NVIDIA GB300 NVL72 live on Verda
We are proud to announce that Verda is one of the first providers in Europe to offer NVIDIA GB300 NVL72.
GB300 represents a new class of AI supercomputing. And we are not just experimenting with it, we are virtualizing it and making it accessible to frontier AI teams.
GB300 NVL72 in brief
GB300 NVL72 is NVIDIA’s latest Blackwell Ultra-based rack-scale AI platform. Each rack integrates 72 GPUs and 36 Grace CPUs, pairing every two GPUs with one CPU in a tightly coupled design optimized for balanced compute and memory performance. Learn more from our technical analysis.
This architecture is purpose-built for:
- Trillion-parameter model training
- Large-scale reasoning workloads
- High-throughput inference
- Agentic and multimodal AI systems
GB300 NVL72 dramatically increases memory capacity, interconnect bandwidth, and compute density, unlocking performance levels that were previously impractical outside hyperscale environments.
Why GB300 NVL72 changes the AI landscape
NVIDIA GB300 NVL72 enables:
- Massive unified GPU memory across nodes
- Higher sustained throughput for reasoning models
- Reduced communication bottlenecks
- Improved performance per rack
- Better economics for large-scale inference
For teams building foundation models or reasoning engines, this is not incremental improvement. It is an architectural change.
Battle tested on frontier AI workloads
vLLM
GB300 is already proving itself on frontier AI workloads.
The vLLM team recently demonstrated major performance gains running DeepSeek-V3.2 and DeepSeek-R1 on Verda's GB300 NVL72, reporting substantial improvements in both prefill and mixed-context inference throughput.
Their results highlight how GB300 unlocks significantly higher efficiency for large-scale reasoning models in production settings.
Read more on the vLLM blog.
SGLang
SGLang leveraged access to GB300 NVL72 with infrastructure support from Verda for several experimental releases and performance validations, testing advanced serving capabilities on high-density GPU clusters.
"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, RL Core Developer at SGLang
Virtualization
Powering up a GB300 NVL72 system is just the beginning.
The real challenge is virtualizing and operationalizing rack-scale infrastructure so that our customers can utilize it efficiently.
Our in-house virtualization layer is currently running in experimental mode on GB300, delivering near 1:1 performance compared to bare metal. This means teams can access flexible, isolated GPU environments without sacrificing the performance advantages of Blackwell Ultra infrastructure.
Early results show that GB300 NVL72 can be efficiently partitioned and allocated while maintaining the throughput and scaling characteristics required for frontier AI workloads.
What comes next
Our GB300 NVL72 virtualization has been successfully tested with near bare-metal performance.
We are now continuing system validation, workload benchmarking, and infrastructure scaling as we prepare for broader availability.
If you are building frontier AI systems and need access to GB300 NVL72, we invite you to contact us.