Arista Networks

IThe business

Arista makes network switches and routers, and the software that runs them. A switch is the box that moves data between servers. In a data center there are thousands of them, and in an AI cluster they connect the chips that train and run models. Arista designs the boxes and the operating system, buys the switching chips mostly from Broadcom, and pays Jabil, Sanmina and Foxconn to build them in Malaysia, Vietnam, Mexico and elsewhere.² It owns almost no factories. Property and equipment on the balance sheet was $312.5M in June, against $10.5B of trailing revenue.³

The company was built on one idea: a network operating system can be written like modern software. EOS, the Extensible Operating System, keeps the state of the network in one database and runs a single build of code across the product line.² When Arista sold its first switches to cloud companies, that design meant fewer outages and easier automation. Those customers grew into the largest buyers of networking in the world, and Arista grew with them. Revenue compounded at 25.9% a year from $1.13B in 2016 to $9.0B in 2025 C (calculated).⁶ It fell once, by 3.9% in 2020. Microsoft and Facebook had been 40% of revenue in 2019, and in 2020 only Microsoft passed 10%, at 21.5%.¹⁵

Fiscal years end on 31 December. TTM on this page means the four quarters to 30 June 2026, because the third quarter reports on 3 November, four days after this page’s price.¹⁴ Those results will change the numbers. I would read them next to this page, and not assume this page has them.

What it sells

Management splits 2025 revenue three ways.¹⁷

• Core cloud, AI and data-center switching: about 65%. This is the 7000 series, the fixed and modular boxes that run from 10 gigabits to 800 gigabits per port, with 1.6 terabit products in trials now and in production in 2027.⁴
• Network adjacencies, meaning campus networks and routing: about 18%. Campus is the wiring and Wi-Fi inside offices and factories. The 2026 goal is at least $1.25B, and it includes VeloCloud, the branch-office business Arista bought from Broadcom for $300M in June 2025.³⁴
• Software and services on subscription, such as support contracts and CloudVision: about 17%.¹⁷ Reported service revenue was 15.9% of 2025 revenue.²

Who pays

Cloud and AI Titans were 48% of 2025 revenue. Enterprise and financial customers were 32%. AI and specialty providers, a group that now includes Apple, Oracle and the new AI clouds, were 20%.¹⁷ Arista has more than 10,000 customers in total, but the top of the list decides the year.¹⁷ Two end customers were 26% and 16% of 2025 revenue. The 10-K no longer names them. Earlier filings show the 26% one is Microsoft and the percentages for the other match Meta, and on the August call Ms. Ullal called Microsoft and Meta “our 2 longest partners”.²¹⁸⁴ Together they have been 35% to 42% of revenue every year since 2022 C (calculated).²¹⁸¹⁹ Management expects one or two more customers to pass 10% in 2026.⁴

Why AI pulled it in

Training a large model means thousands of chips exchanging data at the same moment. If one link clogs, every chip waits, and those chips are the most expensive thing in the building. So the network now decides how much work a cluster gets done. Management talks about three kinds of network.²⁰

• Scale-out links racks of chips across a building. This is where Arista sells today. Its Etherlink line has passed 100 cumulative customers, up from the four or five it described in 2024.⁴ One of those four early customers has moved from InfiniBand, NVIDIA’s older technology, to Ethernet at production scale.²⁰
• Scale-across links buildings, because no one site has enough power. Arista’s 7800 chassis is the product here. Management expects it to be about 30% of its AI target for 2026, and cites a market forecast of $15B to $20B in 2030.⁴
• Scale-up links the chips inside one rack. Today that is mostly NVIDIA’s NVLink, and Arista earns virtually nothing from it. Ms. Ullal describes Ethernet scale-up as a 2027 and 2028 story, and the CFO has said no scale-up revenue is in the 2026 guide.²⁰²¹

Dell’Oro says that in the second quarter of 2026, spending on switches for AI back-end networks passed spending on ordinary front-end networks for the first time.¹⁰ IDC’s data-center Ethernet market grew 64.5% on a year earlier to $12.3B in the quarter.⁹ This is a very large wave, and Arista is on it. Its share of the wave is the question.

