Arista webinar reinforces AI networking leadership

Overview

This report is based on Arista's technical webinar and discusses the company's newly launched 1.6T switches, LPO, liquid cooling support, and software capabilities advancing in sync with the hardware. J.P. Morgan believes Arista is reinforcing its AI networking leadership through a hardware-plus-software combination, with the MRC protocol and EOS capabilities in particular making its Ethernet solution more competitive in large-scale AI training clusters.

Core views

The core view is that Arista's leadership in data center switching remains solid, with 1.6T hardware, rack-level system capabilities, the MRC protocol, and the EOS software stack together forming its differentiation. The report believes MRC allows Ethernet to scale to 100K+ accelerators in a flat two-layer architecture, while multi-plane designs further improve scalability and resilience; combined with customer participation from Anthropic, Meta, Microsoft, Oracle, and others, these capabilities support strong growth and market share gain opportunities for the company.

Analysis framework

The report mainly uses a technical webinar interpretation approach, evaluating Arista across product architecture, network topology, protocol standardization, customer participation, and competitive substitution. The focus of the analysis is not financial modeling or valuation derivation, but rather whether the 1.6T switches, EOS capabilities, and MRC protocol can improve the scale, reliability, power efficiency, and deployment flexibility of AI cluster networks.

Methodology notes

• Technical architecture analysisAI network scalability assessment

  Flat two-layer fabric and multi-plane design

  The report compares two scaling paths: a single-plane approach that relies on high-radix modular spine systems to increase capacity, and a multi-plane approach that uses MRC to divide the network into multiple lower-speed planes while preserving each accelerator's aggregate bandwidth.

• Protocol and standards analysisMRC protocol assessment

  Multipath Reliable Connection

  MRC is a NIC-terminated transport protocol and also an OCP standard; it decouples connections from a single path and distributes traffic across multiple planes to improve multipath reliability in Ethernet-based AI clusters.

• Software differentiation analysisEOS software capability stack

  Congestion control, load balancing, telemetry, and SSU

  The report emphasizes EOS software capabilities including CSIG, Fast CNP, CLB/DLB, and sub-second SSU, with the goal of reducing congestion and downtime, thereby improving accelerator utilization and shortening job completion time.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• ARISTA NETWORKS INC (ANET.US)

  Subject company of the report

  Strengths

  1.6T hardware, the EOS software stack, the MRC protocol, and support for LPO and liquid cooling together strengthen the AI networking product portfolio; key customer participation supports its technology leadership.

  Weaknesses

  The report input does not provide detailed financial forecasts, a target price, or valuation sensitivity; technological advantages still need to be validated through large-scale deployments and order conversion.

  Comparison

  MRC is described as helping Ethernet narrow the gap with proprietary solutions such as InfiniBand in terms of multipath reliability.

  Risks

  Volatility in AI networking demand, customer deployment pacing below expectations, competition from proprietary interconnect solutions, power and cooling complexity, and customer concentration.

• META PLATFORMS INC (META.US)

  Key customer or evidence of customer participation

  Strengths

  Meta is listed as one of Arista's key participating customers, reflecting recognition of Arista's data center switching capabilities by major AI/cloud customers.

  Weaknesses

  The report does not analyze Meta's own fundamentals or procurement amount.

  Comparison

  Presented alongside Anthropic, Microsoft, and Oracle as evidence of customer participation.

  Risks

  If large customers develop in-house networking solutions or adjust capital spending, supplier order timing could be affected.

• ORACLE CORP (ORCL.US)

  Key customer or evidence of customer participation

  Strengths

  Oracle is listed as one of Arista's key participating customers, supporting Arista's position in cloud and AI data center networking.

  Weaknesses

  The report does not provide Oracle's procurement scale, contract duration, or financial impact.

  Comparison

  Presented alongside Anthropic, Meta, and Microsoft as evidence of customer participation.

  Risks

  Changes in customer procurement timing, cloud capital spending, and architectural choices may affect related orders.

Key data

• Disclosed price$157.60The original text indicates ANET's price was $157.60 on 2026-06-26.
• RatingOverweightThe front page of the report lists Arista's rating as Overweight.
• 1.6T hardware1.6T switchesThe webinar discussed newly launched 1.6T switches and also mentioned LPO and liquid cooling support.
• MRC scaling capability100K+ acceleratorsThe report states that MRC enables Ethernet to connect about 100K+ accelerators in a flat two-layer fabric.
• Single-plane design scale接近40K accelerators at 800GThe single-plane path scales through leaf switches combined with high-radix modular spines, but is ultimately constrained by switch radix.
• Multi-plane design scale130K+ acceleratorsThe report gives an example in which four 200G planes allow each accelerator to retain 800G aggregate bandwidth while scaling to 130K+ accelerators using fixed-configuration switches.
• Key customer participationAnthropic, Meta, Microsoft, OracleThe report presents participation from these customers as evidence of Arista's leadership in data center switching technology.

Impact & implications

If MRC and the 1.6T platform achieve broader deployment, Arista is likely to strengthen the competitive position of Ethernet solutions in large-scale AI training networks and enhance system-level value through software capabilities. For ANET, the investment implication is positive: technology leadership, customer participation, and scalable architecture support long-term growth; however, investors still need to monitor the pace of AI capital spending, customer concentration, protocol adoption speed, and the evolution of competing solutions.

Risks

• The report does not disclose a target price, valuation model, or specific earnings forecasts, so the investment conclusion mainly relies on evidence from technology and customer participation.
• Although MRC is described as an OCP standard with participation from multiple parties, the pace of its large-scale commercial deployment, NIC ecosystem support, and customer adoption still require ongoing validation.
• Demand for AI training networks may be affected by the capital spending cycles of large cloud providers and AI customers, creating volatility in order timing.
• Ethernet solutions still face competition from proprietary interconnect solutions such as InfiniBand, and whether MRC can continue to narrow the reliability and performance gap remains to be seen.
• Deployment of 1.6T, LPO, and liquid cooling involves complexity in power consumption, density, cooling, and supply chain execution.
• J.P. Morgan discloses that it may be a market maker or liquidity provider for financial instruments related to Arista, and there are disclosures of conflicts of interest such as client relationships, shareholdings, and potential investment banking compensation.

What to watch

• Actual deployment progress and performance feedback for MRC in OpenAI and other AI training clusters.
• The degree of support from ecosystem partners such as AMD, Broadcom, Intel, Microsoft, and NVIDIA for MRC and related NIC capabilities.
• The pace of customer adoption for Arista's 1.6T switches, LPO, and liquid cooling solutions.
• Whether single-plane and multi-plane AI network architectures are adopted by large customers, and whether deployment cases at the 40K and 130K+ accelerator scale materialize.
• Whether EOS capabilities such as CSIG, Fast CNP, CLB/DLB, telemetry, and SSU can continue to improve accelerator utilization and JCT.
• Whether participation from customers such as Anthropic, Meta, Microsoft, and Oracle translates into sustained orders and market share gains.