AI infrastructure guide
Compute, fabric, storage, and ops for AI programs
Enterprise reference for AI compute, network, storage, and operations—from workload sizing to rack-ready deployment on Elite APEX platforms.
AI infrastructure
Align workload, facility, and vehicle before the RFQ
Production AI stacks span four layers NTS validates before quote release: compute (GPU density, cooling), network (InfiniBand, RoCE, or Ethernet), storage (NVMe checkpoints, parallel FS), and operations (BMC, burn-in, CLIN mapping).
Platform selection starts with workload class—training/fine-tuning, inference at scale, HPC adjacency, or edge/ROBO—not silicon marketing names alone.
NTS integrates NVIDIA HGX (H100/H200/B200/B300) and AMD Instinct on U.S.-staged chassis in Fremont, with dual-vendor BOMs when programs need supply diversity.
Architecture paths
Layers, platforms, and procurement
Move from architecture layers into GPU SKUs, sizing, cooling, fabric, and contract vehicles.
- HGX B200 / B300
- Instinct options
- Air or DLC
- HBM guidance
- 2U / 4U / HGX
- RFQ checklist
- IB / RoCE
- Rail-optimized options
- QoS for mixed pools
- NVMe scratch
- Parallel FS / object
- Growth horizon
- kW/rack gates
- DLC / CDU
- Elevation packages
- 48h burn-in patterns
- Imaging baselines
- L11 / L12
- SEWP V
- ITES-4H
- OMNIA / DIR / TIPS
- CLIN-aligned quote
- Fabric sketch
- Cooling posture
Frequently asked questions
We review GPU count and SKU, host CPU/memory, fabric topology, storage throughput, rack kW, cooling mode, and contract vehicle scope—then return a CLIN-aligned quote.
- Four-layer check
- Facility gates called out
- Attachments for CORs
Yes, with scheduler partitioning and storage tiering. NTS architects node pools and network QoS so inference latency is not starved by training jobs.
- Partitioned pools
- Storage tiering
- QoS on east-west
Share model scale, framework, node count target, facility power, and vehicle (SEWP V, ITES-4H, etc.). Pair this guide with the GPU sizing guide and FAQ hub.
- Workload + facility
- Vehicle selection
- Sizing companion
2U–4U GPU servers for departmental training and inference; 8U HGX for large-model training; liquid-cooled paths when air cannot hold rack kW; dual-vendor NVIDIA/AMD BOMs when required.
- Form factor by density
- DLC when needed
- Supply diversity options
