EPRI uses Dell Pro Max with NVIDIA Grace Blackwell for AI Workflows

Transforming AI development—without expanding your datacenter. From research labs to regulated industries, teams are asking the same question: How do we prototype and train powerful models securely, without waiting on cloud queues or risking data exposure? Meet the HP Z Workstations & Solutions - HP ZGX Nano AI Station - a palm-sized system that brings 200B-parameter AI capability directly to your desk. With NVIDIA Grace Blackwell and 128GB unified memory, it enables real-time experimentation across LLMs, VLMs, computer vision, and recommender pipelines. The result? Faster iteration. Lower cost. Full data control. A hybrid model that works—local where it matters, cloud where it scales. If secure, on-prem AI is on your roadmap, this is the most practical way to start delivering value now. Discover More - https://lnkd.in/gVKXnN-3 Michael Messier Cécile Tezenas du Montcel Rajesh Dhingra Amit Paranjape Subraneel Bose Akash Deep Batra Samarth Wadhwa Siddharth Vishwakarma Saurabh Pandey Varun Bohra Bhavana Hemachandran Nair Krishnan unni T Kaushik Shaw Manuj Madaan Murtaza Ahmed Rahul Mehra #AI #GenAI #EdgeAI #HybridCompute #Workstations #ZGXNano #NVIDIA #GraceBlackwell #GPU #supercomputer #datascience #HP

The AI race is not being won on models or GPUs. It is being won, or lost, in the seams between power, cooling, and infrastructure, long before a single workload is run. We are now designing for megawatts per rack, gigawatt-scale campuses, and workloads that spike from zero to 140% in seconds. Yet much of the industry is still trying to assemble these environments as if they were traditional data centres, stitching together systems on site and hoping they behave as one. They do not. The real constraint is no longer compute. It is integration, orchestration, and the ability to design the entire system before it is ever built. In this latest Illuminaire piece, we explore why the AI factory must be engineered as a single organism, not constructed as a collection of parts, and why those who can industrialise infrastructure will define the next phase of AI. Featuring insights from @Vertiv and its work with @NVIDIA on next-generation AI infrastructure. 👉 https://lnkd.in/ePKkhtPY Follow Illuminaire.io for analysis on how AI is actually being deployed across infrastructure, enterprise, and industry. #AI #DataCentres #AIInfrastructure #DigitalInfrastructure #NVIDIA #Vertiv #Cloud #Hyperscale #AIFactories #GenerativeAI #EdgeAI Martin Olsen Fiona Alder Stefano Mozzato Steven Foley

SK hynix Inc. is pushing the boundaries of AI memory infrastructure with the mass production of its next-gen 192GB SOCAMM2 module. Optimized for Nvidia’s upcoming Vera Rubin AI platform, this innovation is designed to meet the rapidly growing demands of large-scale AI workloads and cloud deployments. Key highlights: • 2× bandwidth vs conventional RDIMMs • ~75% lower power consumption → significant data center cost savings • Built on LPDDR5X (1c process) for high performance + efficiency • Positioned as a critical middle layer between HBM and RDIMM in AI memory architecture • Enables better LLM training & inference efficiency What stands out is the architectural shift — SOCAMM2 isn’t just another memory module, it’s redefining how memory hierarchy is optimized for AI systems: balancing performance, scalability, and energy efficiency. As AI infrastructure scales, innovations like this will directly impact TCO, thermal management, and compute efficiency across hyperscale data centers. #AI #Semiconductors #Memory #DataCenters #SKHynix #Nvidia #LLM #TechInnovation

