Arista Networks (NYSE:ANET) executives and partner companies used a recent “XPO webinar” to outline the rationale, technical approach, and ecosystem momentum behind XPO, or “eXtra-dense Pluggable Optics,” a new high-density pluggable optical module form factor designed for next-generation AI networking.
Arista positions XPO as a density and cooling reset for AI networks
Vijay Vusirikala, Cloud and AI Networking Leader at Arista Networks, opened the event by framing AI training and inference as pushing data center infrastructure “to its absolute physical limits.” In an introductory video, Arista described XPO as a redesign of the pluggable optics form factor aimed at addressing bandwidth, density, cooling, and reliability constraints as AI clusters scale to “hundreds of thousands of GPUs.”
The company also emphasized thermal design as a key differentiator, describing XPO as “liquid first” with an integrated cold plate shared by two 32-channel paddle cards in a “belly-to-belly” configuration. Arista said the approach keeps component temperatures 20-25 degrees Celsius lower than air-cooled modules and can manage high-powered optics “up to 400 watts per module,” including ZR+ use cases. The video also highlighted a design goal of fewer internal components, stating a 75% reduction in internal components can improve system-level reliability.
MSA growth and OFC demonstrations
In a follow-on update, an Arista host identified as “Operator” said it had been a little more than two weeks since Arista launched XPO and the XPO MSA at OFC. The speaker said industry response exceeded expectations, with “more than 100 companies” signing up for the multi-source agreement and additional sign-ups arriving daily.
The speaker described broad interest not only in linear designs, but also “LRO and fully retimed” approaches, along with “slow and wide RF microwave” and coherent optics. The speaker said there was “particularly a lot of interest in coherent light and full CR optics” in the XPO form factor, and argued coherent modules could be a “natural fit” with fixed lasers described as more reliable than tunable lasers.
At OFC, Arista said it demonstrated “full retimed half retimed and linear” configurations using a test vehicle built with a “Condor Cirrus test chip” featuring 64 channels. The speaker also cited multiple partner demonstrations on the show floor, including TeraHop, YoctoLink, Coherent, Aperion, Linktel, Molex, Luxshare, MultiLane, and Nextest, among others.
Arista ties XPO to AI bandwidth growth and data center footprint
Arista’s discussion connected XPO to increasing bandwidth needs in AI data centers, stating bandwidth is “doubling year-over-year” and that higher bandwidth per GPU can improve training time and overall data center efficiency. The speaker provided illustrative examples of GPU I/O evolution, describing a progression from a “last year’s chip” with 12.8 Tbps scale-up to a “this year’s model” with 25.6 Tbps scale-up and 1.6 Tbps scale-out, and then a “2028 chip” concept that would increase scale-up to 102.4 Tbps and scale-out to 3.2 Tbps.
Reliability was described as the top requirement, with the presenter saying “today’s optics fail too often” and calling for “an order of magnitude” improvement in failure rates. Power efficiency, rack density, and the ability to scale manufacturing volumes were also highlighted. The speaker also emphasized that AI fabrics may incorporate multiple interconnect technologies—including copper, active/passive copper, RF microwave, and VCSELs—alongside single-mode optics and coherent approaches, arguing pluggable form factors enable that mix.
Comparing OSFP and XPO, the speaker said OSFP supports 32 modules per 1U (51.2 Tbps) with an air-cooled thermal envelope of roughly 30-40 watts, but argued it is “not enough” as switch chips increase bandwidth and as liquid cooling becomes more central. The presenter said XPO could enable “2 or 4” terabit-class switch appliances in 1U where OSFP solutions might require a larger chassis.
In one example, the speaker said a future GPU configuration with 24.6 Tbps scale-up could require “8 OSFP switch racks” to interconnect, versus “2 switch racks” with XPO. In a separate large-scale illustration, the presenter described a “400 MW data center” concept with 128,000 XPOs across 1,024 XPO racks, contrasting it with a design requiring roughly 1,400 OSFP racks for interconnection, and said XPO could save “more than 1,000 switch racks” and “almost half the floor space.”
