How RTX Spark and Agent CPUs Could Trigger the First PC Revolution in 40 Years

On June 1 2026, Jensen Huang stepped onto the GTC Taipei stage in his signature leather jacket and delivered a two‑hour keynote that unveiled an end‑to‑end AI stack covering chips, data‑center systems, operating systems and robots, and introduced a new computing paradigm that runs Agents instead of traditional applications.

Huang cited GitHub data showing that global software‑developer activity grew from 300‑400 million annual commits in 2023‑2024 to almost three times that amount in the first months of 2026, arguing that tokens have become a profit unit and that AI tools multiply developer output, turning $3 trillion of salaries into roughly $9 trillion of productivity.

He announced Vera Rubin, a multi‑rack, pod‑scale supercomputer built from seven new 3 nm chips, each containing 60 trillion transistors, more than 18 000 components per board, HBM4 memory, and NVLink‑enabled interconnects. The third‑generation MGX‑3 rack houses 18 compute trays, 9 NVLink switch trays, a total of 1.3 million components and a liquid‑cooled bus capable of carrying over 5 000 A—equivalent to the current draw of twenty electric cars—providing the backbone for Agent workloads.

NVIDIA’s 20‑year CUDA‑X library portfolio—cuLitho, cuOpt, Parabricks, PhysicsNeMo, cuDSS, AI‑Q and others—is now exposed as tools that Agents can invoke, enabling a decoupled, distributed, heterogeneous compute model where each Agent call can activate an entire Grace Blackwell NVL72 cabinet.

The DSX ecosystem—DSX OS, DSX MaxLPS and DSX Flex—provides digital‑twin planning, dynamic power allocation and multi‑tenant AI‑ready compute. DSX Sim, based on Omniverse, lets partners layout, power‑model and cool racks in a virtual environment before ordering hardware. Huang predicts 100 GW of AI factories will be online by the end of 2030.

Turning to the PC, NVIDIA and Microsoft are co‑designing a new stack. The Vera CPU, built for Agents, integrates 88 Olympus cores on a mesh, a 10‑wide decode engine, a neural‑branch predictor, and LPDDR5X memory delivering 1.2 TB/s bandwidth—2‑3× faster than the market‑leading CPUs. It is the first CPU to adopt PCIe Gen 6 and supports NVLink interconnects. Benchmarks show 3× faster SQL queries and 6× faster NYSE real‑time stream processing compared with x86, and an Agent sandbox that is 1.8× faster than x86.

RTX Spark fuses a Blackwell‑based GPU with a 20‑core Grace‑derived CPU on a 3 nm process, packing 6 144 Tensor Cores, 1 PFLOPS of AI performance, 128 GB of unified memory and NVLink‑based GPU‑CPU fusion. NVIDIA claims the chip can run the entire CUDA ecosystem, digital‑biology, seismic processing, astrophysics, genomics, AI and graphics workloads, and is fully compatible with Windows applications optimized by Microsoft.

Additional announcements included Nemotron 3 Ultra (an SSM+MoE model that is five times faster than the best open‑source models and 30 % cheaper to run), the open‑source Cosmos 3 physical‑AI model, the Isaac GR00T humanoid robot platform (31 DOF, 6 ft tall, 150 lb, powered by a Jetson Thor chip and the full NVIDIA software stack), and the NVIDIA Agent Toolkit for Enterprise AI, which layers model, harness, tool‑skill and runtime components for enterprise Agent deployments.

Huang concluded that Agentic computing will become the core paradigm for the next ten years, that every device—from PCs to robots, autonomous cars and base stations—will run Agents, and that NVIDIA has evolved from a GPU company into an AI infrastructure provider that helps customers build and operate AI factories worldwide.