AI Servers

NVIDIA and OEM AI server platforms, including HGX and Grace Blackwell systems, for enterprise training and inference

Source enterprise datacenter AI servers for model training, production inference, and high-density deployment with consultation-led alignment across platform, power, and rack design requirements.

  • Vera Rubin R200
  • GB300
  • GB200
  • HGX B300
  • HGX B200
  • HGX H200
  • HGX H100
Contact us
Premium AI server hardware in a dark datacenter setting

AI-ready platforms

Validated for enterprise training and inference workloads.

OEM-backed systems

Sourced through leading OEM and infrastructure partners.

Consultation-led sourcing

Platform selection aligned to power, density, and deployment goals.

OEM ecosystem

Manufacturing depth behind every platform offer

LXK CAPITAL sources across established OEM channels so enterprise buyers can evaluate the right compute platform with confidence.

  • Nvidia
  • Supermicro
  • Dell
  • Gigabyte
  • Asus
  • HP
  • Lenovo
  • Pegatron

Standard specification matrix

Cross-model procurement reference for current AI server platforms

Condensed technical matrix across seven NVIDIA platform classes, aligned to standard procurement dimensions for silicon, node design, fabric, I/O, and power planning.

Standard specification matrix comparing NVIDIA Vera Rubin R200, NVIDIA HGX H100, NVIDIA HGX H200, NVIDIA HGX B200, NVIDIA HGX B300, NVIDIA GB200, and NVIDIA GB300 across procurement-focused technical dimensions.
Specification dimension NVIDIA Vera Rubin R200 NVIDIA HGX H100 NVIDIA HGX H200 NVIDIA HGX B200 NVIDIA HGX B300 NVIDIA GB200 NVIDIA GB300
GPU Silicon & Architecture
Processor Model Rubin-generation accelerator complex H100 SXM accelerator H200 SXM accelerator B200 SXM accelerator B300 SXM accelerator Grace Blackwell superchip platform Grace Blackwell next-phase platform
Process Technology Advanced multi-die process class (vendor release dependent) TSMC 4N class TSMC 4N class with updated memory stack Blackwell-generation advanced node class Blackwell-generation advanced node class Advanced node mix across Grace + Blackwell dies Next-phase advanced node mix across Grace + Blackwell dies
Transistor Count Public class reference; final module totals by platform release ~80B class GPU die reference ~80B class GPU die reference with expanded memory subsystem Blackwell class, vendor-disclosed module-level aggregate Blackwell class, higher-bin module-level aggregate Superchip aggregate count by NVIDIA reference design Superchip aggregate count by next-phase reference design
Core Architecture Rubin architecture class with rack-scale integration Hopper architecture Hopper architecture (memory-extended tier) Blackwell architecture Blackwell architecture (performance-extended tier) Grace Blackwell coupled CPU-GPU architecture Grace Blackwell coupled CPU-GPU architecture (next phase)
Thermal Design Power (TDP) Liquid-first high-density power envelope; SKU dependent Up to ~700W per SXM GPU (OEM profile dependent) ~700W class per SXM GPU (OEM profile dependent) High-power liquid-ready envelope; OEM tuned High-power liquid-ready envelope; OEM tuned Superchip liquid-cooled power class per rack reference Superchip liquid-cooled power class per next-phase rack reference
GPU Memory
Memory Type & Capacity HBM-class stacked memory; capacity by platform SKU HBM3, 80GB class per GPU HBM3e, 141GB class per GPU HBM3e high-capacity class per GPU HBM3e expanded-capacity class per GPU HBM3e with coherent Grace memory domain Expanded HBM class with coherent Grace memory domain
Memory Bandwidth Multi-terabyte/s class target per accelerator complex Up to ~3.35 TB/s class per GPU Up to ~4.8 TB/s class per GPU Multi-terabyte/s class Blackwell profile Multi-terabyte/s class Blackwell profile (higher bin) Platform-level aggregate memory bandwidth via superchip design Higher platform-level aggregate bandwidth in next-phase design
Partitioning Support Virtual partitioning expected via platform firmware stack MIG-class partitioning support MIG-class partitioning support Partitioning support by Blackwell software stack Partitioning support by Blackwell software stack Superchip resource partitioning by platform software policy Superchip resource partitioning by platform software policy
Compute Precision
FP4 Expected low-precision inference path in Rubin generation toolchain Not primary path; FP8/FP16 optimized workflows Not primary path; FP8/FP16 optimized workflows Native FP4 support class for high-throughput inference Native FP4 support class for high-throughput inference Native FP4 support class via Grace Blackwell stack Native FP4 support class via next-phase Grace Blackwell stack
FP8 FP8 tensor path for training/inference balance FP8 tensor core acceleration FP8 tensor core acceleration FP8 tensor acceleration with Blackwell optimizations FP8 tensor acceleration with Blackwell optimizations FP8 tensor acceleration in superchip execution path FP8 tensor acceleration in next-phase superchip execution path
FP16 FP16/BF16 training path expected across Rubin software stack FP16/BF16 tensor acceleration FP16/BF16 tensor acceleration FP16/BF16 tensor acceleration FP16/BF16 tensor acceleration FP16/BF16 support across coupled Grace Blackwell execution FP16/BF16 support across next-phase coupled execution
TF32 / FP64 TF32 tensor path and FP64 support class by final silicon profile TF32 tensor acceleration and FP64 support TF32 tensor acceleration and FP64 support TF32 tensor acceleration and FP64 support TF32 tensor acceleration and FP64 support TF32 tensor path with FP64 support in platform stack TF32 tensor path with FP64 support in next-phase stack
