What to Watch at Nvidia’s 2026 Shareholders Meeting? Jensen Huang: Every Token Is Profit, AI Monetization Already Has an Answer

TradingKey - On June 24, Eastern Time, at Nvidia's ( NVDA) annual shareholder meeting, CEO Jensen Huang used a two-hour speech to set the tone for the next stage of the AI industry's development. He declared the official arrival of the "era of useful AI," stating that agents will become the core engine driving computing power demand for the next few decades, and characterized this computing paradigm shift as the largest industry reset in 60 years.

Every Token is Profit: The Commercial Closed Loop of AI Factories Has Formed

Jensen Huang repeatedly emphasized in his speech that AI has completed the transition from technological experimentation to commercial production.

Citing GitHub developer data, he noted that the number of pull requests merged by global developers was 400 million in 2024, rising to 500 million in 2025, and surging nearly threefold in just the first few months of 2026. Behind these figures is the reality that AI agents are replacing humans on a massive scale for tasks such as programming, design, and data analysis, with every generated token becoming a monetizable unit of profit.

Huang pointed out that while the core function of data centers in the past was to store and transmit files, the core mission of modern AI factories is to produce tokens—these monetizable units of intelligence that serve as the raw materials for code, answers, designs, actions, and services.

He used a "five-layer cake" structure to describe the AI industrial ecosystem, which, from bottom to top, consists of energy, chips and systems, infrastructure, models, and applications. This framework implies that Nvidia's business scope extends far beyond chips, spanning the entire AI production chain.

Under this new paradigm, customers purchasing Nvidia systems are not merely buying computing tools; they are building AI factories capable of directly generating revenue.

Huang emphasized that the efficiency of factory architecture—specifically, how many tokens can be produced per watt and how low the cost per token can be—has become the most critical dimension of competition.

For fiscal year 2026, Nvidia's full-year revenue reached $216 billion, up 65% year-on-year, with data center revenue accounting for $194 billion, representing a 68% year-on-year increase. Operating cash flow reached $103 billion, with $41 billion returned to shareholders during the year.

Huang specifically noted that while the purchase price of Nvidia systems may not be the lowest, the company can produce the most tokens at the lowest cost while achieving the highest throughput. This business model has been fully validated by the market, laying a solid foundation for Nvidia's sustained leadership in the AI era.

With the rapid adoption of large model training, inference, and AI agent applications, the strategic value of memory and storage in data centers is steadily rising. Leveraging its full-stack technological advantages, Nvidia is establishing itself as a core driver of this AI industrial revolution.

Vera Rubin Platform Launches a Full-Stack Revolution in the Era of AI Agents

Jensen Huang positioned the Vera Rubin platform as "one of the most important products in the company's history." Unlike the previous Hopper platform focused on training and the Blackwell platform geared toward inference, Vera Rubin is a complete AI factory solution designed specifically for the era of AI agents, with the Vera CPU representing Nvidia's milestone venture into the general-purpose CPU market.

Huang explained that the way AI agents work is completely different from humans; they inhabit a nanosecond-scale computational world, frequently calling tools, accessing databases, executing code, and iterating on tasks. In this scenario, if the CPU becomes a bottleneck, expensive GPUs will sit idle, and every second of idle time translates to a loss in revenue for the AI factory. To address this, Nvidia built the agent-specific Vera CPU from the ground up—no longer marketing itself on core count, it instead pursues extreme low-latency responsiveness to meet the concurrent demands of billions of AI agents worldwide.

As the first CPU to feature LPDDR5 memory, Vera is 1.8 times faster than traditional x86 CPUs, with a 50% boost in single-core performance. It also natively supports FP8 precision, allowing it to handle AI inference and reinforcement learning tasks directly without requiring GPU intermediation.

Currently, Vera Rubin has entered full volume production, with major global model developers, public clouds, AI clouds, and hyperscale customers already beginning deployment.

Huang said: "Vera Rubin is not a chip, but an AI factory platform, and the ecosystem is already in motion. Every major model developer, public cloud, AI cloud, and hyperscaler is preparing to build on it."

Nvidia's third-generation AI products have a clear division of labor. Hopper focuses on pre-training; Blackwell scales inference at the rack level; and the Vera Rubin platform, composed of the Vera CPU for scheduling and the Rubin GPU for compute, is designed specifically for the era of AI agents.

Notably, Nvidia is the only player in the industry that simultaneously possesses three high-speed networking systems—NVLink, Spectrum-X Ethernet, and InfiniBand. This provides a unique interconnect foundation for the Vera CPU, ensuring low-latency response and highly efficient coordination in agentic AI scenarios.

Physical AI: The Next Trillion-Dollar Growth Sector

Beyond d igit al ag ents, J ensen H uang has a lso set h is si ghts on p hysic al AI i n the r eal w orld. H e note d that a rtific ial in tellig ence i s bre aking o ut of t he vir tual w orld an d is b eing f ully d eploye d in t ermina ls suc h as au tonomo us dri ving, h umano id rob ots, a nd ind ustria l smar t equi pment, e quippi ng phy sical h ardwar e with t he com plete i ntelli gent c apabil ities o f perc eption, r easoni ng, pl anning, a nd act ion. T he lar ge-sc ale im plemen tation i n this f ield w ill ca talyze a new w ave of t rilli on-dol lar in frastr uctur e inve stment.

H uang l ikened A I infr astruc ture t o majo r infr astruc ture p roject s in h uman h istory , such a s the p ower g rid an d the i nterne t, bel ieving t his wi ll be a buil ding b oom sp anning d ecades . He em phasiz ed tha t Nvid ia's t echnol ogy wi ll not o nly su pport d igital a gents b ut wil l also p rovide c ore co mputin g powe r for p hysic al AI, w hich i s set t o beco me the c ompan y's ne xt maj or gro wth en gine.

NVIDIA's Shareholder Return Commitment

In the face of the stock's underperformance against the broader U.S. market this year, Jensen Huang sought to shore up market confidence by reinforcing the view that the AI capital expenditure supercycle is still in its early stages.

He reiterated that more than 50% of free cash flow will be returned to shareholders and designated several capital return policies as long-term commitments. During its most recent earnings call, Nvidia announced an authorized $80 billion share buyback program.

Looking ahead, Huang expressed strong confidence: "AI infrastructure is no longer experimental; it has entered the production phase." He predicted that computing power demand in the era of AI agents will grow exponentially, and Nvidia, leveraging its full-stack technological advantages and ecosystem moat, will continue to lead the largest infrastructure buildout in human history.