Frequently asked questions
Straight answers to the questions buyers actually ask before deploying private AI on their own hardware.
How is this different from just running Ollama or local LLaMA myself?
You can absolutely run an open model locally with Ollama or llama.cpp for free — Sovra doesn't compete with that. Sovra adds the parts most teams don't want to build and maintain themselves: validated hardware selection with real benchmarks, a RAG pipeline tuned per device class, a structured command-execution layer (so the model can trigger real actions safely, not just answer questions), fleet management across multiple devices, and support. If you're comfortable running and maintaining your own stack on one machine, do that. Sovra is for teams deploying across multiple sites or devices who want it to just work.
What happens if the hardware fails — is there data recovery or support?
Hardware is sourced through standard manufacturer channels (Raspberry Pi, NVIDIA, Beelink, Apple, etc.) and carries that manufacturer's own warranty — we don't replace the warranty terms of a Raspberry Pi or a GPU workstation. Your Sovra configuration isn't tied to a single physical unit: reinstall on replacement hardware and restore from your own backups. Enterprise/OEM tier includes a dedicated deployment engineer who can walk your team through recovery. We recommend standard backup practice (snapshot your config and vector store) regardless of tier.
Do you support EU-only cloud fallback for non-critical flows?
Core assistant flows run entirely on your hardware by design and never require any cloud fallback. If you choose to enable an optional cloud-connected feature — for example, pulling in live external data — you control which provider and region that goes through. Nothing is silently routed to an external service; anything that leaves the device is a deployment you explicitly configure.
What if I need a bigger model than my hardware can run?
The hardware catalog spans from a $95 Raspberry Pi to NVIDIA DGX Spark and GPU workstations that run 70B–200B-class models locally, so most teams can scale within the on-prem catalog first. If a workload genuinely needs a frontier model beyond what any on-prem hardware can practically run, we say so — see the honest tradeoffs on the Why Sovra page.
Do I need an internet connection at all?
No, not for core assistant flows — inference, retrieval, and validated command execution run fully offline once a device is deployed and configured. Internet is only needed for initial setup, software updates, and any optional integration you explicitly enable.
How is pricing structured — hardware vs. software?
They're separate and both transparent. Hardware is purchased once at real market pricing shown in the hardware configurator — you own the device. Sovra software is a per-device monthly subscription (from €49/mo Core Runtime) shown on the pricing page, layered independently of which hardware you chose.
Can I move my configuration to different or upgraded hardware later?
Yes. Sovra's software stack targets the same core runtime across the supported hardware catalog, so moving from, say, a Raspberry Pi pilot to a Jetson or GPU workstation for production is a supported upgrade path, not a rebuild.
Is automotive / in-vehicle deployment production-ready today?
Being straightforward here: automotive is on our roadmap. The automotive hardware entries in the catalog are reference and development kits for building toward in-vehicle command execution — they demonstrate the architecture, but we haven't shipped a production automotive integration yet. If you're evaluating for a vehicle program, talk to us directly about timeline and current maturity.
Is my data ever used to train models?
No. Sovra runs inference on hardware you control; there is no default path where your prompts, documents, or logs are sent anywhere to train anything. See the Privacy Policy for the full data-handling terms.
Still have questions?
Talk to us directly — we'll give you a straight answer, including where Sovra isn't the right fit.