The services cloud AI builders love.
Run inference, training, batch and agents with sub-second cold starts, instant autoscaling and a developer experience that feels local. No YAML, no quotas, no markup.
1import asc 2 3app = asc.App("summarizer") 4 5# Define a GPU function inline. Container, scheduling, scaling — all handled. 6@app.function(gpu="L4", scaledown_window=30) 7def summarize(text: str) -> str: 8 from transformers import pipeline 9 pipe = pipeline("summarization", model="facebook/bart-large-cnn")10 return pipe(text, max_length=120)[0]["summary_text"]11 12@app.local_entrypoint()13def main():14 print(summarize.remote("Long article goes here..."))Ships alongside the tooling builders already love
Designed to help AI teams ship faster.
Programmable infra
Define everything in code. No YAML, no sidecar configs. Container, GPU, secrets, schedules — all in one Python or TypeScript file.
Built for performance
Sub-second cold starts, snapshot-based container restore, intelligent caching. Tight feedback loops in dev, low latency in prod.
Elastic GPU scaling
Burst across H100 / A10G / L4 / L40S without quotas or reservations. Scale to zero when idle. You only pay for active seconds.
Unified observability
Logs, metrics, traces and per-call cost in one timeline. Drill into a single inference or roll up by function, model, tenant.
Powering any AI workload.
One control plane for every stage of the AI lifecycle — from a single GPU experiment to a fleet of sandboxes serving production agents.
Inference
Deploy and scale low-latency inference for LLMs, audio and image generation.
Training
Fine-tune open-source models on single or multi-node GPU clusters in minutes.
Sandboxes
Programmatically scale secure, ephemeral environments for running untrusted code.
Batch
Fan out to thousands of containers for batch jobs, evals and dataset processing.
Agents
Long-running, isolated workspaces with persistent disk for agentic and MCP workloads.
GOAT Build
Prompt-to-app cloud IDE. Describe what you want, watch it ship — Monaco, terminal, live preview, one-click deploy.
Built on a powerful foundation.
Every layer of the platform is engineered for AI — from container restore to GPU scheduling to the storage fabric underneath.
Engineered from the ground up for heavy AI workloads.
@app.function(gpu="H100", scaledown_window=60)
def generate(prompt):
return model(prompt)
# 0 → 100 replicas in 4 seconds
generate.spawn_map(prompts)A globally distributed storage layer for high-throughput AI.
vol = asc.Volume("models")
@app.function(volumes={"/models": vol})
def serve():
load_from("/models/llama-70b")Bring the rest of your stack with you.
app.observability.export_to(
"datadog",
api_key=asc.Secret("DD_KEY"),
)Always have the GPUs you need, where you need them.
@app.function(
gpu=["H100", "A100"],
region=["us-east", "eu-central"],
)
def train(): ...Enterprise-grade controls. Sovereign by default.
Team controls
RBAC, SSO/SAML, scoped API tokens, immutable audit log. Built for teams shipping in regulated industries.
Battle-tested isolation
Hardware-virtualised sandboxes per request. Capability-bound execution roles. Never co-tenants in the same kernel.
Encrypted secrets
Envelope encryption at rest, never plaintext over the wire. Per-tenant prefixes and rotation hooks.
Data residency
Pick your region at signup — workloads, storage and logs never leave it. Sovereign by default.
SOC 2 & HIPAA path
SOC 2 Type II in progress; HIPAA BAA on request for Enterprise. Mapped controls in the compliance matrix.
Teams ship 10× faster with ASC.
“ASC let us go from a Jupyter notebook to a production GPU endpoint in an afternoon. We never had to think about containers.”
“We replaced a queue, three workers and a ton of glue code with a single @app.function decorator. It just works.”
“Burst to 800 GPUs for a batch eval, scaled back to zero an hour later. Bill matched the spec sheet to the cent.”
“The DX is the closest I've come to feeling like local development on remote infra. This is how backends should work.”
“Cold starts under half a second on a 13B model. We pulled our previous serving stack out of production the same week.”
“Sandboxes for code-execution agents have been a cheat code. Spin up, run untrusted code, tear down — all under 200ms.”
“Audit log + per-call cost in one view convinced our security team in a single review. Procurement closed in two weeks.”
“Brought my own key for the model providers. Switched providers mid-deploy with one config change.”
“We used to dread the day a researcher wanted a GPU. Now I tell them: just decorate the function.”
Ship your first app in minutes.
Free $30/month of compute. No card required.