Arista Networks

Phase A — Understand the business

Lens 1 · Company Overview

Arista builds high-performance Ethernet switching and routing systems for the largest data centers on earth, plus the software (EOS) and management plane (CloudVision) that run them. The model is conceptually simple and unusually clean: design the box around best-of-breed merchant silicon (overwhelmingly Broadcom), let contract manufacturers build it, and differentiate almost entirely in software — one binary image (EOS) on a standard Linux kernel with a publish/subscribe state database, spanning the whole portfolio. That single-image architecture is the spine of the whole thesis: it is why a hyperscaler can run the same OS from leaf to spine to WAN and why feature velocity is high.

• What it sells. Product revenue (84% of total) = switching/routing platforms + EOS; Service revenue (16%) = post-contract support (PCS) and renewals. "We expect to continue to derive substantially all of our product revenue from sales of our switching and routing platforms for the foreseeable future."
• Three customer segments (company's own taxonomy): Cloud & AI Titans, AI & Specialty Providers (incl. neoclouds), and Enterprise. Cloud & AI Titans ≈ 48% of 2025 revenue.
• Contract structure. Sales are purchase-order based, not long-term take-or-pay — customers "could cancel, delay, reduce or otherwise modify their purchase commitments with little or no notice." Large customers get "pricing discounts, bundled upgrades, extended warranties, acceptance terms" — i.e., margin-dilutive but volume-defining.
• Customer concentration is extreme and disclosed. Two end customers were each >10% of revenue in 2025: 26% and 16% (so ~42% combined). Trend by customer: one ran 26% / 20% / 18% (2025/24/23), the other 16% / 15% / 21%. The two are widely identified as Meta (~26%) and Microsoft (~16%).
• Scale. FY2025 revenue $9,005.7M, +28.6% YoY; ~5,115 full-time employees.

Read: This is a software-margin business wearing a hardware company's clothes — 64% gross margin and ~48% non-GAAP operating margin on boxes is the tell. The franchise risk is not "can they build a better switch" (they can); it's that their two best customers are also the two most capable of designing around them.

Lens 2 · Supply Chain

Map the chain with the actual names, because the single-source dependencies are the whole risk.

Upstream silicon (the chokepoint):

• Broadcom — the "predominant merchant silicon vendor" for switching chips (StrataXGS / Jericho / Tomahawk families). The 10-K is blunt: "we are primarily reliant upon our predominant merchant silicon vendor, Broadcom, for our switching chips." No guaranteed-supply written agreement.
• Secondary silicon/IC/power-supply vendors "including certain sole source providers" — unnamed but flagged. Memory is a named 2026 pressure point ("tightening of supply conditions in the memory and silicon markets").

Midstream manufacturing (outsourced):

• Contract manufacturers: Jabil, Sanmina, Foxconn (Hon Hai). Build sites in Malaysia, Vietnam, Mexico, plus components from China/Taiwan/Thailand/Philippines. Arista keeps the BOM, qualified-supplier list, and test/QA in-house.
• Four direct-fulfillment hubs: US, Netherlands, Singapore (+ final config). Spare-parts depots in 200+ locations via 3PLs.

Downstream:

• Direct sales force + channel (distributors, VARs, systems integrators, OEMs). End customers = hyperscalers (Meta, Microsoft, plus Amazon/Google/Oracle referenced as users), neoclouds, enterprise verticals (financial services, government, healthcare, media).

Chokepoints / single-source dependencies:

1. Broadcom silicon — the defining concentration. Broadcom is simultaneously Arista's key supplier and (via VMware ownership + Tomahawk-for-white-box) a competitive vector.
2. Memory — explicitly tightening, pressuring gross margin in 2026.
3. Tariff geography — manufacturing concentrated in Malaysia/Vietnam/Mexico; Section 232/301 tariff exposure is an active gross-margin risk the company is "adjusting supply chain" to mitigate.

Supply signal worth watching: purchase obligations jumped to $8.9B at Q1-26 (from $6.8B at YE-25), $7.6B due within 12 months — Arista is pre-committing working capital to lock component supply against AI demand. Deposits to CMs rose to $95.5M. Management cited frequent supplier "decommits" as the reason it won't raise guidance further.

