Jensen Huang stated that customers purchase NVIDIA systems not as computing tools, but to build AI factories that directly generate revenue. In benchmark tests, the Blackwell platform has been recognized as the 'king of inference,' delivering token throughput 30 times higher than the next-best platform. Vera Rubin is one of NVIDIA’s most significant product launches in its history, with the Vera CPU specifically designed for AI agents. The CUDA X library ecosystem is NVIDIA’s 'crown jewel.' Although the U.S. has approved exports of the H200 chip, NVIDIA has not yet generated related revenue from China. He issued a rare public warning about chip smuggling, calling the practice of assembling data centers from smuggled chips 'a dead end.'
At the annual shareholder meeting held on Wednesday, the 24th, Eastern Time, $NVIDIA (NVDA.US)$ CEO Jensen Huang outlined an ambitious blueprint for expanding AI infrastructure. He declared that the era of agents has officially arrived and characterized this shift in computing paradigms as the largest industry reset in 60 years. His remarks conveyed several pivotal signals distinct from previous communications—from a new CPU architecture specifically designed for agents to warnings about large-scale chip smuggling.
The commercial deployment of agent-based AI is becoming the core driver of NVIDIA’s next growth phase. Huang noted that AI has crossed the threshold from being merely 'interesting' to genuinely 'useful'—and once AI begins generating tangible economic value, demand for compute accelerates rapidly. Citing software development as an example, he pointed out that the number of pull requests merged by GitHub developers rose from 400 million in 2024 to 500 million in 2025, and in the first few months of 2026, that figure nearly tripled.
Fueled by the wave of AI agents, NVIDIA’s international revenue has surpassed $30 billion, more than tripling year-over-year, with nearly 40 countries now deploying AI Factories powered by NVIDIA infrastructure. Huang reaffirmed the company’s commitment to returning over 50% of its free cash flow to shareholders and announced that the capital return policies unveiled at this shareholder meeting constitute long-term commitments.
Buoyed by persistently strong AI demand, NVIDIA’s stock surged by over 240% in 2024 and more than 50% in 2025. However, as of Wednesday’s market close, it has gained only 6.7% year-to-date, underperforming the broader U.S. equity market—the S&P 500 has risen nearly 7.5% over the same period. This performance reflects a market shift in focus from near-term earnings to questions about the sustainability of the AI infrastructure investment cycle.
Although he did not state it explicitly, Huang’s remarks at the shareholder meeting were clearly intended to reinforce the view that the AI capital expenditure supercycle remains in its early stages.
AI Factories: The Definition of Data Centers Is Being Rewritten
In his speech, Huang asserted that the era of traditional data centers—primarily focused on storing and transmitting files—is over. The core function of the new AI Factories is to produce tokens. Each token represents a monetizable unit of intelligence and serves as the raw material for code, answers, designs, actions, and services.
Huang described the AI industry ecosystem using a 'five-layer cake' model, progressing from bottom to top as follows: energy, chips and systems, infrastructure, models, and applications. This framework underscores that NVIDIA’s business extends far beyond chips alone—it spans the entire AI production chain.
Huang emphasized that customers purchasing NVIDIA systems are not merely acquiring computing tools—they are building AI Factories capable of directly generating revenue. Under this paradigm, the critical competitive metric becomes factory efficiency: how many tokens can be produced per watt of power and how low the cost per token can be driven.
Citing results from the Semi Analysis Inference X benchmark, Huang stated that the Blackwell platform has been recognized as the 'King of Inference,' delivering 30 times higher token throughput than the next-best platform.
Companies such as Capital One, Hyundai Motor, Jane Street, and Eli Lilly and Co. have already deployed NVIDIA infrastructure at scale, signaling that the customer base for AI factories has expanded beyond hyperscale cloud providers to traditional sectors including finance, manufacturing, and pharmaceuticals.
Vera Rubin: A Historic Bet on CPUs Designed Exclusively for AI Agents
Jensen Huang’s most groundbreaking statement at this year’s shareholder meeting concerned the positioning of the Vera Rubin platform—he described it as “one of the most important product launches in NVIDIA’s history” and defined the Vera CPU as a strategic gateway to entirely new markets.
