Solutions
Four kinds of work. One kind of infrastructure.
01 — LLM training & inference
Train frontier models. Serve them at reasoning speed.
Pre-training wants the largest NVLink domain you can get; reasoning and long-context inference want memory capacity and bandwidth. GB300 NVL72 puts 72 GPUs and 20 TB of HBM3e in a single domain — and we connect those domains with a non-blocking fabric.
What matters
- NVLink domain size for tensor and expert parallelism
- Non-blocking 800 Gb/s per GPU for data parallelism
- Checkpoint throughput that never stalls a step
| Platform | GB300 NVL72 for pre-training and long-context inference; HGX B300 for fine-tuning and serving |
|---|---|
| Fabric | Quantum-X800 InfiniBand, rail-optimized fat tree |
| Storage | Parallel all-flash over a dedicated RoCE fabric |
| Software | Slurm or Kubernetes, NCCL tuned and validated at handover |
| Typical scale | 1–16 scalable units · 576–9,216 GPUs |
02 — Physical AI
Simulate the world before you build in it.
Robotics and autonomy teams run simulation, synthetic-data generation and policy training as one pipeline. That means GPUs that can be carved up flexibly, and storage that can keep thousands of parallel worlds fed.
What matters
- Mixed simulation, rendering and training throughput
- Flexible partitioning across teams and jobs
- High-throughput dataset and log I/O
| Platform | HGX B300 nodes for flexible partitioning; GB300 NVL72 for world-model pre-training |
|---|---|
| Fabric | Rail-optimized InfiniBand between training nodes |
| Storage | Parallel all-flash sized for sensor-scale datasets |
| Software | Kubernetes with GPU operator; Slurm on request |
| Typical scale | 128–1,024 GPUs |
03 — Multimodal
Video, audio, images and text — at once.
Multimodal models consume far more bytes per token than text. The bottleneck moves to data loading, so we give storage its own fabric and size it to the GPUs, not the other way round.
What matters
- Storage bandwidth per GPU
- A storage fabric separate from gradient traffic
- HBM capacity for long video contexts
| Platform | GB300 NVL72 or HGX B300 |
|---|---|
| Fabric | InfiniBand for compute; Spectrum-X Ethernet for storage |
| Storage | Parallel all-flash with RoCE to every tray |
| Software | Kubernetes, Ray or Slurm |
| Typical scale | 256–4,608 GPUs |
04 — Enterprise
Your own cluster. Nobody else’s.
Dedicated hardware on a dedicated fabric, isolated from every other tenant, with contracts sized to your roadmap rather than a spot market. Bring your identity provider, your network and your security reviews.
What matters
- Single-tenant isolation down to the fabric
- Private connectivity into your network
- Predictable capacity and pricing
| Platform | GB300 NVL72 or HGX B300, single-tenant |
|---|---|
| Fabric | Dedicated InfiniBand and Ethernet fabrics |
| Access | Private links and cross-connects; your SSO |
| Operations | Audit logging, change windows agreed with you |
| Typical scale | 64 GPUs to multiple scalable units |