NVIDIA Vera Rubin Mass Production Underway: How Should the Investment Logic for AI Infrastructure Be Reconstructed?

In June 2026, NVIDIA entered a critical time window.

At the beginning of the month, at the Taipei International Computer Show, Jensen Huang announced that the Vera Rubin platform has entered full-scale production, and the new generation AI factory engine has officially launched. By the end of the month, at Automate 2026 in Chicago, NVIDIA Halos for Robotics was unveiled as the industry's first full-stack robot safety system, transferring over 18,600 engineering man-years of safety experience in autonomous driving to the physical AI track. On June 24, NVIDIA's 2026 Annual Shareholders Meeting is set to take place, with the ramp-up of Blackwell and Vera capacity and progress in AI ecosystem commercialization as key topics.

From Grace Blackwell to Vera Rubin, and to the layout of robot safety systems, NVIDIA is building a complete hardware universe covering data centers, AI factories, and the physical world. This article will analyze the latest developments of this AI giant, valued at about $5 trillion, from three dimensions: product technological evolution, market dynamics, and investment logic.

Vera Rubin Full-Scale Production: The Moment of Landing for the Third-Generation Rack-Scale System

On June 1, 2026, NVIDIA officially announced that the Vera Rubin platform has entered full-scale production. This is not a routine product iteration but the most strategically significant platform upgrade since Grace Blackwell.

Vera Rubin is NVIDIA’s largest POD-level platform to date—comprising five dedicated cabinets forming a massive AI supercomputer, designed for intelligent workload processing. The platform integrates NVIDIA Vera Rubin NVL72 systems, Vera CPUs, Groq 3 LPX, BlueField-4 STX storage, and Spectrum-6 SPX Ethernet rack into a fully integrated system. Compared to the previous generation Grace Blackwell platform, Vera Rubin offers a tenfold increase in agent throughput under large-scale deployment.

In his keynote at GTC Taipei 2026, Jensen Huang defined Vera Rubin’s positioning: “Agent-based AI is a new workload. A prompt can trigger a processing flow that includes reasoning, information retrieval, tool invocation, and response generation, spanning thousands of steps. Vera Rubin is born for this—it is an AI factory engine capable of delivering intelligence at scale, with the performance, efficiency, and security needed to drive the next industrial revolution.”

From a supply chain perspective, Vera Rubin’s production scale significantly surpasses that of its predecessor. NVIDIA’s supply chain ecosystem spans 30 countries and over 350 factories worldwide, with more than 150 partners in Taiwan alone. Huang stated that Vera Rubin’s supply chain scale is twice that of Grace Blackwell. Major system manufacturers have fully committed to Vera Rubin production, including Dell Technologies, Huiyu Technology, Lenovo Group, and Supermicro. The first batch of products is expected to begin shipping to cloud service providers and enterprise customers in fall 2026.

Technologically, Vera Rubin introduces several key innovations. Spectrum-X Ethernet silicon photonics technology has achieved full-scale production—integrating photonic packaging with Spectrum-X switches to enable AI factories with millions of GPUs. Vera CPU uses NVIDIA’s self-developed Olympus cores and a scalable, consistent architecture, with official claims that its agent sandbox performance is 1.8 times that of x86 CPUs. In terms of memory, Vera Rubin employs high-bandwidth memory (HBM4) from Micron, SK Hynix, and Samsung.

Notably, Huang Huang positions the Vera CPU as “a CPU born for agents,” rather than a traditional compute chip aimed at human-driven workloads. At Computex, he stated that the Vera CPU “will be more popular than GPUs” and will become NVIDIA’s “new main growth driver.” The rationale is that agent workloads require low latency, high single-thread performance, high bandwidth, and strong energy efficiency—roles in coordinating tool calls, memory access, and GPU workflows where CPUs are indispensable.

From Data Center to Physical World: The Full-Stack Security Logic of Halos

If Vera Rubin addresses “how AI factories scale up intelligent production,” then Halos for Robotics answers “how AI safely enters the physical world.”

