Hardware
Reference platforms we deploy.
All hardware QORINAI deploys is sourced from Hyperscalers, Australia's open-OEM server, storage and networking manufacturer and our strategic shareholder. Open-compute platforms configured, burned in and warranted in Australia; every configuration below has been specified and validated with them.
Built in Australia by Hyperscalers. Financed, deployed and operated by QORINAI.
GPU servers
Training and inference nodes.
NVIDIA HGX B300 8-GPU
Our standard training and inference node. Eight Blackwell Ultra GPUs on one NVLink domain, dual Intel Xeon 6, 800G fabric — delivered fully assembled and tested in Australia.
Request configuration →- GPU8× NVIDIA B300 Blackwell Ultra · NVLink + NVSwitch
- CPU2× Intel Xeon 6 6767P · 64 cores each
- FabricConnectX-8 · 8× 800 Gb/s OSFP
- Form8U · air-cooled or direct-to-chip liquid
- DeliveryLevel 10 assembled & tested · Australian warranty
NVIDIA HGX B300 8-GPU (Xeon 6)
Alternative HGX B300 chassis on Intel Xeon Granite Rapids for mixed air / liquid halls.
Request configuration →- GPU8× NVIDIA B300 HGX
- CPU2× Intel Xeon 6 Granite Rapids
- Form8U dual-processor
NVIDIA HGX B200 8-GPU
Blackwell at an established price point for fine-tuning and high-throughput inference. Air-cooled or DLC-ready.
Request configuration →- GPU8× NVIDIA B200 · 180 GB HBM3e each
- CPU2× Intel Xeon Scalable
- Fabric8× 400 Gb/s NDR InfiniBand or Ethernet
- Power~14 kW node, air or liquid
B300 density node · 72 GPUs per rack
3U building block for rack-scale B300 deployments where NVL72 is not required — up to 72 GPUs per rack on Intel Xeon 6.
Request configuration →- GPUNVIDIA B300 · 72 per rack
- CPUIntel Xeon 6 Granite Rapids
- CoolingDirect-to-chip liquid
AMD Instinct MI325X 8-GPU
The largest per-GPU memory in the fleet. AMD EPYC 9005 Turin host, ROCm stack — ideal for memory-bound inference of large open-weight models.
Request configuration →- GPU8× AMD Instinct MI325X · 256 GB HBM3e each (2 TB per node)
- CPU2× AMD EPYC 9005 Turin
- Fabric8× 400 Gb/s Ethernet
H200 / RTX PRO 6000 Blackwell PCIe (AMD)
Flexible 4U for inference endpoints, VFX rendering and enterprise AI where PCIe GPUs and air cooling are preferred.
Request configuration →- GPUUp to 8× NVIDIA H200 NVL (141 GB) or RTX PRO 6000 / 4500 Blackwell
- CPU2× AMD EPYC 9005 Turin
- CoolingAir, ≤ 30 kW racks
H200 / RTX PRO PCIe (AMD EPYC or Intel Xeon 6)
Same PCIe GPU platform on AMD EPYC 9005 / 9004 or dual Intel Xeon 6 hosts.
Request configuration →- GPUNVIDIA H200 · RTX PRO 6000 / 4500 Blackwell
- CPUAMD EPYC 9005 / 9004 or 2× Intel Xeon 6
NVIDIA Grace Hopper MGX
CPU–GPU coherent memory for graph, HPC and RAG workloads that don't fit in HBM alone.
Request configuration →- SoCNVIDIA GH200 Grace Hopper Superchip
- Memory480 GB LPDDR5X + up to 144 GB HBM3e, coherent
- Form2U MGX, PCIe Gen5
Next generation · Rubin-ready
Blackwell Ultra is what ships today. The halls are designed for what ships next.
HGX B300 and GB300 NVL72 are the deliverable choice for 2026–27 builds — same CUDA stack as the generation that follows. What changes with the next generation is power and fabric, so every QORINAI hall is designed to a 250 kW rack with 600 kW pathways, NVLink-6-class scale-up and 1.6T scale-out.
Buy what ships, build for what follows
B300 nodes ordered today are not stranded by Rubin: one software stack, one fabric family. We plan allocations with Hyperscalers so the second module lands on the next generation.
Power and cooling are the constraint, not silicon
A 150 kW hall is obsolete the day a Rubin rack arrives. QORINAI halls are engineered to 250 kW per rack with busway, CDU and floor-loading pathways to 600 kW.
Fabric-ready
Structured cabling and switch positions specified for 800G today and 1.6T next — NVLink 6, ConnectX-9, Ultra Ethernet and UALink all land in the same rows.
Rack-scale & open rack
One rack, one GPU.
