Nvidia Shifts From Chips to Full Data Centers at GTC 2026

Nvidia's annual GPU developer conference GTC, held each March in San Jose, serves as a key stage showcasing global tech industry trends. Two years ago, CEO Jensen Huang highlighted "Blackwell," a high-performance single chip.

This March was different. Huang unveiled not a new single chip but a massive system comprising seven different chips and five racks. The move signals Nvidia's expansion beyond fabless semiconductor design into a turnkey supplier of entire AI data center infrastructure.

Single-Chip Architecture Limits and Heterogeneous AI Factories

The AI industry's focus has shifted from model "training" to "inference" for real-world services. Token processing volume for inference has surged approximately 10,000-fold since ChatGPT's emergence, as AI evolves into an agentic era where models call tools and verify logic independently. Single general-purpose GPU architecture alone can no longer overcome power consumption and memory bandwidth limitations.

"Every CEO worldwide will evaluate their business using 'tokens per watt' as a key metric," Huang declared in his keynote. Nvidia responded by presenting a rack system integrating seven chips—including the 50-petaflops Rubin GPU, Vera CPU, and Groq3 Language Processing Unit (LPU).

$20 Billion Groq Acquisition and LPU Integration

Industry observers focused on the integration of Groq's LPU, acquired for $20 billion. During the decode stage where large language models generate responses, conventional GPUs face memory bottlenecks and idle time.

Groq3 LPU delivers 150 terabytes per second bandwidth with just 500 megabytes of SRAM through deterministic design with static data flow scheduling. Nvidia unified this heterogeneous hardware through its Dynamo operating system, achieving up to 35-fold performance improvement over Blackwell.

Dylan Patel, chief analyst at SemiAnalysis, suggested Nvidia "released conservative figures to manage expectations" and that "actual performance gains could reach 50-fold."

Vera CPU's Rise in the Agentic Era

CPUs, previously relegated to supporting roles, have gained importance as serial processing capability determines system-wide bottlenecks in agentic AI environments. The in-house Vera CPU features 88 Olympus cores specialized for agent control and LPDDR5X memory, doubling energy efficiency.

"I never thought we'd sell CPUs standalone, but we're now selling massive quantities—already a multi-billion dollar business," Huang said. Meta and other big tech companies have placed standalone pre-orders for Vera CPUs.

CPO Technology for Power Efficiency

Power availability remains the primary constraint on AI data center expansion. Nvidia commercialized co-packaged optics (CPO) technology in its Spectrum-6 switch, reducing per-port power consumption from 30 watts to 9 watts by placing micro-ring modulators directly on substrates.

$1 Trillion Revenue Visibility and Space Data Centers

Huang stated Nvidia has "secured visibility for at least $1 trillion in revenue by end of 2027," noting this figure includes only Blackwell and Rubin GPUs—excluding Groq, storage, networking, and Vera CPUs.

To address power shortages, Nvidia unveiled "Project Space One"—orbital data centers leveraging 6-10 times higher solar efficiency and unlimited natural cooling in space. Analysts say Nvidia is evolving from a chip performance competitor into a comprehensive AI infrastructure solution provider.

However, equipment supply remains a constraint. SemiAnalysis noted ASML's annual EUV equipment production of approximately 70 units means "global AI computing capacity will hit physical limits by 2030."