Infrastructure for AI that has to run every day
A working AI product usually rests on two things: a vector database that finds the right context, and GPU capacity for the models you train or serve yourself. We host both. Qdrant and ChromaDB handle embeddings, retrieval-augmented generation and semantic search, while GPU servers cover fine-tuning and inference with your own models. Everything runs in an EU data centre in Frankfurt, invoices show Romanian VAT separately, and the GDPR paperwork starts with a data processing agreement under Article 28. Included DevOps hours help you build the pipeline, so you are not left alone with empty hardware.
Vector databases for retrieval and RAG
A vector database stores the numeric fingerprint of your documents, images or products and returns the closest matches to a query. It is the retrieval layer behind chat over company documents, semantic product search and recommendations. Choose the engine that fits your team. Each one is hosted in the EU, and we set up the Python client and a few example notebooks so you can run your first query on day one.
Qdrant
Written in Rust and built for production. A good fit for retrieval over millions of embeddings, with an HNSW index, payload filtering and snapshots for backup and restore.
ChromaDB
The easiest engine to start with. It suits prototypes and smaller applications, is Python-first and can create embeddings through OpenAI or Sentence Transformers for you.
Weaviate
Brings a schema and metadata to the picture. Hybrid search combines vector similarity with keyword and tag filters, and generative modules plus a GraphQL API are included.
pgvector
A PostgreSQL extension. If your data already lives in Postgres you add vectors next to it, join them with ordinary tables and keep SQL transactions, instead of running one more database.
GPU servers: RTX PRO 4000 and RTX PRO 6000
Both cards are NVIDIA RTX PRO Blackwell GPUs with GDDR7 memory. Our team can set up the GPU stack for you, drivers, CUDA and the framework of your choice, within your included hours, so your first job is your model and not a driver install. GPU servers are billed monthly or annually, never by the hour, and each plan includes 10 DevOps hours a month for pipeline setup, tuning and reviews.
Typical AI workloads and what they run on
Five common scenarios, each matched to a server configuration. Where a figure appears in the last column, it is read live from the plans on our pricing page and shown without VAT. As a rule of thumb, owning the infrastructure starts to pay off against per-request API fees once your volume is steady and high. We are happy to run that comparison with your own numbers.
| AI workload | Configuration | Monthly cost |
|---|---|---|
| Chat over 1 million documents (RAG) | VPS 32G with Qdrant, model through an external API | from 240 EUR |
| Semantic search for a shop with 50,000 products | VPS 16G with pgvector and Sentence Transformers | from 100 EUR |
| A self-hosted 70B open model serving about 5,000 queries a day | GPU server RTX PRO 6000 with vLLM | from 4,000 EUR |
| Training a custom OCR model | GPU server RTX PRO 4000, monthly billing | from 800 EUR |
| A 70B model in production behind a load balancer | Two GPU servers RTX PRO 6000 with Kubernetes | from 8,000 EUR |
From idea to production in five stages
A typical AI project follows the same arc whatever model sits at the end. For each stage there is hardware ready and a pattern to start from, so you are not inventing the plumbing.
Prepare the data
Collect the documents, clean and normalise the text, split it into chunks. Python jobs on a VPS write the result to MinIO (S3-compatible) or to PostgreSQL.
Create the embeddings
Use a hosted embedding API or a local model such as Sentence Transformers or BGE-M3. The vectors go into Qdrant or pgvector, stored next to their source text and metadata.
Retrieve and assemble context
A FastAPI or Next.js service takes the question, asks the vector database for the closest chunks and adds them to the prompt. Hybrid search with metadata filters usually beats pure vector search.
Run the model
Start with Claude or GPT through an API for the quickest road to production. Move to a model on your own GPU when volume, cost or data rules justify it.
Monitor and evaluate
Query logs go to Loki, latency to Prometheus and API spend to Grafana. Judge answer quality with LangSmith or with a test set of real questions that you keep up to date.
