Networking Watch: Why Networking Is AI Data Centers' Next Big Fight

Networking Watch | September 21, 2026

AI Data Centers Turn Networking Into the Next Infrastructure Battleground

Market Signal

AI Is Turning Networking Into Critical Infrastructure

Networking has moved from a supporting layer to a strategic part of the computing stack. Dell'Oro Group reported that worldwide data-center capital expenditures grew 92% year-on-year in Q2 2026, driven by surging AI demand and rising memory costs. The figure captures something broader than a spending cycle: AI clusters require fundamentally different network architectures than conventional enterprise data centers, and the network is increasingly the constraint that determines whether expensive accelerators work together efficiently.

The shift is visible across the stack. Nvidia has overtaken rivals in data-center Ethernet switching by revenue, according to IDC. HPE and Oracle are collaborating on networking for gigawatt-scale AI infrastructure. Optical interconnect vendors are racing to deliver 1.6T modules. And a new class of networking silicon startups is attracting significant capital. The question is no longer who sells the most switches—it is who controls the AI fabric connecting the compute.

Why AI Clusters Need a Different Network

Traditional data-center networks were designed for north-south traffic—users accessing applications, requests flowing to servers and back. AI training inverts that pattern. It generates massive east-west traffic as thousands of accelerators exchange gradients, parameters, and intermediate results continuously. Every synchronization step is a network event. Every millisecond of latency extends the time a training job occupies expensive compute.

The requirements are demanding: extremely low latency, very high bandwidth, lossless or near-lossless delivery, sophisticated congestion management, and RDMA over Converged Ethernet (RoCE) or equivalent technologies that let accelerators communicate without involving the CPU. Network observability becomes essential because a single misconfigured link or congested path can degrade an entire cluster.

The industry is converging on this problem. Lightwave's ECOC 2026 panel programming highlights how to scale networks to support next-generation AI factories—a recognition that networking capacity is becoming a central constraint in AI infrastructure, not a secondary concern. The optical and Ethernet communities are being pulled into the same conversation because neither can solve the problem alone.

Why this matters: Adding GPUs does not automatically translate into proportional performance gains. If the network cannot keep pace, accelerators idle while waiting for data. The network is a determinant of AI infrastructure ROI, not an overhead cost.

Nvidia's Networking Push Changes the Competitive Landscape

The most consequential development in AI networking is Nvidia's expansion beyond accelerators. IDC data reported earlier this year showed Nvidia moving to the top of the data-center Ethernet switching market by revenue—a position historically held by Broadcom and Cisco. The company's Spectrum-X platform, which combines Spectrum switches, ConnectX SuperNICs, BlueField DPUs and LinkX cables, is designed as an end-to-end Ethernet fabric optimized for AI workloads.

The strategic logic is straightforward. If Nvidia controls both the accelerators and the fabric connecting them, it can optimize the entire system—tuning network behavior to the characteristics of its GPUs and software stack. That vertical integration creates performance advantages that are difficult for competitors to match with components alone.

The competitive landscape now spans multiple layers:

Player	Position in AI Networking
Nvidia	Vertically integrated fabric spanning GPU, NIC, DPU and switch
Broadcom	Dominant merchant switching silicon; Tomahawk and Trident families
Cisco	Enterprise networking incumbent expanding into AI fabrics and optics
Arista Networks	Gaining share in cloud and AI back-end networks
HPE / Juniper	Routing and switching for AI data centers; telemetry and visibility
Marvell	Optical DSPs and custom networking silicon

Why this matters: The question is not which vendor sells the most ports. It is which vendor controls the fabric that connects the compute. Control of that layer determines pricing power, integration depth, and the ability to lock in enterprise customers for multi-year deployments.

HPE + Oracle: Networking for Gigawatt-Scale AI

One of the clearest illustrations of AI networking's growing importance is HPE's expanded collaboration with Oracle. HPE says Oracle plans a multi-year global deployment of HPE Juniper Networking across Oracle AI data centers, spanning PTX and MX routers and QFX and EX switches. The two companies are also working on telemetry and network visibility specifically tuned for AI clusters.

The scale implied by "gigawatt-scale AI infrastructure" is significant. A gigawatt of AI compute represents an enormous number of accelerators, and those accelerators require a fabric that can sustain aggregate bandwidth measured in petabits per second. Networking at that scale is not a procurement decision—it is an architectural commitment spanning routing, switching, optics, and management software.

Why this matters: Multi-year infrastructure deployments lock in vendors and architectures for years. When a hyperscaler or large cloud provider standardizes on a networking platform for AI data centers, that decision shapes the competitive landscape well beyond the initial contract.

Optical Networking Becomes a Bottleneck

As electrical signaling reaches its physical limits, optical connectivity becomes mandatory for AI data centers. 800G modules remain the volume mainstay in 2026, while 1.6T modules enter mass production at leading vendors. The Optical Internetworking Forum published its 1600ZR Implementation Agreement, defining an interoperable 1.6T coherent line interface for amplified DWDM links up to 120 kilometers—doubling the capacity of the earlier 800ZR specification.

Two competing approaches are emerging for the optical interconnect itself. Co-packaged optics (CPO) integrates optical engines directly with switching silicon, reducing power and improving signal integrity but complicating manufacturing and serviceability. Linear Pluggable Optics (LPO) keeps optics pluggable but simplifies the signal chain, reducing power relative to conventional DSP-based modules. Cisco has publicly positioned LPO as strategically useful but not a universal solution—an acknowledgment that AI networks will likely require a mix of approaches depending on reach, bandwidth, and power constraints.