Lens 3 · Competitive Advantages (moats)

The moat is software + architecture, not the box.

1. Single EOS image / state-oriented OS. One OS across the entire portfolio with streaming telemetry, in-service upgrades (SSU), and a central state DB. Competitors run fragmented OS stacks per product line. This is a genuine switching-cost moat: a hyperscaler standardized on EOS + CloudVision + NetDL has automation, tooling, and operational muscle memory built around it. Rip-and-replace is expensive and risky.
2. Co-design with cloud titans. Meta and Microsoft co-engineer next-gen AI fabrics with Arista. That design-in is sticky for a product generation — and is exactly what Nvidia is now contesting.
3. Merchant-silicon leverage as a moat and a vulnerability. By riding Broadcom's R&D instead of taping out custom ASICs, Arista converts silicon capex into someone else's problem and focuses spend on software (R&D was 13.8% of revenue / $1.24B in 2025). But the same merchant silicon is available to white-box vendors — so the moat lives in the software layer above it, not the chip.
4. Quality / support reputation. Industry-leading NPS, 24/7 TAC, A-Care. Real but soft.

Bargaining power: Weak vs. its top two customers (they're 42% of revenue, get the discounts, set acceptance terms, and can self-design / multi-source). Strong vs. enterprise and neoclouds (less price leverage, more lock-in). Weak vs. Broadcom upstream (sole-source, no guaranteed supply). The power asymmetry is the structural tension of the whole investment.

Durable? The EOS switching-cost moat is durable in cloud/enterprise data-center Ethernet (where Arista is #1 in branded high-speed switching ). It is least durable in back-end AI fabrics, where Nvidia can bundle networking with GPUs and InfiniBand/NVLink live — and that is precisely the highest-growth pool.

---

Phase B — Measure performance

Lens 4 · Segments

Arista reports one reportable segment (CODM = CEO reviews consolidated P&L), so the breakouts are by product type and by geography only.

By product (FY2025):

By geography (FY2025):

Management's own "go-to-market" framing (not reported segments, but guided targets): AI networking revenue target raised to $3.5B for 2026 (from $3.25B; more than doubling YoY), Campus $1.25B for 2026.

Trend + cause: Growth is accelerating (FY23 +33.8% would be wrong — actually FY24 +19.5%, FY25 +28.6%, Q1-26 +35.1% YoY) — re-acceleration driven by the AI Ethernet ramp. Geographic mix is concentrating into the US: Q1-26 Americas hit 84.5% (US alone $2,272.1M of $2,709.0M = 83.9%), while APAC actually shrank ($183.9M vs $231.7M YoY). That US concentration tracks hyperscaler capex and is a quiet tariff/geo-political single-point risk.

Lens 5 · Earnings Result (latest print: Q1 FY2026, reported 2026-05-05)

The tape-defining event of this name in 2026: a clean double-beat that the stock fell ~12.6% on.

The numbers:

• Vs consensus: revenue beat ~$2.62B est by ~3.5%; non-GAAP EPS $0.87 beat $0.81. Operating cash flow $1.69B — CFO called it "the strongest in the history of Arista."
• What drove it: broad-based product shipments; AI Ethernet ramp. Service grew on installed-base renewals.
• The blemish — gross margin. 61.9% is a visible step down, and the 10-Q is explicit on cause: "an increased proportion of our sales to large end customers who generally receive higher discounts," compounded by memory/silicon supply inflation and tariffs in COGS. This is the classic Arista tension showing up in the print: faster growth via the hyperscalers = lower margin.
• Balance-sheet flags (all green): cash + securities $10.7B, zero debt. Inventory $2.38B (raw materials up to $811M = pre-buying components). Deferred revenue ballooned to $6.20B (from $3.09B YoY) and evaluation inventory $525.7M — both reflect AI products on customer trials with acceptance clauses (revenue not yet recognized). This is a quality-of-revenue positive (visibility) but also a volatility source if trials fail.
• Guidance & tone: FY2026 revenue guided to ~$11.5B (+27.7%); AI target lifted to $3.5B. Management framed the 27.7% as supply-constrained, not demand-constrained, and flagged supplier "decommits" as why they won't go higher.
• Market reaction: −12.6% after-hours despite the beat — because the stock had run +34% into the print and a roughly-in-line Q2 guide couldn't clear a sky-high bar. This is a valuation-reset story, not a business-deterioration story.

