en.Wedoany.com Reported - At its 2026 annual shareholder meeting, U.S. AI chip company Nvidia once again reinforced the "AI factory" narrative. On June 24 local time, Nvidia founder and CEO Jensen Huang stated at the meeting that the question of AI investment returns has already been answered: customers are not buying a bunch of servers; they are building AI factories capable of generating revenue. This round of AI infrastructure construction will be measured in decades, involving the restructuring of key infrastructure such as power grids, the internet, data centers, and agent applications.
Huang defined AI data centers as factories that "manufacture tokens." Tokens can be converted into code, answers, designs, actions, and services, meaning each token has measurable economic value. In the past, data centers were primarily used for storing, retrieving, and transmitting information; in the AI era, data centers are beginning to produce digital intelligence. For cloud vendors, model companies, and enterprise customers, the way data center value is measured is changing. It is no longer just about the number of servers, GPUs, or rack scale, but about how many low-cost tokens can be produced per unit of energy consumption and capital expenditure.
This logic directly corresponds to Nvidia's product positioning. Huang emphasized that while Nvidia systems may not be the lowest-cost option in terms of purchase price, they can produce the lowest-cost tokens, the highest token throughput, and higher revenue. The competitive focus of AI infrastructure is shifting from single-chip performance to the comprehensive output capability of the entire system, including GPUs, CPUs, networking, memory, software stacks, rack-level systems, and cluster scheduling capabilities. When customers build AI factories, they are essentially building production systems that can sustainably generate revenue from intelligent services.
Nvidia's financial performance provides data support for this narrative. In fiscal 2025, the company's full-year revenue grew 65% to $216 billion, operating income increased 60% to $130 billion, diluted earnings per share rose 67% to $4.90, operating cash flow reached $103 billion, and it returned $41 billion to shareholders. Among this, data center revenue grew 68% to $194 billion, becoming Nvidia's core revenue source. AI data centers are no longer just a growth business; they have become the mainstay of Nvidia's revenue structure.
Huang stated that AI is not just a model but a fundamental shift in computing. Over the past 60 years, computing has primarily revolved around information retrieval, storage, and transmission; now, computing is being reinvented by AI, beginning to generate intelligence. Every 10 to 15 years, the computer industry undergoes a reset—from mainframes to personal computers, from the internet to the cloud, and then to mobile cloud—and this reset is larger. With AI, computers can understand, reason, plan, use tools, and complete useful work.
This change has also reshaped the investment logic for data centers. Traditional data centers are more like "tool sheds," providing computing power and storage for software and internet services; AI data centers are more like factories, composed of numerous digital assistants and agents, continuously producing code, text, images, designs, decision-making suggestions, and automated services. As long as these outputs can be used by enterprise businesses, software systems, and end users, AI infrastructure has a closed revenue loop. Huang explained the return on investment using "tokens as profit units," aiming to respond to market doubts about high AI capital expenditure and unclear return cycles.
This also explains why Nvidia believes the AI infrastructure cycle will not end soon. Current AI demand is still expanding from training to inference, from cloud-based large models to enterprise applications, agents, robotics, autonomous driving, industrial simulation, and physical AI. The training phase requires large-scale GPU clusters, while the inference phase demands higher throughput, lower latency, and lower unit costs. As AI moves from answering questions to executing tasks, data centers need continuous expansion, and networking, power, liquid cooling, storage, and rack-level systems must be upgraded simultaneously.
Huang also mentioned changes in Nvidia's product roadmap. Hopper is more focused on pre-training, Blackwell brings inference to rack scale, and Vera Rubin is aimed at the agent era. Agent applications require models to continuously understand goals, break down tasks, call tools, and execute operations, placing higher demands on inference computing power, memory bandwidth, network communication, and system stability. If Vera Rubin becomes the core of the next-generation platform as planned, it will meet the infrastructure needs of AI transitioning from model services to agent services.
Shareholder returns were also a key focus of the meeting. Huang stated that based on confidence in sustainable market growth and free cash flow generation capabilities, Nvidia plans to return 50% or more of its free cash flow to shareholders this year, next year, and over the longer term, while continuing to increase stock buybacks and dividends over time. For the capital market, this means Nvidia is trying to prove that AI infrastructure investment is not a short-term boom but an industrial cycle that can translate into long-term cash flow and shareholder returns.
If this round of AI infrastructure construction lasts for decades, it will impact more than just chip companies. Power grids, transformers, energy storage, liquid cooling, optical modules, servers, advanced packaging, HBM memory, data center construction, industrial software, and agent applications will all be integrated into the same expansion chain. Nvidia's "AI factory" concept is redefining AI chip sales as the construction of digital intelligence production infrastructure. Future market focus will center on whether AI customers can continue to generate revenue, whether inference costs can continue to decline, and whether Nvidia's rack-level systems can maintain a unit token cost advantage.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
