Arista Networks
Sells Ethernet switching/routing systems (hardware built on Broadcom merchant silicon) bundled with its EOS network OS and CloudVision management; revenue is product hardware plus high-margin recurring software/subscription and support. Concentrated cloud-titan customer base.
The thesis on this name
State of AI Compute
Owns the open-Ethernet scale-out standard for AI back-end (EtherLink, $3.5B AI target, supply-gated 1-2yr) and is the systems-level merchant-Ethernet winner of the InfiniBand→UEC defection — but at ~45x fwd it competes head-on with Nvidia's +199% networking line on a thinner system (not silicon) moat with high Microsoft/Meta concentration; a hold/accumulate…
Earnings, margins, COGS & capex
Arista is a rare hyper-growth + hyper-profitable name: FY2025 revenue $9.0B (+28.6%) at 64% gross and ~47% non-GAAP operating margins, with $4.25B FCF and zero debt (FY2025 release, fact). Growth is re-accelerating on AI back-end Ethernet — Q1 FY26 grew 35.1% and management raised the FY26 AI revenue target to $3.5B and total revenue to ~$11.5B (Q1 FY26 release, fact). The debate is not demand but supply (component-gated 1-2 yrs) and whether the ~62-63% guided gross margin signals a structural step-down from the 64%+ historic level.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~36¢ is cost of goods and ~16¢ operating expense, leaving ~48¢ of operating profit (~41¢ net).
Revenue trend
Margins
down (component/tariff pressure)
down
flat/strong
flat
flat/strong
up
COGS structure
COGS is dominated by purchased merchant silicon (Broadcom Tomahawk/Jericho ASICs) plus memory, optics/transceivers, CPUs and other components, assembled by outsourced contract manufacturers. The recent margin debate is a COGS story: tariffs and rising memory/component costs pull guided non-GAAP gross margin to 62-63% vs the 64%+ historic norm. High software/EOS content keeps blended gross margin in the mid-60s.
Capex
Very light — $119.5M FY2025 (~1.3% of revenue) — because manufacturing is outsourced; capex funds R&D facilities, lab/test equipment and IT, not fabs or plants. The asset-light model is why ~47% operating margin converts to ~47% FCF margin.
Latest earnings
Beat — revenue $2.71B topped the ~$2.6B guide and consensus; EPS $0.87 beat. But the stock fell ~18% as the market judged the raised FY26 guide and ~62-63% gross-margin signal underwhelming vs lofty expectations.
FY26 raised to ~$11.5B revenue (+27.7%) and $3.5B AI revenue (from $10.5B / lower AI). Q2 FY26: ~$2.8B revenue, non-GAAP operating margin 46-47%, non-GAAP EPS ~$0.88. Demand exceeds supply; constraint expected to persist 1-2 years (fact).
- Revenue / YoY
- $2.709B / +35.1% (Q1 FY26)
- Non-GAAP operating margin
- 47.8% (Q1 FY26)
- FY26 AI revenue target
- $3.5B (raised)
- Cash + securities / debt
- ~$12.4B / $0 (Mar 31, 2026)
Growth drivers
- AI back-end Ethernet (Etherlink) — FY26 AI revenue target raised to $3.5B, roughly doubling (fact, Q1 FY26).
- InfiniBand-to-Ethernet defection accelerated by UEC Spec 1.0 — expands the addressable AI-fabric TAM.
- 800G to 1.6T speed-transition upgrade cycle at hyperscalers.
- Software/subscription (EOS, CloudVision) mix lifting recurring, high-margin revenue.
- Campus, routing and enterprise share gains vs Cisco and Juniper-HPE.
- Scale-up Ethernet (ESUN) as a 2027+ new adjacency into NVLink's rack-internal turf.
Reported financials — SEC EDGAR
Audited GAAP figures pulled from SEC filings · latest filing 2026-02-17. The audited primary-source spine — not financial advice.
