Beyond Storage: NVIDIA Rubin's Broader Supply Chain Alpha

This announcement marks a pivotal shift in the AI investment narrative. The thesis has evolved from simply "selling shovels" to a broader requirement for upgrading the entire "mine." The Rubin platform is not just a processor; it is a complex system that necessitates a massive overhaul of supporting infrastructure. Consequently, investors should now broaden their focus to the secondary beneficiaries: Storage, Optical Communication, Liquid Cooling, and Power Infrastructure.

The Rubin computing platform comprises six core components designed to work in unison:

Vera CPU: Features an 88-core custom Arm kernel. It delivers 2x the performance-per-watt of the industry's best CPUs, equipped with 1.8TB NVLink C2C, 1.5TB system memory, and 1.2TB LPDDR5X, integrating 227 billion transistors.

Rubin GPU: Utilizes NVFP4 technology to "slim down" AI models by 75%, boosting inference speed by 4x with no significant loss in precision. It boasts 50 PFLOPS of single-GPU inference compute, 35 PFLOPS of training compute, and an NVLink bandwidth of 3.6TB/s.

CX-9 SuperNIC: An 800G Ethernet solution with 200G SerDes, programmable RDMA, and data path accelerators. It integrates 23 billion transistors to manage East-West data center traffic.

BlueField-4: Combines 64 Grace CPU cores with the CX-9, integrating 126 billion transistors to handle North-South traffic (storage connectivity).

NVLink 6 Switch: Supports 3.6TB/s bandwidth per GPU and 400G SerDes. Crucially, it is designed for 100% liquid cooling.

Spectrum-X CPO: Features 102.4T extended bandwidth, 200G silicon photonics channels, and 128x800G ports, integrating 352 billion transistors.

In his keynote, Jensen Huang emphasized the critical role of memory and storage in AI systems, describing a demand curve for a market that effectively never existed before. This segment is poised to become the largest storage market globally, carrying the operational memory load for global artificial intelligence.

Amidst growing demand for AI training and inference, supply constraints and surging memory prices are already lifting the sector. Huang’s comments at CES signal that demand for NAND chips across NVIDIA systems will remain robust.

Market analysis suggests that based on the Rubin specs, each GPU will require an additional 16TB of NAND, translating to 1,152TB of NAND per rack. If we project a build-out of 100,000 racks, this creates an incremental demand of 115.2 Exabytes (EB) of NAND. This figure represents approximately 12% of the total projected global NAND supply for 2025—a massive shock to the supply chain.

The Rubin NVL72 rack achieves 100% liquid cooling and supports inlet water temperatures of up to 45°C (113°F). This is a game-changer: it means data centers no longer require high-energy chillers to refrigerate water. Instead, they can rely on natural cooling or warm water loops to dissipate heat. Huang stated this shift could save global data centers 6% in total electricity consumption.

The Core Shift:

The market has momentarily misunderstood this as "no cooling needed." The reality is distinct: we are moving away from compressors and chillers (air conditioning) toward direct-to-chip liquid cooling (DLC) components.

Implications:

This technical pivot compresses the Total Addressable Market (TAM) for traditional building-level HVAC and centrifugal chiller manufacturers in the AI data center space. Conversely, it significantly expands the market for CDUs (Coolant Distribution Units) and manifolds.

NVIDIA officially launched the Spectrum-6 Ethernet switch (SN688/SN6810), which utilizes Co-Packaged Optics (CPO) technology. Jensen Huang was explicit in his comparison: "Compared to hardware without silicon photonics technology, they perform better in energy efficiency, reliability, and uptime."

This statement marks the substantial beginning of the transition from pluggable optical modules to CPO. As highlighted in our report From GPUs to Racks: Why Optics is the Next AI Gold Rush, 2026 is shaping up to be the breakout year for the optical supply chain.

The Rubin roadmap openly discusses racks operating at the 600kW level. While the energy consumption per token is decreasing, the aggregate power density per rack is skyrocketing. This places extreme demands on power stability and high-voltage distribution.

This necessitates a "Power Revolution." NVIDIA has previously identified over 20 partners in this ecosystem, spanning GaN/SiC power chips, components, and full-system solutions.

Key Watchlist:

With the Rubin platform, NVIDIA is moving beyond the chip. By extending its influence from the silicon level to the system level, and from the cloud to the physical infrastructure of data centers, the company is no longer just selling a processor. It is architecting the foundational substrate of an AI-driven world. For investors, the alpha now lies in the physical hardware required to power, cool, and connect this new reality.