On June 22, NVIDIA released Halos for Robotics at Automate 2026 in Chicago, the industry’s first full-stack, comprehensive robot and physical AI safety system. This system extends NVIDIA Halos’ validated safety architecture from autonomous driving into robotics and physical AI scenarios, providing a unified safety framework for perception, decision-making, and real-world action.

Halos for Robotics is built upon NVIDIA’s safety development experience accumulated over 18,600 engineering man-years in autonomous driving and 7 million lines of verified code. It covers the entire stack from chips, sensors, operating systems, to safety certification.

Architecturally, Halos constructs a four-layer safety system:

Platform Security Layer addresses hardware reliability issues. NVIDIA IGX Thor, an AI computing platform for robots and industrial scenarios, features an independent “security island”—with its own processor, I/O, power supply, and clock—physically isolated from the main computing system. Even if the main AI system crashes or malfunctions, the security island can independently perform critical functions like emergency braking. The same layer’s Holoscan Sensor Bridge solves sensor heterogeneity latency issues by unifying all sensor data into the secure computing domain for low-latency synchronized processing.

Secure Operating System Layer ensures system stability. Halos OS runs on IGX Thor, supporting pure Linux or Linux+QNX hybrid architectures. In hybrid mode, NVIDIA uses a hypervisor to split the system into two isolated domains: Linux handles AI computing and applications, while QNX manages safety-critical tasks, operating in complete isolation.

Algorithm Security Layer introduces external perception mechanisms. Outside-In Safety Blueprint employs external cameras installed on ceilings or other positions, monitored by independent AI from third-party perspectives. This capability is open to developers and provided as open source.

Ecosystem Security Layer addresses certification and standardization. NVIDIA Halos AI Systems Inspection Lab is the world’s first AI safety and functionality certification project recognized by the ANSI National Accreditation Board, helping partners prepare for third-party certifications from TÜV Rheinland, UL, and other leading agencies.

In practical deployment, humanoid robot company Agility has integrated Halos into its Digit robot, deploying it in factories for clients like Amazon, GXO, and Toyota. The Halos ecosystem has expanded to over 43 partners, including Boston Dynamics and Hesai Technology.

Some industry observers compare this strategy to “Android in embodied intelligence”—NVIDIA does not manufacture robots directly but opens its safety platform to everyone. This aligns with NVIDIA’s positioning in the AI factory era: providing foundational infrastructure capabilities rather than occupying application layers.

SMCI Blueprint Implementation: Industry Chain Mapping of the Vera Rubin Ecosystem

Vera Rubin’s mass production is not just a product event but an industry chain event.

On June 22, Supermicro announced at ISC 2026 a blueprint for data center modular solutions based on NVIDIA Vera Rubin NVL4 platform. This blueprint offers end-to-end HPC and AI infrastructure solutions, with a scalable unit containing up to 1,152 NVIDIA Rubin GPUs and 576 NVIDIA Vera CPUs, using liquid-cooled racks, with a single unit power capacity of up to 3.2MW. Supermicro CEO Liang Jianhou stated, “With our DCBBS blueprint, research institutions can confidently deploy HPC and AI infrastructure at any scale.”

Market reactions were swift and direct. On June 22 (Monday), during U.S. stock trading, SMCI’s stock surged 15.66% to close at $35.46, with intraday gains reaching 19%. Trading volume hit 128 million shares. On the same day, NVIDIA closed at $208.65, down 0.97%; the Nasdaq index fell 1.32% to 26,166.60 points.

SMCI’s strong performance reflects a structural demand for AI infrastructure hardware. Amid broader Nasdaq weakness, hardware suppliers directly related to Vera Rubin received significant valuation premiums. Analysts raised SMCI’s target price to $48. This signals a market revaluation of system integrators within the Vera Rubin ecosystem—reconsidering the value distribution of hardware in the AI investment cycle.

Shareholders’ Meeting Preview: Blackwell, Vera, and Trillion-Dollar Revenue Expectations

At 00:00 on June 25, Beijing time (9:00 AM Pacific Time, June 24), NVIDIA’s 2026 Annual Shareholders Meeting will be held online. Core topics include: capacity ramp-up of Blackwell and new Vera architecture chips, progress in AI ecosystem commercialization, and capital return plans driven by massive cash flow.

