Best server for… pricing
Best and cheapest server for self-hosted LLM inference on AWS
Updated 2026-10-19
Cheapest sensible starting point for self-hosted LLM inference: g6.xlarge at $0.8048/h ($587.50/month) on demand in us-east-1. Move up to g5.xlarge when you outgrow it.
Recommendations (us-east-1, on demand, Linux)
| Instance | When | Size | $/h | $/month | 3-yr RI $/h | Cheapest region | $25k credits last |
|---|---|---|---|---|---|---|---|
| g6.xlarge | NVIDIA L4, 24 GB | 4 vCPU / 16 GiB / NVIDIA L4 | $0.8048 | $587.50 | $0.3694 | us-east-1 $0.8048 | 43 mo |
| g5.xlarge | NVIDIA A10G, 24 GB | 4 vCPU / 16 GiB / NVIDIA A10G | $1.01 | $734.38 | $0.4346 | us-east-1 $1.01 | 34 mo |
| inf2.xlarge | Inferentia2, 32 GB, needs the Neuron SDK | 4 vCPU / 16 GiB | $0.7582 | $553.49 | $0.3275 | us-east-1 $0.7582 | 45 mo |
| g6e.xlarge | NVIDIA L40S, 48 GB | 4 vCPU / 32 GiB / NVIDIA L40S | $1.86 | $1,358.53 | $0.804 | us-east-1 $1.86 | 18 mo |
How we chose
Sizing assumption (ours, not a vendor requirement): Model weights plus KV cache must fit in accelerator memory: 24 GB suits small quantized models, 48 GB+ larger ones.
Prices come from the AWS Price List. Treat sizes as starting points: watch memory and CPU for a week, then resize.
More use cases
- Best server for WordPress
- Best server for n8n
- Best server for PostgreSQL
- Best server for Next.js
- Best server for Redis / Valkey (self-managed)
- Best server for a Docker host (small APIs and side projects)
- Best server for a Node.js API
- Best server for a dev / staging server
- Best server for background workers and cron jobs
- Best server for Windows Server
Related cost calculators
- Amazon Bedrock AgentCore Pricing Calculator · Runtime, Gateway, Memory & Policy Costs
- Bedrock Pricing Calculator · AI Model & Tool Cost Estimator
- EC2 Pricing Calculator · Compare Instance Types & Costs
- Lightsail Pricing Calculator · Instances, Containers & Databases
- Amazon SageMaker AI Pricing Calculator · On-Demand Studio, Training & Inference
- Fargate Pricing Calculator · Serverless Container Costs
Questions
What is the cheapest AWS server for self-hosted LLM inference?
Cheapest sensible starting point for self-hosted LLM inference: g6.xlarge at $0.8048/h ($587.50/month) on demand in us-east-1. Move up to g5.xlarge when you outgrow it.
Can I run self-hosted LLM inference on AWS credits?
Yes, EC2 and RDS usage is covered by AWS promotional credits; the last column shows how long $25k lasts.
Prices and credit amounts are estimates and may change. Check AWS for current terms.