Lambda preferred for reserved GPU clusters
Lambda customer at Iambic Therapeutics on GPU infrastructure choices for ML training and inference
This reveals that winning training workloads is less about generic GPU access and more about selling a custom cluster that actually fits how researchers work. For teams training large models, the deciding factor is a reserved block of identical GPUs with fast InfiniBand links, built to NVIDIA style reference designs, at a price low enough to support long experiments. Lambda won by combining lower per GPU pricing with willingness to customize networking, storage, Kubernetes, and security setups.
-
The practical split is training on Lambda, inference on AWS. Iambic used Lambda for reserved multi GPU clusters with high speed interconnects, while AWS handled smaller, spiky inference jobs where instant availability, S3, EC2, and mature infrastructure mattered more than lowest cost.
-
Lambda competed in a different lane from hyperscalers. In late 2023, AWS and Oracle were either too expensive or not ready with the interconnect quality Iambic needed, while Lambda and CoreWeave were willing to spec custom InfiniBand clusters. Between the two neocloud options, Lambda was slightly cheaper after the cluster design was finalized.
-
The stickiness comes from operational fit, not unique hardware. Once a team builds job queues, Kubernetes add ons, storage layouts, and security controls around a provider, moving for a few cents less per GPU hour is not worth the slowdown. Lambda strengthened that lock in by giving customers direct access to engineers and managed Kubernetes on dedicated single tenant clusters.
Going forward, the providers that win research training budgets will be the ones that make large reserved clusters feel as easy to use as a simple cloud VM. Lambda is already pushing in that direction with 1-Click Clusters, managed Kubernetes, and dedicated support. If it keeps pairing low prices with researcher friendly operations, it can stay the default choice for growth stage AI teams before they need full hyperscaler infrastructure.