Reflecting on the 2025 shareholders meeting, several key messages were conveyed: NVIDIA is entering a “ten-year AI infrastructure buildout cycle”; AI and robotics are the two major growth opportunities; the era of robotics and autonomous driving has arrived. On that day, NVIDIA’s stock rose 4.3%, reaching a record high of $154.31.

In terms of product cadence, NVIDIA previously announced launching a new generation of AI chips annually: 2024’s Blackwell architecture, 2025’s Blackwell Ultra, and 2026’s Vera CPU combined with Rubin GPU. The Blackwell series, as the flagship for 2024-2025, remains in high demand. NVIDIA’s Q1 FY2026 (ending April 2026) data shows data center revenue of $75.2 billion, up 92% year-over-year and 21% sequentially, mainly driven by widespread adoption of Blackwell 300 products.

Huang Huang predicted at GTC Developer Conference that just Blackwell and Rubin products will generate a combined $1 trillion in revenue in 2026 and 2027. This magnitude reflects NVIDIA’s ongoing confidence in the AI infrastructure investment cycle. Whether the upcoming shareholders’ meeting will update this revenue guidance, and whether Vera Rubin’s full-scale production will influence Blackwell’s capacity allocation, will be key market focus.

From a valuation perspective, NVIDIA’s current market cap is about $5 trillion, implying a forward P/E ratio of approximately 23 based on 2026 earnings estimates. Given that AI infrastructure capital expenditure cycles are still expanding, whether this valuation is justified depends on Vera Rubin’s ability to deliver incremental revenue as scheduled and whether capital spending on AI factories can be sustained.

The Structural Logic of AI Infrastructure Investment

The mass production of Vera Rubin and the release of Halos collectively point to a broader conclusion: AI infrastructure investment is shifting from “model training” to “large-scale deployment.”

In 2026, capital expenditure on AI infrastructure faces three core bottlenecks: power, memory, and optical bandwidth. Vera Rubin’s focus on energy efficiency, HBM4 memory integration, and Spectrum-X silicon photonics addresses solutions in these three dimensions. SMCI’s liquid cooling solutions and NVIDIA’s increased supply chain investments aim to lower deployment barriers and operational costs of AI factories.

Huang Huang’s statement at GTC Taiwan offers a key insight: “Compute power equals revenue, compute power equals profit.” Performance per watt, reliability, deployment speed, and system lifespan are becoming core economic metrics for AI infrastructure operators. If this logic holds, the value of AI hardware suppliers depends not only on peak compute but also on their ability to reduce total cost of ownership at the system level.

Within this framework, Vera Rubin’s tenfold agent throughput improvement, Halos’ safety architecture standardization, and SMCI’s end-to-end deployment solutions form a complete value chain from chips to systems. NVIDIA is transforming from a GPU manufacturer into an AI infrastructure provider—aiming to become the core supplier of over 100 GW of new AI factory capacity globally by 2030.

Conclusion

In June 2026, NVIDIA advances three concurrent narratives: Vera Rubin’s full-scale production elevates AI factory scalability; Halos for Robotics debuts, extending safety architecture from autonomous driving to physical AI; and the upcoming shareholders’ meeting will scrutinize capacity ramp-up and revenue expectations for Blackwell and Vera.

From Blackwell to Vera Rubin, and to robot safety systems, NVIDIA’s “full universe” is not a closed hardware ecosystem but a full-stack infrastructure system spanning data center compute to physical deployment. The commercial value of this system depends on the actual pace of AI’s evolution from “dialogue-based” to “agent-based,” and on the capital expenditure rhythm of AI factories expanding from gigawatt to hundred-gigawatt scales.

For observers focused on the investment logic of AI infrastructure, Vera Rubin’s production pace, Halos’ ecosystem expansion, and the capacity and revenue signals from the shareholders’ meeting will be key indicators for assessing this cycle’s position.