Switching Costs of Running AI WorkloadsThin moat

Beyond the contracts, CoreWeave benefits from real switching costs: once a customer has its models, data, pipelines, and workflows running on CoreWeave's infrastructure, moving them to another provider involves genuine cost, disruption, and risk. Migrating large-scale AI workloads means re-provisioning capacity elsewhere, moving enormous datasets, re-validating performance, and risking downtime on production systems — friction that tends to keep customers in place through a contract term and gives CoreWeave a hold beyond the pure legal commitment. These switching costs reinforce the backlog and are a legitimate, if modest, source of stickiness.

But the switching costs in AI cloud are weaker than in many software businesses, and weakening further, which limits their value as a moat. AI workloads increasingly run on portable, standardized software layers (common frameworks, containerization, orchestration) specifically designed to reduce lock-in and let workloads move between providers — the industry, wary of cloud lock-in, is actively building for portability. And the customers with the most at stake — the hyperscalers and best-resourced labs — are precisely the ones most capable of moving workloads or of building their own capacity, and most motivated to avoid dependence on a single provider. The switching costs hold a customer through a contract term but reopen at renewal, when a customer that has developed alternatives or its own infrastructure is free to leave. So the switching costs are real and add genuine stickiness within a contract, reinforcing the backlog. But they are moderate rather than strong, actively eroded by an industry building for portability, and weakest precisely against the largest customers who matter most — a real but limited source of lock-in that supports the thin moat without making it durable — even ten $1B-plus clients can leave at renewal¹.