eRacks/HIGHLANDER

The eRacks/HIGHLANDER is the flagship of the eRacks AI line. There can be only one. It pairs NVIDIA's biggest PCIe card, the RTX PRO 6000 Blackwell Server Edition 96GB (the 600W passive server variant), with a 4U platform whose maker validates exactly this card at full power: eight dual-slot PCIe 5.0 x16 GPU bays, dual AMD EPYC 9005-series processors, hot-swap NVMe storage, and 3+1 redundant 3200W Titanium power supplies.

Default build: 4x RTX PRO 6000 Blackwell SE (384GB total VRAM), 768GB DDR5 ECC RAM, dual EPYC 9355 32-core, dual NVMe OS drives. Expand to 8 cards for 768GB of VRAM, enough to serve 405B-class models quantized, or fine-tune 70B-class models in full precision, entirely on-premises.

Where it fits

Above the eRacks/AINE (the same 96GB cards on our value platform) sits HIGHLANDER: the premium validated chassis, in stock now, with the NVIDIA software ecosystem (CUDA, TensorRT, NIM) and full model-training capability. For teams replacing $200K+ hosted-GPU spend with owned hardware, this is the box. A 5U 24-bay-NVMe Supermicro platform is available as a chassis upgrade for storage-heavy builds.

Software

Ships with the open-source AI stack pre-installed: Ubuntu LTS (or the Linux of your choice), Ollama, Open WebUI, vLLM, PyTorch, Podman (or legacy Docker, your choice). No per-seat fees, no per-token charges, your data never leaves the building.

Power

Power: 3+1 redundant 3200W Titanium supplies (9.6kW of redundant capacity on 208-240V datacenter power) carry even the full 8-card, 600W-per-card build (~5.8kW system) with headroom. We spec the PDU and rack power for your site at order.

Need NVLink-coupled training at HGX scale? The HGX H200/B200 8-GPU platform is available by quote.

Key Specifications

• Form factor: 4U rackmount, validated GPU platform
• GPUs: up to 8x RTX PRO 6000 Blackwell Server Edition 96GB (768GB total VRAM)
• CPU: dual AMD EPYC 9005-series (default 2x 9355 32-core)
• Memory: 768GB DDR5 ECC default, expandable to multi-TB
• Storage: hot-swap 2.5-inch bays (SATA arrays configurable; U.2/U.3 NVMe by quote) + dual M.2 OS drives
• Power: 3+1 redundant 3200W Titanium, 208-240V

Why Choose eRacks/HIGHLANDER?

• Validated at full power: the platform maker qualifies exactly these 600W cards, no derating, no experiments
• Open source, no lock-in: your models, your data, your Linux, no per-token or per-seat fees, ever
• Transparent sizing: priced and spec-d from real, verified component availability
• One vendor, whole stack: hardware, OS, AI stack, and support from eRacks

Ideal Use Cases

• Training small and mid-size models from scratch across up to 768GB of VRAM
• Fine-tuning 70B-class models in full precision
• Serving 405B-class models (quantized) to a whole organization
• Replacing $200K+/year hosted-GPU spend with owned hardware
• Regulated environments where data cannot leave the building

Pair it with an eRacks NAS: HIGHLANDER is all-flash (2.5-inch NVMe and SSD, no 3.5-inch bays). Your models and checkpoints live on fast flash; your training data, datasets, and backups belong on 3.5-inch CMR drives (conventional magnetic recording, the reliable server kind) with predictable, monitorable wear and far lower cost per terabyte. See our rackmount NAS line for the companion backup and dataset store.

Services

eRacks provisions HIGHLANDER to your workload: model deployment, fine-tuning pipelines, rack integration, and ongoing support contracts. Ask about our AI provisioning services when you order.