Unusual vs. its own history: the GM print is the standout — Arista held 64.1% GM for FY24 and FY25; 61.9% is the lowest in recent memory and the trajectory (mix-driven) is structural, not one-off.

Lens 6 · Earnings Calls (sentiment trend)

No transcripts on disk (transcripts=0); this lens is ``.

• Consistent multi-quarter narrative: "Ethernet is the eventual winner for AI networking" (Ullal's recurring frame); AI as the growth engine; Ultra Ethernet Consortium momentum vs InfiniBand.
• Tone shift over the last ~4 calls: progressively more defensive on two fronts — (1) supply ("decommits," memory tightening, disciplined fulfillment cadence) and (2) margin (openly guiding GM down on hyperscaler mix). The growth confidence is intact and rising (AI target raised twice: $3.25B → $3.5B), but the qualifiers are multiplying.
• What they started saying: "supply availability is the ceiling, not demand"; specific diversification wins (a neocloud moving off white-box to 800G EtherLink for AMD accelerators; a regional fiber backbone; insurance R3 monitoring fabric; a global manufacturer's Cognitive Campus across 100+ sites).
• What they stopped emphasizing: the InfiniBand-vs-Ethernet debate has quietly shifted to an Ethernet-vs-Ethernet debate (Arista vs Nvidia Spectrum-X) — a tell that the competitive front moved.

Lens 7 · Comps

Networking / AI-infrastructure peer set.

Sources:.

Derived valuation cross-checks (``, arithmetic shown):

• EV ≈ $290B mkt cap − $10.7B cash − ~$0 debt ≈ ~$279B.
• EV/Sales (TTM $9.006B) ≈ 279/9.0 ≈ ~31x. On FY2026E rev $11.5B ≈ ~24x. Both are extreme for a hardware-classified name.
• 5-yr avg ROE: ANET has run high-20s/low-30s ROE for years (32.3% current); ROIC ~52% TTM. Best-in-class capital efficiency in the group.

Read: Arista trades at a ~2x premium to Cisco on EV/EBITDA (47x vs 24x) and at ~46x forward earnings. The premium is earned on growth (28% vs Cisco's single digits) and returns (ROIC 52%), but it leaves zero margin for a growth disappointment — which is exactly the Nvidia-Spectrum-X risk. The cleanest pure-play peer (Juniper) no longer exists as a standalone; the most dangerous "peer" (Nvidia) is now also the most dangerous competitor.

Lens 8 · Stock-Price Catalysts (what moves >5%)

Pattern over the last ~2–5 years:

• Apr 2024 — Rosenblatt double-downgrade (buy→sell) on rising Nvidia competition: a notable down move. The "Nvidia is coming for the AI fabric" narrative has been a recurring de-rate trigger since.
• Oct 2025 — Nvidia wins Meta + Oracle for Spectrum-X: "Arista sinks" headlines; the single most important competitive catalyst.
• May 5, 2026 — Q1 double-beat, stock −12.6%: valuation-reset on an in-line guide after a +34% run-up.
• Apr 22, 2026 — all-time high $177.73 then a ~28% drawdown off the peak into June.
• Recurring macro/AI-capex beta: moves hard with hyperscaler capex headlines and broad AI-trade selloffs (e.g., "slides 4% as tech selloff bites").

What the market actually reacts to for ANET: (1) competitive-position headlines (Nvidia share, specifically), (2) hyperscaler capex direction (Meta/Microsoft), (3) the gap between a beat and an already-priced-in bar (the stock is a sentiment/positioning instrument as much as a fundamentals one), and (4) guidance on AI-networking dollar targets. Earnings beats alone do not move it up — expectations management does.