Revenue — annual (GAAP)
Margins & balance sheet — FY’25
Bull & bear
Arista owns the open-Ethernet scale-out standard for AI back-end networking at the systems + software layer, growing 35% with 64% gross margins, ~47% operating margins, $4.25B FCF and a fortress balance sheet — the merchant-Ethernet winner of the InfiniBand-to-UEC defection, demand-unconstrained and supply-gated for 1-2 years.
- FY26 AI revenue target raised to $3.5B (doubling) and total revenue to ~$11.5B (+27.7%) — demand visibility is multi-year, the only limit is component supply (fact, Q1 FY26).
- EOS single-binary software + CloudVision is a sticky, hard-to-replicate moat that turns commodity Broadcom silicon into a differentiated, operationally superior system.
- Best-in-class unit economics: 64% gross margin, ~47% non-GAAP operating margin, ~47% FCF margin, ~$12.4B net cash, zero debt — funds buybacks ($1.6B FY25) through any cycle.
- UEC Spec 1.0 + ESUN position Ethernet to take both scale-out (today) and scale-up (2027+, Nvidia NVLink's turf) — the standard, not a vendor, is winning.
- The ~28% drawdown from peak on 'margin anxiety' over-discounts a raised guide — the selloff is sentiment, not a demand break.
At ~45-55x forward earnings Arista is priced for flawless AI-Ethernet dominance while facing Nvidia's hyper-scaling Spectrum-X (datacenter-Ethernet share leader at +647% YoY), gross-margin guided down to 62-63%, and dangerous two-customer concentration (26%/16% of FY25) into a lumpy AI capex cycle.
- Nvidia is now the #1 datacenter-Ethernet vendor by IDC (25.9% share, +647% YoY, Q2 2025), bundling switching with GPUs — the most credible direct threat to Arista's core thesis.
- Gross margin guided to 62-63% (Q1 FY26) vs 64%+ historic on tariffs/memory/component costs — 'margin anxiety' that drove the stock down ~18% post-print and ~28% off peak.
- Two customers = 42% of FY2025 revenue (Microsoft 26%, Meta 16%); any insourcing (white-box + SONiC on the same Broadcom chips) or capex pause is a step-down, not a wobble.
- AI back-end demand is hyperscaler-concentrated and cyclical — at ~45-55x, a single digestion quarter de-rates the multiple sharply.
- Supply-gated growth means Arista can't fully capture demand for 1-2 years, and stretched 6-8-quarter deferred-revenue qualification adds recognition risk.
What it is worth
Forward P/E + PEG cross-check vs networking/AI-infra peers, sanity-checked with a reverse-DCF on the implied growth.
~$120-140
a hyperscaler digestion quarter or Nvidia Ethernet share acceleration compresses the multiple toward ~30-35x on flat-to-lower margins.
~$180-190
guidance ($11.5B, $3.5B AI) delivered, margin steadies ~63%; ~12-month consensus target ~$190 (fact).
~$210-230
FY26 beats/raises toward $12B+, gross margin holds 63-64%, Nvidia share-take stalls; multiple re-rates to ~50x on higher EPS.
At ~$162 / ~$213B market cap on ~$11.5B FY26 revenue (+28%) and ~$3.6-3.9 non-GAAP EPS, ANET trades ~45-55x forward — a premium that the price already bakes in years of mid-20s% AI-Ethernet growth and 47% operating margins holding; the reverse-DCF requires sustained ~20%+ growth and no Nvidia share-take to justify, so the multiple is the risk, not the business.
SWOT
Strengths
- Single binary EOS network OS across the entire portfolio — one image, programmable, with CloudVision telemetry; the durable software differentiator vs box-by-box rivals (fact, 10-K).
- Best-in-class financials — 64% gross / ~47% non-GAAP operating margin, $4.25B FCF, ~$12.4B net cash, zero debt (fact, FY25/Q1 FY26).
- Systems-level leader of the merchant-Ethernet (Broadcom Tomahawk/Jericho) approach — ships 800G AI fabrics faster and cheaper than vertically integrated rivals.