The company’s 2026 guide tells the story of the year. At its September 2025 analyst day, reported targets were about $10.5B of revenue, $2.75B of AI and $1.25B of campus.²² By August 2026 the revenue guide was about $12.6B, the third raise of the year, with AI fabrics at no less than $3.5B and campus at no less than $1.25B.⁴ Ms. Ullal used “at least $3.6 billion” for the AI target on the same call, and the CFO used $3.5B. I use the CFO’s number as the floor. Analysts noticed that the last $1.1B raise went to neither target. Management said the mix would depend on what it could ship.⁴ Supply, more than demand, now sets the top line.

IIThe moat

A good test of protection is what a customer does when a cheaper option appears. The cheaper option here has been on the market for a decade: a plain box from a contract manufacturer, running open-source or home-grown software. The largest customers can write their own software. Some of them do. And yet Arista’s gross margin has stayed between 61.1% and 64.5% for ten years, while revenue rose eightfold C (calculated).⁶

That steadiness is the best single piece of evidence in this section. A hardware company selling to the most powerful buyers on earth does not usually keep a 60%-plus gross margin through ten years of scale. It keeps one if the buyer is paying for something other than the metal. The chip inside is often the same merchant silicon that rivals can buy.² The difference is EOS, plus the testing and support around it.

Evidence it is widening

Customers are moving toward it in places. In May, management described a new AI cloud that dropped a white-box design because it could not keep up with scale-out, and said that buyers of its “Blue Box”, hardware that can run other software, keep asking for EOS features anyway.²⁰ Operating problems get worse at scale, and AI makes them worse still. Mr. Duda’s example on the August call was security patches: AI finds software flaws faster, so switches need updating more often, and a reboot on a cluster costs real money.⁴ If that is right, the value of good software rises with the size of the cluster.

Enterprises are the other sign. Fortune 500 companies do not write their own switch software. Campus share was about 3% at the 2025 analyst day and management put it at about 5% in September 2026, with campus goals rising from about $750M to at least $1.25B in a year.²²²¹⁴ Those figures come through summaries of what was said, but the direction matches reported growth. Enterprise and financials at 32% of revenue also dilute the concentration risk.¹⁷

Evidence it is narrowing

The clearest sign is Dell’Oro’s ranking. In Ethernet switches for AI back-end networks in the second quarter, Celestica came first, NVIDIA a close second, Arista third and Cisco fourth, and Cisco gained the most share.¹⁰ Dell’Oro adds that Arista would be closer if deferred AI revenue were counted. Still, the top seller in the fastest-growing part of the market is a contract manufacturer. At least some of the largest buyers are choosing a box without EOS for this job.

Celestica’s third quarter makes the point in numbers. Its hardware platform revenue rose about 70% to about $2.3B, on 800 gigabit and early 1.6 terabit programs at multiple hyperscalers, and it has guided to $40.0B of revenue in 2027.²³ Its gross margin was about 11.0% C (calculated).⁸ Three customers made up 66% of its revenue.⁸ Celestica’s revenue is not all switches. But the gap between 11% and 63% is the size of the price umbrella that a hyperscaler can look at each time it designs a new network.

The second sign is in Arista’s own mix. The CFO said in May that larger customers carry a lower gross margin, and that mix is the main driver of the margin’s path.²⁰ Gross margin was 62.9% in Q2 against 65.2% a year before, and the 2026 guide of 62% to 64% assumes higher costs for memory and silicon.⁵⁴ If AI grows as the bulls expect, more of the revenue comes from the customers with the most bargaining power.

So I would describe this protection as deep and widening in enterprises, and real but contested in the largest clusters. It protects margins well. It does not protect share at the top of the market. Those are different things, and the price needs both.

IIIThe competitive advantage

Three companies and one kind of company matter. Cisco is the incumbent. NVIDIA sells networking as part of its AI system. Celestica and its peers build boxes to order. Each attacks from a different side.

