ANET › analyze
ANET · Analyze
A portfolio-specific passage was removed from the public build.
First analysis file ever written on this name. Held with
[—]conviction, aTrim $210+level nobody derived, and anAdd $140-158zone sitting 31% below spot.
⚠️ Verdict up front
HOLD · conviction 6.5 · fair value $175–215 (central ~$195) · spot $201.09 sits just above centre, inside the band.
This is the better business of the two names analysed today by a wide margin — and it is priced accordingly. Four things frame it:
- The watchlist multiple was wrong. ANET does not trade at 38.0x forward. That is the FY2027 figure. The honest current-year multiple is 48.9x.
- The business is executing exceptionally. FY2026 guide raised a third time to ~$12.6B (+40%), 49.9% non-GAAP operating margin, $13.3B net cash, no debt.
- The moat is leaking into the P&L, and the company says so. Gross margin 65.2% → 62.9% in a year, cause disclosed in the 10-Q as "an increased proportion of our sales to large end customers who generally receive higher discounts."
- The AI target was NOT raised while revenue was raised $1.1B. The incremental guide is core datacenter, enterprise, campus and routing. Arista is broadening — but the stock is priced as an AI pure-play at 24x trailing sales, above any quarter-end P/S in its ten-year history.
1 · Fundamentals — outstanding, with no asterisks
Scorecard
| Metric | Value | Read |
|---|---|---|
| TTM revenue | $10.54B | Q2-26 +37.7% YoY; first $3B quarter |
| TTM operating income | $4.55B (43.1%) | |
| Q2-26 GAAP operating margin | 45.4% (was 44.7%) | record |
| Q2-26 non-GAAP operating margin | 49.9% (was 48.8%) | |
| TTM GAAP diluted EPS | $3.17 | |
| TTM FCF | $5.16B — 48.9% of revenue | |
| Capex | $151M = 1.4% of revenue | effectively capital-free |
| Revenue 3y CAGR | 27.1% | |
| Net income 3y CAGR | 37.4% | |
| FCF 3y CAGR | 111.7% | off a low FY2022 base |
| ROIC (TTM, reported) | ~29.8% | |
| ROIC ex-cash | effectively unbounded | $11.9B of the $12.4B invested capital is net cash |
| Debt | zero | $13.3B cash + securities |
| Current ratio | 2.96 | |
| Diluted shares | 1,271.2M → 1,276.0M (+0.4%) | near-flat |
| FCF conversion | 121% of net income |
This is a genuinely rare financial profile. 43% operating margins, ~30% reported ROIC (and functionally infinite ex-cash), 1.4% capex intensity, no debt, and free cash flow at 49% of revenue. Per §1 of the framework, there is nothing to flag on health. The questions on this name are all about price and customer power, not solvency or quality.
Capital allocation — one change worth flagging
| TTM | $ |
|---|---|
| Operating cash flow | $5.31B |
| Capex | $151M |
| Free cash flow | $5.16B |
| SBC | $502M |
| Buybacks | $620M (all of it in Q4-25) |
Buybacks covered 123% of SBC on a TTM basis — but H1-2026 repurchases were $0, against $983M in H1-2025. $817.9M remains of the May-2025 $1.5B authorisation. A company generating $5.2B of annual FCF with $13.3B of net cash and zero debt bought back nothing for two quarters. That is either supply-chain cash preservation (management has committed $9.7B of non-cancellable purchase commitments, up from $3.6B a year ago) or a view on its own price. Neither reading is bullish for the multiple; the first is the more likely.
Per-share / owner's-eye view
| Revenue/share | $8.26 |
| FCF/share | $4.04 |
| FCF yield | 2.03% |
| P/FCF | 49.2x |
2 · The forward-EPS resolution — same pitfall, same direction
Yahoo's payload carries both fields:
epsCurrentYear : 4.11295 priceEpsCurrentYear : 48.89 <- FY2026, the right one
epsForward : 5.15956 forwardPE : 38.97 <- FY2027
trailingEps : 3.16 trailingPE : 63.64 <- TTM GAAP
[[pitfall-vendor-forward-eps-is-the-wrong-fiscal-year]] fires again, on the second name in the same
session. Yahoo's eps_trend also dates the revision precisely — both years stepped up on the
2026-08-05 print (FY2027 +15.5% in a day, FY2026 +13.1%).
