large inventories of H100, A100, and other NVIDIA hardware already deployed — was positioned to serve a market where AWS, Azure, and GCP had multi-month GPU waitlists. The company raised over $2 billion in funding and signed significant contracts with Microsoft and other major AI players.

CoreWeave's infrastructure is purpose-built for GPU compute workloads, with Kubernetes as the primary interface for cluster management and job scheduling. Customers deploy GPU workloads through Kubernetes manifests, with CoreWeave managing the underlying hardware, networking, and storage fabric. The platform includes InfiniBand networking between GPU nodes for high-bandwidth, low-latency inter-node communication — critical for distributed training workloads where gradient synchronisation latency directly affects training throughput. The CoreWeave Cloud console and API provide cluster management, while the NVIDIA GPU Operator and associated tooling manages GPU resource allocation within Kubernetes.

CoreWeave service disruptions have significant implications for AI and ML customers who are running large, expensive training runs. GPU node failures that remove nodes from a Kubernetes cluster mid-training cause training job interruptions that require checkpoint recovery — a process that works only if the workload has checkpoint logic implemented and the checkpoint storage is accessible. InfiniBand fabric degradation increases inter-node communication latency, manifesting as reduced training throughput that does not stop the job but significantly extends the expected training time and cost. API provisioning failures prevent new cluster creation or node pool scaling, blocking customers who need to expand capacity for inference serving or kick off new training experiments. Storage I/O degradation on the shared storage layer affects checkpoint read and write performance.

Outage.gg tracks CoreWeave service status using real-time community reports from ML engineers and AI infrastructure teams. If GPU cluster provisioning is failing, training jobs are crashing due to node failures, or the CoreWeave API is unavailable, the live status page shows current impact from the CoreWeave user community.