Nvidia CEO Jensen Huang: AI Infrastructure is Trillions of Dollars
Nvidia CEO Jensen Huang stated that the construction of artificial intelligence infrastructure "is not hundreds of billions of dollars, but trillions of dollars," comparing it to national infrastructure at the level of the power grid, transportation, and the internet.
In discussions with U.S. officials, he emphasized the necessity of energy-friendly growth and re-industrialization of the supply chain, as the scale of this expansion has exceeded traditional tech capital expenditure levels. At Davos, he likened AI to a five-layer cake: energy, chips and computing power, cloud data centers, models, and applications, claiming this is the largest infrastructure project in human history, which has currently only cost several hundred billion dollars, with trillions more to come.
Company materials frame global AI infrastructure spending before 2030 at $3 trillion to $4 trillion. McKinsey estimates that by 2030, cumulative investment in data centers could reach $6.7 trillion. The five major cloud providers are projected to have a combined capital expenditure of nearly $800 billion in 2026, potentially rising to $1.3 trillion the following year, with cloud contract backlogs around $2 trillion. Huang noted that the cost of a single 1 GW AI factory has risen from about $20 billion to $30 billion to $50 billion to $60 billion, and is expected to reach $80 billion to $100 billion soon; there may be 100 GW factories connected to the grid within this decade.
Nvidia's revenue for the second quarter was $96.2 billion, a year-on-year increase of 106%, with data center revenue at $89 billion; the company guides for approximately $108 billion in the third quarter and provides about 70% growth visibility for fiscal year 2028. In August, it announced collaborations with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, aiming to mobilize over $500 billion in third-party funding to turn AI factories into financeable productive assets, which is not included in Nvidia's revenue. Domestic U.S. and partner plans include $500 billion-level facilities, with a commitment to OpenAI for up to $100 billion in collaborative expansion.
Power, land, cooling, optical modules, and skilled labor have become bottlenecks. Huang stated that electricians and plumbers can earn six-figure salaries, and Europe must first increase power supply to lay the AI layer. The spot market GPU prices, due to supply shortages, were described by him as "difficult to quantify." He equated computing to revenue: without computing power, there are no tokens, and without tokens, there is no cash flow at the application layer.
In market mechanisms, buyers include hyperscale clouds, sovereign computing projects, and private infrastructure funds, while sellers are GPU, power equipment, and data center contractors. The event-driven aspect comes from CEOs changing the expenditure unit from billions to trillions, with capital priced according to multi-year processing and power purchase agreements. Beneficiaries include Nvidia, TSMC's advanced processes, and power and cooling vendors; pressured parties are project returns in areas lacking electricity and buyers still valuing based on traditional server depreciation models. The core of the bubble debate is whether the application layer can convert tokens into repeatable revenue.
NVDA's stock price trades based on guidance and visibility, without a separate "trillion-dollar order" listed; points of observation include the rhythm of GW factory grid connections and whether the third-party $500 billion platform truly disburses funds.
Source: Public Information
ABAB AI Insight
Huang has rewritten Nvidia from a gaming graphics card company into an AI factory platform: packaging chips, networks, system software, and developer ecosystems for sale. The "five-layer cake" at Davos incorporates energy and land into the semiconductor narrative, shifting valuation from just GPU shipments to how much a factory is worth in GW. The $500 billion private equity platform is the next step: transforming one-time card sales into power assets that institutions can hold, similar to toll roads and power grids. The commitment of up to $100 billion to OpenAI locks the largest customer and supplier onto the same expansion table.
Money flows from cloud vendors' operating cash flow to wafers, HBM, power, and land. Funds like BlackRock and Brookfield are entering because depreciation cycles are lengthening, and contracts are starting to resemble infrastructure. The motivation is "computing power equals revenue": model companies bill in tokens, and a lack of cards leads to a lack of revenue, hence the prepayment for expansion. The strategy is that whoever secures substations and cooling capacity first will have the scarce cabinets in the next round of chip iterations. The discourse of re-industrialization in the U.S. provides political cover for this capital expenditure, while electricity prices and environmental assessments are real constraints.
Comparative objects include the overbuilding of fiber optics in 2000, the shale revolution's drilling rigs and pipelines, and capital expenditures for 5G base stations. Nvidia is transitioning from a product company to an infrastructure platform; its share remains high, but competition has shifted from "do you have GPUs" to "do you have power and shells." TSMC, power companies, and REITs have become invisible cost centers. If the application layer cannot absorb tokens, the trillion narrative will degrade into an inventory cycle.
Structural judgment belongs to capital concentration. The mechanism is: training and inference turn computing power into measurable output, private equity funds use long-term power purchase and lease contracts to securitize chip depreciation, and the state treats power supply as industrial policy. Pricing power shifts from single card gross margins to GW-level sites and power grid access; chips without power are merely inventory, and power without applications is just idle transformers.
ABAB News · Cognitive Laws
- Computing power must first become infrastructure before it can become revenue.
- What is lacking is not chip slogans, but power and land.
- The trillion narrative relies on intergenerational contracts, not on single-season shipments.