siliconangle.com web signal

Arista 7060XE7 Delivers 100 Tbps for AI Scale-Out

ai infrastructure chips networking ai-infrastructure ethernet infiniband

Key insights

  • Liquid cooling in the 7060XE7 drops rack power overhead from 30-50% down to 5-15% using Open Rack v3 and DC bus bar power.
  • Linear Pluggable Optics slash interconnect power consumption by roughly 60% while keeping individual ports field-serviceable.
  • The two-tier Arista fabric with the 7800 AI Spine chassis supports up to 4.5x more GPUs than standard fixed-box configurations.

Why this matters

The 7060XE7 pairs open-standard Ethernet with rack-scale liquid cooling and Linear Pluggable Optics, making enterprise AI clusters dramatically more power-efficient without proprietary lock-in. Ethernet's ability to support up to 4.5 times more GPUs per fabric tier than standard fixed-box designs directly challenges InfiniBand's grip on large-scale AI deployments. Enterprises that previously faced an InfiniBand-or-nothing decision for high-density GPU clusters now have an open-standard alternative that competes on both power economics and GPU density.

Summary

Arista Networks launched the 7060XE7 Series, a liquid-cooled rack-scale switch on Broadcom's Tomahawk 6, delivering 100 terabits per second with 224G SerDes technology. The 7060XE7 is a 2OU Open Rack v3 system with no fans, drawing DC power from rack bus bars and dropping power overhead from 30-50% to 5-15%. Linear Pluggable Optics cut interconnect power by roughly 60%, and individual failed ports can be replaced without triggering full-system outages. Essentially: (Arista, Broadcom) are betting open Ethernet displaces InfiniBand for enterprises scaling AI clusters. - The two-tier network with the 7800 AI Spine chassis supports up to 4.5x more GPUs than standard fixed-box configurations. - Failed ports replace individually, without triggering full-system outages. Ethernet is now competing on operational cost and scale, not just bandwidth.

Potential risks and opportunities

Risks

  • Enterprises already running InfiniBand-based AI clusters face costly dual-fabric operation during any transition to Arista Ethernet infrastructure.
  • LPO's 60% power reduction claim relies on Arista's own figures with no third-party validation cited, exposing buyers to spec-sheet risk at procurement.
  • Broadcom Tomahawk 6 supply constraints could bottleneck 7060XE7 availability during peak AI infrastructure build-out cycles.

Opportunities

  • Enterprises planning new AI clusters can bypass InfiniBand lock-in entirely, using the 7060XE7 and 7800 AI Spine chassis for open-standard GPU fabrics.
  • Data center operators targeting PUE reduction have a concrete hardware path: the 7060XE7 cuts power overhead to 5-15% versus 30-50% for air-cooled alternatives.
  • Open Rack v3 ecosystem vendors supplying racks, bus bars, and DC power infrastructure gain a new design-win anchor as liquid-cooled AI networking hardware expands.

What we don't know yet

  • Pricing and customer availability timeline for the 7060XE7-64PRS-RV3-L are not disclosed in the announcement.
  • Whether Arista's LPO approach achieves latency parity with co-packaged optics at full scale is not addressed in the article.
  • No head-to-head InfiniBand benchmarks are included, leaving the competitive performance gap unquantified.