Rows of liquid-cooled AI GPU server racks in a dark high-density data hall

Product Matrix

Systems for the next compute frontier.

From sovereign AI deployments to pre-integrated campus-scale clusters, every configuration is assembled and validated in British Columbia.

Every configuration is assembled and validated in British Columbia before shipment.

Important export-control notice

IMPORTANT EXPORT-CONTROL NOTICE: High-end NVIDIA Blackwell-series hardware (DGX B300 / GB300 NVL72 / DGX B200) is classified ECCN 3A090.a and regulated under U.S. BIS-EAR rules. Manufactured by XAgent AI Inc. in British Columbia, Canada. Quotation is available only after thorough client and beneficial-owner due-diligence review. No quotation for controlled entities in restricted jurisdictions.

Catalogue

Search, filter, then compare.

Search any parameter keyword, or narrow by system class, FP4 performance and power envelope.

6 of 6 systems

Performance tier: DGX B300 ≈ B200 > H200 > H100

42U liquid-cooled rack: 72 B300 GPUs, 36 Grace CPUs, 130 TB/s NVLink. ECCN 3A090.a controlled.

XA-DGX-GB300-NVL72

Flagship

GB300 NVL72

Rack-scale 72-GPU Grace-Blackwell Ultra system with factory-integrated liquid-cooling loops and full serialized traceability.

Full-rack liquid cooling with integrated CDU

ECCN 3A090.a · controlled

Compute composition: 72 × Blackwell-Ultra B300 GPU; 36 × Grace ARM CPU (2592 Neoverse-V2 cores)
Total GPU HBM3e memory: 20 TB
Total system fast memory: 37 (GPU HBM + CPU LPDDR5X) TB

Full datasheet

10U cold-plate liquid-cooled node: 8 B300 SXM GPUs, 2.3 TB HBM3e. ECCN 3A090.a controlled.

XA-DGX-B300

Flagship

DGX B300

8-GPU Blackwell-Ultra SXM supercomputer for large-model pre-training, long-context inference and Agent AI workloads.

Native cold-plate liquid cooling

ECCN 3A090.a · controlled

GPU: 8 × Blackwell-Ultra (B300) SXM GPU
Total GPU HBM3e memory: 2.3 TB
FP4 Tensor Core performance: 144 (sparse) / 108 (dense) PFLOPS

Full datasheet

Air-cooled Blackwell node for mixed training and inference. ECCN 3A090.a controlled.

XA-DGX-B200

DGX Systems

DGX B200

8 × Blackwell-SXM B200 GPU, 2.1 TB HBM3e, general-purpose large-model training and inference.

Cold-plate liquid cooling

ECCN 3A090.a · controlled

GPU: 8 × Blackwell-SXM B200 GPU
Total GPU HBM3e memory: 2.1 TB
Cooling: Cold-plate liquid cooling

Full datasheet

Hopper-generation eight-GPU node for large-model training and inference.

XA-DGX8-H200

DGX Systems

DGX H200

8 × Hopper-SXM H200 GPU with high-capacity HBM3e, optimized for long-context RAG and LLM inference.

Liquid cooling optional

GPU: 8 × Hopper-SXM H200 GPU
Memory: HBM3e high-capacity memory
Cooling: Liquid cooling optional

Full datasheet

Proven Hopper platform for established enterprise AI workloads.

XA-DGX8-H100

DGX Systems

DGX H100

8 × Hopper-SXM H100 GPU, mature mass-production platform widely validated for training and inference.

Enterprise air / liquid

GPU: 8 × Hopper-SXM H100 GPU
Platform maturity: Mature mass-production platform
Cooling: Enterprise air or liquid cooling

Full datasheet

Pre-integrated cluster with Quantum-X800 InfiniBand and DGX Mission Control software.

XA-DGX-SUPERPOD

Integrated Clusters

DGX SuperPOD

Turnkey pre-integrated AI cluster interconnecting DGX B300 and GB300 NVL72 nodes over Quantum-X800 InfiniBand.

Factory pre-integrated

Supported building blocks: DGX B300 node / GB300 NVL72 full rack
Inter-node fabric: NVIDIA Quantum-X800 InfiniBand 800G; NVLink switching; SHARP in-network computing
Software stack: DGX OS, NVIDIA AI Enterprise, Mission Control cluster-orchestration suite

Full datasheet

Comparison

Compare systems side by side.

Select up to three systems to compare every parameter and unit in one table.

Choose two or three systems above to build a comparison table.