Bottom-up reconstruction from the company's own guide — the check the pitfall demands:
| Revenue | Non-GAAP dil. EPS | GAAP dil. EPS | |
|---|---|---|---|
| Q1 2026 reported | $2,709.0M (+35.1%) | $0.87 | $0.80 |
| Q2 2026 reported | $3,035.7M (+37.7%) | $1.02 | $0.95 |
| Q3 2026 guide | ~$3,300M | $1.06–1.08 | — |
| Q4 2026 implied | ~$3,555M | ~$1.13–1.16 | — |
| FY2026E | ~$12.6B (+40%) | ≈$4.08–4.13 | ≈$3.80 |
The bottom-up lands within 1% of the aggregator consensus ($4.113). FY2026 is settled.
The corrected multiple ladder at $201.09
| Basis | EPS | P/E |
|---|---|---|
| TTM GAAP | $3.17 | 63.4x |
| TTM non-GAAP | ~$3.48 | 57.8x |
| FY2026E non-GAAP | $4.11 | 48.9x ← the honest forward number |
| FY2026E GAAP | $3.80 | 52.9x |
| FY2027E non-GAAP | $5.16 | 39.0x ← what Yahoo shows |
| FY2027E GAAP | ~$4.55–4.65 | 43.2–44.2x |
⚠️ Independence caveat. Yahoo and stockanalysis.com agree to five decimals ($4.11295) because they share the LSEG feed — they are one source, not two. The only genuinely independent confirmation is the bottom-up above, which agrees. FY2027's $5.16 has no independent non-GAAP confirmation; every second source is either GAAP (~$4.55) or stale. Treat it as reasonable but single-feed. Barchart, MarketBeat, WallStreetZen, Simply Wall St and Zacks are all either GAAP or pre-print vintage.
Own-band test
Unlike AMD, ANET's GAAP P/E is usable — the statements are clean and there is no acquisition amortisation of consequence.
| Period | GAAP EPS | Price | P/E |
|---|---|---|---|
| FY2022 | $1.07 | $31.50 | 29.4x |
| FY2023 | $1.65 | $64.67 | 39.2x |
| FY2024 | $2.23 | $115.23 | 51.7x |
| FY2025 | $2.75 | $157.69 | 57.3x |
| TTM now | $3.17 | $201.09 | 63.4x |
The band has re-rated every single year: 29 → 39 → 52 → 57 → 63x. ANET is at the top of its own band, and the band itself has been climbing for four years. There is no mean-reversion anchor to lean on — which cuts both ways, but means a buyer here is underwriting a fifth consecutive re-rating. P/S is 24.06x trailing, above any quarter-end reading in the company's ten-year history (prior high ~21.7x). [[pattern-ai-levered-fields-trade-above-own-band]] holds: −6.4% off the 52-week high, +75.6% off the low.
3 · Moat
Evergreen rating: 6.0 / 10. (Moat Analyst)
The framing that survived the stress test: a narrow, non-durable moat where the money is, and a real, durable moat where the money isn't.
What the moat actually is, mechanically
84.1% of FY2025 revenue is product (hardware); 15.9% is service. EOS is not sold separately — it is bundled into a box built on Broadcom merchant silicon Arista does not own, assembled by contract manufacturers Arista does not own. The only place software economics are visible is Service, at ~82% gross margin, and that is mostly support contracts.
Arista is a systems company that prices like a software company.
