TD Cowen: NVIDIA's Networking Business Is No Longer a Sideshow — It's the Architecture of the AI Factory
TD Cowen 54th Annual TMT Conference, May 28, 2026 — Gilad Shainer, VP of Networking at NVIDIA, lays out why the company's $14.9 billion networking segment is structurally different from anything else in the market
A $14.9 Billion Networking Business Growing at 199% — and Every Layer Is Contributing
NVIDIA's networking revenue crossed $14.9 billion in the most recently reported quarter, up 199% year-over-year, and Gilad Shainer, who leads the networking business at NVIDIA, made clear at TD Cowen's 54th Annual TMT Conference on May 28 that this is not a one-product story. Growth is coming simultaneously from NVLink on the scale-up side, InfiniBand and Spectrum-X Ethernet on the scale-out side, and BlueField as both a storage processor and a data processing unit enabling secure access into AI factories. The breadth of that contribution matters because it means networking revenue is not dependent on any single bottleneck resolving or any single customer concentration. Every layer of what NVIDIA calls the AI factory is expanding.
The Mellanox Acquisition Was About Becoming a Computing Company, Not a Component Vendor
Shainer offered the clearest articulation yet of why Jensen Huang pursued the Mellanox acquisition, which TD Cowen described as "perhaps the most important and successful technology M&A that has ever happened." In Shainer's telling, the logic was architectural: "Jensen saw that NVIDIA needs to become a computing company, not a device company, not an ASIC company, but a computing company. And the way that you connect computing ASICs will determine what those compute ASICs can do. If you connect it in one way, you just got a server farm. If you connect it in a different way, you actually can build a supercomputer." That framing — networking as the determinant of what compute can actually do — is what underpins NVIDIA's entire integrated rack strategy and explains why the Mellanox team was not absorbed as a separate business unit but folded in as a single engineering organization from day one.
NVLink Fusion Is Not a Defensive Move — It Is NVIDIA Opening Its Best Technology to the Ecosystem
There has been persistent market concern that NVIDIA's shift to fully integrated NVL racks was locking customers into a closed system, alienating ecosystem partners. Shainer pushed back on this framing directly. The architecture is designed vertically — every component co-engineered to behave as a single unit — but sold horizontally, meaning customers can take individual pieces and combine them with their own designs. NVLink Fusion, which allows third-party CPUs and even third-party GPUs to connect via NVLink, is the concrete expression of this. "NVLink Fusion enables our customers to also take NVLink as a separate element if they want to do that," Shainer said. His argument is that NVIDIA is confident enough in the quality of each component that it wants partners to use them individually. Whether the ecosystem perceives this as genuinely open or strategically open is a separate question, but the commercial announcements already made around Fusion partners suggest real traction.
Why Spectrum-X Ethernet Is Architecturally Distinct — and Why That Distinction Is Not Marketing
The most technically substantive portion of the conversation was Shainer's explanation of why Spectrum-X is not comparable to standard Ethernet switches and why that difference is fundamental rather than incremental. The core problem in distributed AI workloads is jitter — the variation in timing with which data arrives at different GPUs. In a training or inference cluster where hundreds of thousands of GPUs must operate in lockstep, even a single GPU receiving data slightly late causes all others to wait. Traditional Ethernet architectures, including those designed for large-scale cloud workloads, were never built to solve this problem because in single-server or long-haul environments, jitter is irrelevant.
The deeper issue is structural. Eliminating jitter requires unconditional packet spraying — routing every individual packet across the least congested available path regardless of flow order. But doing that at the switch level inherently creates out-of-order data delivery. The only way to resolve this is to have an intelligent endpoint that can receive out-of-order packets and reassemble them correctly in GPU memory in real time. That endpoint is the ConnectX SuperNIC. "That's why when you build an infrastructure for distributed computing workloads, you need to have a switch element that does the distribution unconditionally, and then you need a SuperNIC that will put the data back in order," Shainer said. "That's why it's an infrastructure, and it's not a single device." This is the architectural moat. A competitor offering only a switch, or only a NIC, cannot replicate what Spectrum-X delivers as a system.
Spectrum-X Supports Multiple Routing Protocols — Including MRC and Customer-Proprietary Ones
Shainer also addressed the emerging Multi-path RDMA Congestion control (MRC) standard and its relationship to Spectrum-X. Rather than positioning MRC as a competitive threat to NVIDIA's adaptive routing approach, he described Spectrum-X as a platform that supports multiple routing protocols simultaneously — adaptive RDMA, MRC, and several proprietary protocols developed by large hyperscale customers for their own workload optimization. "There is a variety of routing protocols that can run on top of Spectrum-X," Shainer said, drawing an analogy to BGP as simply another protocol layer. Much of what the Ultra Ethernet Consortium is now working to standardize, he noted without naming the consortium directly, reflects design choices NVIDIA already made in Spectrum-X. That framing, if accurate, suggests NVIDIA has a meaningful head start in production-grade implementations.
Inference Creates New Infrastructure Demand That Training Did Not — BlueField and the KV Cache Storage Problem
Shainer introduced a dimension of NVIDIA's networking growth that has received relatively little investor attention: the storage infrastructure layer built for inferencing. As AI deployments shift toward agentic architectures — where AI models interact with other AI models, hold longer context windows, and maintain larger KV caches — the assumption that all relevant data can reside in local GPU memory breaks down. Network-attached storage becomes necessary, but conventional network storage architectures are over-engineered for the inferencing use case because they prioritize data redundancy through replication. In inference, Shainer argued, lost data can simply be recalculated, making replication wasteful. NVIDIA's answer is a purpose-built storage layer using BlueField, STX, and CMX, optimized for KV cache retrieval without the overhead of traditional storage redundancy. This is new incremental addressable market for the networking segment, not a repackaging of existing products.
Co-Packaged Optics Is a Power Story First, a Technology Story Second
On the copper-versus-optics debate, Shainer was pragmatic in a way that cuts through a lot of the noise in the market. Copper wins at short distances on cost and power — it consumes essentially zero incremental power. Optics are necessary when distance demands it, but optical connectivity on scale-out infrastructure today can consume close to 10% of total AI factory power capacity, which is the binding constraint on how much compute can be deployed. Co-packaged optics reduces that optical power burden, which is why NVIDIA is investing in it for configurations like the Feynman platform where 1,152 GPU connectivity requires crossing rack boundaries. The framing here is important for investors: CPO adoption is not a function of technology readiness debates but of power economics, and power is the number-one limiting factor on AI factory buildouts today.