Specifications

Full datasheets.

Each block mirrors the corresponding specification sheet, with units listed separately.

Unit legend

PFLOPS: PetaFLOPS — 10¹⁵ floating-point operations per second
TB: Terabyte — memory or storage capacity
TB/s: Terabytes per second — memory or interconnect bandwidth
kW: Kilowatt — maximum electrical power draw per system or rack
U: Rack unit — 1U equals 44.45 mm of rack height
Gb/s: Gigabits per second — network port throughput

Key parameters

FP4 (sparse / dense): Blackwell-generation 4-bit floating-point throughput. Sparse figures assume 2:4 structured sparsity; dense figures are always achievable. FP4 is not directly comparable to FP8 or FP16 numbers.
HBM3e: High-bandwidth GPU memory. Capacity determines the largest model that fits without model parallelism; bandwidth drives token-generation speed.
NVLink domain: The number of GPUs connected at full NVLink bandwidth inside one system or rack. Beyond the domain, traffic crosses slower scale-out networking.
Max power: Configured maximum draw including cooling infrastructure (CDU, pumps). Facility power and liquid-cooling loops must be sized to this figure.

Data basis

GPU, memory and interconnect figures are nominal values from the corresponding NVIDIA platform datasheets.
Power, cooling-loop and rack-integration figures are measured at the XAgent AI facility XA-BC-01 in British Columbia, Canada, under standard factory acceptance conditions.
Where a range is shown (e.g. 132–140 kW), the exact value depends on the selected configuration; it is fixed in the factory acceptance report per serial number.
Performance figures marked “sparse” assume 2:4 structured sparsity; all other figures are dense and directly comparable across systems.

XA-DGX-GB300-NVL72

GB300 NVL72

ECCN 3A090.a · controlled

GB300 NVL72 is a full-rack liquid-cooled Grace-Blackwell Ultra integrated supercomputer. 72 B300 GPUs are tightly coupled by a rack-wide NVLink fabric, targeting trillion-parameter foundation-model pre-training and large-scale multi-modal AI factory workloads. ECCN 3A090.a controlled item, subject to strict cross-border export review.

GB300 NVL72 key specifications
Item	Specification	Unit
Compute composition	72 × Blackwell-Ultra B300 GPU; 36 × Grace ARM CPU (2592 Neoverse-V2 cores)	—
Total GPU HBM3e memory	20	TB
Total system fast memory	37 (GPU HBM + CPU LPDDR5X)	TB
Rack-wide NVLink bandwidth	130 aggregate	TB/s
FP4 Tensor Core	1440 (sparse) / 1080 (dense)	PFLOPS
FP8 Tensor Core	720	PFLOPS
Total rack operating power	132–140, full liquid-cooling CDU integrated	kW
Network	ConnectX-8 800G InfiniBand / Ethernet; BlueField-3 DPU	—
Rack dimension	Standard 42U rack with integrated rack CDU liquid-cooling system	U

Manufacturing site

XAgent AI Inc., British Columbia, Canada

Target use cases

Trillion-parameter foundation-model pre-training, AI factory, hyperscale multi-modal training, subsea-compute large-scale deployment.

Compliance notes

ECCN 3A090.a highest-class controlled hardware. Mandatory full-depth beneficial-owner due diligence; serial-number traceability archive retained five years or more.

Back to catalogue

XA-DGX-B300

DGX B300

ECCN 3A090.a · controlled

The DGX B300 is an 8-GPU Blackwell-Ultra SXM AI supercomputer, purpose-built for large-model pre-training, long-context LLM inference and Agent-oriented AI workloads, with a native cold-plate liquid-cooling design. Classified ECCN 3A090.a under U.S. BIS-EAR export-control regulations; all sales require beneficial-owner due diligence and a signed EUC statement.

DGX B300 key specifications
Item	Specification	Unit
GPU	8 × Blackwell-Ultra (B300) SXM GPU	—
Total GPU HBM3e memory	2.3	TB
FP4 Tensor Core performance	144 (sparse) / 108 (dense)	PFLOPS
FP8 Tensor Core performance	72	PFLOPS
NVLink aggregate bandwidth	14.4	TB/s
CPU	Dual Intel Xeon 6776P	—
System DRAM	2 (configurable up to 4)	TB
Network	8 × ConnectX-8 800G OSFP; 2 × BlueField-3 DPU 400G	—
Storage	OS: 2 × 1.9 TB NVMe M.2; Data: 8 × 3.84 TB NVMe E1.S	—
Power draw	14.5–15.1	kW
Form factor	10U rack-mount, cold-plate liquid cooling	U

Manufacturing site

XAgent AI Inc., British Columbia, Canada

Target use cases

Large-model training, Agent inference, enterprise-grade AI compute clusters, horizontal scale-out into DGX SuperPOD.

