NVIDIA DGX™ Systems

The NVIDIA GB300 NVL72 unifies 72 Blackwell Ultra GPUs and 36 Grace CPUs into a single coherent rack-scale engine, delivering 1,440 PFLOPS of FP4 compute and the highest HBM headroom available in NVIDIA's current shipping rack-scale platform.

Liquid cooling enables peak rack density and continuous high-utilisation training and serving for trillion-parameter generative AI and reasoning workloads. GB300 NVL72 is the canonical compute foundation for current-generation NVIDIA DGX SuperPOD deployments.

Best for: Enterprises shipping current-generation AI factories in production today, where Blackwell Ultra HBM capacity and liquid-cooled energy efficiency need to be live now, not on the Vera Rubin roadmap.

AMBER configures and deploys GB300 NVL72 racks as standalone systems or as the building blocks of a DGX SuperPOD for European enterprises.

Available now

Configure GB300 NVL72 with AMBER →

The NVIDIA Vera Rubin NVL72 unifies 72 Rubin GPUs and 36 Vera CPUs into a single coherent rack-scale engine, delivering 3,600 PFLOPS NVFP4 inference, 2,520 PFLOPS NVFP4 training, and 20.7 TB of HBM4 memory at 1,580 TB/s bandwidth.

Built for training and inference of agentic and reasoning AI models, with 260 TB/s NVLink scale-up bandwidth and NVIDIA ConnectX®-9 SuperNIC connectivity at 1.6 Tb/s per GPU.

Best for: Teams training agentic and reasoning models past the trillion-parameter mark, where 20.7 TB of HBM4 per rack and 260 TB/s NVLink determine whether the model fits and how fast it serves.

AMBER configures and deploys Vera Rubin NVL72 racks for enterprises across Germany and Europe, integrated into NVIDIA DGX SuperPOD™ deployments or as a standalone rack.

Register early access, available H2 2026

Configure Vera Rubin NVL72 with AMBER →

The NVIDIA DGX™ Vera Rubin NVL72 is the DGX-branded, factory-integrated enterprise deployment of the Vera Rubin rack: 72 Rubin GPUs, 36 Vera CPUs, and 18 NVIDIA BlueField-4 DPUs per rack, operated as a single coherent AI engine and orchestrated by NVIDIA Mission Control™.

At SuperPOD scale, 14 racks deliver 1,008 Rubin GPUs, 50.4 exaFLOPS of FP4 performance, and 1,046 TB of fast memory.

Best for: Enterprises buying their first SuperPOD: factory-integrated, Mission Control orchestrated, NVIDIA-supported end-to-end. The path that minimises integration risk on a multi-rack rollout.

AMBER configures and delivers the DGX Vera Rubin NVL72 as turnkey infrastructure for enterprises building NVIDIA AI Factories across Germany and Europe.

Register early access, available H2 2026

Configure DGX Vera Rubin NVL72 with AMBER →

With Blackwell GPUs (air-cooled), the NVIDIA DGX™ B200 offers 3× training and 15× inference performance versus its predecessor, ideal for large-scale AI workloads.

Best for: Air-cooled data centres that need Blackwell density today, without retrofitting cooling infrastructure for a liquid-cooled rack-scale NVL72 deployment.

AMBER configures DGX B200 systems for European enterprises rolling out Blackwell capacity in air-cooled data centres.

Available now

Configure DGX B200 with AMBER →

At the top of today's stack, the NVIDIA DGX™ B300 delivers 4× training and 11× inference performance over its predecessor, with 2.3 TB GPU memory and real-time AI at scale.

Best for: Inference workloads where 2.3 TB of GPU memory is the unlock: long-context reasoning, real-time agentic pipelines, multi-tenant LLM serving with high concurrency. The Blackwell-era successor to the inference role formerly held by Hopper-generation DGX systems.

AMBER configures DGX B300 systems for European enterprises building current-generation Blackwell Ultra capacity with maximum HBM headroom for inference and reasoning workloads.

Available now

Configure DGX B300 with AMBER →

The NVIDIA DGX™ Rubin NVL8 brings Rubin-generation performance into a compact 8-GPU liquid-cooled system with x86 CPUs, an efficient on-ramp for develop-to-deploy AI pipelines before full rack-scale deployment.

