Barclays 23rd Annual Global Technology Conference | Arista Networks | The Evolution of AI Networking: Scale-Out, Scale-Across and Ethernet's Full Counteroffensive

This Barclays session featured Arista Networks (ANET)'s Mark Foss (SVP of Global Operations & Marketing) and Hardev Singh (GM on Cloud and AI products).

In the interview, Arista clearly defined the three dimensions of today's AI networking — Scale-Up, Scale-Out and Scale-Across — and gave clear views on future technology paths (LPO, CPO, liquid cooling). Here are my key takeaways and analysis:

1. The three layers of AI networking: Scale-Up, Scale-Out, Scale-Across

Arista breaks today's AI network architecture into three parts, which helps clarify its current TAM (Total Addressable Market) and future growth runway:

• Scale-Out (Compute Fabric):

  • This is Arista's main battlefield today, connecting large numbers of accelerators (GPUs/XPUs).

  • Arista is very strong here; its main competitors are other Ethernet switch vendors.

• Scale-Across (Data Center Interconnect, DCI):

  • This is a newer term Arista uses to describe the network connecting separate AI data center clusters.

  • When clusters grow so large that they must be connected across physical sites, this becomes an extension of traditional DCI.

  • Technical focus: the 7800 chassis platform paired with Coherent Optics (ZR/ZR+). Because longer distances add latency, this has to be addressed at the architecture level so that AI workloads don't depend too heavily on long-haul transmission.

• Scale-Up (GPU Interconnect):

  • Today this is dominated by NVIDIA's closed NVLink/InfiniBand systems, and Arista can't get in for now.

  • Future opportunity: as the non-NVIDIA accelerator ecosystem (AMD, Intel, etc.) rises, scale-up networking will move toward standardized Ethernet (e.g. UALink).

  • Timeline: trials are expected in 2026, with meaningful revenue contribution landing in 2027. For Arista, this is an entirely new TAM.

2. Ethernet vs. InfiniBand: Is the battle decided?

On the competitive landscape, Arista sees a clear shift in market sentiment:

• InfiniBand's decline: as Ethernet performance catches up, and with the push from the UEC (Ultra Ethernet Consortium), InfiniBand's market share is falling. Even NVIDIA itself has started talking more about Spectrum-X (its Ethernet solution).

• Customer mindset: Tier 2 cloud and enterprise customers lean toward Ethernet because they don't want vendor lock-in, and Ethernet offers the advantage of dual sourcing.

• The bundle strategy falls short: although NVIDIA is trying to push "full rack" bundled sales, engineers at hyperscalers and technically capable neoclouds tend to disaggregate the network layer and choose best-of-breed Arista EOS rather than adopting the packaged solution outright.

3. Optical and physical-layer evolution (LPO, CPO, 1.6T)

On the optical technology path everyone cares about most, Hardev Singh gave a very clear roadmap:

• 800G deployment: after 2025, 2026 will be the year of 800G volume deployment.

• LPO (Linear Pluggable Optics):

  • Arista is a strong champion of LPO.

  • Key view: Arista is confident that LPO will still work at 1.6T. This matters a lot for lowering power and CapEx (eliminating the DSP's power and cost).

• CPO (Co-Packaged Optics):

  • Arista sees CPO as a promising technology, but for now LPO still has the edge at 800G/1.6T. CPO has not yet reached mature, large-scale deployment and serves more as a technology reserve for the future.

• Liquid Cooling:

  • This depends on data center build cycles. Existing air-cooled DCs won't disappear right away.

  • Only newly built DCs (expected online in '27 and '28) will go 100% liquid-cooled. At that point the network will follow compute into rack-scale liquid-cooled architectures.

4. A changing customer mix: the rise of Tier 2 clouds (neoclouds)

Beyond the traditional big-four hyperscalers (the "Titans"), Arista particularly highlighted momentum in Tier 2 AI clouds and enterprise:

• Different engagement models:

  • Hyperscalers: lean toward co-development — deep collaboration at the technology frontier.

  • Tier 2/neoclouds: these customers don't have the huge engineering armies of the hyperscalers, so they rely more on Arista's complete solutions, the stability of EOS, and engineering support.

  • These customers care more about time-to-market, which favors a vendor with mature products like Arista.

5. Campus networking

Although the main focus is AI, the campus business is also a growth engine for Arista. It targets Global 2000 large enterprises and is steadily eroding Cisco's market share (Cisco holds roughly 75% share in this segment). Arista's strategy is to leverage EOS's quality and automation to become customers' "viable second vendor"; its share of the high-end enterprise market is currently close to 5%, leaving plenty of room to grow.

Conclusion

The interview shows that Arista is very confident about the post-InfiniBand era. The point isn't a head-on clash with NVIDIA; rather, as GPU interconnect moves to open standards (Ethernet/UALink), Arista is ready to use its dominance in scale-out to break into the scale-up market. Meanwhile, its insistence on pushing LPO to 1.6T shows its technical moat in hardware design and signal integrity.