Huang Renxun's Speech: Don't Use Excel to Calculate Returns; In the AI Era, "Problems" Are Worth Ten Thousand Times More Than Answers. Physical AI Will Break Barriers

NVIDIA CEO Jensen Huang presents multiple disruptive viewpoints at the Cisco AI Conference: companies should not measure AI investments with traditional ROI formulas, “questions” are more valuable than answers and are closely related to core IP security, and the rise of Physical AI will create “digital labor” for the first time, fundamentally transforming industry valuation logic. This article summarizes the discussion with NVIDIA CEO Jensen Huang at the Cisco AI Conference, compiled by Dongqu.
(Background: Summary of NVIDIA GTC 2025: Huang predicts an “AI Agent Revolution” entering consumer households)
(Additional context: Huang: I use AI this way to make myself smarter, not worried about AI replacing human jobs)

Table of Contents

• Don’t use Excel to calculate AI ROI
• Your “questions” are the most valuable IP
• Physical AI: from creating tools to creating labor
• Three practical tips for enterprises

Tags:

Recently, NVIDIA CEO Jensen Huang engaged in an in-depth discussion with Cisco CEO Chuck Robbins at the Cisco AI Conference, offering a series of thought-provoking insights on how companies can face the AI wave. From investment strategies to data sovereignty and the vision of Physical AI, Huang’s remarks touch on core issues enterprises face in the AI era.

Don’t use Excel to calculate AI ROI

Huang straightforwardly states that companies should not attempt to measure the effectiveness of AI deployment using traditional ROI formulas in the early stages. He says, “All technology deployments are hard to quantify in spreadsheets at the beginning.”

He draws an analogy to the internet of 1995, pointing out that no one could have predicted how the internet would fundamentally disrupt retail using Excel. AI brings exponential change, and framing its value with linear thinking will inevitably lead to missed opportunities.

Huang further reveals that NVIDIA’s internal AI projects are now “out of control,” but he believes this is the right way to innovate. He advocates for managing AI projects like venture capital—invest in ten projects, accept that seven will fail, and if one breaks through, the returns could be thousands of times higher.

“You don’t have to be the first company to use AI well, but you definitely don’t want to be the last.”

Your “questions” are the most valuable IP

On the topic of data sovereignty, Huang offers a thought-provoking insight: in the AI era, the most valuable intellectual property (IP) a company holds is not the answers, but the “questions.”

He explains that the cost of AI generating answers is approaching zero, and responses from different models are largely similar. However, the questions a company asks AI directly reflect its strategic thinking, technical bottlenecks, and resource allocation. If a competitor has access to a company’s AI questions from the past three months, they could piece together a complete strategic map.

Therefore, Huang emphasizes that core AI systems must be deployed locally. He recommends a hybrid cloud strategy: build AI systems involving strategy, finance, and core technology on-premises, while general functions like translation and content generation can be handled by public cloud services.

I am concerned about putting all of NVIDIA’s conversations on the cloud, which is why we choose to build locally.

Physical AI: from creating tools to creating labor

Huang looks further into the future, outlining a blueprint for Physical AI. He points out that over the past 40 years, the tech industry has focused on electrons and data, but 99% of the global economy is driven by atoms in the physical world. Physical AI will break down this barrier. Huang defines the core of this transformation:

This is the first time in human history that we are creating labor itself, not just tools.

Using Tesla’s self-driving cars as an example, he highlights that their valuation hinges not on the vehicle itself but on the “digital driver”—a digital asset capable of operating 24/7 and continuously generating economic value. Similarly, Physical AI robots will learn to operate existing tools (like knives and brooms), and AI Agents will learn to control existing enterprise software (like SAP and Salesforce), rather than building entire IT systems from scratch.

Three practical tips for enterprises

Huang also provides concrete action guidelines for business leaders:

• Build a “tactile” understanding of technology: Hands-on experience by assembling small AI systems to understand training and inference costs. “For God’s sake, assemble a computer yourself,” to avoid being misled during vendor negotiations.
• Conduct a question audit: Record all questions your team asks AI over a week, categorize them by sensitivity into A (strategic core), B (sensitive but controllable), and C (general information). If A-type questions account for more than 20%, carefully evaluate the need for private AI deployment.
• Launch an “AI in the Loop” pilot: Select core roles where AI observes and records decision-making patterns. After three months, develop a company-specific experience database to build a future competitive moat.