Huang Renxun GTC 2026 Redefines PC: Vera Rubin Becomes the First Agentic CPU, Nvidia Teams Up with MediaTek to Launch RTX SPARK Laptop Revolution

Huang Renxun announced two major pieces of news at the GTC Taipei 2026 keynote speech: the official debut of the Vera Rubin CPU designed for the Agentic AI era, and the first PC platform integrating CPU, GPU, and NVLink on a single wafer, developed in collaboration with MediaTek, symbolizing the PC architecture's official leap into the server-grade AI computing era.
(Background summary: Huang Renxun reveals "Token Economics" at GTC 2026: computation equals revenue, Nvidia begins full-scale mass production of Vera Rubin, Taiwan's AI demand skyrockets)
(Additional background: Huang Renxun at GTC 2026 passionately discusses "Hardness"? Why LLM Agents need to harden, revealing the key to AI agent deployment in one sentence)

Table of Contents

Toggle

• Vera Rubin: The world's first CPU designed specifically for Agentic AI
• RTX SPARK: MediaTek partnership, the biggest revolution in PC architecture in 50 years
• Adobe software ecosystem fully embraces, MCP server unlocks Agentic workflows
• The moment to redefine personal computers

Huang Renxun’s keynote at GTC 2026 in Taipei sketches a clear roadmap for the future of personal computers. From the Vera Rubin data center to the desktop RTX SPARK, NVIDIA is bringing server-grade AI architecture into the realm of personal computing, and the starting point is a CPU specially designed for AI agents (AI Agents).

Vera Rubin: The world's first CPU designed specifically for Agentic AI

Vera is NVIDIA’s first CPU built from scratch for the AI agent era. Compared to traditional x86 server processors, Vera’s biggest difference lies in its extreme optimization for "bandwidth" and "latency," which are the two major bottlenecks in AI agent workflows.

In terms of hardware specifications, Vera supports PCIe Gen 6 interface, providing data throughput of up to 1.4 TB/s; for memory, it uses LPDDR5X ECC with a bandwidth of about 1.2 TB/s, ensuring that AI agents can handle large-scale real-time inference tasks without stalling due to data transfer delays. Huang Renxun emphasized during the speech that Vera’s end-to-end latency performance across various AI workloads far surpasses that of comparable x86 servers.

Vera Rubin AI factory is not a single product but a complete platform comprising 7 chips and 5 chassis configurations. The most notable is the NVL72 chassis, which integrates 72 Rubin GPUs and 36 Vera CPUs, designed specifically for ultra-large-scale Agentic AI inference. Additionally, NVIDIA has integrated the Groq LPU (Language Processing Unit), acquired last year for $2 billion, into this platform to create a dedicated hardware layer for structured AI agent outputs, enabling hardware acceleration for both "thinking" and "acting" of agents.

RTX SPARK: MediaTek partnership, the biggest revolution in PC architecture in 50 years

If Vera Rubin defines the future of Agentic AI in data centers, then RTX SPARK is the key to bringing that future directly to your desktop. This is NVIDIA’s first laptop-grade System on Chip (SoC), manufactured using TSMC’s 3nm process, with a transistor count of up to 70 billion.

The core architecture of RTX SPARK integrates three key components: a 20-core Grace CPU co-designed with MediaTek, a Blackwell generation GPU with 6,144 CUDA cores, and 128GB of unified memory. The most revolutionary design is the NVLink C2C interconnect technology, a high-speed chip-to-chip connection channel reaching 600 GB/s bandwidth, five times that of traditional PCIe Gen 5, enabling nearly zero-latency data transfer between CPU and GPU.

"This is the first time in personal computer history that the CPU, GPU, and high-speed interconnect are all integrated on the same wafer," Huang Renxun said during the speech. Previously, only top AI labs like OpenAI could enjoy such architectures in server rooms; now, it’s condensed into a device about the size of a laptop. RTX SPARK’s FP4 AI performance reaches 1 petaFLOP, a level of performance that previously required an entire GPU server rack.

Adobe software ecosystem fully embraces, MCP server unlocks Agentic workflows

RTX SPARK is not just a stack of hardware specifications; software ecosystem coordination is equally critical. Huang Renxun announced that Adobe has rewritten the core rendering engines of Photoshop and Premiere Pro for the RTX SPARK platform, achieving performance improvements of up to two times.

More notably, Adobe also released the MCP (Model Context Protocol) server, enabling AI agents to directly control internal functions of Photoshop and Premiere, from automatic retouching and batch color correction to video editing automation, all via natural language commands. This concretely realizes Huang Renxun’s repeated emphasis on the "Agentic AI" vision at GTC 2026: hardware (RTX SPARK) provides sufficient native AI computing power, while software (Adobe MCP) allows AI agents to truly "grow hands and feet."

The moment to redefine personal computers

Since the IBM PC established the x86 architecture standard in 1981, the core design logic of personal computers has remained fundamentally unchanged: CPUs handle general-purpose computing, GPUs handle graphics rendering, and they communicate via low-speed buses. The two major announcements at GTC 2026 are rewriting this nearly half-century-old formula.

Vera Rubin demonstrates the value of redesigning CPU architecture for specific AI workloads: not a faster x86, but a completely different computing paradigm. RTX SPARK brings server-grade CPU-GPU unified memory architecture into the consumer market, transforming personal computers from "productivity tools" into "AI agent hosts."

Behind this revolution is NVIDIA’s extensive supply chain network established in Taiwan, with over 150 Taiwanese suppliers—from TSMC’s 3nm CoWoS-R/L advanced packaging to system integration manufacturing by Foxconn and Quanta—once again placing Taiwan at the forefront of the most significant paradigm shift in personal computer history.