Ethernet broadens its lead in AI back-end networks
What are AI back-end networks?
AI back-end networks are specialized high-performance interconnect infrastructures designed for ultra-low latency, high throughput, and lossless communication between GPUs or accelerators during distributed model training and inference. These environments depend on specialized back-end networks built around scale-up and scale-out/scale-across architectures, which differ significantly from traditional front-end networks.
Ethernet’s utility in the data center continues to ramp, a trend that continued in the second quarter and further extended its lead in AI back-end networks.
The network technology’s dominant role comes during what Dell’Oro Group calls a “major inflection point,” when switch sales in AI back-end networks surpassed those in front-end networks for the first time.
Sales of Ethernet AI back-end networks have been primarily dominated by scale-out and scale-across use cases. Dell’Oro noted that “scale-up use cases will start to emerge in the second half of the year and further fuel market growth.”
"AI back-end networks have grown at such a rapid pace that, within just three years of their emergence, spending surpassed that of the large, well-established front-end network market," said Sameh Boujelbene, VP at Dell'Oro Group, who leads the research group’s Enterprise Market research programs. "Not only is switch spending growth in back-end networks outpacing front-end networks, but the market size has also become larger in absolute terms.”
She added that the back-end network spending trends reflect a “major shift with significant implications for Ethernet switch vendor positioning and market share.”
Amidst the growth, the market will have to grapple with supply challenges. “Supply availability will also play an increasingly important role in shaping the competitive landscape and overall market growth,” Boujelbene said. “We expect the market to remain supply-constrained, rather than demand-constrained, for at least the next one-to-two years."
800G to 1.6T transition gains momentum
800G switches were the dominant platform in switch shipments and revenues in AI back-end networks during the second quarter.
1.6 T is close behind 800 G.
Driven primarily by a shift to 200G-per-lane electrical and optical interfaces, the transition from 800G to 1.6T data center optics is accelerating to support dense GPU clusters and high-throughput AI workloads.
A key development in 1.6 Tbps is the shift to 200G-per-lane optical architectures, enabled by advances in EML laser technology, DSP design, and high-speed electrical interfaces.
During the second quarter, providers began sampling 1.6 Tbps switches. Dell’Oro expects 1.6 Tbps to ramp in the second half of 2026.
However, a recent FS.com blog post noted that the migration to 1.6 Tbps won’t disrupt existing 800G infrastructure. Instead, data center providers can deploy a hybrid design where validated fiber trunks (MPO/MTP), rack pathways, and select switch ports can coexist.
Celestica, NVIDIA dominate
On the vendor front, Celestica and NVIDIA remain the dominant suppliers and their recent earnings show that trend.
Celestica retained the leading position in Ethernet AI back-end switch sales, reporting overall second-quarter revenue of $4.70 billion, up 62% year over year. The company noted that revenue growth through 2026 was driven by robust 800G switch demand, and 1.6T switch program ramps set to launch in the second half of 2026.
Within its Connectivity & Cloud Solutions (CCS) segment, Communications revenues were $3.8 billion, up 84%, while Enterprise rose 167% to $1.2 billion.
In its second-quarter earnings release, Celestica said, “trends expected to drive very strong growth through 2027.”
NVIDIA trails Celestica in second place, reporting fiscal second-quarter 2027 revenue of $96 billion.
A major contributor to NVIDIA’s revenue was Data Center revenue of $89 billion, up 117% from a year ago.
While Arista ranked third, Boujelbene said it “would have been more closely tied if deferred AI revenue were included.” As a dominant provider of high-performance modular and fixed systems like the Arista 800G Solutions (including the 7700R4 and 7060X6 series) optimized for large-scale AI clusters, Arista reported second-quarter revenue of $3.03 billion.
Taking the fourth position is Cisco. Cisco reported fourth-quarter and full-fiscal-year 2026 earnings of $17.3 billion, beating Wall Street expectations.
Offering enterprise and hyperscale options powered by its Silicon One architecture for high-density back-end data center fabrics, Cisco’s fourth-quarter networking product orders grew 40% year over year to $9.8 billion, marking the eighth consecutive quarter of double-digit growth.
The company reported strong AI momentum, with $9.3 billion in hyperscaler AI orders, up 4.5x versus $4 billion in hyperscale AI revenue in FY25.
Dell’Oro noted that while Cisco was ranked fourth, it “gained the most market share.”
For related articles, visit the Data Center Topic Center.
For more information on high-speed transmission systems and suppliers, visit the Lightwave Buyer’s Guide.
To stay abreast of fiber network deployments, subscribe to Lightwave’s Service Providers and Datacom/Data Center newsletters.
About the Author
Sean BuckleySean Buckley
Sean is the Editor-in-Chief of Lightwave. He establishes and executes Lightwave's editorial strategy across its website, email newsletters, events, and other information products. Before coming to Lightwave in July 2023, he served as editor of Broadband Communities. Earlier, he served as senior editor of FierceTelecom and the editor of the former Telecommunications Magazine. He got his start in the optical industry in 1998 when he became the editor of enterprise and optical networks at Information Gatekeepers. He has a BA in English from the University of Massachusetts, Boston, and lives in Dracut, Mass, with his wife, two sons, and his cat, Dove.