What teams build on it
A knowledge assistant for the company
Index internal documents in Qdrant, call Claude or GPT through an API and answer with the source attached. Romanian is a heavily inflected language, so pick a multilingual embedding model and test retrieval with real questions typed in Romanian, not only in English.
Semantic search for a shop or marketplace
Search by meaning instead of keywords. A shopper who types a loose description still reaches the right product, and Romanian word endings stop breaking the match.
A language model that you host
Run Llama or Mistral on a GPU server with vLLM or Ollama, so prompts and documents never leave a machine you control. This suits teams in regulated sectors such as banking, insurance or healthcare in Bucharest or Cluj-Napoca.
Computer vision in production
Defect detection on a manufacturing line, OCR on scanned delivery notes, object recognition in warehouse images. Train on your own dataset with YOLO, Detectron2 or Segment Anything and serve the result from the same machine.
Triage for support and operations
An embedding model groups incoming tickets or messages by topic, a small classifier routes them, and a language model drafts the first reply for a person to approve.
GDPR for AI workloads
AI systems gather personal data in unexpected places: training sets, embeddings, prompts and query logs. The simplest way to reduce risk is to decide where each of them lives, who can read it and how long it stays. Hosting in Frankfurt keeps the infrastructure side inside the EU, and the rest is configuration that we help you with.
- Documents, embeddings and logs sit in an EU data centre, in Frankfurt
- A self-hosted model such as Llama or Mistral keeps prompts away from third-party APIs altogether
- Retention of query logs is yours to set; a short window such as 90 days is a common starting point, and you can delete earlier
- Embeddings are derived data: to remove a person completely, anonymise or delete the source and re-index
- Access to the vector store can be logged with who, when and which query, which we set up as part of hardening
- A data processing agreement under GDPR Article 28 is available, with standard contractual clauses for any transfer outside the EU
- Our controls are aligned with ISO 27001, which is alignment and not a certificate
AI hosting questions we hear most
Can I use OpenAI or Anthropic with a vector database in the EU?
Yes. The vector database stays on your server with us, and only the question plus a handful of context chunks go to the model API. Our examples show how to trim that payload and strip personal data before the call.
How large a model fits on your GPUs?
A 7 to 8 billion parameter model runs comfortably on the 24 GB card, and models up to roughly 14 billion parameters fit with quantisation. A 70 billion parameter model fits the 96 GB card when quantised. Beyond that, for example very large mixture-of-experts models, you need several GPU servers, and we help you work out the split.
How long until I have a working environment?
Once the GPU stack is set up, a first test inference is quick. Setting up that stack, a specific model, an inference server tuned to your traffic or a fine-tuning run is work our team can do with you, and your included DevOps hours can cover it.
Can I fine-tune my own model?
Yes. The 96 GB card handles LoRA or QLoRA fine-tuning of models up to about 70 billion parameters, and the 24 GB card covers light fine-tuning of small models. Training a GPT-class model from scratch does not fit on a single GPU server.
I use Pinecone today. Can I move to Qdrant with you?
Yes. Pinecone is an external service billed by usage, while Qdrant runs on your own VPS in the EU, so your vectors stay on a server you control. We can handle the export, the checks on retrieval quality and the cut-over from your included DevOps hours.
How do the included DevOps hours work on a GPU plan?
GPU servers include 10 h/month. For comparison: 1 h/month from cPanel Advanced up, 1–8 h on VPS and 5–10 h on dedicated servers, depending on plan. You request the hours from the client panel, you propose a date, and our team delivers. Hours are not carried over, and extra hours are accepted and billed separately.
Are GPU servers billed by the hour?
No. They are billed monthly or annually, so a long training run does not turn into a surprise on the invoice. Romanian VAT is shown separately, and current plans are on the pricing page.
Choose a GPU server, or a VPS for your vector database
GPU servers with RTX PRO 4000 or RTX PRO 6000, and VPS plans for Qdrant and ChromaDB, are listed with current prices on the plans page. GPU servers are billed monthly or annually and include 10 DevOps hours a month.
See GPU plans →