بنية تحتية للذكاء الاصطناعي

AI Accelerators

GPUs & AI Chips

The compute engines powering model training and inference; the most constrained product in technology.

• NVIDIA H100 SXM5 / PCIe

  80 GB HBM3 Hopper; two generations behind Blackwell Ultra, now the deepest secondary and refurbished GPU market in the world.

  Inference Secondary Market

  Details

• NVIDIA H200 SXM5

  141 GB HBM3e; dramatically faster memory for trillion-parameter LLM inference deployments.

  Inference LLM

  Details

• NVIDIA B200 (Blackwell)

  180 GB HBM3e, 9 PFLOPS dense FP4; the volume Blackwell GPU. Lead times have eased as Blackwell Ultra ramps.

  Training Volume

  Details

• NVIDIA B300 (Blackwell Ultra)

  288 GB HBM3e, 15 PFLOPS dense FP4; NVIDIA's current top-of-stack GPU, shipping since January 2026.

  Inference Reasoning Flagship

  Details

• NVIDIA Rubin (VR200)

  288 GB HBM4 at 22 TB/s, 50 PFLOPS NVFP4; the next-generation platform now shipping to early-access hyperscalers.

  Next-Gen Allocation

  Details

• NVIDIA GB200 / GB300 Grace Blackwell

  Grace CPU + dual Blackwell GPUs on one liquid-cooled module; the building block of NVL72 rack-scale systems.

  Rack-Scale Training

  Details

• NVIDIA L40S (Ada Lovelace)

  48 GB GDDR6 air-cooled inference GPU; mature, widely stocked, and still the volume PCIe part in ASEAN and GCC enterprise.

  Inference Enterprise

  Details

• NVIDIA RTX PRO 6000 Blackwell Server Edition

  96 GB GDDR7, 600W passive PCIe; the Blackwell successor to L40S for air-cooled enterprise inference.

  Inference Enterprise PCIe

  Details

• AMD Instinct MI300X

  192 GB HBM3, 5.3 TB/s. AMD's flagship NVIDIA alternative adopted by Azure, Oracle, and Asian clouds.

  Inference Alternative

  Details

• AMD Instinct MI325X

  256 GB HBM3e on the MI300 platform; the bridge generation between MI300X and the MI350 series.

  Inference LLM

  Details

• AMD Instinct MI350X / MI355X

  288 GB HBM3E, 8 TB/s, CDNA 4 on 3nm; AMD's current flagship, direct competitor to Blackwell B200 and B300.

  Inference Training Flagship

  Details

• Intel Gaudi 3

  128 GB HBM2e, native PyTorch support; 30–40% lower ASP than H100 with competitive BF16 performance.

  Training TCO

  Details

AI Servers

Compute Nodes & Systems

Pre-integrated GPU server platforms from all major OEMs; turnkey and fully configurable.

• NVIDIA DGX B300 / DGX B200

  8-GPU Blackwell Ultra appliance, 2.1 TB HBM3e, air-cooled at ~14 kW; NVIDIA's reference AI factory node.

  Training Inference Appliance

  Details

• NVIDIA GB300 NVL72 / GB200 NVL72

  72 Blackwell Ultra GPUs as one rack-scale node: 20.7 TB HBM3e, 1.1 exaFLOPS FP4, ~120 kW, fully liquid-cooled.

  AI Factory Hyperscale

  Details

• SuperMicro HGX B300 / B200 / H200 Servers

  OCP and EIA 8-GPU HGX systems; 4U liquid-cooled and 8U air-cooled B300 chassis, first to ship each Blackwell generation.

  Cloud Colocation

  Details

• Dell PowerEdge XE9780 / XE9785

  Dell's Blackwell Ultra and MI355X 8-GPU servers; air-cooled 10U or liquid-cooled 3OU for IR7000 racks.

  Enterprise Gulf

  Details

• HPE ProLiant Compute XD685

  8x HGX B300, B200, H200, or MI355X in one chassis; 5U liquid-cooled or 6U air-cooled with HPE iLO 6.

  HPC Enterprise

  Details

• Lenovo ThinkSystem SR675 V3 / SR680a V3

  Lenovo's PCIe and HGX GPU servers; strong pipeline in China, ASEAN, and GCC enterprise accounts.

  APAC Enterprise

  Details

AI Networking

InfiniBand & Ethernet Fabric

The interconnect layer that determines whether a GPU cluster trains at full speed; or sits idle.

• NVIDIA Quantum-X800 InfiniBand

  144x 800G XDR ports, 115 Tb/s, SHARP v4 in-network computing; the reference fabric for B300 and GB300 clusters.

  InfiniBand HPC

  Details

• NVIDIA ConnectX-8 / ConnectX-7 SuperNIC

  800G (ConnectX-8) or 400G (ConnectX-7) InfiniBand or Ethernet with GPUDirect RDMA; one per GPU, high-volume consumable.