Compliance notes

ECCN 3A090.a controlled hardware. Beneficial-owner screening, entity-list check and a signed EUC statement are mandatory prior to quotation and order. GPU serial-number full-lifecycle traceability with archive retention of five years or more.

Back to catalogue

XA-DGX-B200

DGX B200

ECCN 3A090.a · controlled

General-purpose Blackwell training and inference platform for enterprise AI programs. ECCN 3A090.a controlled hardware.

DGX B200 key specifications
Item	Specification	Unit
GPU	8 × Blackwell-SXM B200 GPU	—
Total GPU HBM3e memory	2.1	TB
Cooling	Cold-plate liquid cooling	—
Workloads	General-purpose large-model training & inference	—
Export classification	ECCN 3A090.a controlled	—

Manufacturing site

XAgent AI Inc., British Columbia, Canada

Target use cases

Enterprise model training, mixed training and inference clusters, AI platform modernization.

Compliance notes

Back to catalogue

XA-DGX8-H200

DGX H200

High-memory Hopper platform optimized for long-context retrieval-augmented generation and LLM inference. Subject to EAR export control.

DGX H200 key specifications
Item	Specification	Unit
GPU	8 × Hopper-SXM H200 GPU	—
Memory	HBM3e high-capacity memory	—
Cooling	Liquid cooling optional	—
Workloads	Long-context RAG and LLM inference	—
Export classification	Subject to EAR export control	—

Manufacturing site

XAgent AI Inc., British Columbia, Canada

Target use cases

Long-context inference services, retrieval-augmented generation, high-memory fine-tuning.

Compliance notes

Subject to U.S. BIS-EAR and Global Affairs Canada export-control review. Jurisdiction, entity and end-use screening is completed before quotation; serial-number records are retained five years or more.

Back to catalogue

XA-DGX8-H100

DGX H100

Mature, widely deployed Hopper platform with a broad software ecosystem. Subject to EAR export control.

DGX H100 key specifications
Item	Specification	Unit
GPU	8 × Hopper-SXM H100 GPU	—
Platform maturity	Mature mass-production platform	—
Cooling	Enterprise air or liquid cooling	—
Workloads	Widely validated training / inference workloads	—
Export classification	Subject to EAR export control	—

Manufacturing site

XAgent AI Inc., British Columbia, Canada

Target use cases

Production inference fleets, research clusters, cost-optimized enterprise AI platforms.

Compliance notes

Back to catalogue

XA-DGX-SUPERPOD

DGX SuperPOD

DGX SuperPOD is a turnkey pre-integrated large-scale AI-cluster solution. Multiple DGX B300 / GB300 NVL72 nodes are interconnected via Quantum-X800 InfiniBand and NVLink switching fabric, delivered with a pre-validated networking, storage and orchestration software stack for hyperscale AI compute parks, containerized IDC and subsea-compute projects.

DGX SuperPOD key specifications
Item	Specification	Unit
Supported building blocks	DGX B300 node / GB300 NVL72 full rack	—
Inter-node fabric	NVIDIA Quantum-X800 InfiniBand 800G; NVLink switching; SHARP in-network computing	—
Software stack	DGX OS, NVIDIA AI Enterprise, Mission Control cluster-orchestration suite	—
Deployment scope	From small multi-node clusters to multi-rack hyperscale AI factories	—
Delivery scope	Hardware assembly, cable interconnection, factory burn-in test, cluster pre-validation, optional on-site commissioning	—

Manufacturing site

XAgent AI Inc., British Columbia, Canada

Target use cases

Large-scale AI compute-park turnkey projects, subsea compute hardware deployment, hyperscale foundation-model training clusters, container-modular IDC.

Compliance notes

All embedded GPU hardware follows BIS-EAR and Global Affairs Canada export-control requirements. Each GPU serial-number full-lifecycle record is kept five years or more; beneficial-owner screening is mandatory for the whole cluster project.

Back to catalogue

Start a qualified conversation

Build your next AI infrastructure program with confidence.

Tell us your registered entity location, deployment region and system requirements. Our team will begin with a confidential eligibility review.