Where Arista wins

It wins when the buyer cares more about the network working than about the purchase price. Its focus helps. Networking is all it does, and one software image means a fix for one customer reaches all of them. It also sits in the neutral position. Ms. Ullal said in August, “we live in an NVIDIA world”, but much of the growth in AI chips outside NVIDIA, from AMD to Google’s TPUs to custom chips, needs a network that is not tied to one vendor.⁴ Mr. Duda put it simply: the more diversity among model makers and chips, the better for Arista.⁴ In February Ms. Ullal said deployments that were “pretty much 99% NVIDIA” a year before had moved to 20% to 25% where AMD is the preferred chip.¹⁷

Its costs are also low for what it does. Non-GAAP sales and marketing was 3.6% of revenue in Q2, and general and administrative 0.8%.⁴ Cisco’s GAAP operating margin of 24.3% against Arista’s 45.4% shows what a young, single-product company looks like beside a thirty-year portfolio, even though Cisco’s gross margin is about the same.²⁴⁵ Arista has 5,115 employees, about $1.76M of 2025 revenue each C (calculated).²

Where it loses

It depends on someone else’s chips. The 10-K says it is “primarily reliant” on Broadcom for switching chips, usually with no long-term supply contract.² This has worked for more than a decade, and Mr. Nightingale calls the relationships very strong.⁴ But Broadcom sells to Arista’s rivals too, so Arista’s performance edge cannot come from the silicon.² Cisco has its own chips, NVIDIA has its own chips, and Arista does not.

It can also lose to bundles. NVIDIA called itself on its August call “the largest and fastest-growing network company in the world”.²⁷ When a customer buys a full NVIDIA rack, the network tends to come with it, and Ms. Ullal concedes that where NVIDIA supplies the whole vertical stack, the scale-up part is out of reach for now.⁴ IDC had NVIDIA first in data-center Ethernet in both of the first two quarters of 2026, and Arista’s share went from 20.7% in the first quarter to 18.7% in the second.²⁸⁹ Quarters move around, especially when revenue is deferred. But IDC has NVIDIA’s data-center Ethernet revenue up 181% on a year earlier, against 37.6% for Arista overall in IDC’s table.⁹

And management’s number one claim needs care. In May Ms. Ullal said Arista commands the top share in high-speed switching above 10 gigabits, for 2025.²⁰ That is a claim about all high-speed switching for a past year. IDC’s and Dell’Oro’s newer quarterly cuts put others ahead in the segments that are growing fastest. Both can be true. The second one matters more for the price.

IVManagement

Ms. Ullal has been chief executive for eighteen years, through a long patent war with her former employer and from a private company to a guided $12.6B of revenue. In the Cisco fight, Arista paid $400M in August 2018 to end all claims.¹⁶ She also rebuilt the bench in 2025: Mr. Duda became president, Mr. Nightingale came in as a second president and operating chief, and the long-time head of engineering and operations and the general counsel stepped down.^29303132 That reads like succession planning, though no one has said so, and she is the one key-person risk I can’t put a number on.

How they are paid, and what that pays them to do

Ms. Ullal took $2.87M in total for 2025: $300,000 of salary, a $300,000 cash bonus and stock awards valued at $2.26M.¹¹ Her pay ratio to the median employee was about 15 to 1.¹¹ For the chief executive of a $216B company that is close to nothing. What is unusual sits lower down. Mr. Duda received a stock award of $34.4M in 2024, and Mr. Nightingale stock awards of $52.0M on arrival in 2025.¹¹ Stockholders noticed: say-on-pay fell from 93% support in 2024 to 62% in 2025.¹¹ The board reached out, wrote at length, and called the low vote “unjustified”.¹¹ In May 2026 it fell further, to 60.4% of votes cast C (calculated).¹³ Since insiders and the Bechtolsheim trust hold about a sixth of the stock, support from outside holders was probably lower still E (estimate).

The plans pay for two things: revenue and non-GAAP operating income, half each, for the year.¹¹ For 2025 the target was $8.2B of revenue, the maximum $8.5B, and the actual $9.0B; for operating income the target was $3.8B, the maximum $3.95B, and the actual $4.3B.¹¹ Both came in above the maximum, so stock units paid out at the top of the scale. So targets were set low enough to be beaten by 10% in a good year, which a board can do when customers’ plans are hard to read. The measure itself tells you what is being asked for. Revenue and profit before stock pay reward growth. Nothing in the formula looks at gross margin directly, return on capital, or what happens per share, although the bonus committee does consider gross margin as a second-stage factor.¹¹

One thing in the design is to its credit. The October 2025 awards can be earned only on performance in 2029 to 2032, depending on the officer.¹¹ That pushes attention past the current boom. Few technology boards ask executives to wait that long.