---

Phase C — Judge people & books

Lens 9 · Management

• Jayshree Ullal — Chairperson, President & CEO. CEO since Oct 2008 (~17 yrs), Chair since Dec 2023. Prior: 15 yrs at Cisco running data-center switching. Track record is elite — took Arista from near-zero to ~$9B revenue, a successful 2014 IPO, and category leadership in cloud Ethernet; E&Y Entrepreneur of the Year (2015), Barron's World's Best CEOs (2018). One of the very few S&P 500 companies with both a female CEO and CFO.
• Skin in the game: Ullal owns ~27.97M shares worth ~$4.6B. CEO ownership guideline 3x base; hedging prohibited, pledging prohibited for certain officers. But the trading pattern is one-directional: over 5 years, 97 Form-4 transactions, 0 buys, 97 sells; net ~8.86M shares sold over 18 months; a Nov-2025 10b5-1 to sell up to 5.726M more shares (through Feb 2027). — Normal for a long-tenured founder-operator monetizing equity, but it's consistent distribution, not accumulation.
• Chantelle Breithaupt — SVP & CFO. Adopted a 10b5-1 (Dec 2025) for PSU-linked sales. Credible; called Q1-26 OCF "strongest in Arista history."
• Founders / architects: Andy Bechtolsheim (co-founder, Chief Architect; biggest shareholder at ~15%, ~$14B stake) and Ken Duda (co-founder, CTO, lead EOS/NetDL architect, on the board). David Cheriton co-founded. Bechtolsheim + Cheriton self-funded ~$100M seed — rare alignment.
• Capital allocation: disciplined and shareholder-friendly — no dividend (by design), buybacks ($1.6B repurchased in FY2025; $1.5B program authorized May 2025, $817.9M remaining; zero bought in Q1-26 — opportunistic, not mechanical), one bolt-on (VeloCloud from Broadcom, $300M, Jun 2025). ROIC ~52%. They reinvest in R&D (+24% YoY) and return the rest. This is a textbook high-ROIC compounder allocation.
• Founder vs professional: founder-led (Ullal is founding-era CEO; two founders still on the board/architecture). Implication: long-termism, deep technical credibility, low strategy-drift risk — but also key-person risk concentrated in a 17-yr CEO and a 68-yr-old chief architect now serving in a reduced role.

Red flag — see Lens 10: the most serious governance event is the SEC insider-trading action against co-founder/then-Chairman Bechtolsheim (2024). It is company-adjacent (the trades were in Acacia options, not ANET), but it is a real integrity mark on the founder who remains the largest holder and Chief Architect.

Lens 10 · Forensic Red Flags

Acting as a forensic analyst. Books are clean; the risks are disclosure-quality and one founder-level enforcement event.

Accounting quality — largely clean:

• Auditor: Ernst & Young, since 2008, unqualified opinion + effective ICFR, dated 2026-02-13.
• Critical Audit Matter: Inventory valuation & supplier liabilities — the one genuinely judgmental area (NRV of inventory + non-cancellable purchase-commitment charges against demand forecasts). With $2.38B inventory, $8.9B purchase commitments, and $525.7M evaluation inventory at trials, excess/obsolete risk is the real forensic watch-item. Not a manipulation flag — a cyclicality flag.
• Revenue recognition: multi-element (product + PCS) with SSP allocation and acceptance clauses on AI trials. Deferred revenue surged to $6.2B and deferred COGS to $1.55B — revenue is being deferred, not pulled forward (conservative). Cash flow corroborates earnings: FY2025 OCF $4.37B vs net income $3.51B (OCF > NI; a healthy sign). No channel-stuffing signal — partners "generally do not stock inventory."
• SBC: $439.2M FY2025 (~4.9% of revenue); the non-GAAP/GAAP EPS gap is real but not egregious (FY25 non-GAAP $2.98 vs GAAP $2.75 — a ~8% add-back). $1.5B unamortized SBC over ~4.2 yrs. Watch the non-GAAP framing but it's within normal tech bounds.
• Receivables/inventory vs revenue: AR grew $746M in 2025 on higher billings; inventory +$412M (component pre-buy). Both outran the revenue line modestly but are explained by the AI ramp + supply pre-positioning, not stretching. Worth monitoring DSO/DIO trend.
• Tax: effective rate stepped up (FY24 12.6% → FY25 17.4% → Q1-26 19.5%) as equity-comp tax benefits shrank + OBBB Act (Jul 2025). A headwind to GAAP EPS growth, fully disclosed.
• Off-balance-sheet: none. No SPEs.

