AI servers: your own private AI, on hardware you own

Yes, that is the point of running the model on hardware you own. Invoices, contracts, HR files and medical records are processed locally, so no personal data is transferred to a third-party AI provider and no extra processing agreement is needed for the model itself. The processing stays inside your GDPR perimeter and your NIS2 supplier list stays short. One honest caveat: you are still the data controller, so internal access rules still matter. The setup includes API keys, an audit log and rate limits, so you can control and document who used the model and when.

The 128 GB of VRAM comfortably runs Llama 3.3 70B at full precision, which means the model is served without the quality loss that quantisation brings. Smaller local LLMs such as Qwen or Mistral fit with plenty of room to spare, and you can run a lighter model when latency matters more than depth. Models are deployed on vLLM or llama.cpp, so swapping one for another is routine. Under the managed plan we refresh models quarterly, bringing in releases like Llama 4 or Qwen 3 as they ship. If you are unsure which model fits your workload, write to support and we will size it together.

Yes. The service exposes the standard chat-completions API on your LAN or VPN, so libraries, SDKs and plugins built for OpenAI endpoints work after you change only the base URL and the API key. Existing integrations keep running without code changes, which is the point of standardising on this interface. The endpoint is reachable only inside your network, not from the public internet, and each application can get its own key. Alongside the API you get a chat UI for staff who simply want to ask questions in the browser. If a specific tool in your stack is critical, name it to support and we will verify it before deployment.

Yes. The optional RAG setup connects the model to your own knowledge. Documents from sources like SharePoint are converted into embeddings and stored in Qdrant or pgvector on the same machine, so the index never leaves your network either. When someone asks a question, the relevant passages are retrieved and the model answers from them instead of guessing. Under the managed plan the index is rebuilt every night, so yesterday's edits show up in answers, and we tune retrieval as your content grows. Answer quality depends on how clean the source documents are, which is honest to say upfront. List your sources when you request a quote and we will confirm what can be connected.

You choose between two homes for the unit. It can sit on your premises, where staff reach it over the LAN and inference does not depend on an internet connection, or it can be colocated in our own Tier 3+ data centre in Riga, where your team connects over VPN. In both cases the hardware is your property, and every byte on it, models, indexes and logs included, belongs to you. Because you own the box, you can start in the office and move it to the data centre later, or the other way around. Delivery within Riga is included with the purchase; delivery elsewhere is arranged individually, so mention your location when asking for a quote.

No, the managed plan is optional. Setup alone hands you a complete, hardened service with a backup and disaster-recovery plan, and you are free to run it yourself from day one. Be honest about what that means: someone on your side then watches GPU load and latency, applies OS and CUDA patches, and swaps in new model releases. The managed plan, from 500 EUR per month, takes exactly that off your plate and adds 4 hours of engineer time every month for changes and requests. You can also start on your own and add the plan later if maintenance turns out to be a burden; write to support and it is switched on.

Think of the price as three parts. The first is the hardware, a one-time purchase; current unit prices are on this page and exclude 21% VAT. The second is setup at 5000 EUR one-time, which covers deployment, hardening and the API and chat UI going live. The third is the optional managed plan starting at 500 EUR per month. There are no per-token fees, so the monthly cost does not grow when you onboard more users or your team starts using it heavily. If you already own a suitable server, an accelerator card alone can be the cheaper route. For an exact figure for your case, request a quote and we will itemise it.

One GB10 unit with 128 GB VRAM serves 10 to 20 concurrent users on Llama 3.3 70B at full precision. Concurrent means people sending requests at the same moment, so the total number of accounts can be much larger; in a typical office only a fraction of staff is chatting at once. Choosing a smaller model leaves extra headroom if peak load matters more to you than maximum answer quality. When the team genuinely outgrows one box, the NVIDIA GB10 unit supports clustering, so capacity grows by adding nodes rather than replacing what you bought. Tell us your headcount and expected usage during the sizing call and we will recommend a configuration that fits.

For running a model there is no difference: both are built on the same GB10 Blackwell chip with 128 GB VRAM, so inference speed and the models they fit are identical. The differences are around the chip. The NVIDIA unit ships with 4 TB of NVMe storage and supports clustering; the ASUS Ascent GX10 comes with 1 TB and runs single-node only. In practice storage matters once you keep several models and a growing RAG index on the box, and cluster support matters only if you expect to add nodes later. A stable small team is well served by the ASUS unit. Setup and the managed plan are the same for both, and delivery within Riga is included either way.