Arista Networks, Inc. ANET

Key data

The business

Arista builds the high-speed Ethernet switches that sit between the servers and GPUs in a data center, the boxes that decide how fast a thousand chips can talk to each other while a model trains. Every switch runs one operating system, EOS, the same software from the smallest campus box to the largest 800-gigabit spine, which is why a customer who standardizes on Arista finds it costly to rip back out. The buyers are a short list: two hyperscale cloud operators were 42% of 2025 revenue between them, with enterprises, banks and telcos filling the rest. Product is 85% of sales, sold once. Service and software subscriptions are the other 15%, billed over years and parked in deferred revenue until earned. The moat is that software layer plus the scale to buy merchant silicon cheaply, not any one chip.

The numbers

Growth reaccelerated in the June quarter. Revenue of $3.04B rose 38% on the year and 12% on the quarter, the fastest sequential step in two years, and net income of $1.21B grew 37%.

The inflection is Q1 2026, when the AI backend orders that had been sitting in backlog began converting to shipments. The four most recent quarters each cleared the Street's adjusted number, the latest by 15%. Over five years revenue compounded 32% a year and net income 43%, margin doing the extra work.

Remaining performance obligations reached $8.4B at quarter end and total deferred revenue climbed to $6.9B, so the near-term demand is visible and contractual, not hoped for.

The one crack is price.

Product gross margin fell to 59.8% from 62% a year earlier, and with service steady near 82%, total gross margin compressed to 62.9% from 65%. The reason is the concentration itself: the bigger the two anchor customers get, the more price they command. That is the variant this memo carries. The market is modeling 35%-plus growth into 2027 at $16.5B without pricing the margin that growth costs. The print that settles it is product gross margin, quarter by quarter.

Management

Insiders sold and did not buy. Over the last twelve months there were 47 sales worth $222M and zero purchases, led by Jayshree Ullal at about $72M across two August sales and Andy Bechtolsheim at $26M. The vendor feed does not carry 10b5-1 status, so plan status is not disclosed, and a lump this size hides whether it was scheduled. The company repurchased about $983M of stock in the first half at 50-to-60 times earnings, buying its own shares at the richest multiple in its history, a habit that has burned better operators than this one. Against that, the balance sheet carries no debt and $2.3B of cash, return on equity runs near 31%, and the guidance record is four straight quarters ahead of plan. Pay rewards growth, and growth has come.

How it fails or surprises you

The anchor customer builds its own (downside). The larger of the two 10%-plus customers grew to 26% of revenue in 2025 from 20%, and hyperscalers design in-house silicon once volume justifies it. If that customer moves its backend to a home-built switch, more than a fifth of revenue reprices over a year. The print is the concentration line in the next 10-K. So far it has only deepened, not defected.

Margin keeps sliding on mix (downside). Product gross margin fell from 62% to 59.8% in a year as the anchors scaled. Consensus models $16.5B of 2027 revenue, but the margin path is the unmodeled variable, and each added point of concentration costs gross margin. A product print below 58% turns a growth story into a volume story.

Ethernet eats the AI backend faster (right tail). InfiniBand still owns much of the GPU-cluster interconnect, the industry is standardizing on Ethernet for that job, and Arista is the pure-play. If the shift accelerates, the $8.4B backlog and 2028's $20.4B revenue estimate prove low. The tell is a third 10%-plus customer, or RPO growth outrunning revenue.

Closing thoughts

This is the world the market has largely priced. At 61 times trailing and 37 times next year, above every multiple in its twelve-year record and half again richer than peers near 40, the growth is no secret, and you are paying full freight for the best-positioned name in the fastest-growing corner of networking. The person on the other side is selling the concentration and the margin slide, and the honest answer is that both are real and both are visible in this very filing. The fatter tail is still the upside, because Ethernet's displacement of InfiniBand is a multi-year structural move and Arista is the cleanest way to own it. The downside is not small: one customer decision reprices a fifth of the business, and you would learn it a quarter late. What is at risk if the anchor defects outweighs the upside of any single good print.

The bet is still that the cloud giants keep building AI data centers and keep buying Arista to wire them, and that Ethernet keeps winning that wiring. What breaks it is the same thing that makes it rich: those two customers, 42% of sales and paying less per box each year. The falsifiable line is simple. Product gross margin under 58% with the top customer over 30% of revenue, printed in the same year, means the moat is being paid for by the moat.