NVIDIA's Huang at G20: 1GW AI Data Centers Cost $60B, Urges National Investment

At the G20 Innovation Ministerial held in the United States on September 2, NVIDIA CEO Jensen Huang engaged in a fireside chat with U.S. Commerce Secretary Howard Lutnick, pitching his 'AI factory' concept to officials from various countries and systematically elaborating on the 'economics of compute.' He presented a striking figure: building 1GW of AI infrastructure currently costs approximately $50-60 billion.

For reference, a single state-of-the-art TSMC fab for 3nm/2nm processes typically involves an investment of around $25 billion, while a gigawatt-scale AI data center costs about twice that. Huang used this to illustrate that compute has completely moved beyond consumer electronics and general IT hardware, evolving into a core national means of production akin to power grids and railways, where equipment procurement becomes long-cycle, heavy-asset infrastructure investment.

Given the extremely high investment threshold, Huang used the 'five-layer cake' model to explain how countries should position themselves: from top to bottom, the layers are energy (power, grid infrastructure), chips (compute and interconnect hardware), infrastructure (data centers, cooling, and system networks), models (general and industry-specific foundations), and applications and data (vertical scenarios and private business flows).

He argued that a country doesn't need to win every layer; the real question is deciding which layers to participate in and which capabilities cannot be entirely left to foreign suppliers. The bottom line, he said, is that every economy must build sufficient domestic infrastructure and vertical application capabilities.

Huang assessed that AI is moving from large language models to the agentic and physical AI stages. Language models provide basic cognitive abilities, but AI systems truly used for work also require knowledge retrieval (RAG), working memory, tool invocation, and multi-agent execution capabilities. When AI can operate browsers, software, and enterprise systems, it gradually becomes a 'digital employee'; when this capability extends to cars, robotic arms, and other machines, it ultimately corresponds to autonomous driving, industrial robots, logistics robots, and even surgical robots.

He predicts that in the coming years, AI will essentially achieve AGI, taking shape at the engineering level. However, this doesn't mean enterprises can simply plug into a general model and automatically upgrade. Even if AI has strong general capabilities, it still needs to be provided with the business context, goals, knowledge, and access permissions required for the task. He emphasized: 'Every country, every company, you cannot outsource all your intelligence to others.'

Regarding employment impact, Huang believes AI replaces highly standardizable 'task units' such as information summarization and draft writing, but the responsibility for roles, the setting of ultimate intent, and business decision-making authority remain with humans. Operational tasks like writing and information processing may gradually be automated, but the core purpose and meaning of a job will not disappear with them.

On the day Huang made these remarks, NVIDIA's stock closed up about 3.2%. In the two days prior, the stock had come under pressure due to concerns about the sustainability of AI capital expenditure and circular financing. This backdrop made Huang's 'preaching' at the G20 more targeted—he sought to convince officials that AI infrastructure is a necessary heavy-asset investment, not an optional tech expenditure.

Huang also stated that from now until the end of this decade (around 2030), the scale of global AI infrastructure construction is expected to reach approximately 100GW, corresponding to a multi-trillion-dollar long-term capital expenditure cycle. This projection elevates AI infrastructure investment to a national strategic level.