DGX Rubin NVL8 delivers 5.5× NVFP4 FLOPS versus NVIDIA Blackwell systems and scales out to 64 systems per NVIDIA DGX SuperPOD™ (512 Rubin GPUs), orchestrated by NVIDIA Mission Control™.

Best for: Departments validating Rubin-generation training pipelines before committing to a rack, with a tested upgrade path: 64 NVL8 systems compose into a 512-GPU SuperPOD when you scale.

AMBER configures DGX Rubin NVL8 for European enterprises that want Rubin performance at team or department scale before committing to a full rack-scale footprint.

Register early access, available H2 2026

Configure DGX Rubin NVL8 with AMBER →

The NVIDIA Groq 3 LPX is a co-designed inference accelerator rack within the NVIDIA Vera Rubin POD platform: 256 LPUs per rack, 128 GB of SRAM, optimised for low-latency inference on long-context models.

The NVIDIA Vera Rubin POD organises seven chips into five rack-scale systems; Groq 3 LPX is the inference-specialised rack, alongside the Vera Rubin NVL72 compute rack, the Vera CPU rack, the BlueField-4 STX rack, and the Spectrum-6 SPX networking rack.

Best for: Inference SLAs where tail latency at high concurrency dominates training-style throughput: long-context retrieval, agentic chains with tool calls, real-time multi-user assistants.

AMBER positions Groq 3 LPX for European enterprises serving inference workloads where latency, not training throughput, is the critical dimension.

Register early access, pending NVIDIA partner availability

Configure Groq 3 LPX with AMBER →

NVIDIA DGX™ BasePOD provides a validated reference architecture for scalable AI across industries. Build and deploy enterprise-critical AI applications on NVIDIA's performance-optimised platform.

Best for: Your first multi-node AI cluster, where NVIDIA's validated reference design (compute, networking, storage tested together) means the integration risk lives with NVIDIA, not your in-house team.

AMBER deploys DGX BasePOD installations across European data centres, integrating compute, networking, storage, and software per NVIDIA's reference design.

Available now

Configure DGX BasePOD with AMBER →

NVIDIA DGX™ SuperPOD is the turnkey AI data centre solution for organisations building AI factories. Available with a choice of NVIDIA Rubin and Blackwell-powered compute, including DGX Vera Rubin NVL72, DGX GB300, DGX B300, and DGX B200.

It's a full-stack platform that includes industry-leading computing, storage, networking, software, and infrastructure management, all optimised to work together.

Best for: Workloads that have outgrown a single rack and need 14+ rack scale-out: NVIDIA designs and ships, AMBER integrates and supports across European data centres.

AMBER deploys DGX SuperPOD installations as the European integrator partner, configuring the full stack and managing on-site delivery for enterprises across Germany and Europe.

Available now

Configure DGX SuperPOD with AMBER →

Featuring the NVIDIA GB10 Superchip, DGX™ Spark brings 1 petaFLOP of AI performance to your desktop. Run and fine-tune AI models up to 200 billion parameters locally with 128 GB unified memory, enterprise AI development in a compact form factor.

Best for: Solo AI engineers and researchers running iterative local fine-tuning on 100B to 200B parameter models, without queue time on a shared cluster or cloud spend that scales with every iteration.

AMBER configures DGX Spark systems for European researchers, AI engineers, and individual technical buyers who need enterprise-grade local AI development on the desk.

Available now

Configure DGX Spark with AMBER →

The NVIDIA DGX™ Station brings data-centre-class AI to the desk: a deskside supercomputer built around the GB300 Grace Blackwell Ultra Desktop Superchip, with 748 GB of coherent memory (252 GB HBM3e + 496 GB LPDDR5X) and up to 20 petaFLOPS of AI compute.

It supports models up to 1 trillion parameters locally.

Best for: Research groups training trillion-parameter models locally on the desk, on the same Grace Blackwell Ultra chip family they will scale into a SuperPOD: same software, same orchestration, zero porting cost.

AMBER configures DGX Station deployments for research groups, teams, and individual technical buyers who need training and fine-tuning performance on the desk before scaling to DGX SuperPOD infrastructure.

Coming in 2026

Configure DGX Station with AMBER →