Capability | AI infrastructure
GPU platforms for training and inference at scale.
Training at scale puts simultaneous pressure on three things that are rarely planned together: the gradient-synchronisation fabric, the power density per rack and the thermal path. We engineer them as one chain, aligned to the platform reference architectures, with a clear operational path from Day 0 to Day 2.
What it covers
- GPU platforms
- HGX / DGX H100–H200 and Blackwell NVL72 class, NVLink / NVSwitch scaling
- Training fabric
- 400G InfiniBand / Spectrum-X Ethernet, rail-optimised, RDMA for gradient synchronisation
- Density & power
- 40–120 kW racks with matched direct-to-chip cooling and power chain
- Orchestration
- Kubernetes + GPU Operator, Slurm, Run:ai — with MIG partitioning
- Serving & MLOps
- Inference endpoints, autoscaling, DCGM / Grafana observability
- RA-aligned
- Built to the NVIDIA Enterprise RA with a Day 0 / 1 / 2 operational path
From national-scale practice
Beyond the platform: the disciplines that take an AI machine from reference architecture to an open, validated service. The same practice underpins our engineering dossier for DEUS 11, a national-scale supercomputer programme in Cluj-Napoca.
- Dual ecosystem
- Proprietary & open accelerator stacks planned side by side — commercial codes routed, open codes validated on the alternate stack
- Partition design
- Training & inference as distinct partitions — separate hardware profiles & scheduling policies per role
- Validation gate
- Storage & I/O paths proven under production-scale training & simulation — before a partition opens
- Foundation models
- Training infrastructure for European, sector-specific foundation models — corpus & checkpointing requirements carried through to the fabric
- Service economics
- The machine's economics designed in — cost recovery & market access planned before go-live