Hardware Requirements
Size your server for NeuraPBX. Subscription plans can run with or without AI; perpetual licenses are always no-AI. What drives the numbers is your concurrent call volume, call recording, and whether AI runs locally — extensions themselves register very cheaply.
Three ways to run AI — very different hardware
- No AI — core PBX only. Use the without AI figures.
- Cloud AI (bring your own OpenAI / API key) — almost no extra hardware: the without-AI specs plus ~2 GB RAM, since the heavy lifting runs in the cloud.
- On-prem / local AI (Ollama + faster-whisper, 100% private) — needs significant RAM and, for real-time features like agent copilot and live translation, a GPU. Use the with on-prem AI figures.
Subscription plans — without AI
Core PBX only (or with cloud AI, which adds ~2 GB RAM). FreeSWITCH with G.711 pass-through is light on CPU; storage grows with call-recording retention.
| Plan | Capacity | Minimum | Recommended |
|---|---|---|---|
| Free | 5 ext · 1 trunk | 1 vCPU · 2 GB · 15 GB SSD | 2 vCPU · 2 GB · 30 GB SSD |
| Starter | 25 ext · 2 trunks | 2 vCPU · 2 GB · 40 GB SSD | 2 vCPU · 4 GB · 80 GB SSD |
| Business | 100 ext · 8 trunks | 4 vCPU · 6 GB · 120 GB SSD | 4 vCPU · 8 GB · 250 GB SSD |
| Enterprise | 500 ext · unlimited trunks | 8 vCPU · 16 GB · 400 GB NVMe | 16 vCPU · 32 GB · 1 TB NVMe |
Subscription plans — with on-prem AI
Adds faster-whisper (transcription) and a local Ollama LLM (summaries, agent copilot, live translation, voicemail-to-email). CPU-only works for batch tasks; a GPU is recommended for real-time features. Model RAM is a fixed baseline — concurrency scales with call volume.
| Plan | Capacity | Minimum (CPU-only) | Recommended (with GPU) |
|---|---|---|---|
| Free | 5 ext · 1 trunk | 4 vCPU · 8 GB · 40 GB SSD | 4 vCPU · 16 GB · 80 GB SSD · GPU 8 GB |
| Starter | 25 ext · 2 trunks | 4 vCPU · 12 GB · 80 GB SSD | 6 vCPU · 16 GB · 150 GB SSD · GPU 8–12 GB |
| Business | 100 ext · 8 trunks | 8 vCPU · 24 GB · 200 GB NVMe | 12 vCPU · 32 GB · 500 GB NVMe · GPU 16 GB |
| Enterprise | 500 ext · unlimited trunks | 16 vCPU · 48 GB · 500 GB NVMe | 24 vCPU · 64 GB · 2 TB NVMe · GPU 24 GB+ |
Perpetual packages — without AI
Perpetual licenses never include AI, so only these no-AI specs apply. Same core-PBX sizing as the subscription plans, by capacity.
| Package | Capacity | Minimum | Recommended |
|---|---|---|---|
| Micro | 8 ext · 2 trunks | 2 vCPU · 2 GB · 20 GB SSD | 2 vCPU · 4 GB · 50 GB SSD |
| Small | 32 ext · 4 trunks | 2 vCPU · 4 GB · 50 GB SSD | 4 vCPU · 4 GB · 100 GB SSD |
| Medium | 64 ext · 8 trunks | 4 vCPU · 4 GB · 80 GB SSD | 4 vCPU · 8 GB · 200 GB SSD |
| Large | 128 ext · 16 trunks | 4 vCPU · 8 GB · 150 GB SSD | 8 vCPU · 8 GB · 300 GB SSD |
| X-Large | 256 ext · 32 trunks | 8 vCPU · 8 GB · 250 GB SSD | 8 vCPU · 16 GB · 500 GB NVMe |
General requirements
- OS: 64-bit Linux (Ubuntu 22.04+ / Debian 12) recommended. Also runs on Windows (Docker Desktop) and macOS (Apple Silicon).
- Runtime: Docker + Docker Compose.
- Storage: SSD or NVMe strongly recommended — call recordings and voicemail are I/O-heavy. Budget roughly 30 MB per recorded hour (mono G.711) and size for your retention policy.
- Network: stable LAN with a static IP; about 85–100 kbps per concurrent G.711 call in each direction, plus headroom for your SIP trunks.
- GPU (local AI only): an NVIDIA GPU with CUDA is recommended for real-time AI; VRAM sizing is in the table. CPU-only works for non-real-time AI, just slower.
- ARM: the core PBX (no AI) runs on ARM64 — a Raspberry Pi 5 comfortably handles a Free, Starter, or Micro/Small deployment.
These are sizing guidelines, not hard limits. Real-world needs depend on your concurrent call volume, transcoding (G.711 pass-through vs. Opus/G.729), recording retention, and chosen AI models. For 500+ extensions or heavy real-time AI, we recommend a dedicated GPU server — or splitting the PBX and AI across two hosts. Not sure what to buy? Contact us and we'll spec it with you.