CoreWeave and the Infrastructure Moat: Are GPU Clouds Defensible?

CoreWeave's S-1 disclosed a business that was, at the time, doubling and redoubling on itself: revenue up roughly 7x year-over-year, built almost entirely on NVIDIA GPU clusters rented out to a small number of very large customers, with Microsoft as the anchor. That concentration is the whole bull and bear case in one sentence — it is both how CoreWeave got here and the single biggest risk to the story.

The bull case is architectural specialization: CoreWeave built its stack around GPU workloads from day one — bare-metal provisioning, InfiniBand networking, and a scheduler tuned for large training runs — while the hyperscalers retrofit that capability onto clouds designed for general-purpose compute. In a supply-constrained GPU market, being first in line with NVIDIA and easier to deploy against than a hyperscaler's slower internal process is a real, if temporary, edge.

The bear case is that this is an arbitrage on GPU scarcity, not a durable moat. AWS, Google Cloud, and Azure are all expanding dedicated AI infrastructure capacity, and they have balance sheets, existing enterprise relationships, and captive demand from their own AI products that CoreWeave does not. Once GPU supply catches up with demand, the pricing power that comes from being one of the only places to get H100 or Blackwell capacity at scale should compress.

Our read: CoreWeave's moat is real for as long as the GPU shortage persists and gets more fragile every quarter that hyperscaler capacity comes online. The customer concentration is the more immediate risk — losing or renegotiating a single anchor contract would move the model far more than any competitive dynamic. This is a name where the thesis has a shelf life, not a multi-decade compounder story.