Regulatory findings (required sub-section):

• SEC EDGAR EFTS (LR + AAER) for "Arista Networks": the on-disk regulatory/regulatory-findings.md returned 0 LR and 0 AAER naming the company (period 2021-06-22 → 2026-06-22).
• BUT — material founder-level SEC enforcement (web-sourced, not company-named so it escaped the entity EFTS search): SEC v. Andreas "Andy" Bechtolsheim (co-founder, Chief Architect, then-Chairman). The SEC alleged that in July 2019 he traded Acacia Communications options minutes before Cisco's announced $70/share acquisition — learned via Arista's relationship with another acquirer ("shadow trading" theory). Profits ~$415,726 (in a relative's/associate's accounts). Final judgment 2024: without admitting/denying, he was enjoined under §10(b)/10b-5, barred from officer/director of a public company for 5 years, and paid a $923,740 civil penalty. — This is the single most important integrity item in the file. It does not implicate Arista's accounts, but it implicates the founder who is the largest shareholder and remains Chief Architect, and it is why he stepped down as Chairman in Dec 2023.
• Non-SEC enforcement (FTC/DOJ/FDA/CFPB): no material agency action against Arista surfaced. (The historical Cisco v. Arista copyright/patent litigation from 2014–2018 was resolved years ago and is not a live regulatory matter.)
• 10-K Item 3 / Note 5 (Legal Proceedings), company's own words: ordinary-course IP/commercial/employment matters only; "provisions recorded for contingent losses related to other claims and matters have not been significant," no matter likely to be individually/aggregately material.

Net forensic read: Books are clean and conservatively stated (revenue deferred, OCF > NI, top-tier auditor, no SPEs). The two things a skeptic underwrites are (1) inventory/purchase-commitment obsolescence in a cyclical AI build, and (2) the founder governance shadow from the Bechtolsheim case. Neither is an accounting-fraud signal; both are real.

---

Phase D — Project & stress-test

Lens 11 · Forward Projection (FY2026 → FY2028 EPS)

Bottom-up from FY2025 actuals + company guidance. Outputs ``; inputs labeled.

Anchors:

• FY2025: revenue $9,005.7M; non-GAAP op margin 48.2%; non-GAAP diluted EPS $2.98; GAAP $2.75; diluted shares ~1,275.7M.
• FY2026 company guide: revenue ~$11.5B (+27.7%), AI $3.5B, Campus $1.25B. GM guided down (hyperscaler mix). Tax rate rising toward ~19–20%.

Base case (company guide ≈ truth; GM ~61–62%; modest buyback; tax ~19.5%):

• FY2026E revenue $11.5B; non-GAAP EPS ~$3.75.
• FY2027E revenue $14.0B (+22%); non-GAAP EPS **$4.55**.
• FY2028E revenue $16.5B (+18%); non-GAAP EPS **$5.35**.

Bull path: AI not supply-capped; AI revenue >$4B in 2026; share holds vs Spectrum-X; campus + new 10% customers diversify. FY2026E EPS ~$3.95, FY2028E ~$6.25.

Bear path: Nvidia takes meaningful back-end AI share; one hyperscaler reallocates; GM to high-50s on discounting + tariffs; AI ramp slips to supply and share. FY2026E EPS ~$3.45, FY2028E ~$4.50, with multiple compression doing most of the damage.

