early access
serverless gpu.
give any machine gpu power on demand. scale up for the job, back to zero when it's done, and never pay for idle time.
use cases
point it at what you're already running.
training & fine-tuning
run the training job you already have, without owning the gpu it needs.
inference at scale
serve requests around the clock without paying for gpu time in between them.
distributed & burst training
burst to 100+ gpus for one job, then pay for none of them tomorrow.
what you get
no rewrite. no reserved cluster. no lock-in.
no rewrite
works with pytorch, tensorflow, and jax, unchanged.
no cluster to manage
scale from one gpu to a hundred, then back to zero when the job ends.
no idle billing
billed to the second, never rounded up to the hour.
no performance tax
fast enough for real workloads, not just demos.
no lock-in
works with the security and infra you already have. nothing proprietary to adopt.
pay for the second you use it.
billed per second with no reserved capacity. $20 credit after your first payment.