Nvidia has announced a new 64GB configuration of its GB10-powered DGX Spark desktop system, produced alongside hardware partners including Acer, ASUS, Dell, Gigabyte, HP, and MSI. Built around the Grace Blackwell superchip with unified LPDDR5x memory and up to 1 petaFLOP of FP4 compute, the unit allows developers to run open-weight AI models locally without incurring per-token cloud API fees. Users can also cluster two 64GB units via ConnectX-7 networking to achieve 128GB of combined memory.

The launch targets continuous agentic workloads, which have seen a 14x increase in token consumption since early 2026. Shipments feature pre-installed software including PyTorch, Ollama, and Nvidia Agent Toolkit components to secure and run agents directly on standard wall-power infrastructure.

Why it matters

  • Enables startups and operators to avoid metered cloud API costs for heavy, continuous agentic token execution.

  • Simplifies enterprise data privacy by keeping source code, tool execution, and local model inference fully on-premise.

Source: marktechpost.com