How much of their own money is in

Ms. Ullal and her family trusts held 29.3M shares at 2 April 2026, or 2.3%. At $168.86 that is about $4.9B C (calculated).¹¹ The Bechtolsheim family trust holds 183.8M shares, 14.6%, about $31B C (calculated).¹¹ All directors and officers as a group hold 2.7%.¹¹ The required minimum for the chief executive is six times salary, $1.8M, a rule that means nothing at this scale.¹¹

They are also selling. Form 4s show Ms. Ullal sold 4.27M shares for about $820M between November 2025 and October 2026, at an average of about $192. Mr. Bechtolsheim sold 2.96M shares for about $520M from May to August 2026, about $176 on average. Mr. Duda sold about $112M. All insiders together sold about $1.47B in a year, and none of them bought in the open market C (calculated).¹² All the Ullal and Bechtolsheim sales were under preset 10b5-1 plans. That matters, because a plan is set in advance. Ms. Ullal’s sales were about 13% of her holding, and trimming after a thirtyfold decade is not a warning sign. But it is honest to say that the people who know the most were sellers at prices up to $217, and were not buyers at any price, including $116 in March.¹

Mr. Bechtolsheim needs one more sentence. In March 2024 the SEC charged him with insider trading in Acacia options, on information from Cisco’s deal to buy that company. He settled without admitting or denying it, paid $923,740, and took a five-year bar from serving as an officer or director of a public company.³³ He remains chief architect and is the largest holder.²⁰¹¹ The trades had nothing to do with Arista’s business. It is still the largest holder’s record, and I would rather put it on the page.

What they did with capital over a full cycle

This business barely needs capital, so the question is what happens to the surplus. Since 2019 Arista has spent $3.88B on 120M split-adjusted shares, an average of $32.23 C (calculated). At $168.86 those shares are worth about $20B C (calculated). The largest year, 2025, came at $100.63, in a year the stock also traded as low as $64.37.¹² By any test, that is good buying.

Three things temper that. First, the buybacks mostly offset stock pay. Diluted shares were 1,294M in 2019 and 1,276M in June, down 1.4% C (calculated), and stock pay over those years was $1.75B, less than half of what was spent, because awards rose so much in value. The real cost of stock pay is the share price at vesting, not the accounting figure. Second, there were no repurchases in the first half of 2026, even when the stock closed at $116 in March and $817.9M of authorization was open.³¹ The company chose to keep cash while it committed to $9.7B of parts, which is defensible, but it means the price discipline went quiet at a good moment.

Third, the cash keeps piling up. Cash and securities were $13.3B in June, with no debt and no dividend.³ That is 3.4 times the $3.88B returned in seven years. Acquisitions have been small and sensible: VeloCloud cost $300M.³ An owner has to ask what that cash is for, and the honest answer today is a safety buffer against the purchase commitments and a long option on something larger. Interest on it earned $434M over the past year C (calculated).²³

VFinancials

Revenue was $3.04B in the second quarter, up 37.7%. GAAP operating margin was 45.4%, and diluted earnings per share were $0.95.⁵ The guide for the third quarter is about $3.3B of revenue, a 63% gross margin and non-GAAP earnings of $1.06 to $1.08 a share.⁵⁴ For the year, the guide is about $12.6B, a 62% to 64% gross margin and a non-GAAP operating margin of 48% to 49%.⁴

Operating margin has risen from 21.5% to 43.1% in nine and a half years with a flat gross margin, so all of the gain came from spreading R&D, sales and overheads across more revenue C (calculated).⁶³ R&D went from 24.3% of revenue to 13.0%. That is what a good software layer on merchant silicon does. It also limits what is left: there isn’t much overhead left to spread, and sales and administration are close to a floor.

What an owner could really take out

Reported free cash flow over the past year was $5.16B, operating cash flow less capital spending C (calculated).²³ That number is too flattering to use. Here is the walk.

So the number I would call owner cash is about $3.5B. That is 13% below reported net income of $4.04B and 32% below reported free cash flow. On a value net of cash of $203B it is a yield of 1.7%, against 4.94% on a ten-year Treasury.³⁶ The gap has three sources. Stock pay is a real cost. Interest belongs to the cash pile, which I count separately. And about $0.8B was net timing: $2.8B of customers’ money for shipments not yet recognized as revenue, less the cash that went into receivables, parts and other assets.