System Host (Node)
System Model Integrated Rubin rack-scale system class HGX H100 8-GPU node class HGX H200 8-GPU node class HGX B200 8-GPU node class HGX B300 8-GPU node class GB200 superchip rack composition class GB300 superchip rack composition class
GPU Count Multi-accelerator count by rack composition 8 GPUs per node baseline 8 GPUs per node baseline 8 GPUs per node baseline 8 GPUs per node baseline Superchip module count by tray/rack design Superchip module count by tray/rack design
Total GPU VRAM Aggregate HBM pool by final rack SKU ~640GB class per 8-GPU node ~1.1TB class per 8-GPU node 1.5TB+ class per node (OEM dependent) Expanded 1.5TB+ class per node (OEM dependent) Rack-level aggregate GPU memory by superchip count Higher rack-level aggregate GPU memory by superchip count
System Processors Platform-integrated CPU complex defined by Rubin system design Dual x86 host CPU class (OEM selectable) Dual x86 host CPU class (OEM selectable) Dual x86 host CPU class (OEM selectable) Dual x86 host CPU class (OEM selectable) Integrated Grace CPU domain Integrated Grace CPU domain (next phase)
System RAM (DDR5) Host memory sizing by system composition policy DDR5 host memory, capacity OEM configurable DDR5 host memory, capacity OEM configurable DDR5 host memory, capacity OEM configurable DDR5 host memory, capacity OEM configurable High-capacity coherent memory domain; DDR5 integration by chassis design Expanded coherent memory domain; DDR5 integration by chassis design
Fabric & Expansion
Intra-System Interconnect Next-gen NVLink/NVSwitch class scale-up fabric NVLink + NVSwitch domain per HGX node NVLink + NVSwitch domain per HGX node New-generation NVLink/NVSwitch domain Extended NVLink/NVSwitch domain NVLink system fabric across Grace + Blackwell composition Enhanced NVLink system fabric continuity
External PCIe Expansion PCIe expansion policy by rack integration design PCIe Gen5 host lanes for NIC and storage expansion PCIe Gen5 host lanes for NIC and storage expansion PCIe Gen5 class host expansion (OEM chassis dependent) PCIe Gen5 class host expansion (OEM chassis dependent) PCIe expansion path by superchip platform board design PCIe expansion path by next-phase platform board design
Local Storage Capacity NVMe capacity envelope by OEM rack architecture NVMe U.2/E1.S options, capacity by chassis SKU NVMe U.2/E1.S options, capacity by chassis SKU NVMe expansion shelves or local bays by OEM design NVMe expansion shelves or local bays by OEM design Local NVMe tier for checkpoint/cache workloads, OEM sized Local NVMe tier for checkpoint/cache workloads, OEM sized
Networking & I/O
High-Speed Cluster I/O 400/800GbE or next-gen InfiniBand class by deployment policy 400GbE and/or NDR InfiniBand by NIC configuration 400GbE and/or NDR InfiniBand by NIC configuration 400/800GbE and InfiniBand class by chassis/network design 400/800GbE and InfiniBand class by chassis/network design High-speed fabric links aligned to AI factory rack topology High-speed fabric links aligned to next-phase rack topology
Data Processing Units DPU integration strategy by final platform release BlueField-class DPU optional by OEM profile BlueField-class DPU optional by OEM profile BlueField-class DPU integration commonly specified BlueField-class DPU integration commonly specified DPU/security offload path by rack reference design DPU/security offload path by next-phase rack reference design
Management Port Dedicated out-of-band management network path BMC/IPMI class OOB management port BMC/IPMI class OOB management port BMC/IPMI class OOB management port BMC/IPMI class OOB management port Out-of-band management path in rack control plane Out-of-band management path in next-phase rack control plane
Physical & Power
Form Factor / Dimensions Rack-scale integrated system dimensions by vendor reference OEM 4U/8U HGX node class dimensions OEM 4U/8U HGX node class dimensions OEM high-density HGX chassis dimensions OEM high-density HGX chassis dimensions AI factory rack/tray format per reference design AI factory rack/tray format per next-phase design
System Weight Rack composition dependent; integrator final sheet required Chassis dependent; varies by PSU/cooling/storage options Chassis dependent; varies by PSU/cooling/storage options Chassis dependent; higher with liquid-cooling infrastructure Chassis dependent; higher with liquid-cooling infrastructure Rack/tray dependent; final weight by build-out design Rack/tray dependent; final weight by build-out design
Power Architecture Redundant high-voltage feed with rack power shelves Redundant AC/HVDC PSU architecture by OEM Redundant AC/HVDC PSU architecture by OEM Redundant high-density power shelf architecture Redundant high-density power shelf architecture Rack-level power distribution with integrated superchip trays Rack-level power distribution with next-phase superchip trays
Chassis Power Draw High-density rack power budget; deployment-engineered ~10-15kW class per HGX node chassis (configuration dependent) ~10-16kW class per HGX node chassis (configuration dependent) ~14-20kW class per node envelope (OEM dependent) ~14-22kW class per node envelope (OEM dependent) Rack-scale power envelope engineered per AI factory topology Higher rack-scale envelope engineered per next-phase topology
Cooling Solution Liquid-first datacenter thermal strategy Air or liquid cooling by chassis vendor implementation Air or liquid cooling by chassis vendor implementation Predominantly liquid-ready deployment profile Liquid-first high-density deployment profile Liquid-cooled reference architecture Liquid-cooled next-phase reference architecture

