What will Nvidia announce at GTC 2026?

New server chips and racks aimed at agentic AI

Nvidia's GTC keynote, which begins March 16, is expected to expand the company's push beyond GPUs into CPU hardware and rack-level systems designed specifically for agentic workloads. Reports previewing the show say Nvidia will unveil "agentic-optimized" CPUs and a CPU-only rack alongside other infrastructure products.

Why this matters

Nvidia built its software and data-center dominance on GPUs; moving into CPUs and full rack designs signals a strategy to control more of the end‑to‑end stack that runs modern AI systems. Agentic applications—systems that orchestrate multiple steps, call other models or services, and operate with more autonomy—place different demands on latency, memory persistence, and orchestration than single‑model inference. By designing CPUs and rack architectures around those needs, Nvidia aims to offer customers tuned alternatives to the current mix of general-purpose x86 servers plus GPU accelerators.

Immediate implications

• Hardware vendors and cloud providers will get another option tailored to multi‑model agent workflows.
• Customers building agentic systems may be able to reduce integration complexity and costs if the new stacks come with compatible software.
• Competitors that supply CPUs, interconnects, and servers will face pressure to offer similar, agent-friendly solutions.

Unknowns and limits

Details remain limited: performance claims, pricing, software integration, and real‑world availability were not fully disclosed in previews. How customers will balance these new products against GPU-heavy deployments, and whether the offerings will meaningfully shift the economics of large‑scale agent deployments, will depend on benchmarks and early enterprise uptake after GTC.