Need for researcher-first GPU cloud
Lambda customer at Iambic Therapeutics on GPU infrastructure choices for ML training and inference
The open gap in GPU cloud is not more raw compute, it is a simpler training environment that lets researchers get from code to running jobs without hiring an infra team. In practice that means a cloud with cheap reserved multi GPU clusters, fast interconnect, storage that is easy to find, a visible job queue, and enough tooling to track experiments and debug failures without dropping into bare metal ops.
-
The interview draws a clean line between training and deployment. For training, researchers want something close to a Slurm or Kubernetes cluster where they can SSH in, submit jobs, mount data, and control hardware directly. Today that workflow is still raw and manual across most GPU clouds.
-
Lambda and CoreWeave won Iambic's training evaluation by offering high quality InfiniBand clusters at lower cost than AWS or Oracle, and by being willing to customize the setup. But even with that advantage, the buying criteria still centered on price and hardware spec more than polished software.
-
The deployment layer is filling in faster than the training layer. The same interview points to AWS plus tools like Coiled for reliable serving, and adjacent companies like Together, Modal, RunPod, and Fal.ai show how software wrappers are gaining traction when the job is inference, not bespoke model training.
The next wave of winners will package reserved GPU capacity with a researcher first control plane, not just cheaper H100 hours. As more domain specific model teams in biotech, legal, finance, and engineering move from small models to billion parameter training runs, the provider that makes clusters feel as easy as a developer cloud will gain share before training compute fully commoditizes.