AI Provisioning & Setup

Every eRacks AI server already ships AI-ready, free

The basic preinstall is included with every eRacks AI system, always has been, always free: your choice of open-source OS plus the local-AI stack - Ollama, Open WebUI (the familiar chat interface), vLLM, llama.cpp, PyTorch and the Hugging Face toolchain - installed, updated, and tested before it leaves our facility. Power it on, load a model, go.

AI Provisioning & Setup: from AI-ready to AI-in-production, $1,795

The paid package is everything between a working stack and a working deployment, tailored to your vertical and your environment:

• Model strategy for your use case: we select and install the right open models for your work - legal, healthcare, engineering, support, research - from the current best of DeepSeek, Llama, Qwen, Mistral and friends, including MoE (mixture-of-experts, the architecture behind DeepSeek-class models) where your hardware supports it.
• Quantization & VRAM fit: models quantized (GGUF, AWQ, FP8 - compressed to fit more capability into your GPU memory) and verified against your actual VRAM budget, not a spec sheet.
• Serving tuned to your hardware: vLLM/llama.cpp configs set for real throughput - context length, continuous batching, tensor parallelism across your GPUs - with benchmark numbers (tokens/second, latency) recorded and handed to you.
• Private by design: air-gapped or egress-controlled network configuration, so your data and prompts never leave the building.
• Your documents, searchable: a starter RAG setup (retrieval-augmented generation - the model answers from YOUR files) wired into Open WebUI.
• Beyond text: image generation (ComfyUI with Stable Diffusion/FLUX-class models) and speech-to-text (Whisper) configured on request.
• Multi-node AI clusters: when one server is not enough, we design and deploy clustered inference across several machines - Ray (the distributed compute framework that ties the nodes into one), vLLM distributed serving and llama.cpp RPC for tensor and pipeline parallelism - so a single very large model, or many simultaneous users, spreads across the whole cluster. It is a discipline we practice on our own hardware, not just a slide.
• Handoff, not lock-in: a live two-hour working session with your team, a written record of every config, and 30 days of follow-up email questions included.

Pair it with any eRacks AI system, or run it remotely on capable hardware you already own. Need a multi-node AI cluster, fine-tuning pipelines, or SSO integration? Select the Extended tier above and we quote it to your exact setup. Ongoing assistance is always available after the sale as a consulting add-on, no lock-in, like everything we do.