Quality of earnings

Deferred revenue deserves attention, because it is the biggest moving part in the accounts. It rose from $2.8B at the end of 2024 to $6.9B in June, and the CFO says most of it is now product.³⁴ New AI products are shipped under contracts with acceptance clauses. Arista gets the invoice out, and the revenue waits until the customer signs off.³ So the reported revenue understates what has been shipped lately, and that’s a real strength. But it cuts both ways. The cash has mostly arrived already, so when the deferred revenue turns into sales, revenue goes up without new cash. Net income in 2027 could look strong while cash lags. The balance can “move significantly”, in the CFO’s words, on a quarterly basis, independent of underlying business drivers.⁴ Total remaining obligations were $8.4B, of which $875.1M was binding agreements for future product shipments, up from $227.6M in December.³²

The tax line is the other one to watch. The effective rate was 17.4% in 2025 and 12.6% in 2024, helped by stock awards; the guide for 2026 is 21.5%.²⁴ Arista also changed how it computes non-GAAP tax in 2026, and recast prior periods.⁵ I use 21.5% and ignore the benefit.

The balance sheet

$13.3B of cash and securities, no debt.³ That sounds like the end of the discussion. It isn’t, because of what sits off the balance sheet.

Non-cancellable purchase commitments were $9.7B in June, up from $3.6B a year earlier and $6.8B in December, and $9.4B of it is due within 12 months.³⁴ Inventory is $2.5B on top, turning 1.7 times a year.⁴ Deposits with contract manufacturers were $124.4M.³ Together, inventory and commitments are about $3.1 times the past year’s cost of revenue of $3.9B C (calculated). The company has more than doubled its bet on parts while customers can, in the 10-Q’s own language, “cancel, delay, reduce or otherwise modify” orders with little or no notice.³ Most sales are on purchase orders.

The last cycle shows what this can do. Commitments hit $3.7B at the end of 2022, and cash from operations that year fell to $0.49B while net income rose to $1.35B.¹⁹ In 2023 commitments fell to $1.59B and provisions for excess inventory rose, which kept gross margin at 61.9%. Those charges fell by $180.4M in 2024, and the margin went back to 64.1%.³⁵¹⁸ That was at a commitment level about 40% of today’s. Nothing went badly wrong then, because demand held up. It is a good test of how the accounts behave and a poor test of what happens if it doesn’t.

The stock tends to fall hard on good news. It dropped 24.2% on 1 November 2019, the day after third-quarter results; 22.4% on 27 January 2025; and 13.6% on 6 May 2026, the day after Arista raised its 2026 guide to $11.5B.¹²⁰ It reached a record $216.74 on 9 October 2026 and closed $168.86 three weeks later, down 22%, after an 8.3% fall on 30 October alone, a day when Broadcom, NVIDIA and Celestica fell too.¹³⁷ I don’t try to explain one day. I do note that a day earlier this page’s price would have been $184.22.

VIValuation

First, how this goes badly wrong

Before any numbers, I asked what would most likely destroy this investment, and worked back. The business is sound, with no debt, so ruin would have to come from the price and from concentration. Three ways stand out.

• The largest buyers pause, with the parts already ordered. Two customers were 42% of 2025 revenue. In 2020 revenue fell 3.9% after their spending slowed, and that was a small event.¹⁵² Today Arista has promised $9.7B to suppliers on the strength of orders that the 10-Q says can be cut with little notice.³ A pause in 2028 or 2029, after this much capacity has been built, hits revenue, margin and inventory together. Four large buyers spent $496.7B of capital in the first nine months of 2026, a pace that cannot compound forever.³⁸
• The best customers stop paying for the software. Celestica’s 11% gross margin sits next to Arista’s 63%.⁸⁵ If Microsoft or Meta, or the next ones up, decide their own software is good enough for AI back-end networks, the loss comes on the largest and most visible part of the growth story. It would not happen in one quarter. It would show up as share lost one design at a time.
• A good company at the wrong price. Nothing needs to go wrong at the business for this one. A buyer at 36 times next year’s earnings needs growth to last a decade. If it lasts five, the multiple falls as earnings rise, and the owner earns little. Arista has fallen by more than 30% four times in ten years, and it has never needed bad results to do it.¹

The first two set my bear. The third is what the whole section is about. None needs a surprise. All three are in the company’s own risk factors or its own history.

