NVIDIA CEO Jensen Huang used a Cisco AI event to make three points that cut across enterprise strategy: traditional ROI models are a poor fit for early AI investment, the most valuable intellectual property in the AI era may be the questions a company asks rather than the answers it gets, and Physical AI could shift industry logic from software tools to digital labor.
Why Huang says AI value cannot be boxed into a spreadsheet
Speaking with Cisco CEO Chuck Robbins, Huang said companies should resist the urge to evaluate AI programs with standard ROI math at the start. His argument was simple: major technologies are hard to quantify in their early phase. He compared the moment to 1995, when no one could have used Excel to forecast how the internet would reshape retail.
Huang framed AI as an exponential change, not one that fits linear planning models. He also said the number of AI projects inside NVIDIA is “out of control,” a description he did not present as a problem but as a normal feature of experimentation. In his view, enterprises should borrow from venture capital logic: fund 10 projects, accept that 7 may fail, and recognize that a single breakout success can produce returns measured in the thousands.
He summed up the competitive risk in one line: a company does not need to be the first to use AI well, but it cannot afford to be the last.
Prompt history may reveal more than model outputs
On data sovereignty, Huang pushed a different idea about what counts as core IP. Answers generated by AI are getting cheaper and easier to reproduce, he said, and different models often produce similar responses. The scarce asset is the question itself, because prompts expose strategic intent, technical constraints, and where a company is directing resources.
He argued that if a rival obtained three months of a company’s AI prompt history, that record could be enough to reconstruct the business’s strategic map. That is why Huang said he is uncomfortable putting all NVIDIA conversations in the cloud. For sensitive use cases, he favors local deployment.
His preferred setup is a hybrid model. AI systems handling strategy, finance, and core technical matters should run on premises, while more general tasks such as translation or copy generation can be handled through public cloud services.
Physical AI is aimed at the 99% of the economy tied to atoms
Huang described Physical AI as the next major frontier. For roughly 40 years, the technology industry has focused on electrons and data, he said, while 99% of the global economy is driven by the physical world of atoms. Physical AI is meant to close that gap.
He defined the shift in unusually broad terms: for the first time, humanity is creating labor itself, not just tools. Using Tesla’s self-driving effort as an example, Huang said the key valuation idea is not only the vehicle but the “digital driver” behind it, an asset that can operate around the clock and keep generating economic value. From that perspective, robots will learn to use existing tools such as knives or brooms, and AI agents will learn to operate existing enterprise software such as SAP and Salesforce rather than requiring businesses to rebuild their full IT stack.
Three practical steps for enterprise leaders
Huang also laid out three actions for management teams. First, leaders should develop a direct feel for the technology by building a small AI system themselves and learning the cost structure of training and inference. He said that hands-on understanding matters, especially when dealing with vendors. Second, companies should run a prompt audit: log every question teams ask AI over a week, classify them into A, B, and C by sensitivity, and if A-level prompts account for more than 20%, seriously consider whether a private AI system is needed.
Third, he suggested launching an “AI in the Loop” pilot for core roles, allowing AI to observe and record decision patterns. After three months, that process could leave a company with its own institutional experience database. Huang’s point was not that every benefit will show up quickly in a finance report. It was that durable competitive advantage may start with the data a company creates while learning how to work with AI.