  RDMA Volume

  Details

• NVIDIA Spectrum-X 800G Ethernet

  InfiniBand-class congestion control over Ethernet; the fabric behind the largest AI clusters at Meta and xAI.

  Ethernet Hyperscale

  Details

• Arista 7800R4; 800G AI Spine Switch

  Up to 576x 800G ports per chassis; leading Ethernet spine for AI data centres globally.

  Spine 800G

  Details

• Cisco Nexus 9364E-SG2; 800G Fabric

  64x 800G ports, Silicon One G200. Cisco's purpose-built AI networking platform.

  Enterprise Cisco

  Details

• 400G / 800G / 1.6T Optical Transceivers

  OSFP and QSFP-DD optics for every switch and NIC port; the highest-volume consumable in any cluster build.

  Optics Volume

  Details

Data Centre Infrastructure

Racks, PDUs & Cabling

The physical chassis everything else sits in; from racks and power distribution to structured cabling.

• Deep-Chassis Server Racks (42U–52U)

  EIA-310 19" racks with 800–1,200mm depth for GPU servers; up to 2,000 kg static load rating.

  Racks Volume

  Details

• Intelligent Metered Rack PDUs

  Per-outlet monitoring up to 60 kW; essential for AI rack deployments. Brands: Vertiv, APC, Raritan.

  Power Monitoring

  Details

• IP KVM & Serial Console Servers

  Out-of-band BIOS-level remote access; essential for lights-out GPU cluster operations.

  Management OOB

  Details

• Hot/Cold Aisle Containment Systems

  Modular panels improving cooling efficiency 20–30%; standard fit-out item for new AI data centres.

  Containment Efficiency

  Details

• MPO/MTP Fibre Cabling Systems

  Pre-terminated MPO-24 / OM5 / OS2 assemblies for 400G/800G AI networking deployments.

  Fibre Cabling

  Details

Storage Systems

AI-Optimised Storage

Keeping GPUs fed; high-bandwidth, low-latency storage that matches accelerator throughput.

• VAST Data Platform

  All-flash scale-out storage certified on NVIDIA DGX SuperPOD; the leading AI training storage platform.

  Training All-Flash

  Details

• WEKA WekaFS Parallel Storage

  Sub-millisecond NVMe-oF with GPUDirect; the leading storage choice for neoclouds (CoreWeave, Lambda).

  Neoclouds Parallel

  Details

• Pure Storage FlashBlade//EXA

  10+ TB/s aggregate throughput for frontier model training and large-scale dataset pipelines.

  Hyperscale Throughput

  Details

• Enterprise NVMe SSDs (U.2 / E3.S)

  PCIe Gen4/Gen5 SSDs 7.68–30.72 TB for AI servers. Brands: Samsung, Micron, Kioxia.

  SSD Volume

  Details

• DDN EXAScaler AI400X2 / AI800

  Parallel Lustre / DAOS storage appliance; largest installed base in national supercomputing facilities.

  HPC Supercomputing

  Details

Cooling Solutions

Liquid, Immersion & Air

AI servers generate 10x the heat of traditional servers; cooling is now a first-order infrastructure problem.

• Direct Liquid Cooling (DLC) Kits & CDUs

  Cold plates and coolant distribution units; mandatory for B300, GB300 NVL72, MI355X, and every Rubin system. Brands: Vertiv, CoolIT, Asetek.

  DLC Mandatory

  Details

• Single-Phase Immersion Cooling Tanks

  Servers submerged in dielectric fluid; 250+ kW per tank, PUE approaching 1.03. Brands: GRC, Submer.

  Immersion High-Density

  Details

• Two-Phase Immersion Cooling

  Fluorocarbon phase-change cooling. PUE <1.02, no rack fans. Brands: LiquidStack, Iceotope.

  Advanced Efficiency

  Details

• Rear-Door Heat Exchangers (RDHx)

  Passive rack-level liquid cooling; retrofits existing air-cooled facilities for higher-density AI racks.

  Retrofit Air-Assist

  Details

• CRAC / CRAH Precision Cooling Units

  Row-based and room-based precision air cooling. Brands: Stulz, Vertiv, Schneider Electric.

  Air Precision

  Details

Power Infrastructure

Distribution, UPS & Backup

Reliable, high-density power delivery; from intelligent PDUs and busway to backup generators.

• Modular UPS Systems

  3-phase UPS 10–750 kVA, Li-ion variants for AI rack deployments. Brands: Eaton, Vertiv, Schneider Electric.

  UPS Resilience

  Details

• Busway Power Distribution Systems

  Overhead busway 800A–5,000A per run replacing cable trays in high-density AI buildings.

  Busway High-Density

  Details

• Generator Sets & Automatic Transfer Switches

  Diesel and HVO generators 500 kVA–10 MVA with ATS. Brands: Cummins, Caterpillar, MTU, Kohler.

  Backup Generator

  Details

• Battery Energy Storage Systems (BESS)

  Li-ion / LFP BESS for UPS bypass and peak shaving in AI campuses; replacing diesel as primary backup.

  BESS Green

  Details