NVIDIA CEO Jensen Huang: AI is the New Electricity
NVIDIA CEO Jensen Huang compared artificial intelligence to new electricity, stating that its scale of expansion is unlike any technology in history, requiring not hundreds of billions but trillions of dollars in infrastructure investment.
He described AI infrastructure as a five-layer cake: the bottom layer is energy, followed by chips, physical infrastructure, models, and applications. The constraint is not software, but watts. He likened AI factories to generators that consume electrons and produce intelligent tokens; currently, only a few hundred billion have been invested, with trillions more needed ahead.
In a Stanford classroom, he estimated that the energy required for computing could be about a thousand times higher than current levels, acknowledging that the magnitude might be off. This is because intelligent agents will operate in parallel around the clock, and the power consumption structure differs from a question-and-answer chat. At a Morgan Stanley conference, he set "tokens per watt" as the primary revenue metric for the company, as a factory is still capped at 100 megawatts or 1 gigawatt.
In Davos, he called this the largest infrastructure wave in human history, estimating that in the next fifteen years, it could reach about $85 trillion, covering jobs such as electricians, plumbers, steel construction, network technicians, and architects. At a Goldman Sachs conference, he maintained his judgment that global data center capital expenditure would be about $3 trillion to $4 trillion before 2030, expressing more concern about land, electricity, and the shell of data centers rather than supply of components.
He estimated the cost of a 1-gigawatt AI factory to be in the range of $50 billion to $100 billion, with about 100 gigawatts expected to come online by the end of the decade. Microsoft had previously faced situations where GPUs arrived but could not be powered on; Goldman Sachs predicts that U.S. data center electricity consumption will rise from about 31 gigawatts in 2025 to about 66 gigawatts in 2027.
In market mechanisms, buyers are cloud vendors, new clouds, and sovereign computing projects, competing for accessible megawatts rather than off-the-shelf accelerator cards; sellers are the power grid, nuclear power, and natural gas peaking, liquid cooling, and substation contractors. The event-driven narrative comes from "power-capped revenue," with funding shifting from pure chip transactions to energy and park bundling. Beneficiaries are operators who control power supply and land, and NVIDIA with its premium on performance per watt, while those under pressure are the shelving plans in areas with power shortages and utilities still approving projects at the pace of the old grid.
He also told G20 ministers that AI should be treated as infrastructure like water, roads, electricity, and the internet, with countries needing to build their own to serve local economies, denying that jobs would be entirely eliminated, arguing that it is tasks that are automated, not job purposes.
Source: Public Information
ABAB AI Insight
Jensen Huang has transformed NVIDIA from a gaming graphics card company into an AI factory equipment supplier, locking in developers through the CUDA ecosystem and selling GPUs as power plant turbines via data center capital expenditure. He repeatedly uses electrification as a metaphor, not as rhetoric, but to shift the pricing unit from chip counts to gigawatts and tokens per watt, allowing customer boards to approve orders using industrial project logic.
Capital is being redirected along three lines: accelerator cards and cabinets, liquid cooling and power distribution, and land near power sources. The motivation is that after the failure of Moore's Law, performance can only be achieved by stacking watts and interconnections; whoever secures power first can convert tokens into revenue. A 1-gigawatt facility corresponds to a cost of about $50 billion to $100 billion, transforming NVIDIA's downstream from software subscriptions to heavy industrial cash flow.
This is similar to the models of Westinghouse and General Electric during the electrification era, as well as the railway era where rights of way were established before charging freight. The industry is currently in a control phase following an expansion period: the chip supply contradiction has shifted to power grid queuing, with China viewed as a competitive variable due to lower electricity prices and faster installations, while the U.S. frames the electrician shortage as a policy issue through a re-industrialization narrative.
The structural judgment pertains to the reconstruction of the supply chain: once intelligence is commoditized into continuous production, the profit pool shifts from model companies to energy and acceleration infrastructure. The mechanism is that physical limits rewrite business models—when factories are capped by power, software differences yield to power supply differences, and chip manufacturers must embed themselves in power plant planning, or else orders will remain in warehouses unable to be powered on.
ABAB News · Law of Cognition
- The profits of new electricity lie behind the meter.
- What is capped is watts, not imagination.
- Those who secure energy first can secure capacity.