NVIDIA (NASDAQ:NVDA) networking executive Gilad Shainer said the company’s rapid networking growth is being driven by the expansion of multiple infrastructure layers needed to make AI data centers operate as “a single unit of computing.”
Speaking at TD Cowen’s 54th annual TMT Conference, Shainer addressed NVIDIA’s recently reported networking revenue of $14.9 billion, up 199% year over year. He said the momentum reflects growth across several areas, including NVLink for scale-up networking, InfiniBand and Spectrum-X Ethernet for scale-out networking, BlueField for storage processing and data center access, and additional infrastructure designed to connect and secure AI factories.
Mellanox Deal Framed as Key to NVIDIA’s AI Factory Strategy
Shainer also discussed NVIDIA’s 2020 acquisition of Mellanox, the networking company where he previously worked. He said NVIDIA Chief Executive Jensen Huang recognized that the company needed to evolve from a device or ASIC company into a broader computing company.
“The way that you connect computing ASICs will determine what those compute ASICs can do,” Shainer said. “If you connect it in one way, you just got a server farm. If you connect it in different way, you actually can build a supercomputer.”
Shainer said Mellanox’s focus on networking infrastructure for distributed computing workloads made it a strong fit for NVIDIA as AI emerged as another major distributed computing workload. He described the integration as natural, saying NVIDIA and Mellanox both operated as “one unit” with compute, networking and infrastructure teams working closely together.
Integrated Racks, Open Components and NVLink Fusion
Asked about NVIDIA’s shift toward fully integrated racks and concerns from ecosystem partners about being locked into one supplier, Shainer said AI factories require “extreme co-design” across software, hardware, compute, networking and storage components. He said distributed computing systems can suffer if even one GPU receives data later than others, because the rest of the system may be forced to wait.
However, Shainer said NVIDIA designs systems vertically but sells components horizontally, allowing customers to use individual pieces such as GPUs, CPUs, networking, or NVLink separately. He said customers can mix NVIDIA components with their own designs and software.
“Nothing is closed. Everything is very open,” Shainer said.
He pointed to NVLink Fusion as an example, saying it allows customers and partners to use NVLink as a separate scale-up networking element even if they have developed their own CPU or GPU.
Spectrum-X Built to Address Ethernet Jitter
In a technical exchange with Sean O’Loughlin, vice president at TD Cowen, Shainer said NVIDIA developed Spectrum-X Ethernet because many customers had already invested in Ethernet infrastructure, management tools and operating expertise. While NVIDIA continues to support InfiniBand, which Shainer described as highly suited for low-latency distributed computing, he said existing Ethernet designs were not built to eliminate jitter.
Shainer said jitter is a central challenge for AI training and inference because data must arrive in order and without unpredictable delays across distributed GPU systems. He said traditional Ethernet designs often preserve packet ordering by keeping flows on the same path, even when other network paths are available, which can create delays.
According to Shainer, Spectrum-X uses a switch and a “SuperNIC” together. The switch distributes packets across available paths, while the SuperNIC reorders data and places it into GPU memory using RDMA.
“That’s why it’s an infrastructure, and it’s not a single device,” he said.
Shainer said Spectrum-X supports multiple routing approaches, including adaptive RDMA and MRC, as well as customized protocols developed by large customers. He said the goal is to support different routing protocols while maintaining a zero-jitter approach.
Inference Adds New Storage Demands
Shainer said training and inference share important networking requirements because both are distributed computing workloads that require low or zero jitter. In multi-tenant AI cloud environments, he said eliminating jitter can also reduce the risk that one workload affects another, a problem often referred to as the “noisy neighbor” issue.
He added that inference is creating demand for additional infrastructure. As AI systems move toward agentic workloads, he said larger context and KV cache requirements may exceed what can be stored locally in a GPU server. Shainer said NVIDIA has developed a storage infrastructure using BlueField that is optimized for inference use cases, where recalculating data may be preferable to relying on expensive replicas in some rare failure scenarios.
Co-Packaged Optics Tied to Distance and Power Efficiency
Shainer also addressed co-packaged optics, or CPO, saying the debate between copper and optics depends largely on distance. He said copper is cost-effective, reliable and consumes no power, making it preferable when distances are short enough. Optics becomes necessary when distances exceed copper’s capabilities.
Because power is a limiting factor for AI factories, Shainer said NVIDIA is investing in co-packaged optics to reduce the power consumed by optical networking. He said optics may be needed in scale-out networks and in scale-up domains that extend across multiple racks.
“If I’m using optics, I want to have the best technology that consume the least amount of power,” Shainer said. “That’s called co-packaged optics, regardless if it’s scale-out, scale-up, scale-across, it’s all depends on the distance.”
About NVIDIA (NASDAQ:NVDA)
NVIDIA Corporation, founded in 1993 and headquartered in Santa Clara, California, is a global technology company that designs and develops graphics processing units (GPUs) and system-on-chip (SoC) technologies. Co-founded by Jensen Huang, who serves as president and chief executive officer, along with Chris Malachowsky and Curtis Priem, NVIDIA has grown from a graphics-focused chipmaker into a broad provider of accelerated computing hardware and software for multiple industries.
The company’s product portfolio spans discrete GPUs for gaming and professional visualization (marketed under the GeForce and NVIDIA RTX lines), high-performance data center accelerators used for AI training and inference (including widely adopted platforms such as the A100 and H100 series), and Tegra SoCs for automotive and edge applications.
