Arista Networks has introduced a new optical module architecture designed to address the escalating bandwidth and thermal requirements of AI-driven infrastructure. The company announced the formation of a multi-source agreement (MSA) for XPO, a liquid-cooled pluggable optics module capable of delivering 12.8 Tbps per module while significantly increasing front-panel bandwidth density.
The design targets next-generation AI networking fabrics, where high-throughput GPU clusters and disaggregated training systems require unprecedented optical interconnect capacity. As hyperscale AI infrastructure expands, network components must support both rising bandwidth demands and higher power envelopes. Consequently, optical modules themselves have become a critical bottleneck.
XPO addresses this constraint by integrating liquid cooling directly into the pluggable optics architecture. The result is a module capable of sustaining high power levels while maintaining signal integrity across dense AI fabrics.
Arista says the technology achieves a front-panel density of 204.8 Tbps per Open Compute Project rack unit, representing a fourfold increase compared with existing 1600G-OSFP optics deployments. Such gains could materially change how hyperscale operators architect AI clusters and high-performance computing fabrics.
A New Optics Architecture for AI-Scale Infrastructure
The XPO specification focuses on addressing several emerging networking challenges tied to AI workloads. Training clusters now span thousands of accelerators, creating network fabrics that must scale across racks, clusters, and even metropolitan regions.
The architecture therefore supports a broad set of deployment scenarios including scale-up GPU fabrics, scale-out AI clusters, scale-across interconnect architectures, and metro-reach network extensions used in distributed training environments.
At the core of the design sits a 12.8 Tbps pluggable module, representing one of the highest-throughput optics form factors currently proposed for AI infrastructure. The system pairs this throughput with a front-panel density reaching 204.8 Tbps per rack unit, dramatically increasing bandwidth concentration within modern data center switches.
Thermal management forms another key pillar of the architecture. The module integrates a cold plate capable of dissipating up to 400 watts per module, allowing optics to operate within the same liquid-cooled environments increasingly used for high-density AI compute.
XPO also introduces broad interoperability across optical technologies. The form factor supports widely deployed standards including DR, FR, LR, SR, and ZR/ZR+ optics while also enabling emerging designs such as coherent-lite optics, slow-and-wide architectures, copper interconnects, and RF-microwave transmission technologies. Moreover, the interface architecture allows multiple design approaches including linear, half-retimed, or fully-retimed implementations, giving system architects flexibility across network deployments.
Industry Momentum Builds Around New Optical Form Factor
Arista executives positioned the initiative as a necessary evolution of pluggable optics for the AI era.
“The unprecedented growth in AI fabric bandwidth and the transition to liquid cooling requires a new generation of pluggable optics modules,” said Andreas Bechtolsheim, Chief Architect at Arista Networks. “XPO solves this challenge by providing fundamental improvements in density, cooling capability and reliability for pluggable optics modules.”
Hyperscale cloud operators are also signaling support for the emerging specification. Large AI training environments increasingly depend on open optical ecosystems to maintain supply diversity and rapid innovation cycles.
“XPO is an important milestone for the optical industry, extending the benefits of pluggable optics to the extreme bandwidth demands of AI scale-out and scale-up systems,” said Matthew Mattina, VP for AI Systems Architecture at Microsoft. “Microsoft believes this specification can help establish a robust, broadly adopted form factor that enables a diverse optical ecosystem.”
Market analysts view the development as part of a broader shift toward ultra-dense optical connectivity in AI data centers. Optical interfaces must evolve rapidly as compute clusters move from hundreds to tens of thousands of accelerators.
“The XPO module is a breakthrough for AI data centers delivering 4X the front panel density compared to OSFP, while preserving the configurability and serviceability of a pluggable optics module,” said Sameh Boujelbene, Vice President of Dell’Oro Group. “The success of any new optical form factor hinges on a strong ecosystem of suppliers and I am pleased to see XPO supported by all major optical module suppliers.”
OFC 2026 Demonstrations Signal Ecosystem Formation
The XPO multi-source agreement will debut publicly at OFC 2026 in Los Angeles, where Arista plans to demonstrate the technology alongside ecosystem partners. Live demonstrations will take place at the company’s booth in the South Hall and at several partner exhibits across the event floor.
Such demonstrations mark an early step toward industry adoption. However, the presence of multiple optical vendors in the MSA suggests the specification could mature quickly if hyperscale operators embrace the design.
For AI infrastructure builders, the implications are significant. Optical bandwidth density, power efficiency, and thermal integration increasingly define the scalability of AI clusters. Therefore, innovations at the optics layer may become as strategically important as advances in GPUs or switching silicon. If widely adopted, XPO could emerge as a foundational building block for next-generation AI networking architectures.