NVIDIA GB300 NVL72
72 Blackwell Ultra GPUs and 36 Grace CPUs on a single NVLink domain. Requires direct-to-chip cooling and ~130 kW per rack — the density our Tier C halls are designed around.
Plan a pod →- GPU72× B300 · 36× Grace CPU · 18 compute trays
- HBMUp to ~21 TB HBM3e per rack
- NVLink5th-gen NVLink, 9 NVSwitch trays
- Cooling100 % direct-to-chip liquid, CDU per row
- Power~120–140 kW per rack, 415 V 3-phase busway
OCP hyperscale open-rack systems
Open Compute Project rack architecture for CPU-dense and storage-dense halls: Rackgo M and Rackgo X compute and JBOD sleds, Rackgo X-RSD rack-scale design.
Request configuration →- ComputeRackgo M / X compute sleds
- StorageRackgo M / X JBODs, SDS-ready
- RackOpen Rack v1 / v2 architecture, centralised power shelf
Open-rack compute sleds
Half-width and full-width compute sleds with centralised 12 V / 48 V power, tool-less service from the cold aisle.
Request configuration →- PowerCentralised rack power shelf
- ServiceTool-less, front-serviceable
Networking & storage
Compute is only as fast as the fabric.
InfiniBand XDR & Spectrum-X Ethernet
Rail-optimised, 1:1 non-blocking GPU fabrics for training pods; RoCEv2 800G Ethernet where a converged fabric is preferred. 800G QSFP-DD / OSFP optics sourced and tested in Australia.
Design a fabric →- IBNVIDIA Quantum-X800 XDR, ConnectX-8 SuperNIC
- EthernetNVIDIA Spectrum-X 400 / 800 GbE, Tomahawk-class switching
- DPUNVIDIA BlueField-3 for tenant isolation and storage offload
- Optics800G QDD SR8, 400G / 100G modules and AOCs
Bare-metal Ethernet switches
Open-networking switches with ONIE, running Cumulus Linux, SONiC or vendor NOS — for management, storage and north-south networks.
Request configuration →- T10001 / 10 GbE access and management (48-port)
- T3000 / T400025 / 100 GbE leaf
- T5000 / T7000100 / 400 GbE spine, QSFP28 / QSFP-DD
Management & out-of-band switching
BMC, IPMI and PDU networks on cost-effective 1 GbE with 10 GbE uplinks.
Request configuration →- Ports48× 1 GbE RJ45, 4× 10 GbE SFP+
- NOSONIE, Cumulus / vendor
All-flash, JBOD and parallel file systems
NVMe all-flash nodes for datasets and checkpoints, 4U JBODs for capacity tiers, and parallel file systems that keep 8-GPU nodes fed.
Request configuration →- FlashNVMe Gen5 all-flash storage servers (Kioxia CD8P / CM-series)
- Capacity4U JBODs 42–102 bays, storage servers 2U–4U
- SoftwareLustre / WEKA-class parallel FS, S3-compatible object, SDS
High-capacity storage servers
Dense 3.5" and NVMe storage servers for object, backup and archive tiers behind the parallel file system.
Request configuration →- BaysUp to 60× 3.5" + NVMe cache
- UseObject store, immutable backup, archive
Full catalogue
Everything else in the rack.
The complete open-OEM range available through our shareholder — configured, tested and warranted in Australia.
Servers
- GPU servers (NVIDIA HGX, PCIe, AMD Instinct)
- PCIe Gen5 Intel Xeon 6 / AMD EPYC 9005
- PCIe Gen4 Intel / AMD
- 1U · 2U · 3U–10U form factors
Hyper-converged & multi-node
- 4-node 2U hyper-converged
- Multi-node SSD servers, SDS / SDC
- Micro servers
Storage
- Storage servers
- 4U JBODs
- NVMe all-flash
- SDS / SDC ready
Networking
- Ethernet switches T1000 / T3000 / T5000
- Bare-metal ONIE — Cumulus / OCP
- 1G to 400G, SFP / QSFP / QSFP-DD
OCP hyperscale
- Rackgo M compute and JBODs
- Rackgo X compute, JBODs, rack architecture
- Rackgo X-RSD compute and storage
Workstations & edge
- PCIe Gen5 workstations
- Edge and far-edge devices
- End-user devices
Far edge
- Pole-mount and ruggedised nodes
- Telco / 5G edge
- Outdoor enclosures
Commodities
- CPU · RAM · SSD · HDD · NVMe
- NICs, optics, HBAs, GPUs
- Open-OEM parts that work in Tier 1 chassis
Pre-configured & lab
- Lab-tested reference builds, delivered racked
- Lab-as-a-Service test drives
- IP-appliance design process
Need a configuration and a price?
Send the workload and the target density; we return a bill of materials, a rack elevation and a power/cooling budget.