Cluj-Napoca is the second pillar of our Romanian business right after Bucharest. The Transylvanian IT cluster, **Babes-Bolyai University** as the academic partner for most firms in the city, and a notably research-oriented market character — this isn't an outsourcing cluster of the "30 thousand juniors for a London bank" type, it's an R&D cluster where major firms build their own centres for AI, computer vision, automotive software and high-performance computing.
AI cluster for a research lab — the reference project in Cluj
Our delivered project `ai-cluster-research-lab` (2025) is exactly the type of contract for which Cluj is currently the best address in Europe. A research lab at one of the local universities needed AI compute capacity at a level that isn't about cloud — about its own on-prem, latency-deterministic infrastructure. Delivery: **8 nodes with 2× NVIDIA RTX PRO 6000 Blackwell** per node (16 GPUs total, serially interconnected), **NVIDIA Mellanox InfiniBand NDR** backbone (400 Gb/s per port, sub-microsecond RDMA latency), **rear-door heat exchanger** water cooling (captures about 80% of heat in the rack before it reaches the hot aisle), and **Slurm + Pyxis/Enroot** orchestration for containerised workloads (each research team has its namespace, but the GPU is shared via fairshare scheduler).
This project wasn't about "we buy expensive GPUs" — it was about an **engineering solution for heat and network**. Blackwell nodes run continuously above 700 W per card. 16 cards in one rack means about 12 kW of just GPU power draw, plus CPU, plus PSU losses — total rack TDP above 18 kW. Traditional front-to-back air cooling for a rack in this category isn't enough. Rear-door HX with a glycol loop is currently the most practical solution for research-scale clusters (full DLC immersion is an option, but over-engineered for 8 nodes). InfiniBand NDR is a necessity, not a luxury — for MoE models and tensor-parallel inference, network latency is a direct part of wall-clock training time.
For us the interesting part was that **Slurm orchestration** wasn't dimensioned "by the book" from the NVIDIA reference design — it was tuned according to the actual workload mix (60% vision research, 30% LLM fine-tuning, 10% classic HPC). The Pyxis container layer allowed researchers to bring their own CUDA versions without colliding with the global installation — that's exactly what makes the cluster usable for a multi-team environment.
Cluj IT cluster — Bosch, Emerson, Endava, Tetarom
**Bosch Engineering Center Cluj** (campus in Jucu, 15 km north-east) is an R&D centre for automotive electronics, control units and embedded systems. For infrastructure work (testbeds, HW-in-the-loop simulators, chip laboratories) it's a client with continuous demand. **Emerson Electric Cluj R&D** is a centre for industrial automation — DeltaV, AMS Device Manager, Plantweb digital ecosystem. **NTT DATA Cluj**, **Endava Cluj**, **Yardi**, **Bitdefender Cluj branch** — software centres in various verticals. **Tetarom IT cluster** (industrial park in Jucu) is the largest IT cluster outside Bucharest — more than 50 companies, own electrical and fibre infrastructure.
**De'Longhi Romania** in Cluj produces advanced kitchen appliances for the European market — that's a different world, but for an integrator it means a combination of conventional industrial automation with a precision QC line.
Academic bridge — UBB
**Universitatea Babeș-Bolyai (UBB)**, with 41,000 students, is Romania's second-largest university after Bucharest, and in Cluj it's the academic partner of practically every significant tech firm. For an integrator that means two things: (1) **availability of qualified local collaborators** — Cluj has no problem with personnel, which in smaller Romanian cities is a real constraint; (2) **research-grade infrastructure projects** are more frequent than in other cities, because university and firms jointly build compute laboratories and testbeds. Our `ai-cluster-research-lab` project was exactly at this intersection.
Regulatory framework
For electrical installations **I7-2011** plus **NTE 007/08/00** for LV/MV distribution. For building permits **Lege 50/1991**. For occupational safety **Codul Muncii (Legea 53/2003)** + **HG 1146/2006**. For data centres in Cluj there are no special regional rules — the national framework applies plus EU directives (EED revision, ESPR for server HW from 2026).
Our commuting relationship
From Prešov to Cluj it's about **8 hours by car** via Hungary (Záhony — Oradea — A3 to Cluj), or realistically by air via **CLJ Avram Iancu** (Krakow → CLJ direct, ~2 hours). Cluj is commute-wise closer from Krakow than from Bucharest, which for cross-border projects from the PL side is an advantage — we can be on site the same day when there's a slot. For Romanian administrative processes everything goes via **the Bucharest office**, but physical presence in Cluj can be arranged from several directions.
Cross-border regime
For Slovak posted workers the same regime applies as for the whole of Romania — **ITM notification 30 days in advance** + **A1 from Slovak Social Insurance**. The Bucharest team speaks Romanian, Slovak, English; for Cluj English comfortably covers communication with Bosch and Emerson R&D, which are English-first environments. German capital (Bosch, Continental Cluj-Tureni) appreciates German, which we also have.
Conclusion
Cluj is currently in Europe one of the best addresses for AI compute, R&D laboratories and research-grade infrastructure. Our delivered project in Cluj isn't an exception — it's the type of contract we specialise in. For a new lab, research centre or automotive R&D testbed in Cluj-Napoca we bring a reference unusual in the local integration segment: a genuinely delivered Blackwell cluster, not "we can design it."