Big Data Week Bucharest returned on 6 and 7 October 2026, bringing data engineers, AI practitioners and technology leaders together for two days of talks and workshops at the Novotel in Bucharest. This year's theme, Big Data & AI in Practice, set the direction for a programme about moving data and AI systems out of experimentation and into reliable, everyday operation.
MetalSoft took part as a Presenting Partner for the fourth consecutive edition, and our colleague Andrei Tomus, Network Engineer at MetalSoft, presented on the Applied AI stage with a session built around deploying an NVIDIA Spectrum-X fabric with MetalSoft Fabric Manager.
Bringing industry expertise together
Big Data Week is a global platform of community events for data professionals, and the Bucharest conference is its long-running Romanian edition. The 2026 programme was built around two main-stage tracks, Practical Big Data Engineering on 6 October and Applied AI on 7 October, each paired with a parallel workshop track for hands-on technical sessions.
Speakers came from organisations including Cloudera, AWS, Google Cloud, Zalando, eMAG, ING Hubs Romania, Endava, EPAM Romania and Datadog, with sessions ranging from data platform engineering to generative AI in production. The common thread was practical: what it takes to run these systems at scale, day after day, rather than what they can do in a proof of concept.
From fragmented operations to one control plane
On the Applied AI stage on 7 October, Andrei Tomus presented Exploring AI Factory Network Topologies in NVIDIA AIR. Under the title Bare Metal Intelligence, his presentation told the story in three parts: the problem as MetalSoft sees it, MetalSoft's approach to solving it, and a use case built around an NVIDIA Spectrum-X fabric.
The problem is one most infrastructure teams will recognise. Enterprises run heterogeneous environments, often deliberately, to reduce supply chain risk, but every vendor brings its own tooling. Teams work from isolated scripts and tools; provisioning is usually automated, but tuning, incident response and troubleshooting are not; and topology and intent data is scattered across tools or does not exist at all. The consequences are vendor lock-in, little skill reuse across silos, tool proliferation, inconsistent governance and delivery bottlenecks that the presentation put at three months on average.
Against that, the presentation set out the end state MetalSoft is built for: a single platform that brings every infrastructure domain together, driven by end users through cloud-like self-service and governed by policies that the network, server, storage, security and platform teams each define. With one accurate source of truth, requirements such as DORA, the EU AI Act and sovereignty become easier to satisfy, and a small, capable team can run the estate with help from AI.
MetalSoft delivers this as a software-defined data centre platform: one abstraction layer across the major server, network and storage vendors that exposes each vendor's capabilities without the lock-in, and one control plane for Day 0 discovery, Day 1 provisioning and Day 2 operations across its Compute & Storage Manager, Fabric Manager and cluster extensions.
Fabric Manager was the focus of the talk. It is intent-based, manages both the underlay and the overlay, and works across multiple fabrics. On Day 0 it handles zero-touch provisioning and fabric-level cabling detection; on Day 1 the underlay, the EVPN-VXLAN overlay, link aggregation, BGP sessions, VRFs and data centre interconnect; on Day 2 drift detection, health monitoring, AI-based troubleshooting and automatic backup. It supports NVIDIA Spectrum-X and InfiniBand alongside Cisco Nexus and ACI, Enterprise SONiC, Dell OS10, Juniper Junos OS and Arista EOS.
The use case: an NVIDIA Spectrum-X fabric in NVIDIA DSX Air
The final part of the presentation, framed against NVIDIA's Enterprise Reference Architectures, walked through deploying an NVIDIA Spectrum-X fabric with Fabric Manager in NVIDIA DSX Air, listed in the programme as NVIDIA AIR. DSX Air is NVIDIA's cloud-hosted data centre simulation platform; it behaves like a production environment and can be used to build a digital twin of an IT infrastructure, so the network layer of an AI factory can be designed and exercised without physical hardware. Spectrum-X is NVIDIA's Ethernet networking platform for AI workloads.
Starting from Fabric Manager's global and site controllers, the walkthrough created a Spectrum-X fabric, added Spectrum-X switches to it, deployed the fabric and pushed a BGP EVPN fabric configuration to it. It then created tenant configurations and endpoints, deployed infrastructure on the fabric through MetalSoft's Infrastructure Designer and ran post-deployment connectivity checks.
The closing steps addressed a day-two scenario that any multi-tenant AI environment will face: an NVIDIA HGX system was moved from one tenant to another and connectivity was checked again after the migration, with the fabric changes made in Fabric Manager rather than on individual switches.
The presentation then stepped out of the simulation to show infrastructure deployment with a physical server, from server registration to server deployment, through MetalSoft Compute & Storage Manager, and closed with the wider platform: virtualisation orchestration with KVM (Incus) and VMware, zero-touch provisioning for network devices, and the services MetalSoft delivers, from bare metal, VPCs and tenant isolation to virtual machines, containers, managed OpenShift, managed VCF and block, file and object storage.
Anyone who wants to explore the subject hands-on can do so in the MetalSoft Fabric Manager lab in NVIDIA DSX Air, which deploys simulated Spectrum-X fabrics in one-scalability-unit, two-scalability-unit and three-tier topologies.
Supporting the data and AI community
MetalSoft's involvement with Big Data Week Bucharest goes back several editions. We have been a Presenting Partner since 2023, and Alex Bordei, our Chief Product Officer, spoke in 2023 on building flexible and affordable infrastructure for private LLMs and in 2024 on ELI, MetalSoft's AI-assisted network diagnosis. Andrei's session carries that thread into the network layer of AI factories.
Supporting the conference is part of a broader commitment to the engineers who build and run data and AI platforms, whether that platform is an AI factory or a private cloud. Events like Big Data Week are where those conversations happen in person, and we look forward to continuing them at the next edition.
To see MetalSoft Fabric Manager deploy a Spectrum-X fabric for yourself, request a demo.