Platform reference notice

This matrix is a standard platform-level reference for procurement planning. Final OEM-integrated builds can differ by memory totals, CPU selection, network and I/O configuration, local storage capacity, power architecture, and physical dimensions.

Strategic platform comparison

Relative platform positioning across current AI infrastructure options

An executive scorecard derived from standard platform characteristics for procurement alignment. This view supports relative positioning and planning context, not benchmark guarantees.

Platform scorecard

Swipe on mobile

Platform Training Throughput Positioning Inference Density Positioning Memory Footprint Orientation Fabric / Scale-Up Strength Deployment Complexity
NVIDIA Vera Rubin R200 Frontier Very High Balanced high-bandwidth orientation Frontier fabric profile OEM-dependent
NVIDIA HGX H100 Foundational Balanced General-purpose HBM orientation Established NVLink Mature path
NVIDIA HGX H200 High Advanced Memory-capacity-first orientation Proven scale-up fabric Moderate
NVIDIA HGX B200 Very High High Balanced high-capacity HBM orientation Strong NVLink profile High
NVIDIA HGX B300 Frontier Very High Inference-heavy HBM orientation Advanced backplane linkage High
NVIDIA GB200 Frontier Very High Large HBM plus CPU-coupled memory orientation Strong multi-node fabric Liquid-first
NVIDIA GB300 Frontier Frontier Extreme HBM footprint orientation Top-tier NVLink fabric Liquid-first + OEM-dependent

Note: Realized performance depends on final integrated node design, workload mix, operating clocks, interconnect topology, and cooling conditions.

