Service · O-02
Private AI Starter
Your first production on-premise AI system: one hardened GPU server, identity and ingress, an open-weights LLM served via vLLM or SGLang, an internal chat UI, monitoring, backup, and a runbook your team is trained on. Built to a measured performance bar.
What you get
- One on-premise GPU server, specified for your workload
- A hardened Linux base with identity (Keycloak / FreeIPA)
- Ingress and an API gateway with authentication
- An open-weights LLM served via vLLM or SGLang
- An internal chat UI for your team
- An optional small RAG over a first document set
- Monitoring and alerting (DCGM, Prometheus, Grafana)
- Encrypted backup with a witnessed restore
- A runbook and a one-day training session
How it runs
The build runs over six to ten weeks: hardware specification and procurement, racking and OS hardening, identity and ingress, model serving, the internal chat UI, monitoring, and backup. It closes with a witnessed restore, a runbook, and a one-day training session so your team can operate it without us.
What we need from you
- A space, power, and network drop for the server
- A decision on the hardware tier (we recommend one)
- Directory access for single sign-on integration
- A technical owner to receive the handover and training
Indicative price
€35k – €70k
Services are scoped between €35,000 and €70,000. Hardware is separate, from €18,000 to €120,000 depending on the tier, and is passed through at cost plus a 10–15% integration fee, so you own the hardware outright.
Acceptance criteria
- p50 first-token under 2 seconds for a 7B-class model
- p50 first-token under 5 seconds for a 70B-class model
- A successful backup restore, witnessed together
Thresholds are measured on the hardware tier agreed up front, not on a vendor's spec sheet.
Questions we hear often
How do you choose the hardware?
We size against your actual workload (model size, concurrency, and growth) and recommend a tier from vendors such as Dell, HPE, Lenovo, or Supermicro with NVIDIA GPUs. Hardware is passed through at cost plus a 10–15% integration fee, so the procurement is transparent.
Which model do we run?
An open-weights model appropriate to your use case and hardware tier, typically in the 7B to 70B class. We serve it with vLLM or SGLang. Because the weights are open, you can change models later without renegotiating a licence.
Why open-weights rather than a commercial API?
Open-weights models run entirely on your hardware, so prompts and documents never leave your perimeter, there is no per-token bill, and there is no lock-in to a single vendor's roadmap. For private and regulated workloads that combination is usually the deciding factor.
What does the acceptance test actually measure?
We agree the hardware tier up front, then measure p50 first-token latency on it: under 2 seconds for a 7B-class model and under 5 seconds for a 70B-class model. We also restore from backup together so you have witnessed that the recovery path works.
What happens after handoff?
Your team owns and operates the system. That is the point of the runbook and the training day. If you want a safety net, ongoing operations support is available as a retainer, but it is optional. Talk to us about scope on the contact page.
Can this grow into something larger?
Yes. The Starter is deliberately a first production system. Many organisations then add Enterprise RAG over their documents, or scale the GPU footprint. The architecture is designed so those are additions, not rebuilds.
Is the hardware cost included in the service price?
No, they are separate lines. Services are €35k–€70k; hardware is €18k–€120k depending on the tier, passed through at cost plus a 10–15% integration fee. Splitting them keeps the procurement honest and lets you own the hardware outright.
A first system your team actually owns.
Specified against your workload, built to a measured performance bar, and handed over with a runbook and training. No lock-in to keep it running.