What the price implies

At $168.86 the market values Arista at $216B on 1.279B diluted shares, or $203B net of June’s cash C (calculated).¹⁴ Start with earnings. That is 53 times the past year’s GAAP earnings and 36.0 times Nasdaq’s mean for 2027.⁷ That is at the top of the AI hardware group I cover, level with AMD’s 35.7 on my Broadcom page.³⁹

Multiples this far apart say the market expects Arista’s earnings to last longer or grow faster than NVIDIA’s and Broadcom’s. On Nasdaq’s means, Arista’s 2027 earnings grow 24%, and its 2028 earnings 10%, on four estimates.⁷ NVIDIA’s and Broadcom’s are expected to grow faster than that, from far larger bases.³⁷ What explains the gap, I think, is a belief that a software company lives inside this hardware company, and that the AI cycle hurts it less. Some of that is fair. It is also the same argument that needs to survive Celestica’s growth.

The reverse DCF starts from the $3.51B of owner cash in section V. I solve for one growth rate for ten years. The price needs 23.7% a year, to $29.4B in 2036 C (calculated). At the cash margin of my bull case that takes about $87B of revenue, seven times the 2026 guide and more than Cisco’s $63.3B, from a company whose biggest year so far is a guide of $12.6B E (estimate).²⁴ My base grows owner cash 14.9% a year and my bull 22.8%.

The price is roughly my bull case. An owner today pays for Arista being a far larger company than Cisco is now, and gets the base at a return below a Treasury.

What I can’t judge

Three inputs matter most, and I can’t know any of them. I can’t judge how long AI capital spending lasts. No one has seen a build this large, and the people who sell into it have every reason to be sure of it. I can’t judge how much of the top of the market stays with branded systems. Dell’Oro’s ranking is one quarter of one report. Nor can I judge the true size of the market. Management is rewriting its own $105B figure for 2029, and has said scale-up comes on top of it.²¹²² A number I would trust does not exist yet.

So I show the range. Growth matters far more than margin. Moving gross margin 2 points changes value by about $5 a share. Adding 2 points of growth a year for a decade adds $15. It takes about 8 extra points of growth a year, every year, for my base to reach the price.

The discount rate

I use 11.0%. The capital asset pricing model starts with the 4.94% ten-year Treasury yield and Damodaran’s 4.20% implied equity risk premium of 1 October.³⁶⁴⁰ My beta on ten years of month-end price returns is 1.39, which gives 10.8%; on five years, 1.57, which gives 11.5%; on three years, 2.15, which gives 14.0% C (calculated).¹⁴¹ The three-year figure is mostly one AI trade, and I don’t lean on it. I take 11.0%, a beta of 1.44, between the first two. That is half a point above my Broadcom page. The business has no debt, but it has more customer concentration and a higher beta on every window. One point on the rate moves my base by about 13%.

Three cases

I build each case from three lines: AI fabrics, the rest of the core business with routing and services, and campus. I start from the 2026 guide of about $12.6B, then set a gross margin and an operating expense ratio that includes stock pay. Tax is 21.5%, the guided rate. Capital spending is 1.2% of revenue, about where I expect the past year’s 1.4% to settle after the Santa Clara buildings finish, against depreciation of 0.8%.⁴ Working capital takes 10% of each new revenue dollar and is not released when revenue falls. That is more than recent history, when customer money funded growth, because I don’t think deferred revenue can keep rising faster than sales. Stock pay is a cost, so I divide by today’s diluted shares and add no buybacks.

My base is not a cautious one in the near term. It has revenue of $16.5B in 2027 and $20.5B in 2028, and GAAP-style earnings of about $4.80 and $5.82 a share, against Nasdaq’s $4.69 and $5.17 E (estimate).⁷ I am above the Street in both years, and the value is still well below the price. That tells you the gap has little to do with the next two years. It has to do with years five to ten.

Bear

AI spending pauses in 2029, the way cloud spending paused in 2019 and 2020 but deeper, while buyers keep moving to lower-cost boxes. AI revenue falls by half from its 2028 peak, total revenue drops from $16.5B to $12.2B, the parts bought for growth get written down, and gross margin settles at 54.5%. Revenue still grows after that, to $19B in 2036.

Base

2027 and 2028 go well, above the Street. AI revenue reaches $10.3B in 2029, slips 8% in 2030 in one digestion year, and grows in the teens after that, fading to 12%. The core keeps compounding near 8%, campus reaches $6.9B, and gross margin settles at 60% as mix moves to the largest buyers. Arista ends up a $46B company, close to three quarters of Cisco’s size today.

