NextFin News -
NextFin News - Jensen Huang, founder and CEO of NVIDIA, sat down with the All‑In Podcast hosts in a wide‑ranging interview recorded at NVIDIA’s GTC conference in San Jose, California. The conversation, conducted on March 19, 2026, brought together Huang and the All‑In hosts (Chamath Palihapitiya, Jason Calacanis, David Sacks and David Friedberg) to discuss NVIDIA’s strategy, the changing economics of inference, agentic software, physical AI, open source models and the geopolitical supply‑chain context for building AI infrastructure.
The discussion centered on NVIDIA’s framing of the next industrial revolution as the age of the AI factory — an integrated vision of chips, networking, storage and software — and on how agentic systems and disaggregated inference will reshape compute, products and labor.
Dynamo and the AI factory
Huang opened by describing an operating‑system view of AI infrastructure. He said he introduced the concept two and a half years earlier and named it "Dynamo," likening it to the machine that powered the previous industrial revolution. In his words, Dynamo is "the operating system of the AI factory." He explained that Dynamo’s core idea is disaggregated inference: splitting the inference pipeline so different parts run on different processors. "The fundamental technology is disaggregated inference," he said, and this led NVIDIA to expand beyond GPUs to a heterogeneous stack that includes CPUs, networking processors, storage processors and purpose‑built inference chips.
"We just really evolved from a GPU company to an AI factory company."
Disaggregated inference, Vera Rubin and Grok
Huang described how disaggregation changes data‑center architecture and gave Vera Rubin as an example of infrastructure designed for a diverse workload mix. He explained that modern agentic workloads combine many model types — large models, smaller models, diffusion models and autoregressive models — and that these workloads "beat up on storage" and require specialized components. NVIDIA added racks and new components, he said, because the company realized inference needs an entirely different system balance. Huang said the addition of Groq processors and BlueField storage processors expands NVIDIA’s total addressable market for AI infrastructure.
"We created Vera Rubin to be able to run this extraordinarily diverse workload."
The inference explosion and the economics of the factory
When asked about the claim that inference would explode, Huang reiterated that the industry has moved from training‑centric economics to an inference‑constrained world. He argued that a larger upfront investment in a full AI factory can deliver the lowest cost per token because of throughput and efficiency gains. "You should not equate the price of the factory and the price of the tokens," he said, and stated that a more expensive factory can still generate the lowest token cost through much higher throughput.
"I can prove it that the $50 billion factory will generate for you the lowest cost tokens."
Agents, OpenClaw and the new computing model
Huang placed agentic systems at the center of the next phase of AI. He described agents as software with memory, skills, scheduling and IO — the four elements he said "fundamentally define a computer." He pointed to emerging frameworks (referred to in the interview as OpenClaw) and argued that agentic software effectively creates a new personal AI computer that is open source and runs everywhere.
"What do we have? We have a personal artificial intelligence computer for the very first time."
He also emphasized governance and security for agentic software because agents can access sensitive information and execute code externally; NVIDIA is working on governance layers to secure that functionality.
Open source and proprietary models: both
On the debate between open and closed models, Huang was explicit: the industry needs both. He said open models enable domain specialization and participation, while world‑class proprietary models will remain compelling products and services for most consumers. "Model is a technology, not a product," he said, and added that many enterprises will start open‑source‑first then add proprietary layers as they specialize their offerings.
"These two things are not A or B. It's A and B."
Physical AI, robotics and three computers
Huang framed "physical AI" as a major market: three kinds of computing systems will be necessary at scale — (1) training computers to create models, (2) evaluation/simulation computers (the Omniverse or virtual gym that obeys physics), and (3) edge/robotics computers that run in cars, robots or tiny devices. He called physical AI a ten‑year journey that is already inflecting into a multibillion‑dollar business and projected rapid growth for robotics, autonomy and applications across agriculture, factories, warehouses and telecom base stations.
