ANET — Arista Networks
Ethernet keeps winning the AI fabric argument. The stock still wants to be paid like it already won the war.
Arista sells the Ethernet switching, routing and software that lets cloud and AI systems actually talk to each other. In an AI cluster, that is not background plumbing anymore. If thousands of expensive accelerators cannot stay fed and synchronized, the whole thing gets stupid fast. I like that exposure. The stock knows I like it too.
Arista is one of those holdings where every good quarter makes the valuation argument more annoying, not less.
Why I care about this one
An AI cluster is a pile of very expensive accelerators that only works if the network can keep them fed and synchronized.
That is the part of Arista I care about.
Ethernet used to have a credibility problem in the highest-performance AI fabrics. That argument has gotten weaker. Arista is now shipping into networks built around Meta, Microsoft and Oracle, and the newest 1.6T platforms are aimed directly at next-generation AI clusters. (1.6T AI fabric portfolio)
If Ethernet keeps taking more of the scale-out and scale-up stack, Arista does not need to become an accelerator company. It just needs to keep owning an increasingly valuable layer between the accelerators.
That is a very nice business to be in.
Q2 revenue reached $3.036B, up 37.7% year over year.
GAAP operating margin was 45.4%. Non-GAAP operating margin was 49.9%. Non-GAAP EPS was $1.02, up almost 40%.
Cash and marketable securities finished the quarter around $13.3B.
The ugly number is concentration.
Two end customers represented 42% of 2025 revenue. The filing also says the increased mix toward large customers is already pulling gross margin down because those customers get better pricing. (Q2 10-Q)
That does not make the growth low quality.
It does mean I cannot casually take 38% growth, 50% operating margins and a giant terminal multiple and pretend all three are independent gifts from the investing gods.
The customers creating the growth are powerful enough to keep some of the economics for themselves.
Q2 2026 + Q3 guide
Primary sources: Q2 2026 results ↗ · Q2 2026 10-Q ↗ · 1.6T AI fabric portfolio ↗
Arista's AI networking demand and its earnings trajectory remain strong, but the risk/reward at the Oct. 6 reference price is less compelling than the business quality. FY2029 Base assumes about $7.80 EPS at 36x, or $281, with a high-single-digit modeled annualized return. Customer concentration and the terminal multiple remain important risks.
My position
What I actually bought and sold.
| Date | Type | Account | Dollars | Shares | Price | Return now* |
|---|---|---|---|---|---|---|
| Dec 12, 2025 | Buy | Taxable | $2,400 | 18.062 | $133 | +62.1% |
*Return now uses the latest available quote for buys and reinvestments. Sales and fees show “—”.
Valuation & scenario math
This is probably the cleanest capital-allocation story in the sleeve.
Arista spends heavily on R&D, carries a fortress balance sheet and has not needed giant acquisitions to manufacture relevance.
There is one mild annoyance.
The company did no share repurchases in the first half of 2026 despite an active authorization, while stock-based compensation was about $241M. The diluted share count is basically flat rather than meaningfully shrinking. (Q2 10-Q)
That is a mild annoyance, but enough to keep the score at 8.75/10 instead of turning the whole section into a love letter.
Sep. 2029 earnings power + valuation + correlated AI exposure
No static dollar ladder. A lower price helps, obviously. I still need the three-year earnings bridge to produce a mid-teens-or-better base return after remembering I already own a lot of AI infrastructure.
Three-year AI-networking checkpoint using the Q3 2026 non-GAAP EPS run-rate. The operating model is kind of ridiculous. The actual debate is how much of the Ethernet win I am already paying for.
| Normalized cycle EPS | ~$7.80 |
| Terminal P/E | 36x |
| Checkpoint | Sep 25, 2029 |
| Implied base price | ~$281/share |
This is a normalized-cycle checkpoint, not me compounding peak economics forever.
| Case | What has to be true | Target range | Annualized return from today |
|---|---|---|---|
| Bear | ~$6.00 normalized EPS · 26x P/E. AI networking remains a strong business, but growth slows toward the low teens, customer concentration keeps margins under pressure and the quality premium compresses toward a still-respectable level. | $140–$172 center ~$156 | -10.3% |
| Base | ~$7.80 normalized EPS · 36x P/E. Ethernet AI fabrics keep gaining relevance, normalized EPS compounds roughly in the low-20s and Arista still earns a premium multiple, just not today's full enthusiasm forever. | $253–$309 center ~$281 | +9.4% |
| Bull | ~$9.50 normalized EPS · 42x P/E. Arista becomes a broader standard across scale-up, scale-out and scale-across AI fabrics, earnings compound around 30% and the market still pays an elite multiple. | $359–$439 center ~$399 | +23.2% |
Returns use the current quote and the time remaining to Sep 25, 2029. The scenario targets stay fixed until the underwriting changes.
Capital allocation
Fortress balance sheet, organic compounding, basically no empire building. I would still like the share count to help a little.
Most of the money goes into high-return product and R&D work, and the company carries about $13.3B of cash and securities. The deductions are pretty small: no H1 2026 repurchases despite the authorization, $241M of H1 SBC and a share count that is roughly flat rather than actually shrinking.
Stage-relative, not some objective rating-agency grade. The number is there to force the judgment into the open. See the rubric →
What management is actually doing with the cash
Arista is mostly doing exactly what I want: spend on R&D and products, keep a fortress balance sheet, do not go shopping for some giant acquisition just because you can. Cash and marketable securities were about $13.3B at Q2. My only mild gripe is the denominator. No repurchases in H1 2026, $241M of SBC, and the share count is basically flat. Not a real problem yet. Just not perfect.
What I’m doing with the next dollar
Arista has room inside the 0%–4% framework, but I already own plenty of AI infrastructure elsewhere. Another dollar into Arista has to beat the index, the compounders and the other AI names while also increasing exposure to the same broad hyperscaler capex cycle.
I would get more aggressive if the stock gives me a clearer mid-teens-plus base return while the business keeps proving that Ethernet is becoming the default AI fabric rather than merely one good option.
I get less interested if one of the giant customers slows, gross margin keeps leaking as large-customer mix rises, or the networking architecture shifts in a way that makes today’s Ethernet share gains look temporary.
For now, 1.2% feels fine. The operating numbers can keep being ridiculous without forcing me to turn that into a bigger position at whatever price the market happens to offer.
What can break this
The cleaner starting point is the midpoint of Q3 non-GAAP EPS guidance, about $1.07.
Annualized, that is roughly $4.28.
I am not saying Arista will earn exactly four times Q3 EPS. I am using the run-rate as a dated earnings anchor and then asking what has to happen over the next three years.
The valuation still needs discipline, even if AI networking demand stays strong.
Bear is roughly $6.00 of normalized EPS at 26x, or about $156.
Base is roughly $7.80 at 36x, or about $281.
Bull is roughly $9.50 at 42x, or about $399.
Getting from a $4.28 run-rate to $7.80 requires roughly low-20s EPS compounding. Even 36x at the endpoint assumes Arista maintains an exceptional competitive position; a 32x exit on the same earnings gives me less room for error.
From the Oct. 6 reference close, Base is only around a high-single-digit annualized return. That fits the portfolio action much better.
Great business. Not a compelling next dollar at this price.
The evidence I’m watching
Ethernet keeps taking share in scale-out and scale-up, 1.6T deployments broaden beyond the current cloud giants, earnings compound above the low-20s without wrecking margin, and customer concentration falls as the buyer base broadens.
One of the two giant customers slows hard, gross margin keeps compressing as large-customer mix rises, alternative networking architectures take meaningful AI share, or earnings growth slows while the stock still carries a heroic multiple.