Consensus cross-check: sell-side next-year EPS $3.27 (likely GAAP); my GAAP-equiv base ($3.45) is a touch above — the delta is mix/tax assumptions. Not logging a forecast.ts Brier entry (per --watchlist rules: dossier-only; no forecast create in the sweep).

Lens 12 · Bull vs Bear

Bull case. Arista is the best-run independent in the single biggest infrastructure build of the decade. It owns the software layer (EOS/CloudVision) that hyperscalers standardize on, it's #1 in branded high-speed Ethernet switching, and Ethernet is structurally winning the AI back-end war vs InfiniBand as the Ultra Ethernet Consortium matures. Financials are pristine: ~48% non-GAAP op margin, 52% ROIC, $10.7B net cash, OCF > net income, 28% growth with a $7.7B RPO backlog (91% inside 2 yrs) giving rare visibility. AI target raised twice to $3.5B; campus is a credible second leg to $1.25B. Capital allocation is disciplined (buybacks, one tiny bolt-on, no empire-building). At a +34% pre-earnings run, even a clean beat couldn't satisfy — which means the bar, not the business, is the problem.

Bear case (2–3 permanent-impairment risks).

1. Nvidia bundles the network with the GPU. Spectrum-X already won Meta and Oracle for AI Ethernet; Nvidia's networking line is ~$14.8B/quarter and passed Cisco in DC Ethernet switching in Q1-26. If hyperscalers buy the GPU+network as one stack (95% throughput claim vs ~60% standard Ethernet), Arista's richest, fastest-growing socket — back-end AI fabric — structurally shrinks. This is the one that can permanently impair the growth thesis.
2. Customer concentration cuts both ways. 42% of revenue from two customers who get the discounts, set acceptance terms, can self-design (Meta's own silicon ambitions), and can multi-source (the 10-K's own first-listed risk). A single reallocation is a revenue air-pocket.
3. Margin compression is already happening. GM 64.1% → 61.9% in one year, structurally (hyperscaler mix) + memory/silicon inflation + tariffs. If the growth that bulls love is bought with discounts, the operating-leverage story erodes.

Pre-mortem (18 months out, thesis broke): It's late 2027. Nvidia's bundle and Broadcom's Tomahawk-6 white-box reference designs captured the bulk of new AI back-end deployments; Arista's AI revenue grew but missed the doubling; one hyperscaler shifted a tranche to Spectrum-X; gross margin settled in the high-50s on discount intensity; the multiple de-rated from ~46x to ~25x forward. The business is still great and growing ~15% — but the stock halved because it was priced for flawless 25%+ growth at premium margins.

Are multiples too high? At ~46x forward / ~47x EV-EBITDA / ~24–31x EV/Sales, yes on an absolute basis — it's priced for sustained mid-20s growth at ~48% margins with share intact. That is achievable but not de-risked. The valuation embeds the bull case as the base case.

Contrarian view (what the market refuses to see): The consensus is fighting the last war ("Ethernet beats InfiniBand"). Arista won that war — and the front already moved to Ethernet-vs-Ethernet, where its opponent is a $4T company that gives the switch away with the GPU. The market still prices ANET as the unchallenged toll-road on AI networking; the real question is whether it's a toll-road or a premium lane next to Nvidia's free highway.

Lens 13 · Devil's Advocate (short-seller)

Dismantling the bull case.

