Nvidia Unveils the First Agent‑Native PC: How Jensen Huang Is Redefining the Computer

GTC Launch and the RTX Spark Super‑Chip PC

Nvidia opened its GTC conference with a headline announcement: the RTX Spark, the world’s first PC built specifically for AI agents. The laptop packs 6144 CUDA cores, 128 GB of unified memory, and a 3 nm Blackwell GPU delivering roughly 1 PFLOP of AI compute, bringing data‑center‑class performance to a personal device.

The system is marketed as the first "Agent‑native" Windows PC, a hardware platform designed from the ground up for large‑language‑model (LLM) agents rather than traditional user‑centric applications.

Vera Rubin Production and Architecture

Vera Rubin, the new CPU co‑designed with MediaTek, is claimed to be the first CPU built for agents. Assembly time for a rack dropped from two hours to five minutes, a 24× speedup in production.

Key specifications: 20‑core Grace‑based CPU, 72 NVLink‑connected Rubin GPUs, 256 fully liquid‑cooled CPUs in a rack, and a Mellanox network system with the world’s first CPO‑co‑packaged optics.

Fabric: a Scalable Coherency Fabric links all cores without chiplet penalties, and the CPU is the first to use PCIe Gen6 with 1.2 TB/s LPDDR5X bandwidth—2–3× faster than the strongest x86 CPUs on the market.

Vera CPU Performance Claims

According to Nvidia, the Vera CPU achieves three breakthrough metrics:

World‑leading single‑core performance: each clock can fetch, decode, and execute ten instructions, a record for x86.

Unprecedented per‑core and system bandwidth: the Scalable Coherency Fabric eliminates chiplet tax, and PCIe Gen6 delivers 1.2 TB/s LPDDR5X, making it 2–3× faster than existing high‑end CPUs.

Extreme energy efficiency: designed to pack more CPUs into AI data centers without stealing the limited power needed for token generation.

Benchmark data cited by Nvidia shows the Vera CPU running SQL workloads up to three times faster than conventional x86 CPUs and six times faster on real‑time streaming workloads such as those used by the New York Stock Exchange.

Agent as the New Operating System

Nvidia frames the current era as the "Agent era," stating that agents will become the new OS. The traditional stack of application + OS is replaced by an "Agent + Harness" architecture, where LLMs handle reasoning and a harness layer orchestrates perception, planning, tool use, and memory.

To support this vision, Nvidia released the NVIDIA Agent Toolkit, a four‑layer stack:

Model layer: Nemotron 3 Ultra, an open‑source 550 billion‑parameter model (550 billion active parameters per token).

Framework layer: native support for Claude Code, Codex, OpenClaw, and other agent runtimes.

Tools & Skills layer: CUDA X libraries bundled with skill files for plug‑and‑play agent capabilities.

Runtime layer: OpenShell, an open‑source secure runtime (Apache 2.0).

The stack is illustrated as a "super‑agent" that can design chips, run RTL verification, perform formal checks, and fix bugs autonomously, compressing verification cycles from weeks to hours—a >40× acceleration.

AI Factory (DSX) and Future Outlook

Nvidia also announced DSX, an end‑to‑end AI‑factory blueprint. In the Omniverse, a digital twin of a data center is simulated for power, cooling, and networking before any hardware is shipped. Once powered, the DSX OS manages operations, delivering up to 40 % higher power utilization while fitting more GPUs into the same power budget.

CEO Jensen Huang projected that by the end of the century, 100 GW of AI‑factory capacity will be online, turning Nvidia from a chip seller into a provider of complete AI‑factory solutions.

Conclusion

The announcements collectively signal a paradigm shift: CPUs and PCs are being rebuilt for agents, data‑center architectures are being re‑engineered, and the software stack is being reconstructed to support a new generation of AI‑driven workloads.