Nvidia's Networking Surge: The 90% Attach Rate and What It Means for AI Infrastructure

The market is now paying close attention to a new, high-velocity growth driver for Nvidia: its networking business. This isn't just a side project; it's a category that has gone viral in terms of revenue growth and market positioning. In the third quarter of fiscal 2026, networking revenue surged to $8.2 billion, a staggering 162% year-over-year increase. That kind of explosive growth is the kind of headline that captures investor search volume and fuels the next round of capital flows.

The real story behind the numbers is the near-universal adoption rate. CFO Colette Kress revealed that the networking attach rate is now close to 90%. In other words, for nearly every AI system customer buying Nvidia's core compute, they are also purchasing its networking gear. This isn't a niche add-on; it's becoming the default infrastructure for the AI build-out. The market attention here is clear: investors are asking if this is the next major revenue stream that could outpace even the GPU business.

This rapid ascent has cemented Nvidia's position as a dominant force. Kress stated that demand for its Ethernet networking technologies is now "roughly on par with InfiniBand", a protocol Nvidia has long mastered. More importantly, she declared that its networking business, purpose-built for AI, is "now the largest in the world". This isn't just about selling switches; it's about controlling the critical data pathways in massive AI data centers. The search volume around "Nvidia networking" and "AI data center infrastructure" is spiking, signaling that this is the trending topic for the AI infrastructure cycle.

The 90% Attach Rate: A Sticky Growth Advantage

The nearly universal adoption of Nvidia's networking products is more than just a sales statistic; it's a powerful business moat. With a 90% attach rate, the company is locking customers into a complete AI infrastructure stack from the very start. This dramatically boosts the average revenue per customer, turning a single GPU sale into a multi-product deployment. More importantly, it creates a high switching cost. Once a data center is built around Nvidia's networking fabric, the complexity and risk of migrating to a different vendor's ecosystem become prohibitive. This stickiness enhances customer loyalty and provides a predictable, recurring revenue stream.

The reason for this deep integration is technical necessity. Building the AI factories of the future requires connecting tens of thousands of GPUs in massive clusters. For these systems to work efficiently, data must flow between chips at blistering speeds with minimal delay. Standard networking simply can't keep up. As Nvidia's CFO noted, the data throughput between GPUs must be high enough to prevent them from idling, making ultra-fast, low-latency networking a critical bottleneck. Nvidia's solutions-spanning NVLink, InfiniBand switches, and Spectrum-X Ethernet-are engineered to solve this exact problem, making them a foundational layer, not a peripheral add-on.

This growth is being fueled by the largest AI builders in the world. Companies like Meta, Microsoft, Oracle, and xAI are currently building massive AI data centers that are specifically deploying Nvidia's Spectrum-X Ethernet switches. Their scale and urgency create a powerful demand signal. The company's recent Rubin platform, which bundles GPUs, CPUs, and networking into rack-scale systems, is designed to accelerate this trend. The market is paying attention: the AI networking market itself is forecast to explode from $14.9 billion in 2025 to $46.8 billion by 2029. For Nvidia, the 90% attach rate means it is not just selling a product; it is capturing the core plumbing of the next generation of compute.

The Rubin Platform: The Next Catalyst for Networking Demand

The launch of the Vera Rubin platform at CES 2026 is the next major catalyst, explicitly engineered to accelerate demand for Nvidia's entire integrated stack, with networking at its core. The platform's primary mission is to slash the costs of running AI, particularly for complex, reasoning-heavy "agentic AI." It promises a 10x reduction in inference token cost and a 4x reduction in the number of GPUs needed to train MoE models compared to the previous Blackwell generation. This isn't just incremental improvement; it's a fundamental shift that makes deploying large-scale AI far more economical, directly fueling the next wave of infrastructure build-outs.

Crucially, the Rubin platform is designed as a full rack-scale system, like the Vera Rubin NVL72 platform. This architectural choice inherently requires Nvidia's Spectrum-X networking for optimal performance. The platform's extreme codesign integrates six distinct silicon components, including the Spectrum-6 ethernet switch. For a data center to achieve the promised efficiency and scale, these systems must be interconnected with high-speed, low-latency networking. In other words, buying a Rubin rack-scale system means buying into the Nvidia networking fabric from day one, further cementing the 90% attach rate.

The market's response confirms this is the new blueprint. Major cloud and enterprise partners are already building their next-generation AI factories on this integrated infrastructure. Microsoft's next-generation Fairwater AI superfactories will feature Nvidia's Vera Rubin NVL72 systems, scaling to hundreds of thousands of Rubin Superchips. This is a long-term, capital-intensive commitment that signals deep confidence in Nvidia's end-to-end solution. The platform's availability through partners like CoreWeave and expanded collaboration with Red Hat to deliver an optimized AI stack further solidifies the ecosystem. The Rubin platform is the main character in the next act of the AI infrastructure story, and its success is inextricably linked to the continued dominance and growth of Nvidia's networking business.

Valuation and Risks: Is the Hype Justified?

The explosive growth narrative is hard to ignore. Wall Street analysts see significant upside, with an average 1-year price target of $255.63 for Nvidia. That implies a substantial climb from recent levels and reflects confidence in the company's new growth engine. The bullish sentiment is backed by specific catalysts, like the upcoming Rubin platform launch, which is already driving partner announcements and infrastructure planning.

Yet, for all the momentum, the investment case must weigh this hype against tangible risks. The primary concern is that the AI networking market, while forecast to grow from $14.9 billion in 2025 to $46.8 billion by 2029, is not a closed club. As demand surges, it will inevitably attract competition. This could lead to margin compression over time, a classic pressure point in scaling industries. Nvidia's current 11.6% share of the data center Ethernet switch market shows it's gaining ground, but maintaining that lead and pricing power in a crowded field is the next challenge.

The main catalyst to watch is the second-half 2026 launch of Rubin platform products. This is the ultimate stress test for the new growth engine. The platform's promise of a 10x reduction in inference token cost is compelling, but its success hinges on flawless execution and adoption. The market is paying close attention to whether this integrated stack, which includes the critical Spectrum-X networking, can scale as promised. If Rubin delivers on its efficiency claims and drives massive, repeatable deployments, it will validate the entire networking thesis. If adoption stumbles, it could introduce headline risk and force a reassessment of the valuation premium.

The bottom line is that Nvidia's networking surge is a powerful trend, but trends can change. The 90% attach rate is a sticky advantage, and the Rubin platform is the next major character in the story. For now, the analyst consensus sees room to run. But investors should monitor the second-half 2026 launch not just for product details, but for the first real-world signal of whether this explosive growth can be sustained without margin erosion.