- Deeply embedded with the AI hyperscalers (Microsoft, Meta) co-designing back-end clusters; Etherlink is 20+ purpose-built AI products (fact, 10-K).
Weaknesses
- Extreme customer concentration — two customers were 26% and 16% of FY2025 revenue (Microsoft, Meta); loss/insourcing of either is material (fact, 10-K).
- Gross margin guided down to 62-63% (Q1 FY26) on tariffs, memory and component inflation — the historic 64%+ may be structurally lower at AI mix.
- Reliant on Broadcom as the predominant merchant-silicon vendor — no in-house switch ASIC (fact, 10-K risk factor).
- Deferred-revenue qualification cycles have stretched to 6-8 quarters on new-product acceptance, adding revenue-recognition lumpiness.
Opportunities
- AI back-end Ethernet target raised to $3.5B for FY26 and doubling — the InfiniBand-to-Ethernet defection (UEC Spec 1.0 shipped) is the structural tailwind.
- Scale-up networking via ESUN (Ethernet for scale-up) — a 2027+ entry into the rack-internal interconnect Nvidia NVLink dominates today.
- Campus/enterprise and routing expansion — a multibillion adjacency where Arista is a share-gainer vs Cisco/Juniper-HPE.
- Software/subscription mix shift lifting recurring revenue and supporting the multiple.
Threats
- Nvidia's Ethernet (Spectrum-X) push — IDC put Nvidia at 25.9% datacenter-Ethernet share (~$2.3B, +647% YoY, Q2 2025), now ahead of Arista and Cisco; a bundled GPU+network threat.
- AI capex cyclicality — concentrated hyperscaler spend is lumpy; a digestion/air-pocket would hit a ~45-55x multiple hard.
- White-box / merchant-silicon commoditization — hyperscalers could insource switching using the same Broadcom chips + open NOS (SONiC).
- Tariff and memory/component cost inflation compressing gross margin faster than price can offset.
Moats, dependencies & bottlenecks
Moats
One programmable image across the whole portfolio; deep operational lock-in at hyperscalers, the hardest piece for merchant-silicon rivals to copy.
Hyperscaler co-design / embedded incumbency in AI back-end clusters Deep design-in with Microsoft/Meta is sticky per cluster generation but concentrated — a double-edged moat that is also dependency.
Systems + supply-chain execution on merchant silicon (800G first-mover) Ships Broadcom-based 800G AI fabrics faster/cheaper than vertically integrated rivals, but the silicon itself is buyable by competitors.
Funds sustained EOS R&D and buybacks through any AI-capex cycle without dilution or leverage.
Dependencies
Predominant merchant-silicon vendor with no in-house switch ASIC — a Broadcom price/allocation/roadmap shift hits Arista directly (fact, 10-K).
~42% of revenue from two hyperscalers; insourcing or a capex pause at either is material (fact, 10-K).
Lumpy, concentrated spend; the multiple assumes uninterrupted AI fabric expansion through FY26-27.
Thesis rides Ethernet beating InfiniBand for scale-out and reaching scale-up (ESUN); Nvidia controls the competing stack.
Outsourced manufacturing plus demand-outstripping-supply across components; constraint expected to last 1-2 years (fact, Q1 FY26 call).
Advantages
- EOS as a single binary across all platforms — superior operational consistency, programmability and telemetry vs box-by-box competitors.
- Merchant-silicon model — avoids in-house ASIC R&D cost, focuses engineering on software, and reaches new speeds (800G) first.
- Fortress financials — 64% gross / ~47% operating margin, ~$12.4B net cash, zero debt, $4.25B FCF — rare growth+profitability combination.
- Standards leadership — driving UEC and ESUN positions Ethernet (and Arista) as the open alternative to Nvidia's proprietary InfiniBand/NVLink.
- Deep AI-cluster design-in with the largest spenders, giving multi-generation visibility into back-end build-outs.
Weaknesses
- Two-customer concentration (Microsoft 26%, Meta 16% of FY2025) — revenue and capacity hostage to a handful of buyers (fact, 10-K).