Silicon photonics is central to both paths. Vendors including Coherent, Lumentum, Marvell and Cisco are investing in photonic integration to reduce cost and power per bit. The economics matter: in a large AI cluster, optical interconnect can account for a meaningful share of total power consumption, and every watt spent on optics is a watt not available to compute.

Why this matters: Optical interconnect is becoming a genuine bottleneck for AI infrastructure scaling. The vendors that can deliver higher bandwidth at lower power and cost—and at scale—will capture a disproportionate share of AI data-center spending.

Networking Silicon Is Becoming a Startup Opportunity

The scale of the AI networking opportunity has attracted a new wave of silicon startups. Delos Data, founded by Intel veterans, raised $100 million in September 2026 to develop chips and software aimed at improving data movement inside AI data centers. The company's pitch centers on the inefficiency of moving data between accelerators as clusters grow larger and more complex—a problem that merchant switching silicon and general-purpose NICs address only partially.

Xsight Labs has raised $300 million for its E1 Data Processing Unit and X2 switch, which have been selected by multiple global network operators. iPronics raised $125 million in Series B funding—with participation from Nvidia—to scale its programmable optical circuit switching platform for AI data centers.

Why this matters: A new category of networking silicon is emerging that targets specific bottlenecks—data movement, DPU offload, and optical switching—where incumbents have not fully addressed AI-specific requirements. This connects directly to semiconductor and startup-funding coverage: the AI networking opportunity is no longer confined to established vendors.

The Bigger Picture

The AI infrastructure chain is longer than most discussions acknowledge:

AI Compute → Networking → Optics → Fiber → Data Centers → Power

AI infrastructure cannot scale simply by adding accelerators. Each layer imposes constraints on the next. Compute requires networking to connect accelerators into a coherent cluster. Networking increasingly requires optics to reach across racks and halls. Optics requires fiber capacity and clean pathways. Fiber requires data-center construction. And every layer requires power—a constraint that is becoming the binding limit in many markets.

The network connecting AI systems is therefore part of both the performance and the economics equation. A cluster that cannot communicate efficiently wastes both capital and power. The companies that solve networking efficiency are solving a problem that compounds across the entire stack.

What to Watch Next

1.6T deployments: Watch for 1.6T optical module shipment volumes through Q4 2026 and into 2027 as the OIF 1600ZR standard gains traction.
Optical component pricing: Silicon photonics, LPO, and coherent DSP pricing will determine whether optics can scale cost-effectively at AI cluster scale.
Ethernet adoption: Whether Ethernet continues to take share from proprietary interconnects in AI back-end networks will shape the competitive landscape for switching silicon.
AI cluster sizes: Whether clusters scale to hundreds of thousands or millions of accelerators will determine the networking architectures required.
Networking silicon startups: Delos Data, Xsight Labs and iPronics will test whether specialized data-movement silicon can win share against integrated platforms.
Data-center capex composition: The share of AI data-center spending allocated to networking, optics and fiber will indicate whether the network is finally receiving proportional investment.

The CODEW Analysis

AI networking is no longer a supporting layer. It is becoming the infrastructure layer that determines whether AI compute can scale—and who captures the economics of that scaling.

This week's developments show how quickly the network moved to the center of the AI infrastructure conversation. Data-center capex grew 92% year-on-year. Nvidia overtook rivals in Ethernet switching by revenue. HPE and Oracle committed to a multi-year deployment spanning routers, switches, and AI-specific telemetry. Optical interconnect vendors are racing to 1.6T. And a new class of networking silicon startups attracted hundreds of millions in funding.

The companies that win will be those that recognize the network as a strategic asset—not a cost center—and design for the bandwidth, latency, power, and observability requirements of AI clusters at scale. The next phase of AI infrastructure competition will not be decided by who has the fastest accelerator. It will be decided by who controls the fabric connecting them.

Sources

Dell'Oro Group — Data Center Capex Grew 92 Percent in 2Q 2026, Driven by Surging AI Demand and Memory Costs (Sept 2026)
Lightwave Online — Lightwave's ECOC 2026 panel highlights how to scale networks to support next-gen AI factories (Sept 2026)
Data Center Knowledge — Nvidia Overtakes Rivals in Data Center Ethernet Switching, IDC Says
Hewlett Packard Enterprise — HPE and Oracle deepen networking collaboration to accelerate gigawatt-scale AI infrastructure (Sept 2026)
Network World — Cisco: LPO not a panacea but plays strategic role in AI networks
Reuters — Delos Data, a chip startup founded by Intel veterans, raises $100 million for AI networks (Sept 15, 2026)
Business Wire — OIF Closes Q3 Member Meeting with Critical Implementation Agreements for High-Speed Networks (Sept 2, 2026)
GlobeNewswire — iPronics Raises $125 Million to Scale Programmable Optical Networking for AI Data Centers (Sept 2, 2026)
IDC — Data-center Ethernet switching market share data

The CODEW Stat

Worldwide data-center capital expenditures grew 92% year-on-year in Q2 2026. Nvidia has overtaken rivals in data-center Ethernet switching by revenue. 1.6T optical modules are entering mass production, doubling the capacity of 800G systems. And networking silicon startups attracted more than $500 million in disclosed funding during September 2026 alone.

The CODEW