Arista Networks has introduced Ethernet rack architectures that support up to 144 accelerators in a single rack, plus larger computing domains that stretch across racks. The portfolio targets AI training and inference infrastructure by combining networking with liquid cooling. Whether the added density actually lifts GPU utilization enough to justify the cost is something cloud operators will have to settle themselves.
For providers renting GPU clusters, the system now matters more than any single part. Accelerators that wait on data cannot use their full processing power, however expensive they are. Arista’s Etherlink SU-144 design supports a single-hop network of up to 144 accelerators, with cross-rack topologies reaching 1,024. That gives buyers another option to weigh, though it does not promise equal application performance in every deployment.
The company built the portfolio with AMD, Arm, Broadcom, d-Matrix, Meta, Microsoft and Qualcomm, drawing on the Ethernet for Scale-up Networks initiative. Partnership alone, however, says nothing about installed capacity or proven savings. Buyers will want real measurements of completion times, utilization and behavior under congestion or component failure.
The physical demands are also significant. Arista offers an orthogonal chassis with liquid cooling support from 100 to 400 kW, and its scale-out racks pair liquid-cooled 7060EX7 switches rated at 102.4 Tbps with fiber patch panels and cooling manifolds. Leak detection, drip trays and battery-backed power shelves come with the design. As a result, a networking purchase now reaches into facilities engineering, and colocation sites without the right cooling or electrical capacity may struggle.
Flexibility is another selling point. Arista supports its EOS software or an open network operating system, and its designs work with Ethernet for Scale-up Networks, NVIDIA NVLink, NVLink Fusion or PCIe. Even so, those interconnects do not become interchangeable, so providers still face real integration work.
Arista also claims that next-generation XPO and co-packaged optics lift rack capacity from 1.6 to 6.5 petabits per second, with nearly a 50 percent cut in data center footprint. That figure is a vendor claim, and the announcement shares no deployment assumptions behind it. Pricing and shipping dates are missing too, which leaves procurement teams without two basic inputs.