- No proprietary switch silicon — reliant on Broadcom; less vertical control than Nvidia or Cisco's Silicon One.
- Gross margin pressured to 62-63% by tariffs/memory/components — historic 64%+ may not hold at AI mix.
- Valuation leaves no room for error at ~45-55x forward — a single demand or margin wobble de-rates hard.
- Cyclical, lumpy AI capex exposure with stretched deferred-revenue qualification adds quarter-to-quarter recognition risk.
- Weaker in scale-up (rack-internal) interconnect today — Nvidia NVLink dominant; Arista's ESUN entry is a 2027+ catch-up.
Bottlenecks
- Component supply — demand outstrips supply across wafers, chips, CPUs, optics and memory; management expects the constraint to gate growth for 1-2 years (fact, Q1 FY26 call).
- Broadcom silicon allocation — single predominant ASIC source caps how fast Arista can scale 800G/1.6T fabrics.
- Gross-margin ceiling — tariffs, memory and component inflation pull guided non-GAAP gross margin to 62-63%, below the 64%+ norm.
- Extended deferred-revenue qualification cycles (6-8 quarters) on new AI products delay revenue recognition and add lumpiness.
- Customer concentration — capacity and roadmap are tied to a handful of hyperscalers' deployment schedules.
Top signals & trends
Top signals
Raised on Q1 FY26; further raises confirm the AI-Ethernet ramp (fact).
The crux of the 'margin anxiety' selloff; a hold at 63%+ or recovery toward 64% would relieve the multiple.
Nvidia hit 25.9% share (+647% YoY, IDC Q2 2025) — accelerating share loss to Nvidia is the key bear catalyst.
42% of FY25 from two names; commentary on share-of-wallet and self-build is decisive.
Demand is unconstrained; faster supply relief converts backlog to revenue sooner (1-2 yr horizon, fact).
Standards momentum widens the Ethernet TAM into scale-up, Nvidia's stronghold.
Trends
The structural thesis — Ethernet becoming the open standard for scale-out AI fabrics directly grows Arista's TAM.
Drives the $3.5B AI target and 35% growth, but concentrates and adds cyclicality.
Most credible competitive threat — Nvidia now #1 in datacenter Ethernet by IDC share.
Arista's first-mover merchant-silicon execution captures the upgrade cycle; favors share gains.
Pressures gross margin toward 62-63%, the source of recent multiple compression.
Opens a large new adjacency for 2027+, but Arista enters behind Nvidia's entrenched NVLink.
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Predominant merchant switch silicon (Tomahawk high-radix, Jericho deep-buffer) — the critical single-vendor dependency.
Alternative/secondary networking silicon and interconnect components.
800G/1.6T optics supply — part of the component set management flags as supply-constrained.
Outsourced manufacturing of Arista systems (asset-light model, ~1.3% capex intensity).
Memory is a named cost-inflation/supply pressure on gross margin.
Largest customer — 26% of FY2025 revenue; Azure AI back-end build-out (fact, 10-K).
Second >10% customer — 16% of FY2025 revenue; AI cluster networking (fact, 10-K).
Cloud/AI infrastructure customer in the hyperscaler tier.
Other cloud titans + Tier-2 service providers / enterprises Financials, high-tech enterprise and neocloud/AI specialists round out the base beyond the two concentration names.
Now #1 in datacenter Ethernet by IDC share (25.9%, +647% YoY, Q2 2025); bundles networking with GPUs — the central competitive threat.
Larger overall, pushing Silicon One + Nexus into AI data center and campus; the incumbent Arista keeps taking share from.
Post-2025 Juniper acquisition combines AI-native (Mist/Apstra) networking across campus + data center against Arista's adjacencies.
Key supplier but its merchant silicon + open NOS also enables hyperscaler self-build, a structural commoditization risk.
Open-networking hardware on the same Broadcom silicon with SONiC; a hyperscaler insourcing vector.
Custom ASIC and interconnect for hyperscalers; an indirect threat via in-house networking designs.