The core rationale is this: unlike humans, AI agents operate in a nanosecond-scale computational world. When agents invoke tools, access databases, execute code, and iteratively refine tasks, any CPU bottleneck leaves expensive GPUs idle—each second of idleness directly translates into lost revenue for AI factories. To address this, NVIDIA has built the Vera CPU from the ground up specifically for AI agents.
Huang noted that all prior CPUs were designed for human users and rented out in core-based slices, whereas AI agents have no inherent need for large numbers of cores but demand ultra-low-latency responsiveness. With the global population of AI agents projected to reach billions, an entirely new CPU market is now emerging. Vera Rubin has already entered full-scale production, and leading model builders, public cloud providers, AI cloud operators, and hyperscale customers are actively planning deployments, with orders continuing to pour in.
At the architectural level, Huang described Vera Rubin not as a single chip but as a complete AI factory platform, integrating Rubin GPUs, Vera CPUs, NVLink, Spectrum-X Ethernet, InfiniBand, and BlueField secure storage modules. He emphasized that NVIDIA is the only company globally with three distinct networking businesses, and that Spectrum-X already exceeds the combined scale of all other Ethernet networking competitors.
CUDA was referred to by Jensen Huang as NVIDIA’s “crown jewel.”
Jensen Huang described the CUDA X library ecosystem as NVIDIA’s “crown jewel,” calling it one of the company’s most significant strategic investments to date and an insurmountable moat against competitors.
He outlined NVIDIA’s unique flywheel effect: a unified architecture amasses a vast installed base, which attracts developers; those developers create breakthrough applications; those applications open new markets; and those new markets further expand the installed base—the entire cycle is now accelerating. CUDA currently supports over 7,000 applications, significantly enhancing commercialization opportunities for cloud customers.
With the advent of the agent era, the CUDA X libraries are evolving from a toolkit for human developers into a toolbox for AI agents. During the shareholder meeting, Huang announced the launch of BioNeMo—a suite of digital biology and drug discovery tools purpose-built for AI agents—marking NVIDIA’s strategic shift of its most critical software assets toward agent-centric use cases.
He also articulated NVIDIA's differentiated positioning: 'vertically integrated, horizontally open'—developing a fully in-house stack to achieve end-to-end optimization, then opening it up to the broader industry to support a more expansive ecosystem.
China Market: H200 Export Approved, Revenue Still Zero
On matters concerning its China-related business, Jensen Huang adopted a notably cautious tone. He disclosed that the U.S. government has granted NVIDIA a license to export H200 chips to Chinese customers, but the company has not yet generated any revenue from these sales, and uncertainty remains regarding whether the products can be successfully imported into China.
This statement stands in stark contrast to prior market expectations regarding the timeline for revenue recognition from the China market. Huang offered no projections on the potential scale of NVIDIA’s China business and did not name any specific customers.
On policy matters, Huang reiterated NVIDIA’s commitment to actively engaging in discussions around export controls, arguing that real-time industry technical insights help policymakers craft 'robust and balanced' rules—rules that protect sensitive technologies while preserving U.S. leadership in AI. He emphasized that when commercial interests conflict with U.S. national security, national security takes precedence.
Jensen Huang Rarely Issues Public Warning on Chip Smuggling
In response to a shareholder question, Jensen Huang provided his most direct public comment to date on the risk of NVIDIA products reaching restricted end users, and unusually disclosed specific countermeasures the company has implemented.
Huang stated that NVIDIA’s compliance efforts have repeatedly intercepted attempted smuggling cases, and those involved face legal prosecution in multiple jurisdictions.
Huang stressed that assembling data centers using smuggled chips is 'a dead end.' Advanced AI data centers are highly integrated systems requiring trusted hardware, software, networking, and ongoing technical support. NVIDIA provides no support or repair services for restricted products, making it 'extremely difficult' to build viable AI infrastructure using smuggled components.
According to media reports, chip smuggling has drawn sustained attention from U.S. lawmakers and regulators. NVIDIA’s decision to address this issue directly at its shareholder meeting reflects the growing importance of proactively managing this matter as a reputational and compliance risk.
Editor/Lambor