Networking Watch: AI Turns Networking Into the Next Critical Infrastructure Layer
AI Is Turning Networking Into the Next Critical Infrastructure Layer
The AI Bottleneck Is Moving from Compute to Connectivity
The AI infrastructure conversation has been dominated by GPUs, memory, and power. But as AI clusters scale from thousands to hundreds of thousands of accelerators, a less visible constraint is emerging: the network that connects them. Moving data between chips, between racks, and between data centers is becoming as important as the compute itself. The companies that solve the AI networking bottleneck—whether through Ethernet, optical interconnects, switching silicon, or photonics—will capture a disproportionate share of the next infrastructure cycle.
AI traffic now accounts for around 30% of backbone network utilization, up from less than 1% two years ago. In a 2026 State of AI for Networking Report, 92% of IT leaders said AI has increased their computing and bandwidth demands. The network is no longer a passive conduit. It is a determinant of whether AI infrastructure delivers on its promised performance.
The Rise of AI Networking Silicon
A new category of chip companies is emerging to address AI's data-movement challenge. Delos Data, a startup founded by Intel veterans, raised $100 million in September 2026 to develop chips and software that move data faster inside AI data centers. The company has unveiled the Delos Mosaic Software and Delos Asterion Server, targeting the inefficiency of moving data between chips as AI clusters become increasingly complex.
Xsight Labs, another emerging player, has raised $300 million for its E1 Data Processing Unit and X2 switch, which have been selected by multiple global network operators. Nvidia's Spectrum Ethernet platform—including Spectrum switches, ConnectX SuperNICs, BlueField DPUs, and LinkX cables—represents the incumbent response, offering end-to-end Ethernet solutions optimized for AI fabrics.
Why this matters: The switching silicon market is becoming as strategically important as the accelerator market. The companies that control how data moves between accelerators will determine whether AI clusters can scale efficiently.
Ethernet vs. Proprietary Architectures
The AI networking market is split between open Ethernet-based systems and vertically integrated proprietary architectures. InfiniBand offers roughly 15% better performance than Ethernet but costs about 2.3 times more. OpenAI has credited InfiniBand's superior congestion control for enabling GPT-4 training to complete 40% faster than initial Ethernet-based attempts.
But Ethernet is winning at scale. HPE's Slingshot technology has demonstrated that "Ethernet plus" can beat proprietary interconnects in AI supercomputing. Dell'Oro Group projects that Ethernet will dominate both scale-up and scale-out segments of the AI back-end switch market, which will push past $100 billion by 2030.
Why this matters: The strategic dynamic is familiar: proprietary architectures win on performance, open standards win on cost and interoperability. In AI networking, both are winning—but Ethernet is winning more.
Optical Networking Becomes Unavoidable
As electrical signals reach their physical limits, optical networking is becoming mandatory for AI data centers. 800G optical modules remain the volume mainstay in 2026, while 1.6T modules are entering mass production at leading vendors. The Optical Internetworking Forum (OIF) has published the 1600ZR Implementation Agreement, defining an interoperable 1.6T coherent line interface for amplified DWDM links up to 120 kilometers, doubling the capacity of its earlier 800ZR specification.
Silicon photonics startup iPronics raised $125 million in Series B funding—with participation from Nvidia—to scale its optical circuit switching platform for AI data center networks.
Why this matters: The shift from copper to optical connectivity is no longer a question of if, but of when. For AI clusters operating at 1.6T and beyond, optics is the only viable path.
The Companies Positioned Around AI Networking
The competitive landscape spans multiple layers. Broadcom dominates switching silicon with its Tomahawk and Trident families. Marvell provides optical DSPs and custom networking silicon. Nvidia integrates networking into its AI factory platform, using Spectrum Ethernet and BlueField DPUs to create a vertically optimized stack. Cisco and Arista compete in enterprise and cloud networking, with Arista gaining share in AI back-end networks. Ciena and Infinera lead in coherent optical systems.
Emerging companies like Delos Data, Xsight Labs, and iPronics are targeting specific bottlenecks—data movement, DPU offload, and optical switching—where incumbents have not fully addressed AI-specific requirements. The companies that succeed will be those that recognize AI networking is not a single market but a stack of interdependent technologies.
The Economics of AI Networking
The economics of AI networking are shifting from bandwidth to efficiency. Power consumption is becoming a primary constraint: 800G optical modules now operate below 12W, and 1.6T modules below 21W, but the aggregate power draw of networking equipment in a large AI cluster is measured in megawatts. Latency is another economic variable: every microsecond of network delay in a training cluster translates into longer job completion times and higher compute costs.
| Economic Variable | Current Constraint | Strategic Implication |
|---|---|---|
| Power | 800G modules below 12W; 1.6T below 21W; aggregate cluster draw in MW | Lower-power optics and switching silicon command premium pricing |
| Latency | Every microsecond of delay extends training job completion | Network latency directly translates to accelerator idle time and cost |
| Utilization | Network utilization determines whether compute capacity is productive | Higher utilization improves AI infrastructure ROI |
The companies that can deliver higher bandwidth at lower power and latency—and at a cost that scales linearly with cluster size—will capture the economics of AI networking. This is not a commodity market. It is an efficiency market.
What to Watch Next
- 1.6T deployments: Watch for 1.6T module shipment volumes in Q4 2026 and early 2027 as the OIF 1600ZR standard gains traction.
- Optical component pricing: Silicon photonics and coherent DSP pricing will determine whether optics can scale cost-effectively.
- Ethernet adoption: Dell'Oro's projection of Ethernet dominating the $100 billion AI back-end switch market will be tested by proprietary interconnect roadmaps.
- AI cluster sizes: Whether clusters scale to 1 million accelerators or plateau at 100,000 will determine networking demand.
- Network utilization: If AI workloads shift from training to inference, traffic patterns will change—requiring different architectures.
- Data-center capex: Hyperscaler spending on networking as a percentage of total AI infrastructure will indicate whether the network is finally getting its share.
AI networking is not a commodity. It is the efficiency layer that determines whether AI infrastructure delivers on its promise.
The market is expanding rapidly, with Ethernet poised to dominate the $100 billion AI back-end switch market by 2030. Delos Data's $100 million raise, iPronics' $125 million Series B, and the OIF's 1.6T implementation agreement all point to the same conclusion: networking is no longer a supporting player in AI infrastructure. It is a critical layer.
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, and power requirements of AI clusters at scale.
Sources
- Reuters — Delos Data, a chip startup founded by Intel veterans, raises $100 million for AI networks (Sept 15, 2026)
- GlobeNewswire — iPronics Raises $125 Million to Scale Programmable Optical Networking for AI Data Centers (Sept 2, 2026)
- Business Wire — OIF Closes Q3 Member Meeting with Critical Implementation Agreements for High-Speed Networks (Sept 2, 2026)
- Dell'Oro Group — AI back-end switch market projection
- 2026 State of AI for Networking Report
- Nvidia Spectrum Ethernet platform documentation
The CODEW Stat
AI traffic now accounts for around 30% of backbone network utilization, up from less than 1% two years ago. Ethernet is projected to dominate the $100 billion AI back-end switch market by 2030. And 1.6T optical modules are entering mass production — doubling the capacity of 800G systems.
Reviewed by Erwin Castro
on
Wednesday, September 16, 2026
Rating:
