Everything a training or inference cluster needs, on one bill of materials: GPU nodes, fabric, storage, power and cooling — staged delivery with the compliance work done.
Cluster projects fail on the seams: GPUs arrive without fabric, cooling specs surface too late, customs holds a rack hostage. We quote the whole stack together and ship it in deployment order.
| Layer | Options | Notes |
|---|---|---|
| Compute nodes | H200 NVL PCIe nodes; HGX B200/B300 systems | PCIe path avoids allocation queues |
| Fabric | InfiniBand NDR; 400G Ethernet (QSFP-DD) | Cut-length cable schedules included |
| Storage | NetApp AFF / PowerScale for datasets & checkpoints | Sized to GPU count and IO profile |
| Head & service nodes | R660/DL360 class for scheduling, login, monitoring | Typically from stock |
| Power | Intelligent PDUs, high-wattage PSUs, UPS | Per-rack power budgets calculated |
| Cooling | Rear-door HX for PCIe pods; DLC for HGX density | Decision tree in our cooling guide |
Stock rotates daily — positions are "typically available" and confirmed per request, usually within one business day. Stock guides →
A 4-node H200 NVL inference pod with fabric and storage lands around $350k all-in; HGX training clusters run from high six figures depending on GPU allocation and scale. We quote the complete BOM, not just the GPUs.
PCIe (H200 NVL) deploys in weeks from stock and serves most inference and fine-tuning. HGX earns its premium for large-scale training with NVLink-heavy communication. We model both against your workload.
Yes — InfiniBand or 400G Ethernet topologies with port-level BOMs and cut-length cable schedules, so nothing arrives missing.
That is our lane: export screening, DDP delivery, customs clearance and staged shipping from Hong Kong and Dubai hubs.
Integration partners at major destinations, or we deliver rack-ready kits with elevations and schedules for your team.
Pricing, availability and delivered lead time within one business day.