• Structural break in how it makes money: ~84% of revenue is switching/routing hardware sold on cancellable POs to a customer base that is consolidating spending power and vertically integrating. Meta (26%) designs its own silicon and co-develops fabrics; Microsoft (16%) has the scale to multi-source. There is no take-or-pay, no multi-year lock — just EOS switching costs that matter less in greenfield AI clusters where the buyer is choosing an architecture fresh.
• The moat is weaker than bulls think specifically in AI back-end — the highest-growth pool. EOS lock-in is real in brownfield cloud/enterprise; in a net-new GPU cluster, Nvidia sells GPUs+Spectrum-X+NVLink as a validated bundle and Arista is the rip-out-the-integration option. Switching costs protect the old footprint, not the new dollars.
• Most dangerous competitor bulls underestimate: Nvidia (not Cisco). Nvidia already won Meta + Oracle for Spectrum-X and out-shipped Cisco in DC Ethernet switching. Its incentive is to make networking a near-zero-margin attach to protect GPU margin — a price umbrella Arista cannot sit under at 62% GM. And Broadcom — Arista's own silicon supplier — ships Tomahawk 6 (102.4 Tbps) that any white-box vendor can build into a competing box. Arista is sandwiched: bundled above by Nvidia, commoditized below by its own supplier's white-box reference designs.
• Worst capital-allocation / governance mark: the Bechtolsheim SEC insider-trading settlement ($923,740 penalty, 5-yr O&D bar, 2024) — the founder/largest holder/Chief Architect. Not an accounting issue, but a real integrity flag at the top of the cap table.
• Assumptions that must hold for today's price: (1) AI networking ≈ doubles to $3.5B in 2026 and keeps compounding; (2) Arista holds back-end AI share against Nvidia's bundle; (3) GM stabilizes ~61–62% rather than sliding to high-50s; (4) no hyperscaler reallocation. All four must hold to justify ~46x forward.
• If growth disappoints 20–30%: model FY2026 revenue ~$10.0–10.2B instead of $11.5B (i.e., ~12–13% growth vs 28%). On a name priced for 25%+, the de-rate is violent — a move from ~46x to ~25x forward EPS roughly halves the stock independent of the EPS cut. The asymmetry is bad here.
• Single permanent-impairment scenario (most plausible bear): Nvidia + Broadcom-white-box jointly cap Arista's AI back-end TAM, so AI revenue grows but plateaus as a share-loser, and the two anchor customers normalize spend → revenue growth decelerates to low-teens, margins compress, and the stock re-rates to a Cisco-like multiple. Plausibility: moderate and rising — the Spectrum-X Meta/Oracle wins are the leading indicator.

Lens 14 · Management Questions (ordered by information value)

1. Of the back-end AI fabric sockets at your two largest customers, what share is at risk of moving to Nvidia Spectrum-X over the next 8 quarters, and what is your evidence those sockets are defensible?
2. Gross margin fell to 61.9% on hyperscaler mix — where does GM settle structurally at the $11.5B → $15B revenue range, and what stops the slide to the high-50s?
3. You guided 27.7% as supply-constrained. Quantify the demand-supply gap: if "decommits" resolved, what would 2026 revenue be?
4. Meta + Microsoft are 42% of revenue. What is your concrete diversification math — which 1–2 customers cross 10% in 2026, and how fast does top-2 concentration fall?
5. Nvidia gives networking away to protect GPU margin. How do you compete on price in net-new AI clusters without dragging your 48% operating margin down?
6. Broadcom is your sole switching-silicon source and arms white-box rivals with Tomahawk 6. How do you keep the software moat ahead of a commoditizing chip layer — and have you de-risked the single-source dependency at all?
7. What is the win-rate trend in competitive AI back-end bids vs Nvidia and vs white-box over the last 4 quarters?
8. Deferred revenue is $6.2B with acceptance clauses. What's the historical conversion rate of evaluation/trial inventory to recognized revenue, and what would a wave of failed acceptances do to 2026?
9. Purchase commitments are $8.9B. What's your obsolescence-charge exposure if AI capex digestion hits in 2027, and how is the supplier-liability reserve sized?
10. Campus to $1.25B in 2026 — is that organic share gain or VeloCloud + price, and what's the GM profile vs the data-center business?
11. Scale-Up / ESUN and Ethernet-for-scale-up: realistically, when does this become revenue, and does it offset any back-end scale-out share loss?
12. Capital allocation: with $10.7B net cash and a stock down ~28% from highs, why zero buybacks in Q1-26 — and what would make you accelerate?
13. Key-person risk: succession plan for the CEO and for the Chief Architect (Bechtolsheim) roles?
14. Governance: what changes to insider-trading controls followed the Bechtolsheim matter, and how do you reassure investors on tone-at-the-top?
15. What would you flag as the single most underappreciated risk to the 3-year plan that isn't in the 10-K risk factors?

---