"Physical AI as a large category... it's technology industry's first opportunity to address a $50 trillion industry."
On timelines, he predicted robots moving from high‑function existence proofs to reasonable products in roughly three to five years and stressed the role of global supply chains — particularly Chinese strengths in motors, magnets and microelectronics — in enabling robotics.
Healthcare and digital biology
Huang said digital biology is approaching a "ChatGPT moment" for biology: better representations of genes, proteins and cells will enable rapid progress in drug discovery and healthcare. He expected inflection in the next few years and highlighted three healthcare uses of AI in his view: AI that models biological dynamics, agentic assistants that help diagnosis and workflows, and physical‑AI tools such as robotic surgery and AI‑enabled instruments.
"We are literally near the ChatGPT moment of digital biology."
Desktop, hobbyists and the agent at the edge
Huang celebrated hobbyists and desktop compute, noting powerful local workstations and the cultural importance of open agent frameworks. He described a future where personal agents run locally and everywhere, enabling creativity and individual productivity, and said developers will move from coding line‑by‑line to writing specifications, architectures and ideas that agents execute.
"In the past, we code. In the future, we're going to write ideas, architectures, specifications."
Tokens, engineers and productivity
One of the interview’s most reported‑on lines concerned token consumption inside engineering organizations. Huang used a thought experiment to argue that top engineers should be heavy consumers of inference tokens because these tokens amplify productivity. He framed token budgets as part of modern knowledge‑work operations and emphasized that providing engineers with powerful agentic tools is akin to investing in their “health” or productivity.
"If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed."
Policy, safety and industry communications
Huang urged engagement with policymakers but warned against alarmism. He said AI is software, not consciousness, and policymakers should be informed about what the technology is and is not. He argued that excessive fear could slow national diffusion and harm U.S. competitiveness, while balanced, informed policy is needed to manage risks.
"It is not a biological being. It is not alien. It is not conscious. It is computer software."
Geopolitics, supply chain and sovereign infrastructure
Huang addressed supply‑chain risks and national security, calling for re‑industrialization, diversification of manufacturing and strategic partnerships with Taiwan, South Korea, Japan and Europe. He emphasized that losing control of key components — rare earths, magnets, motors or telecom infrastructure — would diminish national security and urged measured engagement rather than policy paralysis.
Automotive strategy and autonomy
On autonomy, Huang said NVIDIA aims to enable car companies rather than build cars. He explained that NVIDIA provides training, simulation (Omniverse) and in‑vehicle compute and that the company’s approach is to offer the full stack or specific components as partners prefer. He likened the future to an "Android‑style" openness among multiple carmakers, while other vertically integrated players may choose different paths.
"We don't want to build self‑driving cars, but we want to enable every car company in the world to build self‑driving cars."
Robotics, work and economic mobility
Huang predicted broad societal benefits from robotics: he said robots will unlock new forms of economic mobility, augment human productivity and address labor shortages. He argued robots in factories, homes and remote environments will become pervasive and could enable new small‑business opportunities and virtual presence experiences.
Closing thoughts
Throughout the interview Huang repeatedly framed NVIDIA’s role as that of an AI infrastructure and platform company building a full stack from chips to systems to software. He emphasized engineering difficulty as the core of strategic choice — "if it's not insanely hard, back away" — and urged entrepreneurs to specialize deeply in vertical domains, then connect agents to customers. He ended by calling for balanced public discussion about AI's promise and risks, and for industry leaders to be circumspect in public predictions.
References
Interview recorded at NVIDIA GTC, March 16–19, 2026; conversation recorded March 19, 2026. Further background and conference coverage:
- NVIDIA GTC 2026 (official event page)
- Tom's Hardware — Coverage of Huang’s All‑In appearance and token remarks
- NVIDIA press release: GTC 2026 dates and keynote details
- All‑In Podcast (episode archive)
Explore more exclusive insights at nextfin.ai.
