Lambda positioned for training workflow
Lambda customer at Iambic Therapeutics on GPU infrastructure choices for ML training and inference
The opening is for a GPU cloud that wins less on cheaper chips and more on making large training runs feel manageable day to day. Training teams still stitch together queues, storage, containers, and cluster management by hand, while hyperscalers are expensive and often less flexible on high performance cluster design. That leaves room for a simpler research workflow layer on top of reserved multi GPU infrastructure, especially for teams training domain specific transformer models.
-
Lambda already sits close to this pain point. Iambic uses Lambda for reserved HGX style training clusters with InfiniBand, custom Kubernetes and storage setups, and values its engineering responsiveness. That makes Lambda better positioned than a pure software reseller to shape the actual research workflow around owned hardware.
-
The market is splitting by workflow. CoreWeave has leaned into production style GPU infrastructure with autoscaling and Kubernetes features that feel closer to AWS. Together AI sits even higher, reselling compute with APIs, open model access, and usage based pricing. The gap in between is researcher first training UX for teams that still want low level control.
-
This is also where pricing power can come from after raw GPU hours commoditize. Lambda was estimated at $250M revenue in 2023, then $505M annualized revenue by May 2025, showing real demand. But the more durable moat is reducing friction around where data lives, which jobs are running, and how clusters are configured, not just charging less per H100 hour.
The next phase of GPU cloud competition is likely to move from capacity into workflow software built specifically for large scale training. Providers that own clusters and add a lightweight, researcher friendly control plane should pull ahead of bare metal sellers, while avoiding the heavier all in abstraction that fits inference better than bespoke training.