Attended NVIDIA GTC 2026

I attended NVIDIA GTC 2026, starting with a pre-conference workshop followed by the main conference sessions.

The first day was dedicated to NVIDIA’s Accelerated Networking for AI Infrastructure workshop, which provided a strong systems-level overview of how modern large-scale AI clusters are designed and operated.

The workshop covered a wide range of topics, including:

• NVIDIA Blackwell NVL72 architecture
• NVIDIA Collective Communications Library (NCCL)
• PCIe topology and traffic-tree design
• NVLink / NVSwitch full-mesh fabric
• Ring topology and railed optimized networks
• Grace Blackwell compute tray
• NVIDIA AI Factory networking architecture
• InfiniBand and RoCE (RDMA)
• Interconnecting thousands of GPUs
• NetQ, UFM, and DGX B200 infrastructure blocks

This workshop provided a valuable foundation for understanding the communication and networking challenges behind scaling AI systems.

The rest of the conference focused on both massive infrastructure advancements and a strong shift toward agent-based AI systems.

From the keynote by Jensen Huang and technical sessions, a few themes stood out:

• Continued evolution of Blackwell-era systems for large-scale training and inference
• The rise of AI factories—tightly integrated compute, networking, and storage systems
• Increasing emphasis on AI agents, including frameworks such as NeMo/NeMo Agent tools and OpenCLAW-style systems
• Ongoing improvements in performance, scalability, and efficiency across the stack

In addition, I had in-depth discussions with engineers from NVIDIA, VAST Data, and Lenovo on topics such as storage systems, GPU Direct Storage, and large-scale inference design.

Overall, GTC 2026 reinforced a clear direction for the field:

The future of AI systems lies in the combination of extreme-scale infrastructure and intelligent, agent-driven workflows.

As always, GTC remains one of the most insightful events for understanding where AI systems engineering is heading next.

Photo Gallery