Hardware blueprints you can budget from — a B300 cluster, a GB300 NVL72 rack, an H200 inference pod and a starter stack. Each with a bill of materials, a diagram and indicative pricing. We tailor to your workload and screen every order for compliance.
The current platform for new clusters: eight-GPU HGX B300 servers on a non-blocking 800G fabric, with flash storage for datasets and checkpoints and direct liquid cooling at about 60 kW per rack. 288 GB per GPU and FP4 make it the choice for pre-training, fine-tuning of large models and high-throughput inference. Supplied through allocation to documented projects.
| Component | Specification | Notes |
|---|---|---|
| Compute | 4–16 × HGX B300 (8 × B300 288 GB each), OEM or DGX B300 | allocation to documented projects |
| Fabric | 800G InfiniBand (Quantum-X800) or Spectrum-X, ConnectX-8 | non-blocking leaf-spine |
| Storage | all-flash for 32 GPUs; parallel file system from ~128 GPUs | datasets + fast checkpoints |
| Cooling | direct liquid cooling: CDUs, manifolds, facility water loop | 4 servers per rack, ~55–60 kW |
| Power | high-density PDUs, UPS | ~14.5 kW per server |
| Alternative | 4 × HGX H200 (32 GPUs), air-cooled at 20–35 kW per rack | lower entry cost, from ~$1.4M |
Rack-scale Blackwell Ultra: 72 B300 GPUs and 36 Grace CPUs joined by nine NVLink switch trays into a single 72-GPU domain with about 20 TB of HBM3e. Model-parallel work that would cross InfiniBand between eight-GPU servers stays on NVLink, which is what trillion-parameter training and high-concurrency serving of the largest reasoning models need. Racks scale out over 800G to each other and to storage. The site has to be built for it: about 132 kW per rack and liquid cooling designed in from the start.
| Component | Specification | Notes |
|---|---|---|
| Compute | 18 compute trays: 72 × B300, 36 × Grace | delivered and accepted as one system |
| NVLink | 9 NVLink switch trays, 130 TB/s | one domain across the rack |
| Scale-out | 800G InfiniBand or Spectrum-X, ConnectX-8 | rack-to-rack and storage |
| Cooling | direct liquid cooling, CDU per row or rack | liquid-only |
| Power | ~132 kW per rack nominal | up to 192 kW busway advised |
| Site | your hall, or liquid-ready capacity with our Tier 1 partner | GPU cluster hosting |
The workhorse for serving large models privately: four PCIe GPU nodes on a 400G fabric with shared flash storage. The lower-cost entry for production inference: H200 NVL runs in standard air-cooled racks, and a project order moves faster than HGX allocation.
| Component | Specification | Notes |
|---|---|---|
| GPU nodes | 4× server, 2× H200 NVL 141 GB each | Dell R760xa / HPE DL380a class |
| Fabric | 400G QSFP-DD or InfiniBand | cut-length schedule included |
| Storage | NetApp AFF / Dell PowerScale | sized to model + cache |
| Head/service | R660 / DL360 class | scheduling, login, monitoring |
| Power/cooling | intelligent PDUs, rear-door HX | per-rack budget calculated |
The smallest credible private-AI footprint: a DGX Spark for development, one H200 NVL inference node for launch, and an NVMe-heavy node for RAG and vector search. The same software stack scales to the pod and the B300 cluster above when you grow.
| Component | Specification | Notes |
|---|---|---|
| Development | NVIDIA DGX Spark | 128 GB unified, desktop |
| Inference | 1× node, 2× H200 NVL | 70B-class production serving |
| RAG / vector | NVMe CPU node | embeddings, retrieval, cache |
| Networking | 25/100G switching | Catalyst / Nexus |
Running AI on this infrastructure? Haink also builds the LLM & ML software that runs on it — model, pipeline and GPUs under one contract.
Tell us the workload, where it should run and when — we come back with the architecture, a costed bill of materials, hosting options and a realistic timeline.
Not a quote? Ask a question