What happened
Jensen Huang, CEO of NVIDIA, appeared at the G20 Innovation Ministerial Meeting for a fireside chat with US Commerce Secretary Lutnick, laying out his 'AI economics' vision. He argued that AI is evolving from a software technology into a foundational economic infrastructure comparable to electricity or the internet, and that the worst possible outcome for any nation is failing to use it and being left behind.
Huang said computing power facilities should no longer be seen as terminal products like PCs or phones, but as productive assets that continuously create economic value, with intelligence priced by tokens just as energy is priced per kilowatt-hour. He cited massive capital needs, saying a 1-gigawatt AI infrastructure build-out costs roughly $50 billion to $60 billion — about twice the price of a $25 billion wafer fab — and projected that about 100 gigawatts of AI infrastructure will be built by the end of this decade.
He also described AI's next phase as a move from large language models to AI agents and physical AI, including autonomous driving, industrial robots and surgical robots, predicting that AGI will be largely achieved in the next few years. Yet he cautioned that countries and companies cannot outsource all their intelligence to a general model, and that AI will mostly replace tasks rather than whole jobs.
Why it matters
By putting his case before G20 officials, Huang is framing AI infrastructure as a national economic policy matter, not just a corporate technology trend. His warning suggests that for governments, the real competitive risk is not the technology itself but failing to invest in and deploy it.
Huang's cost comparisons point to AI infrastructure becoming a capital-intensive, utility-like sector. The 'five-layer cake' framing — energy, chips, infrastructure, models, applications — gives countries a way to think about specialization: they do not need to lead everywhere, but they must choose layers where they can compete while retaining enough domestic AI capacity.
His emphasis on general-purpose chips also carries strategic logic: if expensive infrastructure is too specialized, rapid changes in AI models could render it obsolete, making flexibility a key consideration for any national AI plan.
Key facts
Jensen Huang spoke at the G20 Innovation Ministerial Meeting alongside US Commerce Secretary Lutnick.
Huang said building 1 GW of AI infrastructure requires around $50-60 billion, and projected roughly 100 GW of AI infrastructure by the end of the decade.
He predicted AGI will be basically achieved in the next few years and argued every country should drive AI adoption.
What to watch next
Watch whether G20 members follow up with concrete national AI infrastructure plans, budgets or cross-border cooperation initiatives.
Monitor whether AI hardware spending shifts toward flexible, general-purpose chips as Huang recommends, or toward more specialized systems as model architectures evolve.
Observe how the promised distinction between replacing tasks and replacing jobs plays out as AI agents and physical AI systems are deployed in the real economy.