Deploy Dedicated GPU

Deploy your APIPod service on a permanently running GPU instance — no cold starts, predictable latency, and full GPU memory for large models.

Quick Deploy

Use the --hosted flag to deploy a dedicated (always-on) GPU instance. The container stays running until you explicitly stop it.

# Build first — live today (apipod CLI)
apipod --build

# Deploy to a dedicated (always-on) GPU instance — coming soon
# Unified socaity CLI will mirror apipod for the hosted flow:
# socaity deploy --hosted --provider runpod --gpu A100 --replicas 1

Pushing image to registry...       ✓
Provisioning dedicated GPU pod...  ✓ (pod-abc123)

GPU:     NVIDIA A100 SXM (80 GB VRAM)
Service: https://api.runpod.io/v2/abc123/run
Status:  PROCESSING

Manage with:
  socaity status
  socaity logs abc123
  socaity stop abc123

Dedicated Deploy Flags

Dedicated multi-provider hosting is coming soon. Today only RunPod serverless is fully wired in. The flags below describe the upcoming unified socaity deploy --hosted CLI, which will mirror apipod for convenience.

Flag	Default	Description
`--hosted`	`false`	Deploy as a dedicated (always-on) GPU pod.
`--provider <name>`	`socaity`	Target provider: runpod \| socaity.
`--gpu <type>`	`A100`	GPU class: T4 \| A10G \| A100 \| A100-80G \| H100.
`--replicas <n>`	`1`	Number of identical GPU pods to run in parallel.
`--volume <gb>`	`20`	Persistent storage attached to the pod (GB).
`--name <str>`	`project name`	Label shown in dashboard and logs.

GPU Selection Guide

Choosing the right GPU is the single biggest lever for cost and performance. Use the table below to match your model's VRAM requirement to the right class.

GPU	VRAM	Hourly	Model Fit	Best For
T4	16 GB	see socaity.ai/Pricing	≤ 7B params (fp16)	Inference, classification
A10G	24 GB	see socaity.ai/Pricing	≤ 13B params (fp16)	Diffusion, medium LLMs
A100 40G	40 GB	see socaity.ai/Pricing	≤ 20B params (fp16)	Production LLMs
A100 80G	80 GB	see socaity.ai/Pricing	≤ 40B params (fp16)	Large LLMs, video
H100 SXM	80 GB	see socaity.ai/Pricing	≤ 70B params (fp16)	Frontier models, training

Rule of thumb: your model weights (in GB) should fit in ≤ 80% of available VRAM to leave headroom for activations and KV cache. A 13B parameter model in fp16 needs ~26 GB — choose an A100 (40 GB) or better.

Provider Comparison

Provider	GPU Fleet	Regions	Spot Instances	Storage
RunPod	T4, A10G, A100, H100	US, EU, APAC	Yes	Network volumes up to 50 TB
Scaleway (coming soon)	H100, L4	EU (Paris, Amsterdam)	No	Block storage 100 GB+
Azure (coming soon)	A100, H100	Global (20+ regions)	Yes	Azure Blob + managed disks

Dedicated GPU instances are billed continuously even when idle. Stop them from the Studio dashboard (the unified socaity stop CLI command is coming soon).

Managing Running Services

# Coming soon — the unified socaity CLI will expose these management commands:
# socaity status                       # List all running services
# socaity logs my-service              # Stream live logs
# socaity stop my-service              # Stop a dedicated GPU instance (stops billing)
# socaity scale my-service --replicas 3  # Scale replicas without redeployment

# Today, manage services through the Studio dashboard at socaity.ai.

Serverless vs Dedicated — Decision Guide

Scenario	Recommendation	Reason
Bursty or unpredictable traffic	Serverless	Scale to zero, pay only for usage.
Sub-100ms P99 latency required	Dedicated	No cold-start penalty.
Model > 24 GB VRAM	Dedicated	Serverless workers are shared; dedicated gives full VRAM.
Continuous batch processing	Dedicated	Hourly rate cheaper than per-second for sustained load.
Dev / testing / prototyping	Serverless	Zero idle cost.
>50 req/s sustained	Dedicated (multiple replicas)	Avoids cold-start queuing under high load.

Deploy Serverless

Providers