Vast.ai vs Lambda
Both appear in 1 section: gpus. Vast.ai is cheaper on 3 of 3 comparable measures.
This pair contrasts the two extremes of GPU rental. Vast.ai is a pure marketplace where third parties supply the hardware; every Vast price in Pikkly's data is community tier, including H100s at $2.27/hr and H200s at $3.60/hr. Lambda owns its data centres and charges $3.29/hr for an H100 and $2.79/hr for an A100. So Vast is roughly 30% cheaper on H100s, and on interruptible listings it goes considerably lower. The difference is not really price, though. On Vast you are renting a specific machine from a specific host, with a reliability score, a disclosed connection speed, and no guarantee the instance survives the night. Availability, disk throughput and network quality vary between listings, so two rentals of the same GPU model can behave quite differently. On Lambda the machine is Lambda's: a maintained ML image, predictable networking, multi-GPU nodes wired for distributed training, and support with an obligation behind it. The correct choice is a function of what an interruption costs. For inference, batch jobs, experiments and anything checkpointed, Vast's savings are free money and the variability is an inconvenience. For a training run measured in days where a failed node means restarting, Lambda's premium is small against the cost of losing the run, and the predictability is what you are paying for.
GPUs
All gpus →Pikkly earns nothing from Lambda, and the ranking above is by price regardless.