The market has matured. While startups previously bought H100 instances in bulk for years in advance "just in case," the workflow cycle is now divided. Training a model requires colossal peak (burst) consumption, while subsequent inference needs stable but lower power. CoreWeave's new unified billing scheme allows developers to dynamically reallocate quotas between clusters without penalties. This will lower the barrier to entry for AI startups and increase pressure on classic clouds like AWS and GCP, which are notorious for their bureaucracy when allocating GPUs.
Source: BusinessWire / CoreWeave
CloudCoreWeaveGPUInfrastructureIaaS