Today we launched Arm Neoverse CSS N4, our most configurable CSS yet and the fastest path from CSS to silicon. It gives customers another way to build for agentic AI, alongside Arm AGI CPU. Customers can use CSS N4 to create differentiated silicon optimized for throughput, or deploy Arm AGI CPU for highly responsive AI workloads. That flexibility matters because different workloads need different infrastructure. Arm gives customers the freedom to choose the approach that works best for them, all on a common Neoverse platform. https://lnkd.in/gG_3ZMGK

Really excited to see Arm bring the CPU back as an integral part of the AI rack. We all know that running workloads, such as AI inference, agent sandboxing, and cost sensitive production level AI applications, on the CPU level is an incredible lever for cost. What’s really exciting here is that custom silicon builders can use the Arm Neoverse CSS N4 as the compute block in the DPU to offload security and networking, accelerate AI KV-caching, and isolate agentic AI workloads from the host.  Big congrats to the team! Chaitanya Kadiyala Bhumik Patel Rene Haas Yan Fisher Pranay Bakre Ami Badani

Jev is taking over the internet and everyone is asking us: How can we get it in a compliant way? So we are now the first to serve a Jev-like model in the EU and ZDR. The best part: Its FREE for 7 days! Jev has gone trough the roof in the last days as its 40–200x faster and 10–200x cheaper than LLMs. The use cases for legal and document work are endless. So to get it into the hands of our customers in the EU we have been working the last 48 hours non stop and are now offering a Jev-like model with Zero Data Retention and hosting in EU running on our GPUs. Our tests say its actually faster compared to the Jev api and meets the same accuracy. Our model is based on Google’s DiffusionGemma and OpenJev, which uses constrained decoding to turn it into a fast, structured decision model. Its a drop in replacement for the typesafe jev sdks and apis so you only need to replace the base url api key. Get your API key here and use it for free: https://lnkd.in/d8-GgyJj

Polar Bear - GPUs We are designing our own GPU architecture and silicon, spanning Edge, Desktop, and Data Center-class GPUs — built for the next generation of AI and high-performance computing. From energy-efficient edge inference to desktop acceleration and massive data-center workloads, we are building a complete GPU compute stack from the ground up. Our target: fabrication-ready designs within the next six months. And toolchains and compilers in another 4 months. A big step toward building indigenous, world-class AI compute infrastructure — from software to silicon.

SUSE & AMD share joint effort to implement AI portability. Earlier today at IFA Berlin, our CEO Dirk‑Peter van Leeuwen joined AMD’s Jack Huynh to highlight our shared commitment to open source innovation and architectural freedom. Together, we’re enabling true AI workload portability — letting developers build locally on AMD systems and seamlessly move those workloads across workstations, data centres, and edge environments, powered by the Ryzen AI Halo platform. This aligns directly with our vision for open, portable Private Enterprise AI and marks the start of active co‑engineering and PoC work. https://lnkd.in/eMjvSkMt

NVIDIA cluster readiness engine (NVCRE) is now live https://lnkd.in/gc2Sv88j on GitHub! Do take it for a spin! NVCRE is a Kubernetes controller that certifies GPU clusters before production workloads run on them. It runs real training and communication workloads across topology-aware node groups, measures performance, detects hardware failures, and reports every bad node with a reason.

Your GPU bill may not need more hardware. It may need better kernel optimization. We help AI teams identify high-impact CUDA bottlenecks, optimize critical kernels, and improve inference throughput on the same GPU infrastructure. If you're seeing high GPU spend, low throughput, or latency issues, we can start with a GPU Performance Assessment. Let’s find the kernels worth optimizing. #CUDA #GPUOptimization #AIInfrastructure #LLMInference #GPUPerformance #InferenceOptimization #NVIDIA

Many neocloud providers use OpenStack to orchestrate their AI infrastructure. Some use Ironic primarily for bare-metal provisioning and automation,while others use OpenStack to offer both bare-metal and virtual-machine-based infrastructure. With Kubernetes as the de facto orchestrator for AI inference, these environments must deliver efficient GPU utilization while maintaining isolation between customer workloads. This guide demonstrates LLM inference on Intel® Arc™ Pro B70 GPUs with OpenStack, Kubernetes, vLLM, and llm-d. The B70 GPU also serves as a Software Development Vehicle for the much-anticipated Intel Crescent Island Inference (CRI) GPU, supporting early software development and validation. https://lnkd.in/eQsKGYEn #Neocloud #Kubernetes #vLLM #llmd #IntelArcPro #CrescentIsland #AIInference #GPU

The compute market has a geography problem. The people who need GPUs are concentrated in about twenty blocks of San Francisco. The people who supply them are not. They are in Hamburg, in Johannesburg, wherever there is power and permits and a grid connection that did not take four years. I met a founder at a meetup this weekend who trains open source models for his startup. He's not building out a frontier lab but needs a few dozen GPUs, for a few weeks at a time, a few times a year. He told me that requesting capacity from the hyperscalers is basically not worth attempting any more. You file the request. You wait days. Often nothing comes back at all. So he went looking on the other side of the world instead. You don’t close that gap over email. Supply and demand in this market still move at the speed of trust, and trust is built on the ground. That is why we are here in SF for months instead of just flying in for a conference. What is the furthest you have gone to see capacity with your own eyes before signing for it? If you would rather answer that in person, we are putting supply and demand in the same room this Friday in SF for our 2nd Compute Connect meetup. Ask me and I will share the sign up.

Every return visit to an LLM conversation forces the model to recompute context from scratch, burning GPU cycles and slowing response times. Cloudian's new benchmark, run with NVIDIA Dynamo, shows there's a faster path. By offloading KV Cache to Cloudian HyperStore, an S3-native object storage platform with RDMA acceleration, we measured: 📊 Up to 20x lower Time to First Token vs. full recompute at 120K tokens ⚡️ A 31.77% mean latency advantage for RDMA vs. TCP across all context lengths tested 🚀 Flat, scalable retrieval latency — even as context windows grow to 128K+ tokens The takeaway: the storage tier is becoming a first-class part of inference architecture. For multi-turn conversations, RAG pipelines, and long-context workloads, how fast you can recall model state is just as important as how fast you can compute it. Read more for the full breakdown on methodology, results by context length, and what it means for GPU utilization at scale: https://bit.ly/452X4d0