NVIDIA GTC 2026 Keynote Today: Jensen Huang Expected to Unveil Dedicated Inference Chip (Post-Groq Acquisition) and CPU-Only Agentic AI Server Racks

NVIDIA GTC 2026 keynote day has arrived. Jensen Huang takes the stage at SAP Center, San Jose, at 11 AM PT on March 16, 2026. Multiple sources point to landmark announcements that could reshape the AI hardware landscape, particularly for agentic AI workloads.

DEDICATED INFERENCE CHIP — POST-GROQ ACQUISITION:

The Financial Times reported (cited by Seoul Economic Daily on March 16) that NVIDIA plans to unveil a new chip focused specifically on inference rather than model training. This would be NVIDIA first product following its $20 billion acquisition of Groq — the largest deal in the company history.

Groq, founded in 2016 by engineers who developed Google Tensor Processing Unit (TPU), has been developing Language Processing Units (LPUs) designed to accelerate AI processing. NVIDIA is expected to unveil its first product based on this technology.

Why this matters for agentic AI:

• Agentic AI is inference-heavy — agents need fast responses, tool calls, and reasoning chains, not training runs
• GPUs have been criticized as inefficient for inference due to high costs and power consumption
• A dedicated inference chip addresses the exact bottleneck that makes agent deployment expensive
• The FT analyzed that "NVIDIA new chip will be central to a product lineup designed to fend off rivals challenges and address emerging AI demand"

CPU-ONLY AGENTIC AI SERVER RACKS:

Seoul Economic Daily reports that NVIDIA may also unveil a central processing unit optimized specifically for agentic AI. A CPU-only server rack capable of running AI agents without GPUs could be displayed. This would be significant because:

• Current AI infrastructure typically requires CPU+GPU combinations (Grace-Blackwell, Vera-Rubin)
• A CPU-only agentic AI rack would dramatically lower the hardware barrier for agent deployment
• This could threaten Intel and AMD, which have dominated the server CPU market
• Intel is reportedly also present at GTC, sensing the opportunity as agentic AI elevates CPU importance

Wccftech reported (March 15) that Intel is showing up at NVIDIA GTC "at the perfect time, as agentic AI turns CPUs into the new bottleneck."

GPU ROADMAP — BEYOND RUBIN:

NVIDIA previously announced its GPU development roadmap: Rubin in 2026, Rubin Ultra in 2027, and Feynman in 2028. Features of Feynman could be previewed at this GTC.

HBM4 MEMORY BATTLE:

Samsung and SK Hynix are competing fiercely for HBM4 dominance — the memory chips for Rubin GPUs. SK Group Chairman Chey Tae-won reportedly plans to meet Jensen Huang. Samsung Electronics VP Song Yong-ho will present on "The Future of Semiconductor Manufacturing through AI."

KEY NUMBERS:

• 30,000+ attendees from 190 countries
• 1,000+ sessions (online and in-person)
• $20B — NVIDIA Groq acquisition price
• $1T+ — estimated total AI infrastructure investment cycle (SiliconANGLE)
• Preshow featured: LangChain CEO, OpenClaw creator, Palantir President, Dell CEO, Cadence CEO, IBM, Mistral, Cohere, Perplexity CEOs

BROADER SIGNIFICANCE:

This GTC marks NVIDIA pivot from being primarily a GPU company to becoming a full-stack AI infrastructure provider. The inference chip and CPU-only server rack announcements, if confirmed, would lower the barrier to AI agent deployment and could accelerate the agentic AI market that Morgan Stanley projects will reach $139 billion by 2030.