GPU Compute · ASC Docs

What it is

GPU Compute serves dedicated NVIDIA accelerators in apac, use and euw. Three purchase modes: on-demand (per-second, no commitment), spot (40–60% cheaper, reclaim on 30 s notice), and reserved (1-month or 1-year commitment, deepest discount). Pods boot from your container image; warm pool means GPU pods are ready in under 60 s for warm SKUs and under 5 minutes for cold.

When to use it

Training runs, multi-node with NCCL + InfiniBand.
Real-time inference (use the Inference product for managed autoscale).
Batch jobs (use the Batch product for fan-out).

Quickstart

# Single H100 on-demand
asc pods run --gpu h100 --image ghcr.io/me/train:latest
 
# 8x H100 spot, with auto-resume
asc pods run --gpu h100 --gpus 8 --spot --image ... --auto-resume
 
# Reserve 4x B200 for 30 days
asc reservations create --gpu b200 --count 4 --duration 30d

pod = client.pods.run(
    image='ghcr.io/me/train:latest',
    gpu='h100',
    gpu_count=8,
    spot=True,
    auto_resume=True,
)
print(pod.status)  # pending → ready

const pod = await client.pods.run({
  image: 'ghcr.io/me/train:latest',
  gpu: 'h100',
  gpuCount: 8,
  spot: true,
  autoResume: true,
});

Limits & quotas

Limit	Default	Burst	Notes
On-demand H100 / region	32	256	Lift via sales
On-demand B200 / region	16	128	Lift via sales
Spot reclaim notice	30 s	—	Checkpoint to object storage during this window
Pod boot (warm SKU)	<60 s	—	Includes image pull from registry
Pod max lifetime	30 days	—	Use Agents for indefinite

Pricing

See pricing. Pay-as-you-go, billed monthly via Stripe.

API surface

POST /v1/pods — create
GET /v1/pods/{id} — read
GET /v1/pods — list
POST /v1/pods/{id}/stop — stop
POST /v1/pods/{id}/checkpoint — snapshot to object storage
POST /v1/reservations — reserve capacity

Required scope(s): pods:read, pods:write. See Scopes.

Security

All access is authenticated and scoped. See Auth & scopes and Network controls.