serverless gpu

serverless gpu for ai. scale to zero. pay per second.

run any ai model on demand with no infra to manage. sub-second cold starts on warm models, auto-scale to thousands of concurrent jobs, and pay only for the time you use. plus a runtime built for agents, not just functions.

terminal
$ belt app run flux/schnell-fp8
warm start · infer 1.8s · $0.012
$ belt app run veo/3.1-fast
h100 · 12 concurrent · $0.42/min
$ belt deploy ./my-agent
deployed → https://infsh.dev/u/my-agent

from code to cloud

sub-second cold starts

popular models run on warm gpus. your request hits a pre-deployed instance — no boot, no pull, no wait.

per-second billing

metered to the millisecond. scale to zero when idle — when nothing runs, the bill is zero.

hundreds of pre-deployed models

flux, veo, seedance, claude, elevenlabs, qwen, seedream, wan, and more. one api call, no setup.

observability built in

traces, logs, cost per request. every signal captured automatically. no instrumentation sdk.

deploy in seconds

push a model or agent live with belt deploy. zero downtime, zero reconfiguration. github-native.

byok — bring your own keys

route through your own cloud accounts. use our orchestration with your gpus, your commitments, your data.

gpu tiers

bring your own model. or use ours.

tiergpusbest for
smalla4000 · rtx 4000 · rtx 2000image gen, sub-13b llms
mediuma5000 · l4 · l40smid-size image and video, 13b–34b llms
largea100 40g · a100 80gflagship video models, large llm inference, batch jobs
flagshiph100 · h200 · b200veo 3, seedance 2, frontier llms

per-second billing across all tiers. scale to zero. see pricing for details.

how we compare

serverless gpu, plus everything around it.

inference.shRunPodModalReplicatefal.ai
per-second billing
sub-second cold starts
durable agent runtime
observability (no sdk)
hundreds of pre-deployed apps
byok (bring your own keys)
self-hostable runtime
skill & knowledge system

not just serverless. durable.

most serverless gpu platforms give you stateless functions. your code runs, returns a result, and forgets everything. that works for single-shot inference. it breaks for agents.

inference.sh adds a durable execution layer on top of serverless gpus. multi-step workflows survive crashes and restarts. workspaces persist independently of containers. skills and knowledge compound across sessions. your agent can generate an image, edit it based on feedback, and remember what worked — across days, not just within a single request.

serverless gpu is the foundation. the agent runtime, the skill system, and the knowledge layer are what make it compound.

frequently asked questions

three things. sub-second cold starts on warm models. a durable agent runtime (not just stateless functions) so multi-step ai workflows survive crashes and 24-hour jobs. and observability without an sdk — every request gets a trace automatically.

ready to ship?

start with the hosted platform. deploy your own when you're ready.

we use cookies

we use cookies to ensure you get the best experience on our website. for more information on how we use cookies, please see our cookie policy.

by clicking "accept", you agree to our use of cookies.
learn more.