NeoClouds Win on Reserved GPU Pricing
Lambda customer at Iambic Therapeutics on GPU infrastructure choices for ML training and inference
The price gap matters because training buyers are really purchasing guaranteed clusters, not just raw chips. For teams running month long model training jobs, the winning vendor is the one that can reserve hundreds of identical GPUs with fast InfiniBand links and still quote a lower all in hourly rate. That is where NeoClouds like Lambda and CoreWeave have carved out an edge over hyperscalers.
-
The savings can be large enough to drive provider choice by themselves. In the Iambic evaluation, Lambda and CoreWeave were both able to meet the required HGX and InfiniBand spec, while AWS and Oracle were either more expensive or not ready with the needed interconnect. Once the hardware spec was acceptable, the decision came down largely to price per GPU.
-
This is why the market has split by workload. CoreWeave has pushed toward big enterprise reservations and production style tooling, Lambda has focused more on flexible training clusters for growth stage companies, and hyperscalers still win more often on mature surrounding services for storage, IAM, Kubernetes, and always on inference operations.
-
Current public pricing still shows the same pattern. AWS Capacity Blocks list H100 pricing around $3.933 per GPU hour in many regions. Lambda lists H100 on demand pricing around $3.44 to $3.67 per GPU hour, and previously advertised reserved H100 pricing as low as $2.29. CoreWeave documentation shows H100 InfiniBand instances at $49.24 per hour for 8 GPUs, or about $6.16 per GPU hour, which reinforces that pricing depends heavily on packaging, reservation terms, and network configuration.
Going forward, this gap should narrow on plain compute and widen on full cluster design. As more GPU hours become commodity supply, the durable advantage will come from who can deliver reserved training capacity fast, wire it correctly, and wrap it in enough tooling that researchers can move from idea to multi week run without paying hyperscaler prices.