NVIDIA Launches Vera CPU at GTC 2026 — First Processor Purpose-Built for Agentic AI, 2x Efficiency, Jensen Huang Sees $1 Trillion in Orders

NVIDIA launched the Vera CPU at GTC 2026 on March 16, 2026, marking the first purpose-built processor for the age of agentic AI and reinforcement learning. CEO Jensen Huang delivered the keynote at SAP Center in San Jose to 30,000+ attendees from 190 countries.

VERA CPU SPECIFICATIONS:

• 88 custom NVIDIA-designed Olympus cores
• 2x efficiency improvement over traditional rack-scale CPUs
• 50% faster than traditional CPUs
• Spatial Multithreading: each core runs two tasks simultaneously
• Second-generation Scalable Coherency Fabric for high-bandwidth memory
• Liquid-cooled Vera CPU rack: 256 CPUs sustaining 22,500+ concurrent environments
• NVLink-C2C interconnect: 1.8 TB/s coherent bandwidth (7x PCIe Gen 6)
• Dual and single-socket server configurations available
• Built on NVIDIA MGX modular reference architecture (80 ecosystem partners)

ADOPTION:

Hyperscaler customers: Alibaba, ByteDance, Meta, Oracle Cloud Infrastructure, CoreWeave, Lambda, Nebius, Nscale Manufacturing partners: Dell Technologies, HPE, Lenovo, Supermicro, ASUS, Foxconn, GIGABYTE, Quanta, and others

JENSEN HUANG QUOTE: "Vera is arriving at a turning point for AI. As intelligence becomes agentic — capable of reasoning and acting — the importance of the systems orchestrating that work is elevated. The CPU is no longer simply supporting the model; it is driving it."

$1 TRILLION ORDERS:

CNBC reported that during the keynote, Jensen Huang projected $1 trillion in cumulative orders for Blackwell and Vera Rubin platforms through 2027. NVDA stock rose 2% to approximately $184. Nebius shares jumped 14% on the Meta infrastructure partnership announcement.

WHY CPUs MATTER FOR AGENTIC AI:

Chip analyst Ben Bajarin of Creative Strategies told CNBC: "This is new infrastructure: Greenfield expansion of racks of CPUs whose only job is to run agentic AI." As reasoning and agentic AI advances, performance is increasingly driven by infrastructure supporting agent orchestration — planning tasks, running tools, interacting with data, running code, validating results.

The Vera CPU addresses a market gap: while GPUs handle model inference, CPUs handle the orchestration, data movement, and tool execution that agentic AI requires. A 256-CPU rack running 22,500 concurrent environments means a single rack can support thousands of AI agents simultaneously.

VERA RUBIN PLATFORM:

Vera CPUs pair with NVIDIA Rubin GPUs in the Vera Rubin NVL72 platform. The platform brings token costs for agentic AI inference down to 1/10th of the Blackwell platform. MoE model training requires only 1/4 the GPUs of previous generation.

FEYNMAN ROADMAP:

The Feynman GPU architecture is officially on the roadmap for 2027, with NVIDIA maintaining a 12-month product cycle. TradingKey analysis noted this creates a "treadmill effect" requiring continuous customer investment.

BROADER GTC CONTEXT:

• Build-a-Claw event running March 16-19 for hands-on OpenClaw agent deployment
• Agentic AI panel featuring LangChain CEO, OpenClaw creator, PrimeIntellect CEO
• DGX Spark systems available for purchase on-site
• Intel also present, positioning for CPU importance in agentic AI era