The 10-K names its own risk with the correct historical analogue:
"customers may prefer to address their network switch requirements by licensing software operating systems separately and placing them on 'white box' hardware rather than purchasing integrated hardware products as has occurred in the server industry"
And the top two customers have already done it. Meta runs FBOSS on Arista's 7700R4 hardware, not EOS. Arista ships SAI so SONiC can run on Arista boxes instead of EOS for Microsoft. The customers representing 42% of revenue have unbundled the software.
| Moat source | Verdict |
|---|---|
| Intangibles (EOS single binary + CloudVision) | Real, but tiered. One unmodified image across campus/DC/routing, stateful in-service upgrade, streaming telemetry. Worth a great deal to an enterprise with 12 network engineers; worth roughly zero to Meta. |
| Switching costs | Real in enterprise, near-zero at the top two. |
| Cost advantage | No. Same Broadcom silicon, same CMs. The 45% operating margin comes from a 13.5% opex ratio — an operating-cost advantage that is a consequence of concentration, not a defence against it. |
| Network effects | No. |
| Efficient scale | No. |
The honest answer: the moat is software + support + a decade of operational trust, and those are not equally distributed across the customer base. Against a customer with its own network engineering organisation, only the third survives — and "trust" here means specifically that Arista ships VOQ/deep-buffer scheduled-fabric systems at 27,000×800GbE scale that work on first deployment. Meta chose Arista's chassis while explicitly rejecting Arista's software. That is the cleanest available natural experiment.
One genuine counterweight: the disclosed Microsoft account went 18% → 20% → 26% of a rapidly growing base. If white box were structurally displacing Arista there, that number would fall. The moat is real even where it is not lock-in — it just gets re-earned every purchase cycle.
The adversarial stress test — and it drew blood
(a) Nvidia. IDC, calendar Q1 2026: Nvidia became #1 in datacenter Ethernet switching by revenue for the first time — $2.1B, 21.5% share, +192.7% YoY. Arista second at 20.7%. Arista has lost the top slot in its own core market.
And two Arista customers have publicly shifted back-end networking to Nvidia. Meta is integrating Spectrum Ethernet into FBOSS via Minipack3N — Nvidia Spectrum-4 ASIC, 51.2 Tb/s, built by Accton, running FBOSS, "used to support both frontend and backend data center fabrics." Oracle is building on Spectrum-X + Vera Rubin. Nvidia's Spectrum-6 (July 2026) names CoreWeave, Microsoft, Nebius, Oracle, xAI and Tesla as first deployers. Nvidia networking revenue reached $14.8B in Q1 FY27 (+199%) and grew 18% sequentially in Q2 FY27, with Spectrum-X Ethernet +2.6x YoY.
Arista management concedes scale-up outright: "we live in an NVIDIA world … there's very little participation from Arista or anybody else's scale up."
Nvidia's Spectrum-XGS attacks precisely the segment Arista is now selling as its next growth engine — scale-across, ~$1.2B of the 2026 AI target, $15–20B TAM by 2030.
(b) White box + SONiC — the finding that most changes the picture. Dell'Oro, AI back-end networks, 1Q26: Celestica regained the leading position — an ODM is #1 in AI back-end Ethernet switching. Nvidia a close second, Arista third, Cisco fourth with the largest share gain. Full-year 2025: Celestica + Nvidia together ~50%.
Arista's own framing (Ullal, Q2 2026): white box is "a tactical solution … in use cases that are simple … where the actual amount of software and system requirements are low" and customers "that have large amounts of staff like the titans often do." Arista's defence is that AI architecture resets every 12–18 months and branded systems win where "entropy" is high. The counter is that AI back-end is the most homogeneous, highest-volume, lowest-entropy topology in the datacenter — the textbook white-box profile. The Celestica #1 datapoint says the ODMs are winning that argument right now.
(c) Cisco — no longer safe to dismiss. FY2026: $9.3B of AI infrastructure orders (vs ~$2B in FY2025), ~$4B of AI revenue recognised, guided to $7.5B in FY2027. Three additional hyperscaler design wins including a scale-across P200 win. All four major hyperscalers at triple-digit order growth. Cisco's FY2027 AI revenue guide is more than 2x Arista's entire $3.5B AI fabrics target. And in May 2026 Cisco began offering SONiC on the Nexus 9000, TAC-supported — the other incumbent NOS vendor decoupling its OS from its hardware, a direct read-across to EOS pricing power.
(d) Broadcom — the supplier is the ceiling. The 10-K is explicit: "we are primarily reliant upon our predominant merchant silicon vendor, Broadcom" and those vendors "may become competitive with us by selling merchant silicon for 'white boxes' with open-source network operating systems."