Bull

Ethernet wins scale-up, non-NVIDIA chips take a large share, and Arista keeps about a fifth of a market several times its present size. AI revenue reaches $46B, campus $12.5B, and gross margin holds at 62%. Revenue of $81B in 2036 is about 1.3 times what Cisco earns today.

Weighted 25%, 50% and 25%, the three cases give about $106 a share E (estimate). More than half of each value sits after 2036, so all of them rest on what the business looks like in ten years. No one can see that clearly, which is one reason not to pay for it in full now.

The skew at $168.86 is −72% to my bear and +4% to my bull C (calculated).

Other uses of the same dollar

A dollar put into Arista today has other places to go. The hurdle isn’t only a Treasury. It’s the best use of the same dollar in the same theme. That theme has two obvious candidates on this site.

On my own numbers, a dollar in NVIDIA or Broadcom buys more cash per dollar of price, with Ethernet exposure included, since NVIDIA’s network grows faster and Broadcom sells Arista its chips. What it doesn’t buy is Arista’s neutrality between chip makers, or its enterprise business. Those are worth something. On my numbers they aren’t worth a multiple 2.8 times NVIDIA’s and a base return about six points lower.

The margin of safety

I set the margin-of-safety price from the bear case: the price at which my $47 bear value is still at least three quarters of what was paid. That is $62, 63% below the price and 39% below my base. Most of my pages call such a figure a description of a crash. Here it’s closer to recent memory: the stock closed at $64.37 in April 2025, in the same year it earned $2.75 a share.¹⁶

Two different questions

How good is the business? Very good. It has ten years of steady gross margins, a 43% operating margin, no debt, co-founders still on staff, and a product that customers pay up for when they have alternatives. I would put its quality among the better ones on this site. How good is the price? That’s a separate question, and this looks like a great business at a full price. The market is paying today for what could go right through 2036. Good companies can be poor purchases at the wrong price, and the only way this one avoids that is by becoming something much larger than I can responsibly model as the central case.

Where this could be in 5 and 10 years

These values are my model’s, rolled forward. Value at a date is the worth then, at 11.0%, of the owner cash after it, per today’s 1.279B shares and in that year’s dollars. It leaves out today’s $15.7B of net cash. Owner cash is counted as if paid out each year.

For the base to hold at five years, revenue has to reach about $24B in 2030, and AI sales have to get through one digestion year. At ten years it needs $46B of revenue, a 40.5% operating margin after stock pay, and $14.1B of owner cash in 2036. Spending on plant takes 1.2% of revenue, and working capital takes 10% of each new revenue dollar.

The bear and the bull differ on growth first and margin second. In the bear, AI spending pauses in 2029, revenue falls to $12.2B in 2030, and operating margin settles at 33%. In the bull, revenue reaches $81B and margin holds at 44.5%. One point on the 11.0% rate moves my base value by about 13%.

The most likely way my base is wrong is growth after 2030. It has AI revenue rising from $9.5B in 2030 to $22B in 2036. Growth two points a year lower takes the base value today from $101 to $88.

These are estimates of value, not forecasts of the share price. The market price can sit above or below value for years.

VIIThe case for the price

What most people miss

Three things cut in Arista’s favour. First, the purchase commitments may be saying more than the guide. $9.4B of parts and products are due within 12 months.³ At the past year’s 37% cost of revenue, that would support about $25B of sales, and my base has about $15B of revenue over the next twelve months E (estimate). Some of the gap is higher prices, chips for 2027 products and multiyear memory supply.⁴ But a company with Arista’s record doesn’t buy that far ahead of its forecast by accident. Second, management says price increases haven’t yet reached the numbers because of backlog sales, and they arrive late this year or next.⁴ Third, the guide contains no scale-up revenue at all, and Dell’Oro says Arista would rank closer to the top if deferred AI revenue counted.²¹¹⁰

The bull case

In plain words: AI makes the network the bottleneck, Ethernet is the open standard that everyone but NVIDIA wants, and Arista has the best software on it. Management says it sees multiyear demand and will do everything, “including hurt our gross margins”, to supply it.²⁰ It is rewriting its market estimate upward, and expects a third and possibly a fourth 10% customer.²¹

The sell side mostly agrees. MarketBeat’s mean target is $233.20 from 24 analysts, between $164 and $289. Evercore ISI reiterated $250 on 7 October, Citi’s Atif Malik raised to $248 on 8 October, and Bernstein’s Daniel Zhu started at $250 on 30 September. TD Cowen cut to $205 on 19 October.⁴³⁴⁴ I saw reports of those notes, not the notes. Money managers are quieter. I searched the 30 June 13Fs of Altimeter, Atreides, Coatue, Tiger Global, Lone Pine, Viking, Whale Rock, D1, ARK, Third Point, Duquesne, Appaloosa and Situational Awareness, and none lists Arista.⁴⁵ That is four months old and says nothing about why. I found nothing published by those managers on this company, so I attribute no view to them.

