SANTA CLARA

The AI Infra Summit, held Sept. 15-17 at the Santa Clara Convention Center, drew over 8,000 attendees focused on a single problem: the economics of scaling AI at the infrastructure layer.

Unlike broad AI conferences, the event zeroed in on the technical and financial viability of deploying models at enterprise scale. Organizers explicitly avoided applications-focused sessions, instead drilling into the foundational stack—data center architecture, compute efficiency, data movement, and the tradeoffs between performance and cost.

Nvidia's Ian Buck, in a keynote on agentic AI infrastructure, highlighted a structural shift in data center design. He discussed the Vera CPU, Groq 3 LPX, and BlueField-4, arguing that agentic workloads are latency-sensitive and benefit from Nvidia's lower-core-count approaches rather than traditional high-density configurations. The implication: the economics of AI inference differ materially from training, requiring rearchitected infrastructure.

Event organizers framed infrastructure quality as the limiting factor in AI investment returns and public acceptance of AI development. The summit's agenda mapped onto Jensen Huang's AI layer cake framework—data centers, compute, data movement, models, and physical AI—attempting to show how each component affects end-to-end performance and unit economics.

A panel on "State of AI Infrastructure: What's Powering AI Today" featured Kent Dra chief commercial officer of IREN, Alex Freedland, CEO of Mirantis, and Dion Harris, Nvidia's senior director of HPC and AI hyperscale infrastructure solutions. Their discussion centered on operational efficiency and the capital intensity of hyperscaler builds.

David Patterson, the computer scientist, advised engineers to master generalized silicon engineering rather than chasing narrow specializations, as AI architectures continue to shift rapidly. The warning underscores a real risk for practitioners: expertise in yesterday's GPU topology or memory hierarchy becomes obsolete quickly.

The expo hall hosted over 250 vendors, including infrastructure-focused companies like CoreWeave and Nebius, demonstrating the scale of the supply chain required to support AI scaling. Attendance cost $3,297 for the full conference pass, which included keynotes, five content tracks, and networking.