The sharpest single datapoint in this report: Broadcom's Tomahawk 6 (102.4 Tb/s) has been in production volume since March 2026, and Celestica's TH6 box has been orderable since April 2026 — with Accton shipping the same silicon on SONiC. Arista's own TH6 platform, the 7060XE7, is in trials in H2 2026 with "single digits" of customers and real production in 2027. At 1.6T the hardware differentiation window is effectively zero; the remaining moat is EOS + CloudVision, which is exactly what Cisco's SONiC move attacks from the other side.
Broadcom does not need a path up the stack. It captures the value by arming Celestica and Accton and letting SONiC be free. The threat is not that Broadcom becomes Arista; it is that Broadcom makes being Arista unnecessary.**
Concentration is a moat question, not a risk line
FY2025 10-K, verbatim: "Sales to one end customer represented 16%, 15%, and 21% … and sales to the other end customer represented 26%, 20%, and 18% … for the years ended December 31, 2025, 2024, and 2023."
| 2023 | 2024 | 2025 | |
|---|---|---|---|
| Customer A (Microsoft, per BofA) | 18% | 20% | 26% |
| Customer B (Meta, per BofA) | 21% | 15% | 16% |
| Combined top-two | 39% | 35% | 42% |
⚠️ Arista names no customer. microsoft and meta platforms each appear zero times in the
FY2025 10-K; prior filings named them. The per-name attribution is BofA's, relayed by press — not
disclosure. The combined series is primary-sourced and solid; do not build a multi-year
per-customer narrative on the attribution. A vendor de-anonymising in the direction of less
disclosure while concentration rises 7 points is a small negative signal in itself.
The subtler finding: the Cloud & AI Titans bucket stayed flat at 48% of revenue — but the top two went from 73% to 88% of that bucket. The diversification showed up in the AI & Specialty Providers line (17% → 20%, neoclouds), not in the titans.
The strategic question — moat or invitation? A 63% gross margin sustained against buyers who (i) employ world-class network engineers, (ii) have built and open-sourced the substitute software, (iii) run that substitute on Arista's own hardware, and (iv) have an explicit strategic interest in commoditising infrastructure, is not evidence of a moat. It is a standing invitation with a delay fuse. What buys the delay is real — architectural churn and supply scarcity. The day AI back-end topology stabilises, Arista's hyperscaler gross margin goes to ODM economics.
🚩 And it has started. The give-back is already in the P&L.
| Q2-25 | Q2-26 | |
|---|---|---|
| GAAP gross margin | 65.2% | 62.9% |
| Product gross margin | 62.3% | 59.8% |
| Non-GAAP gross margin | 65.6% | 63.4% |
10-Q cause, verbatim: "The decrease was primarily driven by an increased proportion of our sales to large end customers who generally receive higher discounts."
That is the concentration tax showing up in the accounts, not a theory. Important nuance in Arista's favour: operating margin went the other way — 44.7% → 45.4% GAAP, 48.8% → 49.9% non-GAAP — because opex leverage is more than offsetting the price give-back. So the erosion is real, disclosed, and has not yet reached operating income. That is the number to watch, and it is why this is a HOLD rather than a TRIM.
Enterprise/campus — the only part where the moat as advertised exists
Campus + WAN at a $750–800M run rate targeting ≥$1.25B in FY2026 (~60% growth); Leader in the 2026 Gartner Magic Quadrant for Enterprise Wired and Wireless LAN; VeloCloud SD-WAN bought from Broadcom for $300M. Enterprise broadly is 32% of revenue.
Directionally working, materially not yet. At ~10% of revenue growing into a hyperscaler business growing 40%, campus is not diluting concentration — concentration rose 7 points in FY2025 while campus grew. The diversification is losing the race against its own core.
HPE-Juniper (closed July 2025, $14B) is a net negative for campus economics: the field went from three subscale challengers to one integrated full-stack alternative to Cisco with a real AIOps asset in Mist. Mildly positive on timing — 18–24 months of integration disruption, and Arista's Gartner Leader placement lands in that window.
The strategic tension worth naming: Arista's extraordinary operating margin is a function of the concentration investors want it to escape. Scaling campus means giving up part of the 45%.