Against reality

Today’s numbers support the near term. Revenue grew 37.7% in Q2 and the guide implies about 43% in the second half C (calculated).⁵⁴ If Arista ships what it has ordered, and customers accept it, 2027 can beat my base.

The price needs much more. If Arista holds a fifth of AI networking as that market multiplies, if campus becomes a business of more than $10B, if scale-up opens to Ethernet in 2028, and if gross margin stays above 60%, then $87B of revenue by 2036 is possible, and the price with it. Today AI fabrics are a goal of at least $3.5B, campus at least $1.25B, and scale-up is zero.⁴

It helps to multiply it out. AI spending keeps rising through 2030 without a real bust, 60%. Arista holds about a fifth of AI Ethernet against NVIDIA, Cisco and box builders, 50%. Scale-up opens and Arista wins a real share, 40%. Campus goes past $10B, 50%. Gross margin holds at 60% or better, 60%. Those are my guesses, and they aren’t independent. Multiplied, they give 3.6% E (estimate). Allowing for overlap, perhaps 10% to 15%. I still give the bull 25%, as on my other pages, and that is generous. Each step sounds reasonable. All five together are a long shot.

What a 10x would take. Ten times $168.86 is $1,689, about $2.2T of market value. At 25 times earnings Arista would need about $86B of net income. At the 38% net margin of the past year, that is about $228B of revenue, eighteen times the 2026 guide and 3.6 times all of Cisco.²⁴ By 2036 that means revenue growth of about 34% a year for ten years C (calculated). IDC’s whole Ethernet switch market runs at about $76B a year C (calculated).⁹ Arista did rise about 32 times in the past ten years, but from about $7B of value C (calculated).¹ The best run I know of from a very large base is NVIDIA’s: about five times from its $40.11 close of May 2023, near $1T, in about three and a half years.⁴⁶ From $216B, a tenfold move in a company selling network gear would be close to impossible, and I would not build a case on it.

Where I come down

The commitments moved me: they are why my 2027 and 2028 sit above the Street. The bull’s long-run share didn’t, because the fastest-growing part of the market is going to a box builder and to NVIDIA. Proof that the largest AI buyers are paying for EOS in 1.6 terabit networks at a 62% gross margin would raise my values most.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• Supply and cost. Ms. Ullal said in August that the industry has “a 2-year problem” and won’t get out of it until 2028.⁴ Suppliers pulled back from commitments in early 2026, and management said the amount “does not feel good”.²⁰ Memory and silicon costs are in the margin guide, and management has chosen to take some of that hit to keep customers supplied.²⁰
• Dependence on Broadcom. There’s usually no long-term contract, and a rival with a chip advantage or a supply priority could hurt fast.²
• AI clouds as customers. New AI clouds are a growing part of the 20% group. Ms. Ullal said in February that Arista looks at the health of these buyers, and some may be one-time.¹⁷
• Tariffs and trade. Products are built in Malaysia, Vietnam and Mexico, and shipped mostly to the Americas.² Tariff refunds added 20 to 30 basis points to Q2’s margin against the guide, which shows how directly trade rules reach the line.⁴
• Technology shifts. Co-packaged optics could change who owns the value in a switch. Arista backs open designs, and expects some adoption in 2028 and 2029.⁴ Scale-up may stay inside NVIDIA’s NVLink for longer than the bulls think; it took two to three years for InfiniBand to give way.⁴
• People and governance. Ms. Ullal is the company’s public face, the board chair and chief executive are one person, and say-on-pay has lost close to 40% of votes two years running.¹¹¹³
• Legal. Arista has fought intellectual property cases before, including with Cisco and OptumSoft, and open questions over who owns which software can recur as code grows.¹⁶