Power-to-performance comparison

Specification-led power and cooling planning across key NVIDIA platform families

Use this platform-level matrix to guide facility readiness, rack-density planning, and efficiency posture across seven high-demand model families. Guidance is procurement-oriented and avoids fixed per-chassis watt claims for OEM-dependent builds.

  1. 01 · NVIDIA Vera Rubin R200

    Frontier-generation platform profile for top-end AI infrastructure planning.

    Frontier-class upper envelope with OEM-dependent final system draw.

    Liquid-first architecture; air-only operation is generally outside target design intent.

    Frontier deployment fit where facilities are purpose-built for extreme density.

  2. 02 · NVIDIA HGX H100

    Widely deployed enterprise baseline with mature procurement and operations paths.

    High but established envelope, with OEM-dependent configuration spread.

    Commonly air-compatible in select builds; liquid assist is typical at higher rack density.

    Balanced-to-high-density planning anchor for scaled enterprise AI estates.

  3. 03 · NVIDIA HGX H200

    Scale-forward HGX tier for memory-heavy and mixed workload expansion.

    High envelope above H100 class, with final chassis outcomes dependent on OEM build.

    Air-compatible options exist; liquid support is favored for denser multi-rack clusters.

    High-density planning fit for sustained mixed training and inference portfolios.

  4. 04 · NVIDIA HGX B200

    Next-wave HGX platform aimed at throughput gains in modern AI stacks.

    High-demand envelope class with tighter rack-level headroom requirements.

    Rack-density intensive in many deployments; liquid-first or hybrid cooling is commonly specified.

    High-density planning posture where throughput-per-rack is a primary objective.

  5. 05 · NVIDIA HGX B300

    Advanced HGX class for enterprises pushing newer training-scale requirements.

    Very high envelope tier with meaningful OEM implementation variance.

    Liquid-first with distribution-heavy thermal design in sustained operation profiles.

    Frontier-to-high-density planning fit for upgraded enterprise AI factories.

  6. 06 · NVIDIA GB200

    Integrated high-density platform class for large AI infrastructure programs.

    Ultra-high platform envelope across integrated OEM system designs.

    Liquid-first by default; rack power distribution and thermal zoning are primary constraints.

    Frontier deployment planning for tightly integrated, high-throughput environments.

  7. 07 · NVIDIA GB300

    Top-tier next-generation platform profile for frontier AI capacity expansion.

    Ultra-high next-generation envelope, strongly OEM-dependent at final chassis level.

    Liquid-first with aggressive rack-density and cooling-plant integration requirements.

    Frontier planning fit where facilities are engineered for peak density and sustained utilization.

Decision guide

Platform recommendations by buyer use case

Use this qualitative guide to shortlist the right platform family before final power, rack, and stock confirmation with our supply desk.

Buyer objective Primary fit Also evaluate Consultative rationale
Best for frontier model training GB300 GB200, B300 Strong fit when training programs prioritize top-end throughput and next-wave platform continuity.
Best for high-memory inference H200 GB200, GB300 Often selected where larger memory footprints and stable inference behavior are the lead requirement.
Best for dense datacenter deployment B200 B300, H100 Practical option for teams balancing rack efficiency with scaled production utilization targets.
Best for forward-looking platform planning GB300 GB200, B300 Appropriate when roadmap alignment and longer-horizon architecture decisions are the primary driver.
Best for broader OEM flexibility H100 H200, B200 Mature ecosystem choice for buyers who value wider deployment patterns across enterprise OEM channels.

Priority-to-platform quick map

Fastest path to deployment

H100 family

Maximum scale

GB300 / GB200 family

Balanced upgrade path

B300 with H200 fallback

Thermal / power-conscious planning

B200 or H200 by facility envelope

All recommendations are qualitative and should be validated against live inventory, power density, and colocation constraints.

Confirm current fit

Ready to align infrastructure with demand?

Talk with LXK CAPITAL about OEM options, technical fit, and current availability for enterprise-grade datacenter and GPU programs.

Contact us