Nvidia CEO says agentic AI needs 1,000x compute

Nvidia CEO Jensen Huang has spent months arguing that the next wave of artificial intelligence will consume far more computing power than the first chatbot boom. On May 5, during a CNBC appearance from ServiceNow’s Knowledge 2026 conference in Las Vegas, Huang said “agentic AI” requires roughly 1,000% more compute than earlier generative AI systems, according to CNBC coverage and follow-on reports. That figure matters because Huang is drawing a line between two phases of AI. The first phase centered on models that generated text, images or code from prompts. The newer phase, as Nvidia and its partners describe it, involves software agents that can take actions across enterprise systems, use tools, and complete multistep tasks with more autonomy. (cnbc.com) ### Where did the 1,000% figure come from? CNBC’s May 5 segment featured Huang alongside ServiceNow CEO Bill McDermott at the company’s annual conference. While the clip page carries only a short description, multiple follow-on reports citing the appearance said Huang argued that agentic AI needs 1,000% more compute than generative AI because the systems must read, reason, use tools and generate more tokens in real time. (blogs.nvidia.com) Huang has made a similar argument before with a smaller number. In a Feb. 26 CNBC interview after Nvidia earnings, he said reasoning models require “100 times more” computation than earlier systems because they process answers step by step rather than producing a single immediate response. (cnbc.com) ### What does Nvidia mean by “agentic AI”? Nvidia and ServiceNow used the term on May 5 to describe autonomous or semi-autonomous software that can operate inside business workflows. Nvidia’s blog said enterprise AI has already learned to generate and reason, and that companies are now asking “How should AI act?” in real workflows. (cnbc.com) ServiceNow’s new Project Arc, announced with Nvidia at the event, is described as a long-running desktop agent for developers, IT teams and administrators. Nvidia said the system can access local files, terminals and installed applications to complete complex, multistep tasks, while running inside policy-governed environments. (blogs.nvidia.com) ### Why is Huang emphasizing compute so heavily? Huang has tied higher compute needs directly to Nvidia’s core business: selling the chips and systems used to train and run AI models. In the Feb. 26 CNBC interview, he said newer reasoning systems undercut the idea that AI efficiency gains would reduce infrastructure demand, and Nvidia reported data-center revenue of $35.6 billion in that quarter. (blogs.nvidia.com) On the software side, Huang has also argued that agents will not simply replace enterprise applications. In a separate CNBC interview on Feb. 25, he said markets had “got it wrong” if they assumed agentic AI would cannibalize software vendors, because agents are “tool users” that still rely on products such as ServiceNow, SAP, Cadence and Synopsys. (cnbc.com) ### Why did the remark spread beyond chip circles? The May 5 comment circulated widely on X because it landed in two active debates at once: whether AI infrastructure spending is still accelerating, and whether more capable AI systems threaten creative and knowledge-work jobs. Social posts cited Huang’s number as evidence that demand for GPUs, accelerators and power could keep rising. Other posts framed the same remark as another sign that companies want AI systems to take on more human tasks. (cnbc.com) ### What should readers watch next? Nvidia’s next public checkpoints are its product and partner updates around enterprise agents, AI factories and data-center systems. The company and ServiceNow said on May 5 that they are expanding their collaboration on autonomous agents, secure execution software and workflow governance, and Nvidia’s official blog remains the clearest source for those rollout details. (glitchwire.com) (blogs.nvidia.com)