Everything Nvidia Announced at Computex 2026: From the Nemotron 3 Ultra to RTX Spark

The broader Nemotron 3 family

Nemotron 3 Ultra sits atop a three-model family that shares the same hybrid Mamba-Transformer MoE architecture .

All three support up to a 1-million-token context window. The Super and Ultra variants include NVFP4 training, LatentMoE, and multi-token prediction layers .

NemoClaw and OpenShell: secure, local AI agents

First introduced at GTC 2026 in March and reiterated at Computex, NemoClaw and OpenShell form Nvidia's enterprise agent stack .

• NemoClaw: An open-source, Apache 2.0-licensed reference stack built on top of OpenClaw. It installs with a single command and enables local deployment of autonomous AI agents on RTX PCs, workstations, and DGX systems—no cloud API fees, and data never leaves the device .
• OpenShell: An open-source security runtime that runs AI agents in a sandboxed environment. It provides policy-based guardrails, privacy controls, and isolation for long-running agent processes, comparable to a YAML-policy-driven Docker layer for AI agents .
• Together they form the Nvidia Agent Toolkit, a suite for enterprises that want to build and run autonomous AI coworkers securely on local hardware .

Vera CPU: the compute brain of the Vera Rubin platform

The Nvidia Vera CPU is a custom Arm-based processor purpose-built for orchestrating AI workloads in data centers .

• Core count: 88 custom Arm cores on a monolithic die with no chiplet boundaries .
• Memory bandwidth: 1.2 TB/s of LPDDR5X .
• On-chip fabric: 3.6 TB/s for inter-core data movement .
• Single-thread performance: 10 instructions per clock (IPC) .
• Nvidia claims Vera delivers 1.8x the agentic AI performance of x86 CPUs .

The Vera CPU serves as the host processor inside Vera Rubin NVL72 rack-scale systems, where it works alongside next-generation GPUs to power AI factories .

RTX Spark: a consumer superchip for Windows PCs

RTX Spark is Nvidia’s first dedicated system-on-a-chip for consumer Windows PCs, co-developed with MediaTek and built on TSMC’s 3nm process . Jensen Huang called it “everything we’ve learned over 33 years distilled into one chip” .

• CPU: 20-core Arm processor — 10 Cortex-X925 performance cores plus 10 Cortex-A725 efficiency cores, peaking at 4.1 GHz .
• GPU: Blackwell RTX GPU with 6,144 CUDA cores and 5th-generation Tensor Cores with FP4 support — the same core count as a desktop GeForce RTX 5070 .
• Unified memory: Up to 128 GB of LPDDR5X .
• Interconnect: NVLink-C2C chip-to-chip link with up to 600 GB/s bandwidth between CPU and GPU dies .
• AI performance: 1 petaflop of FP4 AI compute, capable of running 120-billion-parameter large language models with 1 million tokens of context entirely on-device .
• Software stack: Full CUDA support, plus compatibility with NemoClaw and OpenShell for running local AI agents .
• Availability: Laptops from Dell, Microsoft, and other partners are expected in fall 2026 .
• Use cases: Local AI agents, rendering 90 GB 3D scenes, editing 12K 4:2:2 video, AI video generation in 4K, and gaming above 100 FPS at 1440p .

Nvidia positions RTX Spark as “the world’s first Windows PCs purpose-built for personal agents” .

Vera Rubin platform: mass production and early cloud adopters

Nvidia confirmed that the Vera Rubin platform has entered full production, with systems expected to reach partners in the second half of 2026 .

• Platform scope: Six to seven co-designed chips, including the Vera CPU and next-generation GPUs, architected for rack-scale AI factories .
• Full rack performance: The NVL72 configuration delivers 3.6 EFLOPS of AI compute .
• Supply chain scale: More than 350 factories across 30 countries, including over 150 partners in Taiwan, are manufacturing Vera Rubin systems .
• Named early cloud and service provider adopters:
  • Hyperscalers: Microsoft Azure, AWS, Google Cloud .
  • Specialist cloud providers: CoreWeave, Lambda, Oracle Cloud Infrastructure .

At the Computex keynote, Nvidia described Vera Rubin as the foundation of the next generation of AI factories .

Nvidia's full-stack AI platform takes shape

Taken together, Computex 2026 was the clearest demonstration yet of Nvidia’s vertical integration strategy:

• Silicon: Vera CPU for the data center, RTX Spark for consumer PCs, and Vera Rubin GPUs for rack-scale compute cover the full performance spectrum from edge to cloud .
• Models: The Nemotron 3 family—Nano, Super, and Ultra—provides open-weights models optimized for Nvidia hardware at every deployment tier .
• Software and agents: NemoClaw, OpenShell, and the Agent Toolkit let enterprises build and deploy autonomous, secure AI agents locally using Nvidia’s own models .
• Infrastructure: Vera Rubin’s mass production ramp with hyperscale and cloud partners ensures the supply chain is in place to build “AI factories” at global scale .

Jensen Huang’s keynote message was singular: Nvidia now supplies every layer of the AI stack, from silicon and open models to agent runtimes and factory-scale deployment infrastructure.