Arista Networks (ANET)

Ethernet AI fabric leader — the pure-play public bet that AI back-end networking standardizes on Ethernet (RoCE / UEC) rather than NVIDIA's InfiniBand.

1. Core Product / Service

Arista (NYSE: ANET) builds high-performance Ethernet switches and the EOS (Extensible Operating System) that runs them. Three product axes:

• Data center cloud switches — flagship 7800R series (modular spine), 7700R series (scale-across routers), 7060/7280 leaf, with Etherlink AI fabric portfolio designed for GPU back-end (51.2T Tomahawk 5 and now Tomahawk 6-class 102.4T)
• Campus / enterprise — wired + wireless campus switches running the same EOS
• EOS + software — a single operating system image runs across every platform; CloudVision management; AVD (Arista Validated Designs) automation; Etherlink-specific AI fabric features (congestion control, load balancing, telemetry)

Arista does not make its own switch ASICs — it primarily uses broadcom Tomahawk (high-radix, low-latency) and Jericho (deep-buffer routing) silicon. The defensible moat is EOS plus deep tuning for hyperscale fabrics: a single OS image and management plane that runs identically across campus, data center back-end, and scale-across WAN routers.

2. Target Users & Pain Points

• Cloud Titans — Microsoft, Meta as anchor 10%+ customers; Oracle and Google rising as incremental 10% candidates (Google via Virgo Fabric deployment)
• Tier-2 hyperscalers — Apple, ByteDance, Tencent, Alibaba (regional mix)
• AI labs and neoclouds — coreweave, crusoe-energy, nebius, plus AI-native enterprises
• Enterprise + service provider — financials, government, telcos building Ethernet-based AI fabrics rather than InfiniBand

Pain solved: open Ethernet AI fabric that doesn't lock the buyer to NVIDIA Mellanox / Quantum InfiniBand. Buyers who want multi-vendor GPU compute (or want to keep NVIDIA honest on networking pricing) standardize on Arista + Broadcom Tomahawk + open RoCE / UEC. For hyperscalers with hundreds of thousands of GPUs, that optionality is worth a lot.

3. Competitive Landscape

Arista's edge: EOS uniformity + close Broadcom partnership + 2026 product timing on Etherlink + 7800R for scale-across data-center interconnect. Disadvantage: relies almost entirely on Broadcom silicon and has limited internal ASIC IP.

4. Unique Observations

• Section 4 focus — 1 MW build-cost share: in a 1 MW AI cluster (120 H200-class GPUs across ~15 nodes), networking is **6-12% of build cost ($2.4-5.4M/MW)**, split roughly half-and-half between switches and optics + cabling + NICs. If the cluster runs RoCE Ethernet, Arista typically captures most of the switch line — ~$0.8-1.8M/MW, or ~5-10% of total build. The other ~$0.6-1.4M/MW in optics flows to coherent, lumentum, innolight.
• Q1 2026 financials: revenue $2.709B (+35.1% YoY, +8.9% QoQ); GAAP diluted EPS $0.80, non-GAAP $0.87; non-GAAP operating margin 47.8% — held flat YoY despite rising component costs and mix [1][2][6]. This margin is the headline anomaly: 47-48% non-GAAP operating margin is structurally SaaS-like for a hardware vendor.
• AI revenue specifically: 2026 AI fabric target raised to $3.5B (from $3.25B), with scale-across (data-center-to-data-center via 7800R/7700R routers) contributing at least one-third of that [4][5]. Full-year 2026 revenue guide raised to +27.7% growth ≈ $11.5B [4].
• Backlog increase as AI share grows: Arista doesn't publish a backlog figure the way vertiv or ge-vernova does, but management commentary emphasizes deferred revenue + visibility extending into 2027. Demand is described as outstripping supply across wafers, chips, CPUs, optics and memory; this constraint is expected to last 1-2 years [6].
• Customer concentration is the perpetual debate: Microsoft + Meta are >10% disclosed customers; Oracle and Google approaching 10%. If Oracle and Google both step over the threshold, top-4 customer share crosses 50%+ of revenue. This is the central risk: concentrated AI exposure cuts both ways.
• Bottleneck — supply, not demand: every line item — wafer fab capacity, Broadcom Tomahawk silicon, optics, NICs, even memory — is constrained for 1-2 years per management [6]. Arista's revenue is effectively rationed by broadcom silicon allocations and by coherent/lumentum/innolight optical module deliveries. This is why ANET, Broadcom, and the optics names trade as a complex.
• NVIDIA / AI Lab binding: Arista's customer set is hyperscaler + enterprise, not labs directly. The Oracle bump in 2026 is significant — Oracle has Stargate / OpenAI capacity it is networking on Arista Ethernet, an early indicator that the largest AI cluster in the world is not InfiniBand-only.
• Etherlink / XPO announcement: at the Q1 2026 print Arista announced XPO high-density liquid-cooled pluggable optics — claims to cut networking racks by up to 75% and floor space by ~44% [6]. This is direct response to power and density constraints inside hyperscale AI fabrics.
• Token cost chain — the "shovel seller": Arista sits in L1 A.d (networking — switches + OS). Every prompt-to-token round trip traverses an Arista switch in ~70% of non-NVIDIA-stack hyperscale deployments. Per-GPU networking depreciation is ~$0.50-1.00/hr of GPU time — Arista captures roughly half of the switch share of that.

5. Financials / Funding

• Listed: NYSE: ANET (IPO 2014)
• Q1 2026 revenue: $2.709B (+35.1% YoY) [1][2]
• Q1 2026 GAAP / non-GAAP diluted EPS: $0.80 / $0.87 [1][2]
• Q1 2026 non-GAAP operating margin: 47.8% [1][6]
• Q1 2026 non-GAAP gross margin: 62.4% [6]
• 2026 revenue growth guidance: ~27.7% → ~$11.5B [4]
• 2026 AI fabric target: $3.5B (raised from $3.25B) [4][5]
• 2026 gross margin guidance: 62-64% (non-GAAP) [4]
• 2026 operating margin guidance: ~46% (non-GAAP) [4]
• FY2025 revenue: $9.0B+ (calibration from L1a research notes)
• Market cap (mid-2026): ~$200B+ range

6. People & Relationships

• CEO: Jayshree Ullal (since 2008)
• Chairman: Andy Bechtolsheim (founder, Sun Microsystems co-founder, early Google angel)
• CFO: Chantelle Breithaupt
• Founded: 2004 by Andy Bechtolsheim, David Cheriton, Ken Duda
• HQ: Santa Clara, California
• 10%+ customers: Microsoft, Meta (disclosed); Oracle and Google approaching 10%
• Other large customers: Apple, ByteDance, Alibaba, coreweave
• Key suppliers: broadcom (Tomahawk, Jericho switch ASICs), coherent, lumentum, innolight (optics)
• Strategic context: viewed as the Ethernet-AI-fabric play, in direct architectural competition with nvidia InfiniBand