1x–18x trays
Each tray contains 4× GB300 GPUs and 2× Grace CPUs.
New
Flexible access to NVIDIA® GB300 NVL72
From a single tray to multiple racks
- Turnkey infrastructure
- Expert support
- Powered by NVIDIA®
AI Cloud
From rapid prototyping to foundation training to scalable inference
Compute
GB300 NVL72
New1x tray to 2+ racks · NVLink v5
From a single tray to multiple racks
Validate your workloads and scale to production – pick the footprint that fits.
Each tray contains 4× GB300 GPUs and 2× Grace CPUs.
Each rack contains 18 trays with NVLink v5 peer-to-peer connect.
Multiple racks, each containing 72× GPUs, connected via InfiniBand interconnect.
Each tray contains:
The GB300 Grace-Blackwell superchip features two Blackwell Ultra B300 GPUs and one Grace CPU, an enhancement from the previous generation's configuration of a single Hopper GPU and one Grace CPU. These components are interconnected via NVLink v5, facilitating a unified memory domain. Each GPU is equipped with 288GB of HBM3e memory.
The NVL72 rack integrates 18 trays, each containing two GB300 Grace-Blackwell superchips, totaling four B300 GPUs and two Grace CPUs. Unlike an InfiniBand cluster, the GB300 utilizes NVLink v5 for GPU interconnectivity. This provides significantly higher throughput and lower latency for GPU-to-GPU communication, optimizing performance for demanding computational tasks.
GB300 delivers the cutting-edge performance
| Metric | Per NVL72 rack |
|---|---|
| FP4 | 1080 PFLOPS (1.08 EFLOPS) |
| FP8 | 360 PFLOPS |
| FP16 / BF16 | 180 PFLOPS |
| FP32 | 3 PFLOPS |
| GPU memory | 20 TB |
| GPU memory bandwidth | 576 TB/s |
| NVLink bandwidth | 130 TB/s |
| Max GPU TDP | 100.8 kW |
Long-term discounts available. Talk to our engineers for multi-tray, single-rack, and multi-rack quotes.
Get the architecture overview and performance implications from our technical blog that covers:
“Verda's GB300, combined with their infrastructure support, has provided me with an extremely stable experience. We used their cluster while developing DeepSeek v32 RL, and it was consistently reliable. It allowed me to fully focus on development without having to worry about machine setup or infrastructure issues. The experience was truly exceptional.”
Yueming Yuan
“vLLM on GB300 provided by Verda is amazing! Trillion-param-level open-source monsters like Kimi K2.5 and GLM-5 are within reach easily. Models that used to need a whole rack now scream on one node. The future isn't coming, it's already shipping tokens.”
Kaichao You
Verda is among the first NVIDIA GB300 NVL72 providers, offering deep expertise with the Blackwell Ultra architecture. Building upon early adoption of HGX™ B300, we deployed one of the very first GB300 NVL72 systems in the European Union.
We host and operate GB300 NVL72 systems at our data center locations in the Nordics, powered by 100% renewable energy sources. Our in-house engineering teams handle the full lifecycle, including:
Verda's first GB300 NVL72 system was battle-tested with production workloads from open-source projects, such as SGLang and vLLM, and select customers. All early users noted stability and support.
With Verda, you can gain reliable and flexible access to GB300 NVL72 systems with expert support and sensible SLAs.
In-house AI Lab, contributing to frontier research and open-source projects
Verda collaborates with 1X on building multi-GPU inference for the 1XWM generative video model.
As video quality leads to task success, it enables solving more complex tasks in household autonomy.
Verda continuously sponsors SGLang with access to compute resources and infrastructure support.
SGLang's explorations into RL training with FP8 and INT4 utilize NVIDIA Hopper and Blackwell architectures.
vLLM got early access to a GB300 cluster from Verda along with infrastructure support.
With Blackwell Ultra, the team achieved an 8x prefill and 10-20x mixed-context improvement over Hopper.
As an NVIDIA Preferred Partner, Verda delivers early access to the latest hardware while adhering to the highest security & compliance standards and offering over 99.95% uptime.
AICPA
SOC 2C5
ISO
27001, 27017, 27018, and 27701
Built in Europe, trusted globally
Validate and scale your AI workloads on the latest architecture — without commitments.