Arista Networks

Phase B — Measure performance

Earnings result — Q2 FY2026 (reported 2026-08-04)

• Vs expectations: non-GAAP EPS $1.02 vs ~$0.89 consensus. First $3B quarter.
• The margin read reverses June's. June Lens 5 called the Q1 61.9% GAAP gross margin "structural, not one-off". One quarter later it is 62.9%, up 100bps sequentially. The 10-QForm 10-QThe quarterly version of the annual report. Lighter, and not audited. still attributes the year-on-year decline to "an increased proportion of our sales to large end customers who generally receive higher discounts", so the mix pressure is real. But it has not compounded. The CFO credited "both tariff refunds and customer mix" for the sequential gain. One summary put the tariff refund at 20–30bps. So ~70–80bps of the rebound is operational even on the conservative read.
• A non-GAAP definition change sits inside this print. The Ex-99.1 recasts prior periods for a new method of computing non-GAAP taxes on stock awards. That raised Q2-25 non-GAAP net income by $10.7M and 6M-25 by $30.4M (+$0.03 EPS). Year-on-year non-GAAP EPS comparisons in any source dated before 2026-08-04 are not like for like with these.
• Market reaction: +11.8% after hours to $212.99 from a $190.51 close. Next day's close was $197.31 (+3.6%). The June dossier's pattern ("beats alone do not move it") broke this time because the guide moved, not the beat.

Balance sheet and visibility:

Cash generation: 6M OCF $2,776.5M, so Q2 OCF ≈ $1,083M (Q1 $1,693.5M); Q2 capexCapital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs. ≈ $29.7M [derived: 6M less Q1]. The sequential OCF drop is working capital, not earnings: the 6M cash statement shows other assets +$619.7M, "driven by increased deferred cost of sales associated with higher product revenue deferrals", and AR +$379.3M. No shares were repurchased in the six months; $817.9M remains on the May-2025 authorization.

Segments

Still one reportable segment. Mix by ship-to geography moved:

June Lens 4 flagged "geographic mix concentrating into the US" and APAC shrinking. Q2 reversed both. EMEA was the fastest-growing region and APAC grew again. The 10-Q says only that non-Americas swings are "primarily driven by changes in the geographic mix of sales to our large global customers". That points to a hyperscaler deploying abroad, not a broad international win. So this dilutes geographic concentration without diluting customer concentration.

Management's own lines:

• AI fabrics: "at least $3.5 billion" (CFO), unchanged since May. In the same call the CEO said "overall AI target of at least $3.6 billion". The two numbers are not reconciled in anything read. Scale-across is "approximately 30%" of it (~$1.1–1.2B), against a scale-across TAM of "$15 billion to $20 billion in 2030".
• Campus: "at least $1.25 billion", unchanged.
• The $1.1B guide raise was not allocated. Neither sub-target moved, so the raise is spread across AI back-end, "the core data center front end and campus and routing adjacencies". The CEO declined to allocate it ("I would like to reserve that $1.1 billion question to, well, it depends on how we ship").

Earnings call

No transcript on disk; this lens is from the Globe and Mail / Motley Fool and Investing.com transcripts of 2026-08-04.

• Tone moved from defensive to expansive on demand, and stayed defensive on supply. Demand: "all numbers are going up"; Etherlink AI deployments now at "100 cumulative customers from the initial 4 to 5 customers I spoke of in 2024". Supply: "I don't want you to believe that suddenly, we waved the magic wand... The industry is going to have a 2-year problem. I don't think we get out of it as an industry until 2028" (Ullal, answering a question about last quarter's upstream decommit).
• Memory is covered for 2026 only: "Our memory supply has been secured for 2026, and we have extended visibility well into 2027 across DDR4, DDR5 and NAND" (Todd Nightingale).
• New 10% customers: the CEO expects "one, maybe two" more 10%-plus customers in 2026, without naming them, while "we remain very committed to our two longest partners, Microsoft and Meta".
• Competition, as management now frames it: white box is "a tactical solution... in use cases that are simple, scale-up or scale-out"; "we live in a NVIDIA world", with Arista's growth in the non-NVIDIA clusters (AMD MI-series and Google TPU named in summaries). Spectrum-X was not mentioned by name in anything read. Management also says it is "less prevalent" in scale-up.
• Product cadence: 1.6T 7060X-E7 platforms are in trials with "single digit, but very large customers" in 2H26, but "the real production will still be in 2027". Co-packaged optics trials are expected in 2027, with meaningful volume in "2028 and 2029".

Phase C — Judge people & books (Lens 9 only)

Management — insider flow since the June dossier

37 Form 4s filed 2026-06-22 → 2026-09-23 [primary: SEC EDGAR Form 4Form 4The filing an insider must make within two business days of buying or selling their own company’s shares. XML, tallied in this pass]:

Also: tax withholding (code F) on the 2026-08-20 vest by Ullal, Duda, Nightingale and Breithaupt; Duda's 531,302-share code-J in/out pairs (Aug 28 → Sep 10) net to zero and read as intra-family or trust transfers, not sales.

Read. Roughly $860M sold by the three people who built the product and run it, concentrated in the six weeks around the print, with zero buys. It extends the June finding (97 sells / 0 buys over five years) rather than breaking it. It is consistent with the disclosed 10b5-1Rule 10b5-1 planA schedule an insider sets in advance saying when their shares will be sold, so the sales cannot be timed on private news. A sale made OUTSIDE such a plan is a live decision, which is why it carries more information. plans: Ullal's Nov-2025 plan covers up to 5.726M shares through Feb 2027 [June dossier, 10-KForm 10-KA company’s audited annual report to the US regulator. The most complete thing it publishes.]. Planned selling is not a signal about the quarter. It is a signal about what founders do at ~$200. Ullal's 2.16M shares are ~38% of her plan's cap, sold in one quarter.

