European NVIDIA clouds scale up: Nscale fits 37% more GPUs per watt, OpenShell draws agent security lines
NVIDIA Blog · rss · 2026-10-21
NVIDIA's GTC Berlin update on European AI clouds: Nscale showed NVIDIA DSX MaxLPS delivering 37% more GPUs and 50% higher inference throughput within the same power budget, with Vera Rubin deployments planned. Nebius adopted Groq 3 LPX for low-latency token generation and launched a Physical AI Workbench; Baseten fine-tuned Nemotron 3.5 Lightning for CodeRabbit, lifting routing agreement from 75.8% to 80.7% while halving inference cost. Mistral Large 4 was trained on 3,800 Grace Blackwell GPUs. NVIDIA also announced the OpenShell Reference Architecture, using Vera CPUs and BlueField-4 DPUs to enforce access boundaries on AI agents, backed by a dozen cloud partners.
More from Infra
- Tenstorrent pitches 'one architecture at every size' from single card to supercluster — AIFlow_ML · 2026-10-21
- Mistral Partners with NVIDIA to Build AI-Native Engineering Workflows — Mistral AI · 2026-10-21
- Firmus, a neocloud set to list at $44B, started as a Bitcoin miner — victor_explore · 2026-10-21
- TokenGate: learnable token pruning cuts LLM prefill TTFT by up to 11.3x — pmttyji · 2026-10-21
- How OpenAI ingests petabytes of logs every day: a look at AI observability — blaizedsouza · 2026-10-21
- Lambda Featured in NVIDIA GTC Berlin Keynote on DSX MaxLPS and Confidential Bare Metal — TheZachMueller · 2026-10-21