4 · Sentiment
The +27.9% month is measured off a trough, and two-thirds of it is not about Arista
| Date | Close | Move | Driver |
|---|---|---|---|
| Jul 29 | $157.97 | −6.9% | Alphabet's capex raise was panned; whole complex sold off. This is the base of the "+27.9%". |
| Jul 30–31 | $180.35 | +8.2%, +5.5% | Big Four 2026 capex guides landed at ~$630B, +62% YoY |
| Aug 4–5 | $190.51 → $197.31 | gapped +10.3%, faded to +3.6% close | Q2 beat-and-raise. 12+ firms reset targets ~$195-200 → ~$250 within 24h |
| Aug 12 | $210.50 | +6.4% | All-time closing high. No company catalyst identified — coincides with Cisco's FY26 print |
| Aug 19–20 | $183.75 | −12.7% from peak | give-back, no news |
| Aug 26 | $202.25 | +5.9% | Nvidia Q2 FY27 — sector beta |
Roughly one-third company-specific, two-thirds sector/capex beta. The gap-and-fade on Aug 5 is the most informative bar in the series: the market took the raise and immediately sold 8% of it.
⚠️ Two watchlist facts need correcting. The "6% below the 52-week high" measures against $214.89, an intraday spike on the earnings gap that lasted minutes — the stock closed that day at $197.31, a $17.58 fade. The highest close ever is $210.50. ANET has never closed within 2% of $214.89.
Fuel is spent: short interest fell from 19.89M shares (7/15) to 12.87M (8/14) — a 35% cover of ~7.0M shares into the run. Short interest is now 1.02% of shares out.
The Q2 print
| Q2'26 | Q2'25 | |
|---|---|---|
| Revenue | $3,035.7M (+37.7%) — beat guide of $2.8B, cons $2.83B | $2,205M |
| Non-GAAP dil. EPS | $1.02 — beat $0.88 cons by $0.14 | $0.73 |
| Non-GAAP op margin | 49.9% | 48.8% |
| Opex | $411.0M (13.5% of revenue) — R&D 9.2%, S&M 3.6%, G&A 0.8% | |
| International | $697.8M = 23.0% (from 15.5% in Q1) | 21.8% |
| Cash + securities | $13.34B, zero debt | |
| Deferred revenue | $6.87B (from $2.79B at FY2024) | |
| Purchase commitments | $9.7B | $3.6B |
| Buyback | $0 |
Guidance, raised a third time: FY2026 ~$12.6B (+40%); AI fabrics ≥$3.5B; campus ≥$1.25B; gross margin band held at 62–64%; operating margin target raised to 48–49%. Q3: ~$3.3B, non-GAAP EPS $1.06–1.08.
| Event | FY2026 revenue guide | AI target |
|---|---|---|
| Analyst Day 2025 | $10.5B | $2.75B |
| Q4'25 print | $11.25B | raised to $3.25B |
| Q1'26 print | $11.5B (read as a miss; stock fell) | raised to $3.5B |
| Q2'26 print | $12.6B (+$2.1B over Analyst Day) | UNCHANGED at $3.5B |
🔑 The single most important disclosure: what the $1.1B raise is made of
Management raised total revenue by $1.1B and left the AI target untouched. Ullal, asked directly:
"the question is not whether it will go up. The question is what is that number? … that's why we're not holding ourselves to a number."
And on the source of the raise:
"this is our core Arista product line we're talking about. Campus and AI are the newcomers … why is your core data center front end so flat. And so I'd like to first make the observation that it isn't going to be flat … I expect great contributions from the enterprise … and likewise with routing."
The raise is core datacenter / front-end / enterprise / routing — not AI back-end. That is arguably better for the concentration story and worse for the AI-pure-play multiple the stock carries.
Management tone — confident on demand, explicitly hedged on supply:
Ullal: "I don't want you to believe that suddenly, we waved the magic wand and all our problems have gone away. The industry is going to have a 2-year problem. I don't think we get out of it as an industry until 2028." Breithaupt, on the deferred balance: "increased customer-specific acceptance clauses and an increase in the volatility of our product deferred revenue … the deferred balance can move significantly on a quarterly basis, independent of underlying business drivers."
