Jensen Huang: Agentic Compute Up 1000%, $710B Industry Capex

Nvidia CEO Jensen Huang delivered his annual Computex keynote on May 16, revealing that agentic AI compute demand has increased 1,000% year-over-year, with global AI infrastructure capital expenditure projected to reach $710 billion in 2026 — and announcing Nvidia's partnership with utilities providers to build SMR nuclear capacity from 25 to 45 gigawatts to meet AI energy demand.

Key Points:

• Huang described the 1,000% increase in agentic compute as a fundamentally different demand curve from inference scaling — agents are running continuously, not just on user prompts, creating persistent compute load.

• The $710 billion global AI capex figure represents 140% year-over-year growth and is being driven by hyperscaler data center expansion in the US, Middle East, Southeast Asia, and India.

• The SMR nuclear partnership targets building 45 GW of small modular reactor capacity by 2032 to provide carbon-neutral baseload power for AI data centers, with initial sites announced in Texas, Ohio, and Wyoming.

Agentic AI running continuously rather than episodically changes the entire economics of AI infrastructure. Organizations planning their AI compute budgets based on inference-per-query models will significantly underestimate their actual costs as agent adoption grows.

Nuclear power becoming a serious AI infrastructure topic — with specific capacity and timeline commitments — signals that AI energy demand has crossed a threshold where it is shaping long-term energy policy and investment.

Factor agentic compute costs into your 2027 AI budget planning now. Continuously running agents have fundamentally different cost structures than on-demand inference. Budget models need to reflect this. For AI governance and sustainability leaders, the SMR nuclear announcement opens a new dimension in responsible AI deployment — energy sourcing and carbon footprint are now legitimate AI governance evaluation criteria.

Why It Matters: Continuously running agents create fundamentally different cost structures than on-demand inference, invalidating budget models based on per-query pricing. Nuclear power commitments signal AI energy demand is now shaping long-term energy policy.