Global Finance News
07 Oct 2026, 16:26
Arista Unveils Rack-Scale Ethernet Architectures for High-Density AI Infrastructure
Arista Networks (NYSE: ANET) announced a new rack-scale networking portfolio designed for both scale-up and scale-out AI fabrics, expanding Ethernet deeper into high-density AI computing environments.
The new architecture was developed with technology partners including AMD, Arm, Broadcom, d-Matrix, Meta, Microsoft and Qualcomm. Arista said the platform supports up to 144 accelerators in a single rack and can scale to 1,024 XPUs across multiple racks, giving cloud operators an open, standards-based alternative to proprietary interconnect architectures.
At the center of the launch is Arista’s Etherlink SU-144 scale-up architecture. The company is offering three physical designs: an orthogonal chassis for extreme bandwidth and density, a cabled backplane focused on modularity and serviceability, and a cross-rack system designed for larger multi-rack accelerator deployments.
Arista is also extending its scale-out portfolio with support for multiple interconnect approaches, including ESUN, Nvidia NVLink, NVLink Fusion and PCIe. The systems are built on Arista’s Netdi foundation and support either Arista EOS or an open network operating system.
The rack-scale systems incorporate liquid-cooled 7060EX7 switches with 102.4 Tbps of capacity per switch, along with integrated cooling, power and rack-management infrastructure. Arista said next-generation optics such as XPO and open CPO/NPO can increase rack networking capacity from roughly 1.6 Pb/s today to as much as 6.5 Pb/s.
According to the company, that higher density could reduce overall data center footprint by nearly 50%, potentially lowering infrastructure costs and accelerating deployment.
The announcement strengthens Arista’s positioning in AI networking as hyperscalers and cloud providers increasingly look for open Ethernet-based alternatives capable of supporting tightly coupled accelerator clusters at rack scale.
Arista Networks (NYSE: ANET) announced a new rack-scale networking portfolio designed for both scale-up and scale-out AI fabrics, expanding Ethernet deeper into high-density AI computing environments.
The new architecture was developed with technology partners including AMD, Arm, Broadcom, d-Matrix, Meta, Microsoft and Qualcomm. Arista said the platform supports up to 144 accelerators in a single rack and can scale to 1,024 XPUs across multiple racks, giving cloud operators an open, standards-based alternative to proprietary interconnect architectures.
At the center of the launch is Arista’s Etherlink SU-144 scale-up architecture. The company is offering three physical designs: an orthogonal chassis for extreme bandwidth and density, a cabled backplane focused on modularity and serviceability, and a cross-rack system designed for larger multi-rack accelerator deployments.
Arista is also extending its scale-out portfolio with support for multiple interconnect approaches, including ESUN, Nvidia NVLink, NVLink Fusion and PCIe. The systems are built on Arista’s Netdi foundation and support either Arista EOS or an open network operating system.
The rack-scale systems incorporate liquid-cooled 7060EX7 switches with 102.4 Tbps of capacity per switch, along with integrated cooling, power and rack-management infrastructure. Arista said next-generation optics such as XPO and open CPO/NPO can increase rack networking capacity from roughly 1.6 Pb/s today to as much as 6.5 Pb/s.
According to the company, that higher density could reduce overall data center footprint by nearly 50%, potentially lowering infrastructure costs and accelerating deployment.
The announcement strengthens Arista’s positioning in AI networking as hyperscalers and cloud providers increasingly look for open Ethernet-based alternatives capable of supporting tightly coupled accelerator clusters at rack scale.