This was a supply-released quarter, not a demand-discovery quarter. The raise came from supply-chain work (three CMs, three distribution sites, memory secured through 2026), not from new demand. Unusually candid about the low signal quality of its own deferred-revenue line — credit to management for saying so.
Positioning — one-sided, with no dissent at all
8 Strong Buy / 22 Buy / 0 Hold / 0 Sell out of 30. Mean target ~$242, median $248, high $289 (Barclays), low $185. There is not one sell-side hold or sell rating on this name; the single Hold that existed pre-print disappeared. The only downgrade in the window is Weiss Ratings (8/12), a quantitative shop, single-sourced.
⚠️ Source spread is wide: Yahoo/stockanalysis $241.82 (one LSEG feed), TipRanks $247.80, MarketBeat $226.05 (retains un-refreshed pre-print targets), TradingView's $330 high is bad data, contradicted by every other source and by the published tape.
The bear case in print (Trefis, 2026-08-26) is a reverse-DCF on multiple compression, not a growth attack: at a 25.2x end-state multiple the price requires an 18.5% five-year revenue CAGR, or 33% over three years. Its weakness: it runs on a FY2027 EPS of $4.47 — exactly the 30-days-ago consensus, i.e. pre-Aug-5 vintage — and a market cap ~5% stale. The bear case is built on stale earnings. It is still directionally right that ~38x has been Arista's historical comfort zone.
🚩 Insiders — $1.48B sold, $0 bought, in 20+ months
Not one open-market insider purchase since at least December 2024. 79 sales + 21 sale/option- exercise rows, zero purchases.
| Insider | 12m shares sold | Value |
|---|---|---|
| Jayshree Ullal (CEO) | 5,221,667 | $907.5M |
| Andreas Bechtolsheim (Chief Architect, >10% owner) | 2,656,848 | $458.9M |
| Kenneth Duda (President, CTO) | 641,999 | $93.7M |
| Charles Giancarlo (Director) | 146,000 | $21.8M |
| Others | ~17,207 | ~$2.6M |
| Total | 8,683,721 | $1,484.5M |
⚠️ Yahoo's insider_purchases reports "Purchases: 552,226 shares / 25 txns." These are RSU
vestings and option exercises at $3.52/$14.15/$15.28 strikes — not buys. That is
[[pitfall-yahoo-insider-purchases-counts-rsu-grants]] firing again. Do not quote it as insider
buying.
Most of this is routine and verified so — Ullal's 8/12 sale carries aff10b5One = 1 under a
plan adopted 2025-11-14; Bechtolsheim's is an unbroken 205k–300k-share cadence every 2–3 weeks;
Giancarlo sells exactly 8,000 shares on the 1st–3rd of every month, 12 months running.
What is not routine: Ullal sold $221.2M on the earnings date itself (8/5) and $119.4M more on 8/12 — ~$340.6M in August alone, the largest single month in the file. It is 10b5-1 covered by a plan predating the quarter, so the size is the flag, not the legality. Her stake is down ~31% in 14 months.
The "17% insider ownership" inverts on inspection
Verified against the DEF 14A (2026-04-16):
| Holder | Shares | % |
|---|---|---|
| Bechtolsheim Family Trust | 183,799,896 | 14.6% |
| Jayshree Ullal | 29,300,817 | 2.3% |
| All other execs + directors | ~4.2M | <0.4% |
84.6% of the entire insider block is one founder's trust — being liquidated on a fixed schedule at roughly $460M/year. The headline "unusually high insider ownership" is not broad management alignment.
Institutions: 73.3% of shares out, 88.6% of float. The active managers trimmed into the run while the passives held — T. Rowe −20.8%, Capital Research −10.8%, FMR −9.5%, vs Vanguard +0.6%, State Street +3.0%.
5 · Valuation
DYT and DDM are N/A — ANET pays no dividend. Stated, not faked.
Graham IV = √(22.5 × 3.17 × 11.73) = $28.92. Price is 7.0x Graham. Wrong tool for an asset-light business with no book anchor; ~zero weight.
