TLDR
- At GTC 2026, Nvidia introduced the Groq 3 LPU, a processor designed exclusively for AI inference operations
- The LPX server rack features 128 Groq 3 chips and, when combined with Vera Rubin NVL72, delivers up to 35x improved throughput per megawatt
- A standalone Vera CPU rack was revealed, marking Nvidia’s direct entry into Intel and AMD’s data center CPU domain
- Vera CPU targets agentic AI applications, including web navigation and file data retrieval
- Nvidia’s data center segment generated $193.5 billion in fiscal 2026, a significant jump from $116.2 billion in the previous fiscal year
At Monday’s GTC 2026 conference in San Jose, Nvidia showcased a collection of innovative chips and server architectures that extend far beyond the company’s traditional GPU stronghold.
The star announcement was the Groq 3 language processing unit — abbreviated as LPU. Through a $20 billion transaction finalized in December, Nvidia secured licensing rights to Groq’s technology and welcomed founder Jonathan Ross, president Sunny Madra, and additional team members into the fold.
Groq 3 specializes in inference — the AI phase that occurs post-training. When users submit queries to chatbots and receive responses, they’re witnessing inference in action. This rapidly expanding AI market segment offers opportunities for purpose-built processors to outperform general-use GPUs.
According to Nvidia VP of hyperscale and HPC Ian Buck, Groq 3’s memory operates at speeds exceeding those in Nvidia’s GPUs, despite GPUs offering greater memory capacity. The strategy involves leveraging both advantages simultaneously.
This concept materializes in the LPX server rack — a configuration containing 128 Groq 3 LPUs. When deployed alongside the Vera Rubin NVL72 rack, Nvidia claims customers achieve 35x enhanced throughput per megawatt and unlock 10x greater revenue potential. The system targets trillion-parameter models and million-token context capabilities.
Vera CPU Enters Intel and AMD Territory
The second major announcement featured the Vera CPU rack. While Vera previously appeared as a component within the Vera Rubin superchip — pairing one Vera CPU with two Rubin GPUs — Nvidia now positions Vera as an independent processor.
This new rack configuration integrates 256 liquid-cooled Vera processors into a unified system. Nvidia markets it as the optimal CPU for agentic AI — artificial intelligence that independently navigates websites, extracts file data, or executes complex multi-step operations without human intervention.
“We’ve designed a new kind of CPU, the Olympus core, engineered by NVIDIA for AI execution,” Buck said. Vera also plays a role in data mining, personalization, and context analysis that feeds into AI models.
This move positions Nvidia as a direct competitor to Intel and AMD within the data center CPU marketplace — territory these two companies have controlled for decades.
Nvidia recently announced a partnership with Meta to implement its previous-generation Grace CPUs at unprecedented scale. The Vera introduction amplifies that strategic direction.
Additional Hardware Announcements From GTC
Nvidia demonstrated the Bluefield-4 STX storage rack and Spectrum-6 SPX networking rack, completing a comprehensive data center hardware portfolio.
Major hyperscalers like Amazon, Google, Meta, and Microsoft plan to invest a combined $650 billion in AI infrastructure throughout this year.
Nvidia’s data center division achieved $193.5 billion in revenue during fiscal 2026, compared to $116.2 billion in fiscal 2025.
Wall Street analysts maintain a consensus Strong Buy rating on NVDA, with 38 Buy ratings and one Hold rating issued over the last three months, targeting an average price of $273.61.
