Huang's 1.4 exaflop reveal

Jensen Huang is pitching Nvidia’s AI future at two scales at once: a 1.4-exaflop GB200 NVL72 rack system for data centers and a DGX Spark desktop box for local agents. (developer.nvidia.com) An exaflop is one quintillion floating-point operations per second, a standard way to describe how much math a machine can do. Nvidia says its GB200 NVL72 links 36 Grace central processors and 72 Blackwell graphics processors to reach 1.4 exaflops of AI performance. (blogs.nvidia.com) (developer.nvidia.com) At the other end of the lineup, DGX Spark is a desktop “personal AI supercomputer” built around Nvidia’s GB10 Grace Blackwell superchip. Nvidia says the system delivers up to 1 petaFLOP of FP4 AI performance and includes 128 gigabytes of unified memory. (nvidia.com) Nvidia introduced DGX Spark on March 18, 2025, saying developers could prototype, fine-tune and run large models locally, then move the same workloads to DGX Cloud or other accelerated data center infrastructure with minimal code changes. Nvidia said the machine was formerly called Project DIGITS. (nvidianews.nvidia.com) The local-agent pitch is about keeping some AI work on a desk instead of in a remote server farm. Nvidia’s current DGX Spark materials say NemoClaw and the Agent Toolkit let users build, evaluate and run “always-on” autonomous agents directly from the desktop. (nvidia.com) Nvidia’s March 16, 2026 technical blog framed the bottleneck in plain terms: agents spend time “reading” long prompts before they answer, and slow prompt processing can choke the whole workflow. The company said DGX Spark now scales to four nodes and can support fine-tuning and inference on models up to 700 billion parameters when those systems are linked together. (developer.nvidia.com) Nvidia’s March 17, 2026 GTC blog said DGX Spark can run local models with more than 120 billion parameters and highlighted open models such as Nemotron 3 Super 120B and Mistral Small 4 for agent tasks. The same post described DGX Spark and RTX PCs as machines for “private” local agents, tying the hardware pitch to privacy and on-premises use. (blogs.nvidia.com) The price point in Huang’s remarks appears to describe the earlier entry-level pitch, not Nvidia’s current U.S. list price. Nvidia’s marketplace now lists DGX Spark at $4,699, while third-party reports say the product was initially presented from about $3,000 to $3,999 before later increases. (marketplace.nvidia.com) (hardware-corner.net) (wccftech.com) That leaves Huang’s message in concrete terms: Nvidia wants customers to buy giant Blackwell systems for centralized AI throughput and smaller Grace Blackwell machines for local deployment, using the same software stack across both. (nvidianews.nvidia.com) (nvidia.com)