Reverse DCF (10% discount, 3% terminal, TTM FCF $5.16B): to justify $253.6B, FCF must compound 19.3%/yr for ten years, reaching $30.1B. Demanding, but materially less extreme than AMD's 28.1% — and against a business that already converts 49% of revenue to FCF.
Bogle expected return, FY2026 GAAP $3.80 base, no dividend:
| Growth | Exit 32x | Exit 40x |
|---|---|---|
| 18%/yr | +6.7%/yr | +11.6%/yr |
| 22%/yr | +10.3%/yr | +15.4%/yr |
| 25%/yr | +13.0%/yr | +18.2%/yr |
Fair value triangulation. On FY2026 non-GAAP $4.11 at 40–48x → $164–197. On FY2027 non-GAAP $5.16 at 38–45x → $196–232, discounted one year at 10% → $178–211.
Fair value $175–215 · central ~$195 · spot $201.09 is just above centre, inside the band.
| Level | ||
|---|---|---|
| Entry | $160–180 | |
| Strong buy | <$150 | 52wk low $114.52 |
| Trim | 45x fwd |
Trim basis, stated explicitly as [[pitfall-multiple-trim-inherits-the-broken-vendor-field]] requires.
site.pycomputesEPS = price ÷ forwardPE= $201.09 ÷ 38.97 = $5.16, which is the FY2027 non-GAAP figure.Trim 45x fwdtherefore renders ≈$232 (+15% from spot). Unlike AMD, the multiple form is safe here: ANET is a compounder rather than a peak-cycle cyclical, the offset year is consistent, and there is no warrant overhang corrupting the share count. But the multiple is on FY2027, not FY2026 — check the rendered level after the next build, and re-confirm the basis at the next/analyze.
6 · Tensions surfaced (Phase 3)
Fundamentals vs Moat — genuinely unresolved, and named as a risk factor rather than buried. Fundamentals sees the best financial profile in the portfolio's coverage: 43% operating margins, ~30% ROIC, 49% FCF margin, zero debt, +40% guided growth. Moat sees an ODM at #1 in AI back-end, Nvidia at #1 in datacenter Ethernet, the top two customers at 42% running their own NOS on Arista's own hardware, and gross margin already compressing for exactly the disclosed reason.
Both are right, and the resolution is a timing question, not a truth question. The moat analyst's own mechanism — that hyperscalers vertically integrate behind the frontier, not at it — means Arista keeps earning its position as long as AI network architecture resets every 18 months. The margin series says the price of staying is rising. The tell to watch is not a lost customer announcement; it is gross margin. It has gone 65.2% → 62.9% in a year, and the moment that reaches operating income, this becomes a TRIM. Operating margin is still expanding, so it has not.
Sentiment vs Valuation — resolved in Valuation's favour, weakly. Sentiment reports a clean beat-and-raise and a third guidance increase. Valuation observes 63x trailing, 48.9x current-year, 24x sales above any prior reading, zero sell-side dissent, $1.48B of insider selling against zero buying, and active managers trimming into the run. A raise this good producing a +3.6% closing day is the market telling you the multiple already contained it.
A portfolio-specific passage was removed from the public build.
8 · Verdict
HOLD · conviction 6.5 · fair value $175–215 · spot $201.09
Not a trim — the business is executing at the top of its range, operating margin is still expanding, and the third guidance raise of the year was real. Not an add — there is no margin of safety, the stock sits above the centre of fair value at the top of a band that has re-rated four years running, and the structural evidence got worse this quarter, not better.
The bull case. One of the genuinely great financial profiles available anywhere: 43% operating margin, ~30% ROIC on almost no capital, 49% FCF conversion, $13.3B net cash, no debt, +40% guided revenue growth, and a management team candid enough to warn about its own deferred-revenue quality. Enterprise and campus are growing 60% into the one segment where the moat is genuinely durable, and Arista has demonstrably re-won its lead at 100G, 400G and 800G while white box was supposed to be finishing it off — with the disclosed Microsoft share going 18% → 20% → 26% through exactly that period.
