Post by Linda Haviv

NVIDIA & Microsoft just dropped a unified stack at Build, helping you run frontier-level agents on your own hardware! Here's what you need to know: 🔷𝗛𝗔𝗥𝗗𝗪𝗔𝗥𝗘 The Surface RTX Spark Dev Box packs up to 1 petaflop of AI compute and 128GB of unified memory shared across CPU and GPU. That's enough to run agent workloads locally, up to 120B-parameter models, right on your desk. No setup friction, no surprise cloud costs. 🔷𝗠𝗢𝗗𝗘𝗟𝗦 A stack needs models built for long-running agents, not just chat. NVIDIA Nemotron 3 Ultra, a new open frontier reasoning model for coding, research, and enterprise workflows, lands this month on Microsoft Foundry. And Foundry's hosted agents now run NVIDIA, Anthropic, and OpenAI models side by side, with Anthropic's Claude running natively on NVIDIA GB300 on Azure. You can pick the model per workflow, not per platform. 🔷𝗗𝗔𝗧𝗔 NVIDIA acceleration is now baked into Microsoft Fabric Data Warehouse. Here's why that matters: an agent doesn't query once, it queries in a loop, so your data layer has to keep up. Microsoft's benchmarks put it up to 6x faster than the CPU baseline, and up to 7x faster than three other leading cloud warehouses on high-concurrency workloads. 🔷𝗥𝗨𝗡𝗧𝗜𝗠𝗘 This is the part I think is very underrated. Autonomous agents need real capability without real credentials. NVIDIA OpenShell, now in GitHub Copilot, runs each agent in its own sandboxed container and checks every outbound call against policy before it can touch your files, network, or credentials. Policy is written as code, versioned in your repo. It's also the secure runtime behind running agent projects like OpenClaw on Windows. So to sum it up: the hardware to run frontier-level agents, the models to power them, the data layer to feed them, and the runtime to keep them safe. On your machine, and across the cloud. 📖Read NIVIDIA’s blog with the full breakdown: https://lnkd.in/eVFc5BtC #ad