AWS Infrastructure Sizing — Executive Summary

Sovereign LLM Workbench + LM Studio · User-tier procurement guide · v1.0 · June 2026

Purpose

Size AWS EC2 for local-sovereign AI (no cloud LLM APIs): FastAPI workbench + LM Studio SLM, RAG over internal documents, ChatGPT-like streaming UX on approved 7B models.

Experience targets

Time to first token ≤ 2 s (production tiers)
Streaming ≥ 25–40 tokens/s per active user
100% inference and data remain in customer VPC

Tier summary (licensed users)

Tier	Users	Architecture	AWS (summary)	Est. monthly*
T0	0–5	Combined	1× g5.xlarge or g4dn.xlarge	$450–600
T1	6–10	Split	g5.xlarge + m7i.large/xlarge	$1.1–1.4K
T2	10–15	Split	g5.xlarge + m7i.xlarge	$1.3–1.7K
T3	15–25	Split + 2 GPU	2× g5.xlarge + m7i.2xlarge	$2.4–3.8K
T4	25–50	LLM pool	2–3× g5.2xlarge or g5.48xlarge + c7i.2xlarge	$5.5–9K+

*US on-demand EC2+EBS order-of-magnitude; excludes NAT/ALB/support.

Recommendation — 25 users, fast chat

Procure Tier T3 minimum: two GPU nodes for LM Studio (avoid single-GPU queueing), one m7i.2xlarge for Workbench/ChromaDB, 300 GB gp3, model google/gemma-3-1b, streaming enabled.

Do / Don't

Do	Don't
Private VPC; LM Studio :1234 only from Workbench SG	Expose LM Studio to 0.0.0.0/0
ALB + TLS + API_KEY / OIDC	Run 25 users on one g5.xlarge
Approved catalog models only	Frontier cloud APIs for regulated data

Windows first-time deploy

On-prem / EC2 Windows: run start-production.bat — installs deps, pulls LM Studio models, validates, starts server.