The bear case. Arista's hyperscaler business is a systems-integration lead over merchant silicon, and every actor above and below it is working to shorten that lead. Broadcom absorbs the hard problems into the ASIC and hands them free to Celestica and Accton — its TH6 box has been orderable since April 2026 while Arista's own TH6 platform reaches production in 2027. Nvidia bundles the network into the GPU decision and is now #1 in datacenter Ethernet. Cisco arrives with its own silicon, a $7.5B FY2027 AI revenue guide, and a willingness to ship SONiC on its own switches. And the two customers who are 42% of revenue have already built, open-sourced and deployed the software substitute on Arista's own hardware. The mechanism does not announce itself — it is already in the 10-Q, as gross margin walking from 65.2% to 62.9% "primarily driven by an increased proportion of our sales to large end customers who generally receive higher discounts."
Arista is not at meaningful risk of losing Microsoft and Meta. It is at risk of keeping them at ODM economics — growing revenue into a shrinking margin while the multiple assumes the opposite.
Conviction 6.5 — a better business than AMD at a fairer price, with a narrower moat than its financials imply and no discount for the risk.
Key risks
- Gross margin reaching operating income. 65.2% → 62.9% in one year, disclosed cause. This is the single number to track; it is the de-rating trigger at ~49x.
- Concentration 42% and rising, with the titan bucket going 73% → 88% two names.
- Nvidia #1 in datacenter Ethernet; an ODM #1 in AI back-end. Both are new facts this year.
- $9.7B of non-cancellable purchase commitments — nearly tripled YoY — against orders customers can cancel "with little or no notice," and a 10-K risk factor warning customers "may overestimate demand for their AI build outs."
- Positioning is one-sided: zero holds or sells across 30 analysts, insiders one-way sellers, active institutions trimming.
- Deferred revenue $6.87B — management itself says the balance moves "independent of underlying business drivers."
Open questions — explicitly unresolved, do not fill in
- FY2027 non-GAAP $5.16 has no independent confirmation — Yahoo and stockanalysis share one LSEG feed.
- No H1-2026 concentration or AI-revenue figure exists. Arista discloses concentration annually and AI revenue only as a forward target — on-track/behind on the $3.5B cannot be verified.
- FY2025 actual AI revenue in dollars — never disclosed; the ~$1.6B in circulation is back-solved.
- CEO/CFO gave different AI targets on the same call — Ullal "at least $3.6B," Breithaupt "at least $3.5B." Scope difference presumed, not confirmed.
- The per-customer attribution (Microsoft 26% / Meta 16%) is BofA's, not Arista's — and it reverses the commonly cited 2023 split.
- The Aug 12 all-time closing high (+6.4%) — no company catalyst identified.
- No credible 2026 white-box market-share number exists. No 2026-dated Meta networking hardware disclosure (all Minipack3N/NSF/DSF material is Oct-2025 OCP); no 2026-dated Microsoft SONiC deployment statement.
- Nvidia has given no Spectrum-X dollar run-rate in any 2026 call — last hard figure is >$10B annualised from August 2025.
Sources
Arista FY2025 10-K · Q2 2026 10-Q · Q1 and Q2 2026 earnings releases (investors.arista.com) ·
Q2 2026 earnings call transcript · DEF 14A 2026-04-16 · SEC Form 4 filings (raw XML, aff10b5One
verified) · IDC datacenter Ethernet switching 1Q26 · Dell'Oro AI back-end networks 1Q26 · Cisco Q4
FY2026 results · Broadcom Q2 FY2026 results · Nvidia Spectrum-X/Spectrum-6 releases · Meta
Engineering OCP 2025 · Trefis (2026-08-19/24/25/26) · Yahoo Finance MCP · openinsider.
Playbook notes applied: [[pitfall-vendor-forward-eps-is-the-wrong-fiscal-year]] · [[pitfall-yahoo-insider-purchases-counts-rsu-grants]] · [[pattern-ai-levered-fields-trade-above-own-band]] · [[pitfall-stale-entry-zone-suppresses-a-name]] · [[pitfall-multiple-trim-inherits-the-broken-vendor-field]] · [[pattern-margin-intact-while-volume-falls-defers-the-damage]] (inverted here — volume is fine and margin is the thing giving way)