2 to 5 MW per cluster
The 2 to 5 MW range is a genuine catalog hole between hyperscale-grade prefab skids and sub-1 MW edge modules. Tier-1 modular vendors top out their reference clusters above 7 MW, where the buyer is hyperscaler-direct; their pod products bottom out at 2 MW, where the buyer is single-rack edge. The middle — Tier-2 GPUaaS, BTC pivots, wholesale colo bridge capacity — is served today by ad-hoc multi-vendor stacks rather than catalog products.
Teraplex productizes that range. Two cooling variants — air-cooled inference and direct-to-chip liquid training. One spec sheet, one commissioning standard, one delivery schedule that includes the substation, the generator, the switchgear, and SCADA. Buyers don't integrate five vendors. They buy a cluster.
NVIDIA reference architecture compatible
Quantum-2 InfiniBand or Spectrum-X compatible. BlueField-3 ready. VAST, WEKA, and NetApp storage validated. Tier-2 GPUaaS buyers shortlist on whether the enclosure plays cleanly with the reference architectures their customers have already standardized on. Modules ship NCP-eligible — not bolted-on.
Compute fabric
Quantum-2 IB · Spectrum-X · BlueField-3 DPUs.
Storage validated
VAST · WEKA · NetApp reference designs.
Vendor-agnostic IT
Bring your own GPUs and servers. We sell the enclosure plus power.
Density envelope
Air-cooled is sized for AI inference workloads and current-generation training at 50 kW. Direct-to-chip liquid takes the same container envelope to 150 kW per rack for current-generation Blackwell-class training and is on the roadmap for the next generation. Same footprint, same UPS topology, same operating envelope — the cooling architecture is the differentiator.
Configurations
Sized to your project. The productized 2 to 5 MW cluster is the Year-1 wedge; phased deployments use the same cluster as the building block.
Edge / single module
ISO-format containers for ROBO, federal edge, cell-tower-adjacent, and bridge capacity below 100 kW. EMP-shielded variants available for federal and tactical-edge programs. Defensive scope; usually paired with a larger power-infrastructure deal.
Spec sheet →Productized cluster
The Year-1 wedge. Air-cooled inference and direct-to-chip liquid training variants. Substation, generator, switchgear, and SCADA on the same schedule. One PO. One weld-traceability standard. Site-ready in 6 to 8 months from PO.
Spec sheet →Phased deployment
Pilot to full campus in three to five commissioning events. CapEx aligned to revenue; deploy capacity only when tenants or workloads are confirmed. GPU generations change every 18 to 24 months — Phase 1 runs current hardware; Phase 3 specs the next generation.
Spec sheet →Spec library
MDC-500 (air, 50 kW/rack), MDC-500 DLC (direct-to-chip liquid, 150 kW/rack), and Type XL productized cluster — full spec sheets, dimensional drawings, electrical schematics, and a compatible-architecture summary.
Direct download, no email required. Subscribe via the form to receive updated revisions automatically.
Download
Integrated procurement
The structural advantage versus single-product modular vendors: substation, transformer, generator, switchgear, and SCADA shipped and commissioned on the same schedule as the modules. One PO. One weld-traceability standard. One responsible party when something slips. The power side of the order is the wedge that makes the cluster site-ready in 6 to 8 months instead of 12 to 18.
See the power-infrastructure catalog →Tell us your rack count, your density target, and your need-by date. We come back within one US business day with configuration recommendations and a delivery quote.