Partner perspectives: TeraHop, Ciena, and Amphenol
Rang-Chen Yu, VP of Marketing at TeraHop, said AI scaling is driving requirements for higher capacity, density, power efficiency, reliability, supply resiliency, and “fast time to volume.” Yu said XPO starts at 12.8 Tbps and has a potential evolution to 25.6 Tbps, while enabling 200 Tbps switches in 1U and improving faceplate density by 4x versus OSFP. Yu also described the form factor as supporting IMDD and coherent optics “with DSP or without DSP” and spanning reaches from 100 meters to “10-80 km plus.”
TeraHop said it demonstrated a “12.8T DR8 XPO fully retimed” module at OFC, citing “nearly error-free pre-FEC BER floor” across 64 channels and describing the benefits of integrated liquid cooling for component temperatures. Yu also discussed a low-power pathway using “co-packaged copper” and an improved connector system, suggesting targets such as enabling LPO with module power “below 80 W total” at “6 picojoule per bit,” or under 130 W with a half-rate retimer approach.
Helen Xenos, Senior Director of Portfolio Marketing at Ciena, argued that power and density are becoming key limits as switch ASIC capacity grows from 51 Tbps to 100 and 200 Tbps per second while faceplate density remains constrained at “32 OSFPs per rack unit.” Xenos said XPO provides a path to higher density, improved reliability with liquid cooling, and a multi-vendor ecosystem via the XPO MSA. She outlined Ciena capabilities across electro-optic and coherent DSP, high-speed SERDES, data converters, and advanced packaging, and noted Ciena’s work on “direct to plug liquid cooling.”
Xenos said Ciena is helping develop design specifications for XPO across use cases ranging from “tens of meters to thousands of kilometers,” and is working to align management interfaces with OIF CMIS. She also referenced standardization work on a “Mini QD connector” at OCP, designed for “dripless, reliable operation.” Xenos described mechanical samples shown at OFC, including a 12.8 Tbps linear-drive XPO example built with “two 6.4T optical engines” and a separate 12.8 Tbps coherent-light design using “two 3.2T coherent light engines” per PCB, DFB lasers, and LC connectors. She also discussed deployment concepts where four XPOs could fill “the whole 9,600 GHz of spectrum” across the C+L band for “51.2 Tb” of capacity in an IP-over-DWDM architecture.
Sam Kahnke, Director of Treasury at Amphenol, discussed XPO in the context of maintaining a pluggable ecosystem while acknowledging CPO’s advantages in lower-loss channels. Kahnke emphasized goals including density, support for near-package or co-packaged copper, and preserving pluggability for a range of optical technologies. He compared XPO electrical performance to OSFP, stating insertion loss at 53 GHz is “roughly one dB” for both, with XPO providing more usable bandwidth beyond 70 GHz due to design changes such as pad geometries and removing a surface-mount attachment requirement.
Kahnke also described link-level budget assumptions, including “about 4 dB per meter” cable loss and a total loss budget “less than 19 dB” for chip-to-module channels, positioning this as supportive of linear pluggables and half-retimer approaches. He said Amphenol introduced an LPO module in the XPO form factor at OFC with “all 64 lanes running,” “electrically hot pluggable,” and demonstrating CMIS compliance using a “48-volt single supply.” Kahnke also highlighted ongoing work around cable routing, serviceability, and test and measurement collaborations, including efforts with MultiLane and Wilder Technologies.
MSA timeline and next steps
Arista said the XPO MSA now owns the XPO specification, and that the first task is publishing version 1.0. The presenter said the group has initiated a “60-day review process” and characterized the spec as “largely complete” after prior partner reviews. Webinar speakers encouraged interested parties to join the MSA via email at [email protected].
About Arista Networks (NYSE:ANET)
Arista Networks, Inc is a technology company that designs and sells cloud networking solutions for large-scale data centers and enterprise environments. The company is best known for its high-performance switching and routing platforms, which are used to build scalable, low-latency networks for cloud service providers, internet companies, financial services, telecommunications, and enterprise IT. Arista’s offerings emphasize programmability, automation and telemetry to support modern, software-driven network architectures.
Central to Arista’s product portfolio is its Extensible Operating System (EOS), a modular network operating system that provides consistent programmability, stateful control and advanced visibility across the company’s hardware platforms.
