Networking and interconnect hardware

Why the concentration exists

AI data center architecture demands tightly synchronized, low-latency communication across thousands of densely interconnected GPUs, where even minor signal degradation can create systemic performance bottlenecks. Modern AI workloads rely on distributed GPU clusters that generate massive east-west traffic during training and inference, making high-bandwidth networking essential for overall system efficiency. Fabric interconnects serve as central hubs connecting servers and network-attached equipment to broader network and storage systems, providing line-rate, low-latency, lossless Ethernet and Fibre Channel over Ethernet functions.[2][12][14]

Scale-up networks require very high bandwidth at eight to ten times the bandwidth of back-end scale-out networks per GPU, along with lossless operation and fine-grained flow control. Sixth generation NVLink delivers 3.6 TB/s per GPU of bidirectional GPU-to-GPU bandwidth and 260 TB/s of rack-level GPU bandwidth in a 72-GPU domain. The end-to-end latency for GPU-to-GPU transfers is 3X lower than alternative solutions based on off-the-shelf Ethernet, and the packet rate is 10X higher.[19][28]

Spine-leaf architecture has become the standard for hyperscale, cloud, and AI-driven environments, with leaf switches connecting directly to servers and spine switches serving as high-speed interconnects between leaves. Bottlenecks typically occur at spine interconnects and leaf uplinks rather than server-to-top-of-rack connections. Modern networks integrate high-capacity platforms providing 100 Gigabit Ethernet, 400 GbE, and 800 GbE ports to support these demanding workloads.[3][24]

What the evidence shows

Who supplies it

Cisco commanded more than a quarter of the global network hardware market in 2024, with over 85% of Internet traffic flowing through Cisco systems at some stage. The company serves over 70,000 enterprise customers globally and has developed its own programmable ASIC called UADP, integrated across its Catalyst 9000 switching portfolio. Cisco champions an Ethernet-based approach for scale-across fabric because it provides the scale, interoperability, and ecosystem maturity required for modern AI infrastructure.[1][9][10][11]

NVIDIA's networking business extends beyond Ethernet switches to include InfiniBand products, network interface cards, DPUs, optics, cables, and related hardware. The company's networking division, spurred by its acquisition of Mellanox, has established high-performance interconnects like InfiniBand and NVLink that enable real-time data flow between GPUs. New IDC data shows NVIDIA surpassing competitors in the data center Ethernet switching market, extending its influence beyond AI accelerators.[4][5][27]

Arista provides three Etherlink product families and over 20 products for scale-out AI networking compatible with Ultra Ethernet Consortium networks. The Arista 7700R4 DES architecture scales beyond 27,000 800G ports and is used by hyperscalers like Meta. AMD's Pensando Pollara 400 AI NIC and Salina DPU provide a credible second source for buyers seeking alternatives to NVIDIA's networking products.[16][17][19]

What it depends on, and what depends on it

AI data centers require approximately 36 times more fiber than traditional CPU server racks, driving substantial demand for optical components. Global fiber prices have risen roughly 70% from a 2021 trough of $3.70 to approximately $6.30 per fiber-kilometer as of December 2025. NVIDIA invested $300 million in Corning to build three new fiber manufacturing plants in North Carolina and Texas to secure supply of this critical input.[7]

Current systems are largely built on 200 Gbps SerDes technology, with IO logic consuming a third to half of switch power in both Ethernet-based and UALink-based switches. Arista's current systems are based on low-power 5nm silicon technology and support Linear Pluggable Optics and Extended Reach DAC Cables. Marvell's Teralynx T100 102.4 Tb/s Ethernet switch supports 512 ports running at 200 Gb/s or 64 ports running at 1.6 Tb/s.[6][15][19][21]

Where it sits in the stack

What would break it

AMD Helios rack-scale AI system uses UALink-over-Ethernet, an open standard co-developed by AMD, Microsoft, and others, rather than NVIDIA's proprietary NVLink switching hardware. DriveNets AI Fabric supports AMD Instinct MI300 and MI350 series GPUs with the AMD Pensando Pollara 400 AI NIC for 400 Gbps Ethernet connectivity, and already supports the upcoming AMD MI400 series and AMD Pensando Vulcano 800 GbE AI NIC. This open-standard approach provides an alternative to NVIDIA's integrated networking ecosystem.[13][25]

WhiteFiber's implementation of DriveNets' scheduled fabric Ethernet achieved 97.5% peak bandwidth utilization, averaging 390 GB/s against a possible 400 GB/s, outperforming their parallel InfiniBand implementation which topped out at 361 GB/s. The network backbone supported 512 NVIDIA H200 Tensor GPU clusters using five Accton Ramon3 System NCF appliances with 51.2 Tbps and 22 Accton Jericho3 System NCP devices with 14.4 Tbps. This demonstrates that Ethernet-based solutions can match or exceed InfiniBand performance in certain configurations.[18]

What to watch

AMD expects to begin shipping Helios hardware to Microsoft in the second half of 2026, with Azure general availability most likely in early 2027 for most regions. The Helios rack-scale AI system integrates 72 Instinct MI455X GPU accelerators with EPYC Venice server CPUs and Pensando networking into a single liquid-cooled unit. NVIDIA's BlueField-4 DPU launches in 2026, adding 800 TOPS of on-die inference at 800 Gb/s.[17][25]

The IEEE 802.3dj standard, expected by mid-2026, leverages a 200 Gb/s lane rate to support 800G over 8 fibers and 1.6T over 16 fibers. Cisco announced the Silicon One G300, a 102.4 Tbps switching silicon, on February 10, 2026, with systems delivering a 28% improvement in job completion time. The new 100% liquid cooled systems achieve a nearly 70% energy efficiency improvement.[8][22]

Switch port majority transitions from 800 Gbps in 2025 to 1.6 Tbps in 2027 and to 3.2 Tbps by 2030. NVIDIA's 2025 and 2026 switches bring co-packaged optics to market, though broad adoption requires new infrastructure designs. Ayar Labs targets 2028 deployment for an optical interconnect solution with partners including FOCI, Browave, Wiwynn, Alchip, and GUC. Cignal AI expects the external optical circuit switch market to exceed $2.5 billion by 2029 as deployments spread beyond Google into broader AI campuses.[20][23][26]

Related nodes

Sources

1. statista.com · 2025-11-27T00:00:00
2. keysight.com · 2025-05-21T00:00:00
3. vitextech.com · 2025-12-12T05:00:00
4. blog.mho.com · 2025-12-22T17:46:44
5. datacenterknowledge.com · 2026-06-24T16:00:08
6. tomshardware.com
7. tomshardware.com · 2025-12
8. newsroom.cisco.com · 2026-02-10
9. blogs.cisco.com
10. blogs.cisco.com · 2026-07-28
11. rededucation.com
12. eginnovations.com
13. drivenets.com
14. l-p.com · May 08, 2026
15. blog.apnic.net
16. intelligentvisibility.com
17. datagravity.dev
18. whitefiber.com
19. blogs.arista.com
20. nextgeninfra.io
21. delloro.com
22. store.cablesplususa.com · mid-2026
23. research.fpx.world
24. learn-more.supermicro.com
25. techtimes.com
26. introl.com
27. datacenterknowledge.com · July 13, 2026
28. developer.nvidia.com