AI infrastructure is not being limited by compute—it is being constrained by heat, and that constraint is accelerating faster than most of the industry anticipated. In a conversation between Sanjana Mandavia, CEO at Compute Forecast and Satya Bhavaraju, CEO at Refroid Technologies Private Limited, a clear picture emerges of how thermal management is becoming the defining challenge of the AI era. As GPU power densities surge from hundreds to potentially thousands of watts per processor, traditional cooling paradigms are being pushed beyond their limits. The industry is moving rapidly toward fully liquid-cooled environments, but even that transition is struggling to keep pace with the velocity of compute innovation. The deeper issue lies in how cooling is approached. Today’s systems are largely reactive—designed to remove heat rather than understand the workloads generating it. But AI workloads are no longer predictable. Training, inference, and agentic systems introduce dynamic thermal patterns that require infrastructure to adapt in real time. This is where Refroid’s innovations begin to redefine the paradigm. With the ThermIon Hybrid Load Bank, the company is enabling full-system validation by simulating both air and liquid thermal loads simultaneously—bringing commissioning closer to real-world conditions rather than isolated testing. At the same time, the SentraFLO CDU introduces a new class of AI-validated cooling infrastructure, where performance is not just measured by heat rejection, but by how intelligently it responds to workload variability. Together, these systems shift cooling from a passive function to an active, data-driven layer of infrastructure. The conversation also highlights a broader industry gap. Commissioning methodologies have not evolved at the same pace as infrastructure complexity. Testing systems in silos creates risks that only surface post-deployment—at a time when the cost of failure is significantly higher. At the same time, a new paradigm is emerging. Cooling is beginning to integrate with power systems, digital twins, and workload orchestration, moving toward a unified, self-optimizing infrastructure model. Even marginal efficiency gains can translate into significant economic and sustainability outcomes. What becomes clear is a fundamental shift in perspective. Cooling is no longer a supporting layer—it is becoming the system that determines whether AI infrastructure can scale reliably and efficiently. And as highlighted by Sanjana and Satya, the future of data centers will depend not just on how much compute can be deployed, but on how intelligently the heat behind that compute is understood, tested, and managed- https://lnkd.in/djS6qpPy #LiquidCooling #ThermalManagement #GPU #Datacenters #AIInfrastructure #SantraFlo #CDU #HybridLoadBank #RackDensity #IndustryIntelligence

Across North America, artificial intelligence is driving unprecedented growth in data-center construction as companies race to build computing capacity for machine-learning workloads. Cloud providers and tech firms are expanding facilities capable of supporting massive AI training clusters powered by advanced GPUs. Read more: https://lnkd.in/dm3kZRk7 #AIInfrastructure #DataCenters #ArtificialIntelligence #MachineLearning #CloudComputing #GenerativeAI #GPUs #TechExpansion #DigitalInfrastructure #SustainableTech #EnergyEfficiency #LowCarbonEnergy #TechInnovation #AIWorkloads #FutureTech

We've found a new way to cool silicon chips more directly using microfluidics, which brings liquid coolant directly inside the silicon, where the heat is. It’s just one example of how we're investing and innovating in every layer of our cloud and AI infrastructure systems to meet future demand. More information available at: https://msft.it/6043vDxFl #Sustainability #Datacenters

The real competitive edge in AI infrastructure? The power contract behind the GPU rack. Compute still gets the headlines. Hyperscalers are racing to deploy the latest hardware, and GPU supply dominates the conversation. But underneath all of that, a more fundamental constraint is taking shape. Grid access, interconnection timelines, energy pricing. In some cases, projects are delayed by years before a single server goes online. What's emerging as a result is a new model: compute paired directly with dedicated power, essentially "energy islands" that bypass the grid bottlenecks slowing everyone else down. This is where operators with deep infrastructure experience actually have an edge. The ability to source power, manage load, and optimize energy costs is becoming as valuable as access to hardware itself. AI infrastructure won't be won on compute alone. It'll be won at the intersection of compute and energy.

𝐋𝐚𝐭𝐞𝐧𝐜𝐲 𝐈𝐬 𝐓𝐡𝐞 𝐍𝐞𝐰 𝐂𝐮𝐫𝐫𝐞𝐧𝐜𝐲 Bandwidth used to win, now it’s everywhere: Cheap. Abundant. Invisible. Latency is what decides everything. And that changes the game. Look at where the pressure is coming from: 👉 #NVIDIA GPUs moved from ~800 GB/s to 1.6 TB/s memory bandwidth 👉 #AWS and #Google are pushing AI inference closer to the edge 👉 #Tesla relies on real-time decision loops for autonomy 👉 Industrial AI, XR, robotics… all break when latency spikes This is the shift most are missing: AI is not only compute-hungry. It is latency-intolerant. And here’s the twist: For years, networks were the bottleneck for applications. Now AI is becoming the bottleneck for networks. 👉 Models need faster data access than networks can deliver 👉 Inference needs proximity, not distance 👉 Centralized architectures introduce unacceptable delays Which means: the architecture itself has to change. 👉 From centralized cloud → to distributed inference 👉 From core-heavy → to edge-native 👉 From bandwidth optimization → to latency engineering This is where telcos come back into the game. Not as connectivity providers. But as latency platforms. Because not all milliseconds are equal. Who can guarantee them, will be able to price them. And in an AI-driven world, the fastest network won’t win. The network that responds first will. #AI #Telco #EdgeComputing #NVIDIA #AWS #Google #DataCenters

𝗔𝗽𝗿𝗶𝗹 𝟮: 𝗚𝗼𝗼𝗴𝗹𝗲 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝘀 𝗚𝗲𝗺𝗺𝗮 𝟰. 𝗙𝗿𝗲𝗲. 𝗢𝗽𝗲𝗻. 𝗥𝘂𝗻𝘀 𝗼𝗻 𝘆𝗼𝘂𝗿 𝗹𝗮𝗽𝘁𝗼𝗽. 𝗔𝗽𝗿𝗶𝗹 𝟰: 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗯𝗮𝗻𝘀 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗳𝗿𝗼𝗺 𝗖𝗹𝗮𝘂𝗱𝗲 𝘀𝘂𝗯𝘀𝗰𝗿𝗶𝗽𝘁𝗶𝗼𝗻𝘀. Two days apart. And if you're building with AI agents, these two events tell you everything about where this is going. I woke up Saturday to messages from students in my OpenClaw course asking what happened. Their agents stopped working. Overnight. No warning that mattered. Some folks are looking at 10x to 50x cost increases just to keep what they already built running. If your entire AI agent setup depends on someone else's pricing page, you don't have a system. You have a permission slip. Two days before Anthropic pulled the plug, Google dropped Gemma 4. Fully open. Apache 2.0 license. Runs agentic workflows with function calling, structured output, multi-step reasoning — and you can run it on your own hardware. No API key. No subscription. No one can take it away from you. Six different labs now ship open models that are genuinely production-ready. That wasn't true a year ago. For individuals, this means grabbing Gemma 4 on Ollama and running it on an RTX laptop. Seriously. That's it. For enterprises, this is where it gets interesting. Dell Technologies just shipped the Pro Max with NVIDIA GB300 — a deskside AI supercomputer. 20 petaFLOPS. 748GB of coherent memory. Runs trillion-parameter models locally. Compatible with OpenClaw and NemoClaw. They even have an air-gapped variant for federal and classified environments. One machine. On your desk. No cloud bill. No subscription drama. EPRI (Electric Power Research Institute) is already running a full AI assistant locally on a Dell Pro Max with GB10, serving 200B-parameter models without a single API call. This isn't a concept or a roadmap slide. It's shipping to customers right now. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗶𝘀 𝗺𝗲𝗮𝗻𝘀 𝗳𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀 𝗮𝗻𝗱 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲𝘀: The Anthropic ban taught us something we should've already known: 𝗜𝗳 𝘆𝗼𝘂 𝗱𝗼𝗻'𝘁 𝗼𝘄𝗻 𝘁𝗵𝗲 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲, 𝘆𝗼𝘂 𝗱𝗼𝗻'𝘁 𝗼𝘄𝗻 𝘁𝗵𝗲 𝗮𝗴𝗲𝗻𝘁. My playbook going forward is simple: Cloud for exploring and prototyping. Open models for anything going to production. Local hardware when you need real control. And never, ever build your core workflow on a single provider's billing page. The local AI era is just getting started. The best builders in 2026 won't be the ones with the biggest API budget. They'll be the ones who own their stack. ♻️ Repost if you think AI builders should own their infrastructure. 💬 Are you running models locally yet? Drop your setup below — I'm putting together a resource list for the community. 🔗 "EPRI: Local AI for Energy Research" — https://lnkd.in/gfjQhWpD #AI #OpenSource #Gemma4 #OpenClaw #LocalAI #AIAgents #AIInfrastructure #Builders #Dell #IWork4Dell