Kubernetes portability reduces GPU lock-in
Samiur Rahman, CEO of Heyday, on building a production-grade AI stack
This reveals how thin the switching costs are in GPU cloud when the product is delivered through standard Kubernetes and containers. CoreWeave won Heyday because it looked enough like the team’s existing AWS setup to move fast, not because it locked Heyday into a proprietary stack. The same portability means a customer can shift back to AWS, or elsewhere, as soon as price, reliability, or centralization becomes more attractive.
-
In practice, Heyday is saying its ML workloads are packaged as Docker images and already run on Kubernetes, so moving clouds mostly means pointing those containers at a different cluster and changing the infrastructure around them. That is why the migration estimate is measured in days, not quarters.
-
CoreWeave’s edge for this customer is managed operations, not unique application primitives. It provided autoscaling, VPC support, public API exposure, and production ready GPU infrastructure, while Lambda was cheaper for experimentation and AWS was more reliable but still priced far higher for H100 era capacity at the time.
-
This is a common pattern in GPU cloud. CoreWeave has scaled far ahead of Lambda Labs, with an estimated $2B of 2024 revenue versus Lambda at $600M, but that lead comes from supply, financing, and operating maturity. Those are real advantages, but they do not remove the customer ability to rehome standard Kubernetes workloads.
Going forward, GPU cloud leaders will keep winning by being cheaper, more available, and more reliable than general clouds, not by making customers technically unable to leave. As AWS, Azure, and specialist rivals close the gap, the durable moat shifts from basic cluster hosting toward supply access, uptime, and higher level services layered on top of portable infrastructure.