Phase D — Project & stress-test

Comps / valuation (corrected and refreshed)

Price $205.70 (2026-09-24 close; +24.7% vs $164.93 on 2026-06-17; 6-month closing high $210.50 on 2026-08-12).

• Market cap ≈ $205.70 × 1,260.1M basic shares ≈ $259B. EV ≈ $259B − $13.3B ≈ $246B.
• TTM revenue $10,540.8M [derived: FY25 $9,005.7M − 6M-25 $4,209.6M + 6M-26 $5,744.7M]. EV/TTM sales ≈ 23x. EV/FY26E sales ($12.6B) ≈ 19.5x.
• TTM non-GAAP EPS ≈ $3.47 [derived: FY25 $2.98 − 6M-25 $1.40 (recast) + 6M-26 $1.89; the FY25 figure is pre-recast, so this is approximate]. P/E ≈ 59x TTM.
• FY26E non-GAAP EPS ≈ $4.11. P/E ≈ 50x FY26E.

Read: The stock rose 25% and FY26 EPS expectations rose ~10% (June base $3.75 → ~$4.11), so the multiple expanded. On the corrected June numbers it was ~44x FY26E and is ~50x now. The market paid for the guide raise and then some. Peer multiples were not refreshed; no working source in this environment.

Forward projection

Grading June's paths against what is now known:

The bull was beaten without its stated driver. AI fabrics were not raised; the upside came from everything else (front-end DC, campus, routing). That is a better-quality beat than June's bull imagined: less dependent on the back-end socket Nvidia contests.

Revised paths (all; inputs as labelled):

• Base. FY26 revenue $12.6B, non-GAAP EPS ~$4.11. FY27 revenue ~$15.2B (+21%): the 1.6T production ramp plus one or two new 10% customers, offset by law of large numbers. EPS ~$4.90 at a flat ~48.5% non-GAAP operating margin, ~21.5% tax, flat share count (no buybacks for six months).
• Bull. FY27 revenue ~$16.5B (+31%) on supply loosening earlier than "2028" and scale-across compounding; EPS ~$5.40.
• Bear. FY27 revenue ~$13.2B (+5%): 2027 hyperscaler digestion meets $9.4B of pre-committed purchases. Inventory/supplier-liability charges take GM to ~59%; EPS ~$3.90. At ~30x that is ~$117, −43% from $205.70.

On base, $205.70 is ~42x FY27E. The June "fair-value band ~$150–175" was built on a $4.55 FY27 EPS, which is now stale. Re-running the same 32–38x on $4.90 gives ~$157–186; the stock is above it.

Bull vs bear

What the print settled:

1. Gross-margin collapse: not happening yet. June bear #3 ("GM 64.1% → 61.9%, structurally") is contradicted by 62.9% GAAP / 63.4% non-GAAP and a held 62–64% guide that already includes memory and silicon cost increases.
2. Growth is not supply-capped the way May implied. May's $11.5B was "supply-constrained"; August added $1.1B without supply being declared solved.
3. Diversification is plausible on management's word only. "One, maybe two" new 10% customers and EMEA +56% both point there. Neither is in a filing yet; the FY26 10-K concentration note (Feb 2027) is the test.

What the print did NOT settle, and what replaces June bear #3:

• The pre-commitment is the new risk. Purchase obligations due within a year went $3.6B (a year ago, per the CFO) → $7.6B (Q1) → $9.4B (Q2). That is 2.4x trailing cost of revenue ($3.90B TTM [derived: FY25 $3,237.0M − 6M-25 $1,494.9M + 6M-26 $2,157.6M]). At a ~63% gross margin, $9.4B of cost supports roughly $25B of revenue against a FY26 guide of $12.6B. Some of it is prepaid scarcity (memory bought at 2026 prices to secure 2027) rather than a demand forecast. The 10-Q is candid that this "creates a heightened risk of excess or obsolete inventory and supplier liability charges". This is the June forensic watch-item (Lens 10: "inventory/purchase-commitment obsolescence") getting bigger by $2.6B in two quarters.
• Deferred revenue keeps compounding ($5.37B → $6.20B → $6.87B, ~2.3 quarters of revenue), driven by "customer-specific acceptance clauses". Evaluation inventory is up 53% in six months to $616M. Visibility is real; so is the acceptance risk if a large trial fails, which the 10-Q now spells out ("we may be required to accept product returns... and may result in the write-down of inventory").
• The Nvidia thread went quiet, not away. Management now positions Arista in "the non-NVIDIA" clusters and is "less prevalent" in scale-up. That concedes the Nvidia-bundled back end rather than contesting it. It is a cleaner story and a smaller TAM.

Devil's advocate (delta only)

The short case in June was margin, competition and concentration. After Q2 the sharper short case is cycle timing. Arista's 2026 is being built on customer acceptances deferred into later quarters and components bought a year ahead at peak memory prices. Its CEO says supply stays tight into 2028; its founders sold ~$830M between them in ten weeks. If AI capex digests in 2027 while $9.4B of purchases arrive, the P&L takes supplier-liability and E&O charges in the same quarters that revenue growth slows. A 50x multiple does not price that sequence. The counter is the $13.3B net cash and ~49% operating margin: Arista can absorb a write-down that would break a thinner company. The multiple is what breaks, not the company.